Porto: Towards Solving A Robust and Sustainable Vehicle Routing Problem With Backhauls

Maria João Martins dos Santos
U. PORTO
FEUP FACULDADE DE ENGENHARIA
UNIVERSIDADE DO PORTO
DEPARTAMENTO DE ENGENHARIA E GESTÃO INDUSTRIAL
Towards solving a robust and sustainable

Vehicle Routing Problem with Backhauls
Submitted to Faculdade de Engenharia da Universidade do Porto in partial fulfillment of

the requirements for the degree of Doctor of Philosophy in Industrial Engineering and
Management, supervised by Pedro Amorim, Assistant Professor of Faculdade de
Engenharia da Universidade do Porto, Eduardo Curcio, researcher at INESC TEC and
invited professor at Faculdade de Engenharia da Universidade do Porto, and Alexandra
Marques, researcher at INESC TEC
Department of Industrial Engineering and Management

Faculdade de Engenharia da Universidade do Porto
2020
“Research is to see what everybody else has seen,
and to think what nobody else has thought.”
- Albert Szent-Gyorgyi
v
Acknowledgments
First and foremost, I would like to thank my supervisor Pedro Amorim for his deep com-
mitment throughout my PhD journey. He was always present and available to help me in
my research, and guide me whenever I felt adrift. More importantly, he kept me challeng-
ing to dig in current research and to look ahead for innovative approaches. I grow up more
in these last three years than in my previous academic research, and for that, I am genuinely
thankful.
I would also want to express my gratitude to my co-supervisors Eduardo Curcio and
Alexandra Marques. Eduardo was an exceptional mentor. He thought me that more is
never enough, always challenging me to pursue excellence in scientific research. Alexan-
dra is a generous person. She had provided full support in my research, potentiating new
avenues to explore scientifically solutions for real life problems.
A special gratitude goes to my colleagues Sara Martins and Ricardo Soares, which turn out
to be my very good friends. Sara, with her very kind personality, was always available to
listen and help me overcome my problems, to endure and to succeed. Ricardo, with his
genius mind, had always provided me with the example of hard work and commitment,
and I am truly grateful for working with him as a team.
Finally, I would have to thank my favourite people in the world: Guilherme, Mãe and Pai.
Guilherme is a never ended patience person, always by my side, whichever was my mood
and always given me the best words one can ever want to hear. Mãe and Pai are the bravest
and kindest people I have ever known. They had conveyed me life values and principles
that made the person I am today, and I am forever grateful for their guidance and support.
A special word to Pai, who is not present today in my life, but I am sure I have made him
proud. In my heart, you have never left, and you never will.
vii
Abstract
Transportation is a key process in all supply chains as it constitutes the link to execute all
order fulfillment activities. Nowadays, the challenges imposed by a greener circular econ-
omy require more attention to the environmental impact of transportation and logistics. An
opportunity to improve the sustainability of transport operations is to plan simultaneously
outbound routes (deliveries to customers) and inbound routes (pickups from suppliers) in
order to reduce empty running of vehicles and increase their efficiency in terms of used
capacity. Under the domain of Operations Research, this problem is framed as the Vehicle
Routing Problem with Backhauls (VRPB) and it is used as a base problem throughout this
thesis. The scope of this thesis is the integrated transportation planning within a sustainable
and uncertainty context, addressing several challenges that are currently presented in the
literature. Overall, this thesis aims to develop mathematical models and solution methods
for the VRPB to provide more robust plans, understand the value of inter-firm collabora-
tion, and help decision-makers to perform comparative analysis among plans, considering
economic and environmental aspects. This research was grounded in the analysis of two
case studies, namely in the wood-based panel industry and in the food supply chain. The
results provide valuable managerial insights suitable to be implemented in the case studies,
but also to be extended to other industries and supply chains.
ix
Resumo
O transporte é uma atividade chave em todas as cadeias de abastecimento, uma vez que
constitui o elo necessário para todas as atividades relacionadas com o cumprimento de
encomendas. Hoje em dia, os desafios impostos por uma economia circular mais verde
exigem mais atenção ao impacto ambiental do transporte e da logística. Uma oportunidade
para melhorar a sustentabilidade das operações de transporte é planear simultaneamente
rotas de saída (entregas aos clientes) e rotas de entrada (recolhas dos fornecedores), a fim
de reduzir a circulação de veículos em vazio e aumentar a sua eficiência em termos de ca-
pacidade utilizada. No domínio da Investigação Operacional, este problema é enquadrado
como o Problema de Roteamento de Veículos com Retorno (PRVR) e é utilizado como o
problema de base ao longo de toda esta tese. O âmbito desta tese é o planeamento integrado
do transporte dentro de um contexto sustentável e de incerteza, abordando vários desafios
que são actualmente apresentados na literatura. De uma forma geral, esta tese pretende
desenvolver modelos matemáticos e métodos de solução para o PVRP que permitam defi-
nir planos de transporte mais robustos, perceber o valor da colaboração entre empresas, e
ajudar os decisores a realizar análises comparativas entre planos, considerando diferentes
aspetos económicos e ambientais. Esta investigação foi fundamentada na análise de dois
estudos de caso, nomeadamente na indústria de painéis derivados de madeira e na cadeia de
abastecimento alimentar. Os resultados proporcionaram conhecimentos valiosos de gestão
adequados para a implementação nos casos de estudo, mas também para serem aplicados a
outras indústrias e cadeias de abastecimento.
Contents
1 Motivation and overview 1

1.1 Research objectives and methodology . . . . . . . . . . . . . . . . . . . . 3
1.2 Thesis structure and synopsis . . . . . . . . . . . . . . . . . . . . . . . . . 7
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 Vehicle Routing Problem with Backhauls 11

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 The VRPB and its variants . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4 Literature overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4.1 Descriptive analysis . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.4.2 Classification of VRPB works . . . . . . . . . . . . . . . . . . . . 21
2.4.3 Final remarks on the VRPB . . . . . . . . . . . . . . . . . . . . . 27
2.5 Analysis of VRPBs with sustainability concerns . . . . . . . . . . . . . . . 28
2.5.1 VRPBs with sustainable objectives . . . . . . . . . . . . . . . . . . 28
2.5.2 VRPBs in collaborative networks . . . . . . . . . . . . . . . . . . 30
2.5.3 VRPBs in reverse logistics . . . . . . . . . . . . . . . . . . . . . . 31
2.5.4 Final remarks on the VRPBs with sustainability concerns . . . . . . 32
2.6 Future directions for sustainable VRPB . . . . . . . . . . . . . . . . . . . 34
2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.A Classification of VRPB works . . . . . . . . . . . . . . . . . . . . . . . . 51
2.B Benchmark instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.C Cases studies and applications of VRPB and variants . . . . . . . . . . . . 59
3 A Rich Vehicle Routing Problem with Backhauls 61

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2 Critical review of the state of the art . . . . . . . . . . . . . . . . . . . . . 65
3.3 Problem formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.3.1 Logistics planning strategies . . . . . . . . . . . . . . . . . . . . . 69
3.3.2 Opportunistic backhauling planning (OBP) . . . . . . . . . . . . . 70
3.3.3 Integrated Inbound and Outbound Planning (IIOP) . . . . . . . . . 77
3.3.4 Decoupled Inbound and Outbound Planning (DIOP) . . . . . . . . 78
3.4 Solution approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.4.1 Fix-and-optimise approach . . . . . . . . . . . . . . . . . . . . . . 78
3.4.2 Data pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.5 Computational experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.5.1 Case study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.5.2 Comparison among distinct planning strategies . . . . . . . . . . . 84
3.5.3 Managerial insights . . . . . . . . . . . . . . . . . . . . . . . . . . 86
xii Contents
3.5.4 Performance of the solution approach . . . . . . . . . . . . . . . . 90

3.6 Conclusions and future work . . . . . . . . . . . . . . . . . . . . . . . . . 94
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.A Supplementary material . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
4 A Robust Vehicle Routing Problem with Backhauls 103

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.2 Problem description and formulations . . . . . . . . . . . . . . . . . . . . 107
4.2.1 Deterministic VRPSB . . . . . . . . . . . . . . . . . . . . . . . . 107
4.2.2 Robust VRPSB . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.2.3 Estimates of probabilistic bounds . . . . . . . . . . . . . . . . . . 113
4.2.4 Factor model support for uncertain revenues . . . . . . . . . . . . . 117
4.3 Solution methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
4.3.1 Branch-and-Cut . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
4.3.2 Adaptive Large Neighborhood Search . . . . . . . . . . . . . . . . 120
4.4 Computational tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.4.1 Problem instances . . . . . . . . . . . . . . . . . . . . . . . . . . 124
4.4.2 Evaluation of the VRPSB under uncertainty . . . . . . . . . . . . . 125
4.4.3 Evaluation of solution methods . . . . . . . . . . . . . . . . . . . 132
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.A Chance-constrained VRPSB . . . . . . . . . . . . . . . . . . . . . . . . . 144
4.B Comparison of single and multiple backhaul visits . . . . . . . . . . . . . . 144
5 A Collaborative Vehicle Routing Problem with Backhauls 149

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5.2 Background literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
5.3 Problem description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
5.3.1 The collaborative transportation planning problem . . . . . . . . . 154
5.3.2 Mathematical programming formulation . . . . . . . . . . . . . . . 155
5.3.3 Properties of the bilevel VRPSB . . . . . . . . . . . . . . . . . . . 159
5.3.4 Numerical example . . . . . . . . . . . . . . . . . . . . . . . . . . 160
5.4 Single-level reformulation . . . . . . . . . . . . . . . . . . . . . . . . . . 161
5.5.1 Data sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.5.2 Bilevel versus traditional planning . . . . . . . . . . . . . . . . . . 166
5.5.3 Bilevel versus compensation schemes . . . . . . . . . . . . . . . . 169
5.5.4 Computational limits of the collaborative VRPSB . . . . . . . . . . 173
5.6 Conclusions and future research . . . . . . . . . . . . . . . . . . . . . . . 174
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
5..1 Traditional models . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Contents xiii
6 A Green Vehicle Routing Problem with Backhauls 183

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
6.2 Related literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
6.3 The collaborative transportation planning . . . . . . . . . . . . . . . . . . 190
6.3.1 Illustration of collaborative strategies . . . . . . . . . . . . . . . . 190
6.3.2 Vehicle routing problem . . . . . . . . . . . . . . . . . . . . . . . 191
6.3.3 Profit sharing problem . . . . . . . . . . . . . . . . . . . . . . . . 196
6.4.1 Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
6.4.2 Analysis of collaborative solutions . . . . . . . . . . . . . . . . . . 198
6.4.3 Sensitivity analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 200
6.4.4 Performance of allocation methods . . . . . . . . . . . . . . . . . 202
6.4.5 Case-study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
6.A Stand alone model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
6.B Impact of full collaboration versus limited collaboration . . . . . . . . . . . 211
7 Conclusions and future work 213

7.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
7.2 Answers to the research questions . . . . . . . . . . . . . . . . . . . . . . 216
7.3 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
Chapter 1
Motivation and overview
Transportation is a crucial element in any supply chain, since it establishes the connec-
tions between the different stages comprising an entire chain. In addition, transportation
represents a significant portion of the total logistics costs (Crainic and Laporte, 1997) and
accounts for about 24% of total greenhouse gases (GHG) emissions in the European Union,
where road transportation alone comprises about 17% (Demir et al., 2014). Nowadays, pro-
moting smart, green and integrated transportation solutions are challenges that drive many
ongoing research projects in Europe (see ec.europa.eu/programmes/horizon2020).
Given its importance for the supply chain and the high impact on the economic and envi-
ronmental costs, transportation hides a huge potential to promote efficient and sustainable
supply chains. Among several green initiatives adopted by logistics carriers, optimizing
transportation planning and reducing empty trips of vehicles are becoming more popular
(Lin et al., 2014; Evangelista et al., 2017).
Optimization emerges as a powerful approach to find optimal routing plans for a fleet of
vehicles, considering a set of constraints. This problem is generally known as the Vehicle
Routing Problem (VRP) and was firstly introduced by Dantzig and Ramser (1959) to solve
a real-world logistics problem where a fleet of trucks departing from a bulk terminal had
to visit a series of service stations. The VRP is defined as an optimization problem that
aims to plan routes for a homogeneous fleet, in order to satisfy all customers demands in a
single visit and without exceeding the capacity of the vehicle. Each vehicle departs loaded
from a common depot and returns back empty to the same depot. The most common
objective in the VRP is the minimization of the total routing costs (usually given by the
total traveled distance). Since its first application, several new variants have been created,
attempting to capture real-life aspects of transportation problems, such as characteristics
of vehicles (e.g., size, capacity, weight), drivers (e.g., licenses, resting period, maximum
working hours), customers (e.g., time windows) and products (e.g., compatibility between
products and vehicles), among other aspects (Mancini, 2016).
Empty trips can constitute a significant part of the overall travelled distance and contribute
substantially to the total routing costs (Ubeda et al., 2011). A strategy to reduce empty trips
is called backhauling, which takes advantage of the capacity of the vehicle, when it runs
empty, to pick up some load in the return trip to the depot. Under the domain of Operations
Research, this problem is known as the Vehicle Routing Problem with Backhauls (VRPB)
(Deif and Bodin, 1984) and it has been successfully applied in real logistics problems (Toth
and Vigo, 1999; Koç and Laporte, 2018). Including backhauling in transportation planning
may also be combined to optimize the selection of suppliers. When several suppliers that
share similar raw-materials, with different availability and quality levels, the selection of a
supplier should be related with both travelled distance (cost savings) and quantity/quality
2 Chapter 1. Motivation and overview
of raw-materials obtained (revenue) (Yu et al., 2016). Thus, the selection of a supplier will
consider if the obtained revenue collected is enough to compensate the additional increase
in the total distance traveled by the fleet of vehicles. As an extension of the classic VRPB,
this problem is designated the Vehicle Routing Problem with Selective Backhauls (VRPSB)
(Baldacci et al., 2010). Despite the growing importance of transport optimization and
integrated planning, the literature still lacks models and methods that deal efficiently with
the transportation planning with backhauling while including sustainability issues (Bektaş
et al., 2019; Allaoui et al., 2019; Colicchia et al., 2013). In particular, the VRPB is not
well studied in the literature, although it has been advocated that backhauling provides a
suitable alternative for inducing more friendly-environmental transport operations.
Moreover, most of the optimization tools used in transportation planning have a deter-
ministic nature, disregarding uncertainties that may turn out to be very costly. Although
the uncertainty itself may not be seen as a sustainable issue, providing robust transporta-
tion plans that are immune to uncertain events can indirectly impact the sustainability of
transportation. For example, a routing plan that accounts for possible delays and changes
on demand would allow the vehicles to improve their efficiency of use (e.g., travel lower
distances), instead of reacting to the uncertainty after the routing plan has been estab-
lished. The literature shows that there are two main approaches to deal with uncertainty -
stochastic programming and robust optimization (Adulyasak and Jaillet, 2016). Stochas-
tic programming is a well-known approach that allows to represent the uncertainty as a
probability distribution function and thus creating diverse scenarios with a given proba-
bility. Examples of stochastic programming include the two-stage stochastic optimization
and chance-constrained optimization. Robust optimization is a fairly recent approach to
deal with uncertainty when its probability distribution is not known or when only rough
information may be available (Grossmann et al., 2016). In this case, the uncertainty is rep-
resented by a bounded uncertainty set and a robust counterpart of the deterministic problem
is derived. Although robust optimization models may generate more conservative solutions
than stochastic programming, one of its main advantages is the tractability of the robust
counterpart, which, depending on the uncertainty set, may be similar to its deterministic
version. So far, no study has yet been carried regarding the VRPB under uncertainty (Koç
and Laporte, 2018). Moreover, very few works on VRPs have investigated uncertainties
other than demand and time, possibly because these parameters have a direct impact on
the service level and customer satisfaction. However, uncertainty in revenues, which are
crucial parameters to consider in the selection of suppliers, is not usually addressed in the
transportation planning.
Another setting in which backhauling can be effectively developed is on collaborative trans-
portation. Collaboration between companies allows to optimize their entire network, in-
creasing the efficiency of the vehicles and reducing empty trips (Audy et al., 2012). The
literature shows that most of the works on collaborative vehicle routing investigate col-
laborations between carriers or between shippers (Gansterer and Hartl, 2018). This type
of collaborative networks is known as horizontal collaboration, where the entities are at
the same level of the supply chain and perform similar transportation services. In op-
position, vertical collaboration refers to a collaborative network formed by companies at
different levels in a supply chain. Lateral collaboration allows to combine horizontal and
1.1. Research objectives and methodology 3
vertical collaboration, and involves decisions on similar and complementary transportation

services. In the context of collaborative vehicle routing, research on vertical and lateral
collaborations have not received much attention in the literature as the horizontal collab-
oration, despite its practical relevance in real world contexts (Gansterer and Hartl, 2018).
The impact of collaboration is usually measured in terms of gains obtained by comparing
the cost of the collaborative solution with the cost of the non-collaborative solution, and the
percentage difference is known as synergy value. Afterwards, these gains should be fairly
distributed among the participants in the collaboration, which implies solving a pricing
problem. The fairness criteria are usually based on properties derived from game theory,
such as the individual rationality, which ensures that each participant is allocated a cost that
is lower than its individual cost (i.e., when the participant does enter in the collaboration).
In this context, there is a vast literature on allocation methods to distribute collaborative
gains, ranging from simple proportional allocation methods to more complex methods that
require solving linear programs (Guajardo and Rönnqvist, 2016). Despite the growing re-
search on collaborative vehicle routing, the majority of the published works either focus on
solving the routing problem or on solving the profit sharing problem, but usually not both
simultaneously (Gansterer and Hartl, 2018).
In summary, the focus of this thesis is on proposing transportation planning models, en-
hancing the potential of backhauling while coping with uncertainty and sustainable objec-
tives, and promoting inter-firm collaboration. The models developed are applied in two
case-studies, allowing to demonstrate the real benefits of different backhauling strategies
and to derive a set of managerial insights that may support logistics decisions in different
supply chains.
1.1. Research objectives and methodology

This research aims to investigate mathematical programming models for the VRPB that
are capable of handling sustainability issues and uncertainty, by combining or extending
features of existing approaches.
To handle uncertainty, the concepts and modeling aspects of robust optimization will be
borrowed and adjusted to the problem under research. Robust optimization is an emergent
approach in VRP literature that allows to incorporate uncertainty in optimization models
without previous knowledge on the uncertainty profile, which happens in many real logis-
tics planning problems. The existing research relating VRP and robust optimization will be
analyzed, in order to identify how different uncertainties are being represented and modeled
and which decisions are required for different robust optimization approaches. Moreover,
the uncertainty in revenues and how these affect the routing problems and supplier selection
will be examined.
To handle sustainability issues, current literature on both green VRP and collaborative
VRP will be analyzed in order to identify the parameters that influence sustainable aspects
of transportation, as well as the requirements to form efficient collaborative networks. In
addition, different modelling approaches to tackle environmental concerns and collabo-
ration in transportation will be examined, with the goal of adapting such approaches to
the VRPB. Several backhauling strategies are analysed and tested with instances from the
literature and instances from case-studies.
The VRPB is a NP-hard combinatorial optimization problem, which means that the prob-
lem cannot be solved in polynomial time. Different solutions methods are investigated in
order to efficiently solve the different VRPB models that include uncertainty and/or sustain-
ability issues, namely exact methods (e.g., Branch-and-Cut method), metaheuristics (e.g.,
Adaptive Large neighborhood Search algorithm) and matheuristics (e.g., fix-and-optimize
algorithm). Very few exact methods were developed so far for the VRPB, because they are
not usually efficient to deal with the hard combinatorial nature of the VRPB. Metaheuris-
tics, on the other hand, are the state-of-art methods for the VRPB, but the solutions obtained
should be associated to a lower bound, which is usually obtained with an exact method. A
matheuristic, a hybrid solution method that combines exact algorithms and metaheuristics,
is also build in this thesis using a fix-and-optimize approach.
The scope of the research questions and their relationship is presented in Figure 1.1. Re-
search question 0 (RQ0) provides drivers for all the other research questions. Research
question 1 (RQ1) is related with research question 2 (RQ2), since they both explore the
revenues in the settings of the VRPSB. Research question 3 (RQ3) is related with RQ2
since they both investigate new exact algorithms to solve variants of the VRPB. Research
question 4 (RQ4) is related to RQ3, as they both examine collaborative networks and apply
properties from game theory to solve the profit sharing problem. Finally, RQ4 and RQ1
relate to each other since they are investigated in real case studies.
Figure 1.1 – Research questions framework.

1.1. Research objectives and methodology 5
Research Question 0
What is the role of the Vehicle Routing Problem with Backhauls in terms of sustainability?
This research question is motivated by the potential of the VRPB to be applied in real
logistics problems and to cope with economic, environmental and social concerns, despite
that traditional optimization approaches are mostly driven by economic goals only.
To answer this research question, an extensive review on the VRPB literature with particu-
lar focus on sustainability issues is performed. The main trends in the overall transportation
planning problems are identified and compared with the current VRPB research, in order
to determine interesting research lines for this type of problems, which are further tackled
in this thesis.
Research Question 1
How can transportation with backhauling be enriched for real world contexts?
This research question is investigated with the goal of evaluating new strategies to address
the transportation planning with backhauling other than the typical VRPB, while copping
with real-life operational constraints. In order to provide an efficient solution method that
is able to solve these rich versions of the VRPB in reasonable computation time, a fix-and-
optimize algorithm is developed, taking advantage of the hybridization of exact methods
and metaheuristics.
The answer to this research question is the application of a case-study in the forest supply
chain, which assesses the impact of different backhauling strategies, as well as the impact
of different inputs of the problem.
Research Question 2
How to address uncertainty in the Vehicle Routing Problem with Backhauls through robust
optimization?
This research question is motivated by, on the one hand, the nonexistence of any VRPB
study under uncertainty, and, on the other hand, by the increasing popularity of robust opti-
mization among optimization problems for which the probability distribution of uncertain
parameters is not available. Moreover, uncertainty in revenues is recognized as a major
factor to consider in the supplier selection, but few works in the literature have considered
this source of uncertainty.
The outcome of this research question will be a well defined model to represent the robust
VRPB and different solution methods to solve it. The methods to be developed include
exact methods and metaheuristics. The trade-off between robustness and impact on the ob-
jective function of the nominal problem will be examined. The results will allow to draw
conclusions about robust optimization and its impact on the robustness, efficiency and cost
of the transportation planning problem with backhauling under uncertainty.
Research Question 3
How to efficiently model and solve a collaborative Vehicle Routing Problem with Back-
hauls?
This research question is motivated by the increasing awareness of the potential of the
VRPB to cope with sustainable supply chains and collaborative networks. The VRPB can
reduce empty trips, which represent a large portion of the total routing costs, and allows a
better use of the vehicle’s capacity, which impacts considerably on the energy consumption
and efficiency.
Answering this question will result in a newly mathematical formulation based on bilevel
optimization that tackles the vertical collaboration between a shipper and a carrier, while
ensuring the hierarchical decisions and the conflicting objectives among different stake-
holders. Analysis and discussion of the results will allow to compare the new approach
with typical collaborative models and evaluate the synergy effects between the two players.
Research Question 4
How to address the challenges of a practical sustainable collaborative Vehicle Routing
Problem with Backhauls?
This research question is motivated by three main research gaps in the literature. The first
regards the lack of studies on lateral collaboration for the collaborative vehicle routing,
which allows to combine horizontal and vertical decisions. The second regards the lack
of collaborative networks that incorporate other than economic concerns, such as fuel con-
sumption or GHG emissions. The third regards the lack of works that handle both the
optimization of the routing problem and the fair distribution of collaborative gains among
the players.
The final output of this research question is the report of a case-study in a food supply
chain evaluating the potential gains from collaboration with respect to different objectives
(economic an environmental related) and the analysis of different methods to allocate the
gains among the players. The results of this case-study will allow the definition of a set of
managerial insights to implement collaboration in practice.
1.2. Thesis structure and synopsis 7
1.2. Thesis structure and synopsis

This thesis is composed of a collection of papers, and each chapter (or paper) is aligned
with each research question previously presented. This section provides an overview of the
chapters that compose this thesis.
Chapter 2 provides an extensive literature review on the VRPB works, with particular focus
on the sustainability issues of transportation planning, aiming to answer to Research Ques-
tion 0. Since its first introduction in the literature, the VRPB has been commonly used to
plan routes under an umbrella of economic objectives, such as minimizing distances, costs,
time and number of vehicles. With the growing interest by the scientific community and
companies on developing and implementing sustainable strategies for transportation, new
aspects derived from practical requirements, and environmental and social concerns have
been driving the emergence of alternative modelling approaches and practices for more
sustainable transport operations. In this chapter, we classify all reviewed papers according
to a common taxonomy for routing problems and contextualize the VRPB literature under
the domains of Rich VRP, sustainable goals, collaborative networks and reverse logistics.
In this chapter, we also demonstrate that are still very different avenues to explore with
the goal of bringing the VRPB closer to the theme of sustainable transportation in supply
chains. Therefore, we provide a set of guidelines for future research in this context.
Chapter 3 examines and compares three different approaches to deal with the transportation
planning of inbound and outbound routes in real contexts, addressing Research Question
1. The first approach applies the traditional decoupled planning, where the inbound and
outbound routes are optimized separately. The second approach plans the inbound and out-
bound routes simultaneously, allowing the integration of inbound trips after outbound trips.
The third approach is an opportunistic variant of the second where the primary concern is
the optimization of outbound routes and only if it is cost-effective, inbound routes may
be integrated. The underlying problem considers practical aspects from real operations
in the wood-based industry, namely heterogeneous fleet of vehicles, multiple depots and
open routes, split deliveries and limitation on maximum distances. Moreover, the problem
considers the variable quality degree of raw-materials available at the suppliers, which are
translated into different revenues. Hence, the problem is modelled as a Rich VRPB mini-
mizing fixed and variable transportation costs and maximizing the total revenues collected.
A fix-and-optimize algorithm is developed to solve the problem, combining features of
exact and approximate methods. This work is applied to a case-study in the wood-based
industry, on which a set of mills aims to plan optimal routes to deliver wood panels to
their customers and collect raw-materials from suppliers. A set of managerial insights with
application to the wood-based industry, but possible to extent to other supply chains (e.g.,
retailing), are derived based on a series of sensitivity analysis to different parameters of the
problem and from the comparison of the different planning approaches.
Chapter 4 investigates a robust optimization approach to address the VRPB under uncer-
tainty, and thus answers Research Question 2. In chapter 3, the revenues collected at back-
haul customers are deterministic. In this chapter, the revenues are considered uncertain,
and no knowledge on the probability distribution is available. The VRPB with uncertain
revenues is modelled following a robust optimization approach in the literature, where
the space of uncertain revenues is defined by an uncertainty set and a parameter, known
as budget of uncertainty, controls the size of that space. The problem consists on plan-
ning inbound-outbound routes such that all linehaul customers are visited and a minimum
amount of total revenue is collected at backhaul customers. The goal is to minimize the
total travelling costs minus the sum of expected revenues. Two different solutions methods
are proposed to solve the problem: a Branch-and-Cut algorithm and an Adaptive Large
Neighborhood Search (ALNS) metaheuristic. Robust solutions are evaluated in terms of
probability of constraint violation and price of robustness. The probability of constraint
violation refers to the probability of a generated solution failing to attend the minimum
revenue expected, which can be related to the budget of uncertainty. Three different meth-
ods from existing literature and one novel method proposed in this work are used and
compared to estimate the probability of constraint violation. The price of robustness refers
to the percentage cost difference between a robust solution and a deterministic solution. In
this context, the solutions of the robust optimization model are compared with solutions
obtained with a chance-constraint model. The performance, advantages and limitations of
the models and methods used in this work are presented.
Chapter 5 presents a novel formulation to model a collaborative VRPB, and thus addresses
Research Question 3. The collaborative VRPB considers a vertical collaboration between a
shipper and a carrier, where hierarchical decisions are taken by both entities but their goals
are different. The new formulation of the problem consists of a bilevel optimization model
where the upper level covers the problem of the shipper and the lower level covers the
problem of the carrier. In the upper level, the shipper decides the minimum cost delivery
routes and the incentives to offer to the carrier to perform pickups at different backhaul
customers. At the lower level, the carrier decides the incentives to accept and on which
delivery routes are the backhaul customers inserted. In practice, the shipper first decides
the minimum cost routes and then offers a side payment to the carrier aiming to compensate
the additional travelling costs to perform additional backhauling. In this context, the main
novelty of the bilevel model is that it allows to solve simultaneously the routing and the
pricing (incentives) problems, instead of solving them sequentially. The bilevel problem
is solved with an exact reformulation method and this approach is compared with two
different approaches based on side payments. The impact of collaboration is measured
through the synergy value, and the advantages and limitations of the bilevel formulation
are presented.
Chapter 6 addresses the collaborative VRPB from an environmental perspective, tackling
Research Question 4. The transportation network considers different entities (a retailer, a
3PL and several suppliers), which can perform different transportation related activities,
such as pickups and deliveries in backhaul routes and cross-docking in intermediary fa-
cilities. Due to the existence of diverse entities and strategies, the problem addresses the
case of lateral collaboration, which is not commonly investigated in the context of col-
laborative vehicle routing. The collaborative problem is formulated with three different
objective functions: a pure economic (minimize operational costs), a pure environmen-
tal (minimize fuel consumption) and a holistic function (minimize operational and CO2
emission costs). The impact of collaboration is then measured by the synergy value and
the solutions obtained with the three formulations are compared. Afterwards, three profit
Bibliography 9
sharing mechanisms are tested using concepts from proportional allocation methods, and
compared in terms of some fairness criteria. The problem is applied to a case-study in the
food supply chain and managerial insights are derived from the results.
Finally, chapter 7 summarizes the main contributions of this thesis, providing answers to
the research questions raised and suggestions for future work.
Bibliography
Adulyasak, Y. and Jaillet, P. (2016). Models and algorithms for stochastic and robust
vehicle routing with deadlines. Transportation Science, 50(2):608–626.
Allaoui, H., Guo, Y., and Sarkis, J. (2019). Decision support for collaboration planning in
sustainable supply chains. Journal of Cleaner Production, 229:761 – 774.
Audy, J.-F., Lehoux, N., D’Amours, S., and Rönnqvist, M. (2012). A framework for an
efficient implementation of logistics collaborations. International Transactions in Oper-
ational Research, 19(5):633–657.
Baldacci, R., Bartolini, E., and Laporte, G. (2010). Some applications of the generalized
vehicle routing problem. Journal of the Operational Research Society, 61(7):1072–1077.
Bektaş, T., Ehmke, J. F., Psaraftis, H. N., and Puchinger, J. (2019). The role of operational
research in green freight transportation. European Journal of Operational Research,
274(3):807 – 823.
Colicchia, C., Marchet, G., Melacini, M., and Perotti, S. (2013). Building environmental
sustainability: empirical evidence from logistics service providers. Journal of Cleaner
Production, 59:197 – 209.
Crainic, T. G. and Laporte, G. (1997). Planning models for freight transportation. European
Journal of Operational Research, 97(3):409 – 438.
Dantzig, G. B. and Ramser, J. H. (1959). The truck dispatching problem. Management

Science, 6(1):80–91.
Deif, I. and Bodin, L. D. (1984). Extension of the clarke and wright algorithm for solv-
ing the vehicle routing problem with backhauling. In Didder, A. (Ed.), Proceedings of
the Babson Conference on Software Uses in Transportation and Logistic Management,
Babson Park, pages 75–96.
Demir, E., Bektaş, T., and Laporte, G. (2014). The bi-objective pollution-routing problem.
European Journal of Operational Research, 232(3):464–478.
Evangelista, P., Colicchia, C., and Creazza, A. (2017). Is environmental sustainability a

strategic priority for logistics service providers? Journal of Environmental Management,
198:353–362.
Gansterer, M. and Hartl, R. F. (2018). Collaborative vehicle routing: A survey. European

Journal of Operational Research, 268(1):1–12.
Grossmann, I. E., Apap, R. M., Calfa, B. A., García-Herreros, P., and Zhang, Q. (2016).
Recent advances in mathematical programming techniques for the optimization of pro-
cess systems under uncertainty. Computers & Chemical Engineering, 91:3–14.
Guajardo, M. and Rönnqvist, M. (2016). A review on cost allocation methods in collabora-

tive transportation. International Transactions in Operational Research, 23(3):371–392.
Koç, Ç. and Laporte, G. (2018). Vehicle routing with backhauls: Review and research
perspectives. Computers & Operations Research, 91:79–91.
Lin, C., Choy, K., Ho, G., Chung, S., and Lam, H. (2014). Survey of green vehicle routing
problem: Past and future trends. Expert Systems with Applications, 41(4):1118–1138.
Mancini, S. (2016). A real-life multi depot multi period vehicle routing problem with a het-
erogeneous fleet: Formulation and adaptive large neighborhood search based matheuris-
tic. Transportation Research Part C: Emerging Technologies, 70:100–112.
Toth, P. and Vigo, D. (1999). A heuristic algorithm for the symmetric and asymmetric
vehicle routing problems with backhauls. European Journal of Operational Research,
113(3):528–543.
Ubeda, S., Arcelus, F., and Faulin, J. (2011). Green logistics at eroski: A case study.
International Journal of Production Economics, 131(1):44–51.
Yu, F., Xue, L., Sun, C., and Zhang, C. (2016). Product transportation distance based
supplier selection in sustainable supply chain network. Journal of Cleaner Production,
137:29 – 39.
Chapter 2
Vehicle Routing Problem with

Backhauls
The vehicle routing problem with backhauls towards

a sustainability perspective - a review
Maria João Santos ∗† · Pedro Amorim∗† · Alexandra marques† · Ana Carvalho‡
· Ana Póvoa‡
Published in TOP, 2019
Abstract The Vehicle Routing Problem with Backhauls (VRPB) allows to integrate in-
bound and outbound routes, which is an efficient strategy to reduce routing costs and also
to reduce the environmental and social impacts of transportation. In this paper, we ana-
lyze the VRPB literature with a sustainability perspective, which covers environmental and
social objectives, collaborative networks and reverse logistics. First, in order to better un-
derstand and analyze the VRPB literature, all related works are characterized according to
a common taxonomy provided for routing problems. This taxonomy is extended to differ-
entiate between economic, environmental and social objectives. After identification of all
VRPB papers that include sustainability issues, these are analyzed and discussed in more
detail. The analysis reveals that research on VRPBs with sustainability concerns is recent
and relatively scarce and the most popular aspects investigated are the minimization of fuel
consumption and CO2 emissions. Future research lines driven by sustainability concerns
are suggested for the VRPB as a promoter of green logistics.
Keywords vehicle routing problem · backhauling · sustainability · collaboration · reverse

logistics
2.1. Introduction
In the European Union, transportation is the sector with the fastest growing energy con-
sumption and greenhouse gas (GHG) emissions (Oberhofer and Dieplinger, 2014). Road
∗ Faculdade de Engenharia, Universidade do Porto, Porto, Portugal
† INESC TEC - Instituto de Engenharia de Sistemas e Computadores, Tecnologia e Ciência, Porto, Portugal
‡ Centro de Estudos de Gestão (CEG-IST), Instituto Superior Técnico (IST), Lisboa, Portugal
12 Chapter 2. Vehicle Routing Problem with Backhauls
transportation still accounts for about 18% of GHG emissions and yields 25% of empty
returns (Juan et al., 2014). For these reasons, transportation can be seen as one of the most
promising sectors regarding the development and the implementation of environmental
strategies capable of mitigating the environmental impact and of reducing costs, enforcing
therefore the sustainability of supply chains.
In this context, an optimal transportation planning should be used to design one or more
routes that allow meeting all customers’ needs, covering the entire network, using the full
capacity of the vehicles and avoiding empty return trips, thus reducing costs and mitigating
the environmental impact related with the transportation operations. The classical Vehicle
Routing Problem (VRP) is one of the most famous transportation planning problem and
deals with the distribution of goods from a single depot to a set of customers using vehicles
with limited capacity. The problem assumes that a customer’s demand must be satisfied
in a single visit, that the fleet of vehicles is homogeneous (with similar capacities) and
that each route must start and end at the depot. The goal is to determine optimal routes,
minimizing the total cost (usually in terms of distance travelled) (Solomon, 1987).
In an attempt to introduce real-life aspects into transportation problems, several new vari-
ants of the VRP have been created (see Toth and Vigo (2002b); Golden et al. (2008);
Toth and Vigo (2014)). These add significant complexity to the resulting problems. Such
real-life aspects concern vehicle characteristics (e.g., size, capacity, weight), drivers (e.g.,
licenses, resting period, maximum working hours), customers (e.g., time windows) and
products (e.g., compatibility between products and vehicles), among other aspects (Mancini,
2016). Eksioglu et al. (2009) proposed a methodology to classify a VRP based on the char-
acteristics introduced in the problem, which was latter updated by Braekers et al. (2016).
This taxonomy is used to classify the routing problems reviewed in this paper and extended
to classify the objective functions in three dimensions, namely economic, environmental
and social.
The Vehicle Routing Problem with Backhauls (VRPB) is a variant of the VRP that focus
on the efficiency of the vehicle temporal utilization. The problem considers two different
sets of customers: the set of linehaul customers (locations that receive loads from a depot)
and the set of backhaul customers (locations where loads are collected and sent to a depot).
The goal of the VRPB is to optimize cost-effective routing plans that integrate linehaul and
backhaul customers. Although the goal is still minimizing costs, the VRPB can contribute
substantially to reduce empty running of vehicles and fuel consumption, which in turn
contributes to reduce the environmental impact of transportation (Pradenas et al., 2013).
The first review on the VRPB literature was carried out by Toth and Vigo (2002a), which
presents an overview of the exact and heuristics methods to solve the classic VRPB until
the beginning of the century. The authors describe in detail the mathematical formulations,
relaxations, and heuristics algorithms that are the foundation for many solutions methods
used today. A second review can be found in Parragh et al. (2008a), which presents a
common classification of the main variants of the problem and reviews the research carried
until 2008. Papers considering reverse logistics are cited, but outside the scope of that
review. The authors generally agree that the future direction of VRPB research will entail
the incorporation of more detailed real-life aspects into the problem, as well as the effects
of dynamism and uncertainty. Later on, Toth and Vigo (2014) have revived the attention
2.1. Introduction 13
to several variants of the VRP, namely the VRPB. The authors followed up the previous
reviews and discussed the new contributions on this topic, namely in terms of new solutions
methods and extensions of the problem. In their conclusions, they highlighted that exact
methods for the VRPB should be updated to the current state-of-art algorithms. More
recently, Koç and Laporte (2018) has updated the literature on VRPB research. The authors
provide a comprehensive review covering models, solutions methods, new variants and
applications of the problem. It is concluded that the research towards the VRPB has been
increasing but there is room for improvement, particularly focused on modelling aspects
(e.g., continuous approximations, stochastic parameters) and solution methods (e.g., hybrid
algorithms combining population-based and local search metaheuristics).
None of these reviews, however, provide any suggestions on how to enhance the benefits of
the backhauling strategy in the VRPB, nor they provide examples of the VRPB under more
sustainable contexts, such as reverse logistics or collaborative networks. Reverse logistics
relates to the VRPB since the vehicles used to transport goods to customers are also used
to carry returned products back to the depot (e.g. damaged or non-conform products, used
items to recycle). For instance, the retailer Tesco uses backhauling for the returned prod-
ucts collected at their stores. The returned products are collected in a secure cage in each
store and when they are full, they are transported by the delivery vehicles that supply that
store to the distribution centre (DS Smith, 2017). Collaborative networks can be related to
the VRPB whenever a carrier uses its delivery vehicles to perform additional services in
their backhaul routes, reducing thus the empty trips. In fact, collaboration in routing prob-
lems aims to increase the efficiency of vehicles, by reducing empty backhauls and road
congestion, and by enhancing capacities (Audy et al., 2012; Gansterer and Hartl, 2018).
For example, Nestlé and United Biscuits, competitors in the food market, have arranged
a collaboration system to improve their logistics operations. This system involves coor-
dinating the collection of loads of each other company after delivering its own requests.
This backhauling strategy allowed to reduce empty running from 22% to 13% in four years
(Catherine Early, 2011). In addition, some relevant works that model a VRPB with envi-
ronmental concerns are missing from the review of Koç and Laporte (2018), such as the
one in Ubeda et al. (2011) who first propose a VRPB with such features.
This paper aims to present the research on VRPBs with sustainable concerns, by analyzing
papers that introduce environmental and/or social objectives, and papers that consider the
benefits of backhauling in different contexts, such as in collaborative networks or reverse
logistics. This paper complements the review of Koç and Laporte (2018), allowing to show
in detail and to a further extension the developments, scope and future directions for the
VRPB. Nonetheless, this paper has distinct contributions from that of Koç and Laporte
(2018). Particularly, the focus of this review is not on the solutions methods or models
that have been developed for the VRPB, but rather on how the sustainable impacts of the
VRPB have been analyzed and evaluated. In addition, a future research agenda driven by
sustainability concerns is proposed. The three main contributions of this paper are:
- Present a clear classification and comparison of all VRPB works carried up to the
moment, based on a common taxonomy for routing problems (Braekers et al., 2016)
that is extended to differentiate between economic, environmental and social objec-
tives;
- Review relevant works that investigate the impact of the VRPB in sustainable con-
texts, such as reverse logistics and collaboration;
- Provide insights on how to leverage the sustainable potential of the VRPB.
The paper is organized as follows. Section 2.2 describes the VRPB and presents its main
variants. Section 2.3 defines the methodology applied in this research. Section 2.4 focuses
on the VRPB literature overview, which provides a classification and a comparison of pa-
pers according to a common taxonomy. Section 2.5 analyses the relevant VRPB literature
that focuses on sustainable goals, reverse logistics and collaborative networks. Section 2.6
suggests research lines for the future development of more sustainable VRPB works, based
on the findings of previous sections. Section 2.7 concludes this paper.
2.2. The VRPB and its variants

The VRPB was firstly introduced in literature by Deif and Bodin (1984) as an extension
of the VRP that includes two types of customers: linehaul customers are those who re-
ceive goods from a depot (forward flow - outbound) and backhaul customers are those who
send goods back to the depot (backward flow - inbound). The goal is to create the most
cost-effective routes that simultaneously satisfy the demand of linehaul customers and col-
lect the pickup orders at the backhaul customers. The constraints of the VRPB are the
following:
1. Each vehicle travels exactly one route;
2. Each vehicle starts and ends at the depot;
3. Total demand of linehaul and backhaul customers, considered separately, cannot ex-
ceed the total vehicle capacity;
4. Each customer (either linehaul or backhaul) is visited exactly once;
5. All linehaul customers are visited before backhaul customers (precedence constraint);
6. No routes with only backhaul customers are allowed, but routes can contain only
linehaul customers.
The precedence constraint of the VRPB is a route condition that brings several benefits
from an economic and practical perspective, particularly for the manufacturing industry.
Firstly, the vehicles are often rear-loaded and the load is arranged in such a way that reflects
the sequence of deliveries to linehaul customers. Therefore, the precedence constraint
avoids problems that emerge from rearranging the vehicle loads at the delivery points (Toth
and Vigo, 1996, 2002a). Secondly, linehaul customers have higher service priority than
backhaul customers, which is in line with the constraint that allows designing routes with
linehaul costumers only but not routes with backhaul costumers only (Toth and Vigo, 1996,
2.3. Methodology 15
2002a). It is also worth mentioning that, for many industries where the supply needs are
usually satisfied with a single vehicle trip, i.e. the load to be supplied matches the full
capacity of the vehicle, the VRPB seems to be the most suitable variant to employ in such
instances.
The VRPB has also been called Vehicle Routing Problem with Clustered Backhauls (VR-
PCB) by Parragh et al. (2008a). Following the classification of Parragh et al. (2008a),
the main variants of the VRPB are: the VRP with mixed backhauls (VRPMB), the VRP
with divisible delivery and pickup (VRPDDP) and the VRP with simultaneous pickup and
delivery (VRPSPD). The VRPB and the VRPBM establish that customers must be either
linehaul or backhaul, never both, while the VRPDDP and the VRPSPD assume that each
customer requires a delivery and a pickup. These variants are briefly described below.
In the VRPMB, the precedence constraint is not considered; instead, mixing visits between
linehaul and backhaul customers are explicitly allowed. Consequently, the vehicle load
may decrease or increase during the route, depending on whether the load is delivered or
picked up. This load fluctuation requires checking the vehicle capacity regularly, which
increases the complexity of the problem (Mosheiov, 1994). Early works on the VRPMB
include those of Casco et al. (1988), Mosheiov (1998) and Salhi and Nagy (1999). The
VRPMB can be efficiently applied to situations where multiple loads can fit on the truck,
as is the case in food companies (Oesterle and Bauernhansl, 2016).
In the VRPSPD, each customer is associated to both a load to deliver and a load to pick
up. The delivery and pickup occur simultaneously in a single visit. This procedure may
decrease customer expenses and inconvenience, but can also increase the route duration
(Ropke and Pisinger, 2006). The first work addressing the VRPSPD is presented in Min
(1989). Also, the VRPSPD is highly linked to reverse logistics. For instance, in the soft
drink industry, the vehicle has to deliver the product to the customers and, at the same
time, pick up the empty bottles to recycle and transform them into new products (Privé
et al., 2006).
In the VRPDDP, costumers also require both a delivery and a pickup, but vehicles are
allowed to visit a customer more than once. In fact, two visits are required, one for load
delivery and another for load pickup. One of the first works addressing this variant belongs
Gribkovskaia et al. (2001). When a customer faces two visits, they are referred to as a
“split” customer (Nagy et al., 2013).
A variant not specifically classified by Parragh et al. (2008a) assumes that backhauls are
optionally visited based on a revenue they generate. In opposition to the other variants, the
VRPSB aims to maximize the total revenues (or total profits, after discounting the costs).
This problem is referred in literature as the VRP with deliveries and selective pickups
(Gribkovskaia et al., 2008), the VRP with unrestricted backhauls (Süral and Bookbinder,
2003) or the VRP with Selective Backhauls (Baldacci et al., 2010).
2.3. Methodology
The methodology applied in this work follows a methodological approach similar to the
literature review followed by content analysis described in Seuring et al. (2005). This
is based on the following four steps: material collection, descriptive analytics, category
selection and material evaluation.
Material collection Setting clear boundaries for a literature review is necessary in order
to limit the research (Seuring et al., 2005). For the present VRPB review, the following
considerations apply:
- All the works reviewed in Koç and Laporte (2018) are considered in this paper. The
published VRPB papers, as well as their variants, presented in Parragh et al. (2008a)
and in Toth and Vigo (2014) are also considered. Moreover, an additional search is
carried but limited to scientific papers published in international journals since 2000.
- The additional search was limited to the keywords “Vehicle Routing Problem” and
“Backhaul”. Papers that were found with those keywords, but investigate Pickup and
Delivery Problems (PDPs) were not collected. In the PDP, the linehaul customers are
supplied by the backhaul customers instead of the depot (Parragh et al., 2008b).
- Three library databases were used to search for VRPB papers, namely ScienceDirect
(www.sciencedirect.com), Scopus (www.scopus.com) and Google scholar
(www.scholar.google.com).
Overall, 107 papers were selected - 84 papers already cited in the previous reviews and
23 papers from the additional searching process. We anticipate that most of these papers
does not consider sustainability issues, for which these papers are only included for the
purpose of classification of the VRPB literature (Section 2.4). The papers that consider
sustainability issues are analyzed and discussed in further sections.
Descriptive analysis A preliminary analysis of all papers was undertaken, according to

publication year, journal and number of citations. The aim of this descriptive analysis is to
evaluate the interest of the scientific community in VRPB-related research.
Category selection The categories are selected in order to classify each work and fa-
cilitate the comparison of the VRPB literature. This set of categories are based on the
taxonomy proposed by Braekers et al. (2016), with the following adaptations: i) some cat-
egories were not used, as they were adding no value to the analysis when considering the
purpose of this work; ii) the classification of backhauls and node/arc covering constraints
were included in the same criterion, and iii) the objective function distinguishes between
economic, environmental and social dimensions. Table 2.1 provides the categories selected
from Braekers et al. (2016), marked in bold.
The reasons that motivated the investigation of these specific categories are as follows.
- The trade-off between solution quality and computing time determines the use of
exact or approximate solution methods. In addition, the purpose of the research, for
example, comparing backhauling strategies or solving a large-scale problem, also
influences the selection of the solution method. Thus, it is crucial to analyze the
diversity of methods applied to the VRPB.
2.3. Methodology 17
- The characterization of the problem’s scenario and physical network is important for
describing the real context of the problem and for analyzing the accuracy provided
by the model regarding the problem representation (Caceres-Cruz et al., 2014).
- Transportation activities are often subjected to uncertain events, e.g., traffic and
weather conditions, vehicle failure and newly assigned orders, which can jeopardize
the success of a predefined routing plan. Since research devoted to the VRP under
uncertainty has received increasing attention recently (Grossmann et al., 2016), it is
interesting to analyze the current position of the VRPB in the context of uncertainty.
- The increasing interest on promoting more sustainable transportation has been creat-
ing opportunity to design new VRP models that enable to tackle environmental and
social concerns (Lin et al., 2014). The dimensions covered by the objective(s) of a
VRPB are, in fact, the most important criteria to analyze in this work.
Material evaluation All papers selected for the review were classified according to the
set of categories defined above. Note that the focus of the present paper is not an exhaustive
literature review, since this work was already carried in previous reviews, but rather on the
analysis of VRPB works that address sustainability concerns. In this context, we designate
by "sustainable VRPBs" all the works that include: i) environmental and/or social objec-
tive functions, ii) collaboration, and iii) reverse logistics. The reasons that motivated the
analysis of sustainable VRPBs are as follows.
- Measuring the externalities of transportation (impact of transport operations on the

environment (e.g., GHG emissions) and on the society (e.g., noise pollution) is be-
coming more and more important in the context of sustainable supply chains (Demir
et al., 2015). In addition, green variants of the VRP have been developed in literature
in order to model energy minimization (e.g., Kara et al. (2007)), pollution minimiza-
tion (e.g., Bektaş and Laporte (2011)) or alternative fuel vehicles (e.g., Erdoğan and
Miller-Hooks (2012)).
- Collaboration is a sustainable strategy currently used by many companies in the

logistics sector that aim to increase the efficiency of vehicle use and reduce costs
(Evangelista et al., 2017). Collaboration enables the participants in the transporta-
tion network to detect backhauling opportunities that otherwise are not possible. For
this reason, collaborative networks can be modelled as VRPBs.
- Reverse logistics is also a hot topic in the sustainable supply chains that aims to im-
prove the management, planning and control of all backward flows of products (Lin
et al., 2014). Seen as a closed-loop routing problem, the VRPB can be analogous to a
reverse logistics problem, if the same vehicles are used in the forward and backward
flows.
Table 2.1 – Categories selected from Braekers et al. (2016) for the VRPB classification
1.1 Theory
1.2 Applied methods
1. Type of Study 1.3 Implementation documented
1.4 Survey, review or meta-research
2.1 Number of stops on route
2.2 Load splitting constraint
2.3 Customer service demand quantity
2.4 Request times of new customers
2. Scenario Characteristics 2.5 Onsite service/waiting times
2.6 Time window structure
2.7 Time horizon
2.8 Backhauls
2.9 Node/Arc covering constraints
3.1 Transportation network design
3.2 Location of addresses (customers)
3.3 Number of points of origin
3.4 Number of depot points
3.5 Time window type
3. Problem Physical Characteristics 3.6 Number of vehicles
3.7 Capacity consideration
3.8 Vehicle homogeneity (Capacity)
3.9 Travel time
3.10 Objective
4.1 Evolution of information
4.2 Quality of information
4. Information Characteristics 4.3 Availability of information
4.4 Processing of information
5. Data Characteristics
2.4. Literature overview 19
2.4. Literature overview

In this section, an overview of the VRPB literature is presented, covering a general de-
scriptive analysis of the published papers and the respective classification of all 107 works,
according to the taxonomy provided in Table 2.1. The section concludes with the main
remarks of the literature and the identification of the sustainable VRPB works selected for
the analysis in the next section.
2.4.1 Descriptive analysis

A first glimpse at the research and application potential of the VRPB can be described by
the distribution of the body of literature published in the analyzed period. In Figure 2.1,
this distribution is shown and it can be observed that VRPB research has been effectively
growing, particularly since the beginning of the century. Note also that the two papers
found for the year 2019 only consider the first month of the year.
Figure 2.1 – Distribution of published papers per year.
It can be seen in Figure 2.2 that the journal where more papers were published was the
European Journal of Operational Research, followed by the Expert Systems with Appli-
cations. The first journal aims to contribute to the development of operational research
methodologies and best practices for decision making in order to solve real-world prob-
lems, while the second gives more emphasis to the design, testing and implementation of
intelligent systems, as well as to the practical guidelines and management of such sys-
tems. Nevertheless, several papers were published in journals that focus on the theoretical
and methodological aspects of the studies, such as the Computers & Operations Research,
the Journal of Operational Research Society, or the Transportation Science. The category
"Others" groups all journals which appear only once or twice.
To perform a ranking of VRPB works cited in the literature, Google Scholar Citations was
used to collect the number of citations of each VRPB paper. From Figure 2.3, it can be
seen that the most cited paper belongs to Nagy and Salhi (2005), which is one of the first
Figure 2.2 – Distribution of published papers per journal.
works extending the VRPMB and VRPSPD to the multi-depot scenario. The following
most cited papers, namely Dethloff (2001), Min (1989) and Montané and Galvão (2006)
propose well-know instances benchmarks for the VRPSPD. Concerning the classic VRPB,
the most cited paper belongs to Ropke and Pisinger (2006), which develops one of the most
efficient and flexible solution methods to solve a VRPB, an Adaptive Large Neighborhood
Search (ALNS). The second most cited work is that from Ubeda et al. (2011), which is the
first to investigate the potential of the VRPB to reduce CO2 emissions. The third and final
most cited VRPB paper is the one from Goetschalckx and Jacobs-Blecha (1989), which
defines one of the most used benchmark instances for the standard problem.
Figure 2.3 – Top 10 most cited VRPB papers.

2.4.2 Classification of VRPB works

Figure 2.4 presents the number of VRPB works classified according to the categories de-
picted in Table 2.1. A detailed classification of each work is reported in Appendix 2.A. The
first category classifies the solution method in exact algorithm, heuristic or metaheuristic
(2.4a). The second category classifies the backhauling strategy according to the variant
investigated (2.4b). Note that this category combines the categories 2.8 Backhauls and 2.9
Arc/node covering constraints reported in Table 2.1. The third category classifies the sce-
nario characteristics accordingly if the problem includes load split, service/waiting time,
time-windows and multi-periodic issues (2.4c). The fourth category classifies the physi-
cal characteristics accordingly if it includes multi-depot network and heterogeneous fleet
(2.4d). The fifth category classifies the type of objective function of the VRPB according
to the dimension it covers: economic, environmental and social (2.4e). This classification
differs from the one in Braekers et al. (2016) because they do not provide such separa-
tion of the sustainable dimensions. The sixth and final category classifies the quality of
information in a single aspect - if uncertainty is addressed in the VRPB (2.4f).
The classification of the VRPB works shows clear evidences of the following: i) exact
methods are the least investigated solution methods for the VRPB, whereas metaheuristics
are the most popular ones; ii) the majority of the problems is formulated as the classic
VRPB with precedence constraint; iii) time windows, with or without consideration of
service times, is the most used scenario characteristic (after the classic VRPB); iv) hetero-
geneous fleet and multi-depot VRPBs have been smoothly increasing in the last years; v)
the vast majority of the VRPB literature only considers economic objectives; and vi) no
work has yet addressed a VRPB under uncertainty. Complementary information of the lit-
erature review can be found in Appendices 2.B and 2.C, which report the main benchmark
instances used in the VRPB literature and case studies and applications of the VRPB and
its variants, respectively.
Since the previous literature reviews already present a description of 84 out of the 107
VRPB works selected for this study, only the remaining 23 works are presented below.
From these, only those with pure economic objectives are presented here, since a thorough
analysis of sustainable VRPB works is carried out in Section 2.5.
2.4.2.1 Exact methods

A branch-and-price algorithm is applied in Gutiérrez-Jarpa et al. (2010) to solve a VRPSB
with time windows where the goal is to minimize the routing costs minus the revenue
collected with pickups. The solution method developed in this work can be seen as an
extension of the exact algorithm of Dell’Amico et al. (2006), although a first branch-and-
price algorithm is described in Angelelli and Mansini (2002) to solve a VRPSPD with time-
windows. The problem is formulated as an integer linear programming with minimization
of total net costs (routing costs minus revenues collected with pickups). The instances used
were based on the work of Solomon (1987), but only instances with 50 customers or less
were optimally solved.
(a) Solution method (b) Type of variant
(c) Scenario characteristics (d) Physical characteristics
(e) Objective function (f) Quality of information
Figure 2.4 – Number of VRPB papers classified according to each category.

Davis et al. (2014) developed a two phase exact solution method to solve a weekly trans-
portation schedule that involved collecting food donations from local sources and to deliver
them to a food bank and then to deliver the food from the bank to charitable agencies. In
the first phase, the problem is formulated as a capacitated set covering problem (CSCP),
solving the assignment of food banks to local sources. In the second phase, the VRPB is
formulated as an extension of the Miller-Tucker-Zemlin formulation, solving the routing
and scheduling problem for each workday.
Recently, Granada-Echeverri et al. (2019) proposes a mixed integer programming formula-
tion for a VRPB, which is build based on the characteristics of an Open VRP. The linehaul
and backhaul routes are two subproblems of the VRPB solved as independent Open VRPs
and tie-arcs connecting these routes are used to generate a solution for the VRPB. This
compact formulation is demonstrated to be competitive with the two state-of-art exact so-
lution methods proposed in Toth and Vigo (1997) and in Mingozzi et al. (1999), providing
12 new best known solutions for the VRPB instances of Goetschalckx and Jacobs-Blecha
(1989) and 8 new best known solutions for the instances of Toth and Vigo (1997).
2.4.2.2 Heuristics
A cluster-first, route-second heuristic is proposed in Kumar et al. (2011) to solve a VRPB
with tree networks. A tree network refers to a network covering a main highway used
by the vehicles and customers locations outside the highway. The heuristic first creates
different clusters of nodes (locations) and then determines a near-optimal route in each
cluster. Although a cluster-first, route-second was firstly proposed by Toth and Vigo (1996)
to solve a VRPB, the problem in this work takes the advantages of the tree structure, as
introduced by Labbé et al. (1991).
Another two-phase heuristic is applied to solve the VRPB with heterogeneous fleet and
split deliveries, while minimizing routing and handling costs in Lai et al. (2013). The first
phase determines a feasible solution, using the savings algorithm proposed in Clarke and
Wright (1964), and the second phase improves the solution through local search heuristics,
using two different neighbourhoods for the search space. The algorithm was applied to five
real instances with nearly 40 customers and two trucks with different capacities. The results
were compared with the current policy of the carrier, revealing significant reductions in the
total distance travelled, ranging from 2.89% to 12.61%.
A heuristic solution procedure based on Lagrangian relaxation was presented by Zhu et al.
(2010) for a capacitated plant location problem with customer and supplier matching (CLCSM),
where the VRPB is incorporated to reduce empty distances. The problem is formulated as
an integer linear programming aiming to minimize the total opening costs of plants and the
total cost of distance and number of vehicles. The heuristic first formulates a lower bound
based on the Lagrangian relaxation of the problem and decomposes it into sub-problems
that are as many as the number of plants. Based on the optimal relaxed solution, a feasible
solution is then created for the CLCSM. The algorithm is tested for randomly generated
instances and the results suggest it is suitable for producing efficient and robust solutions,
even for large instances with 400 customers, 60 suppliers and 40 plants.
Ghaziri and Osman (2003) presents a heuristic based on Kononen’s self-organizing feature
map (SOFM) to solve a single-vehicle VRPB The heuristic is based on a neural network
architecture that consists of two separate chains of neurons (one for linehauls and one for
backhauls). The interaction between the depot and these chains and the interaction between
neurons in the same chain lead to create a route, which is further improved with a 2-opt
procedure. The authors compare the performance of the heuristic with the branch-and-
bound algorithm developed by Fischetti and Toth (1992). The results provide that the exact
method is competitive for small instances (up to 50 customers), while the SOFM is better
suited for small to large-sized instances (e.g., with 1000 customers).
2.4.2.3 Metaheuristics
Metaheuristics can be divided into local search and population search metaheuristics (Hertz
and Widmer, 2003). A general difference between these two types of classes is that the for-
mer finds solutions by modifying and improve a single solution, while the latter performs
a similar task using multiple solutions.
Local search metaheuristics Among local search metaheuristics, Tabu Search (TS) al-
gorithms, or methods combining TS with other heuristics, are among the most popular
solution methods for the VRPB, since its first introduction by Duhamel et al. (1997). For
instance, Nguyen et al. (2016) proposed a TS algorithm that integrates multiple neighbour-
hoods, in order to solve the VRPB, with and without time windows, while considering the
minimization of both the total travel distance and the number of vehicles. The algorithm
was tested on VRPB instances from Goetschalckx and Jacobs-Blecha (1989) and Toth and
Vigo (1997), and on VRPBTW instances from Gélinas et al. (1995). The algorithm presents
good performance and can be competitive with the state-of-art metaheuristics, such as the
ALNS of Ropke and Pisinger (2006), the unified hybrid genetic algorithm of Vidal et al.
(2014) or the TS of Brandão (2006), among others. Lai et al. (2015), on the other hand,
proposes an adaptive guidance (AG) metaheuristic that combines TS with the savings algo-
rithm of Clarke and Wright (1964) to solve a VRPB with split deliveries. First, the VRPB
is decomposed into different VRPs, one for linehaul customers and one for backhaul cus-
tomers. Each problem is solved with TS and then the savings algorithm is used to merge the
sub-problems and create a global solution for the VRPB. The iterative process is build upon
an adaptive guidance mechanism that defines simple rules to assess the quality of a current
solution and detects potential improvements. The method applied in this problem enables
a reduction of about 25% in the total travel distance, compared with the current solution of
the carrier. A hybrid metaheuristic that combines TS and simulated annealing (SA) is pro-
posed by Küçükoğlu and Öztürk (2015) to solve a VRPB with heterogeneous fleet. First,
a nearest-neighbour heuristic is used to create an initial solution. Then, the TS is used to
generate neighborhoods and finally the SA is used to select a new solution from the neigh-
borhood, based on an acceptance criterion. The hybrid metaheuristic is tested with Gélinas
et al. (1995) instances with up to 100 nodes, for which 11 new best known solutions are
created. Recently, Reil et al. (2018) makes use of TS to study all four main variants of the
VRPB with time windows and three dimensional loading constraints. The solution method
follows a two-phase approach that solves first the packing problem and then the routing
problem. The packing problem is solved with a TS and the routing problem is solved se-
quentially by a multi-start evolutionary algorithm that minimizes the number of vehicles
and then by a TS to minimize the total travel distance. Several benchmark instances, ei-
ther from literature and generated at random, are used to evaluate the performance of the
metaheuristic. Results from this work indicate that the VRPMB can slightly increase the
number of vehicles required compared with the VRPB, when both time windows and three
dimensional loading constraints are considered.
A reconstruction algorithm is designed in Nikolakopoulos (2015) to solve a bi-level VRPBTW,
based on a theoretic game approach of a leader aiming to minimize the number of vehicles
and a follower aiming to minimize the total duration of the routes. The bilevel problem
is reduced to a single level by aggregating the two objective functions, such that the one
from the leader is prioritized over the one of the follower. An initial solution is created by a
construction heuristic and then six different heuristics are used to partially destroy the cur-
rent solutions. The improvement phase consists of a variation of the Threshold Accepting
method proposed by Dueck and Scheuer (1990) and the search intensification uses local
search heuristics. The reconstruction algorithm is shown to be competitive with state-of-
art metaheuristics for the minimization of number of vehicles, but does not compete for the
minimization of total duration of routes.
To solve the VRPBTW and the VRPMBTW, both with minimization of the number of vehi-
cles and the distance travelled, Tarantilis et al. (2013) propose the Adaptive Path Relinking
metaheuristic. A set of feasible solutions is generated and subsets of intermediate solutions
are subsequently created based on a path generation method. This method uses a procedure
to guide the intermediate solutions according to the recurrence of some of their attributes.
Optimal local solutions are selected from the intermediary solutions and improved by local
search algorithm. Tests on several VRPBTW and VRPMBTW instances are performed and
the metaheuristic was able to provide several new best solutions compared with state-of-art
methods. Further tests suggest that allowing mixed pickups and deliveries leads to less
costly routing plans than those obtaining by forcing pickups after deliveries.
Zachariadis et al. (2015) presented a local search algorithm to optimize both, VRPSPD
and VRPDDP, with two-dimensional loading constraints and different configurations of
loading, namely with fixed orientation, with allowed rotation and with or without LIFO
(last-in-first-out) constraints. The algorithm initiates with a feasible solution and a local
search procedure is used further to solve the routing problem. In turn, the routing algorithm
calls another algorithm to check loading feasibility. Tests on generated instances with
up to 150 customers are performed with all models and the results suggest that allowing
simultaneous pickup and delivery allow for significant reduction in the total costs, and
this benefit increases with the increasing of the contribution of backhaul customers to the
problem size.
Population-based metaheuristics Among population-based metaheuristics, evolution-

ary algorithms are the most popular ones to solve the VRPB. Ganesh and Narendran (2007)
uses a genetic algorithm (GA) to intensify the search for best solutions, which are obtained
using a three-phase constructive process. The construction of an initial solution starts by
clustering the nodes according to a proximity criteria and then using a shrink-wrap algo-
rithm (Sofge et al., 2002) to orient the nodes along a path, and thus by finding a route within
each cluster. Finally, each route is allocated to a vehicle by solving a generalized assign-
ment problem using a heuristic similar to Fisher and Jaikumar (1981). The GA uses a two-
point crossover operator and an insertion operator for mutation in the intensified search.
The metaheuristic is tested for VRPB instances of Goetschalckx and Jacobs-Blecha (1989)
and Toth and Vigo (1996), as well as for VRP benchmark instances in literature. The solu-
tions are obtained with reasonable time and with average gaps of, respectively, 0.06% and
0.42% from the best known solutions for the VRPB instances.
The performance of a memetic algorithm (MA) under different combinations of algorithm
components and parameter values is investigated in Saremi et al. (2007) to study a VRPB
with heterogeneous fleet. The MA combines EA to find global solutions and local search
algorithms to find local optimal solutions. All components of the MA are evaluated, namely
crossover operator and crossover rate, mutation operator, local search improvement method
and selection method. The MA is tested for VRPB instances from Toth and Vigo (1999),
distributed by three sets of increasing number of nodes and vehicles. For the first and
second sets, the component that affects most the solution quality is the mutation operator,
followed by the local search method, while the crossover operator is the most important for
the third set of experiments. However, the local search method is the most significant com-
ponent affecting the computational time. A differential evolution algorithm (DEA) is used
by Küçükoğlu and Öztürk (2014) to solve a VRPBTW with minimization of total distance.
The main distinction between the DEA and the GA is the type of operator used to build
better solutions, the former uses mutation and the latter uses crossover. The population
size, crossover ratio and mutation constant are evaluated for different parameter settings
and the effect of each parameter, as well as their interactions, are also examined. The DEA
is applied to five VRPBTW data sets from Gélinas et al. (1995), which found 23 new best
solutions out of 45. Applied to a real-problem of a catering firm, the DEA allowed to
reduce the total distance in 22.32%.
In Paraphantakul et al. (2012), ant colony optimization (ACO) is used to study a real-world
problem with time restrictions, where the objective is to minimize the total travel time.
The particularity of this problem compared with the classic VRPB is that routes composed
with only backhauls are allowed. The metaheuristic is adapted from Reimann et al. (2002)
and the results demonstrate that improvements in the range of 9.1% to 12.3% on the av-
erage route duration can be obtained with this metaheuristic, compared with the current
method of the company. Recently, Lu and Yang (2019) applied a ACO metaheuristic to
solve a multi-depot VRPSPD that minimizes total costs, covering drivers wages and fuel
costs. The authors investigate the performance of different heuristics for both, creating an
initial solution and improving current solutions, and compare the strategies of mixed and
sequential deliveries and pickups. The authors report an improvement rate of 15% in the
total costs for the case of mixed deliveries and pickups.
2.4.3 Final remarks on the VRPB

The classification of the 107 VRPB papers allows to conclude that there is great agreement
with the findings of Koç and Laporte (2018), namely in what respect the solution methods
used and the main goals pursued in the problems. Notwithstanding, three important aspects
must be highlighted from our review.
The first aspect concerns the backhauling strategy adopted in the problem. Although the
majority of the research is devoted to the classic VRPB, some works present a compari-
son between the classic and the mixed bakhauling strategies. When these two strategies
are investigated, the focus of the study is either on the development of solutions methods
capable of solving efficiently different VRPB variants (e.g., Ropke and Pisinger (2006);
Tarantilis et al. (2013); Reil et al. (2018)) and/or on the evaluation and comparison of the
impact of different backhauling strategies (e.g., Tarantilis et al. (2013); Nagy et al. (2013);
Turkensteen and Hasle (2017); Reil et al. (2018); Lu and Yang (2019)). However, which
backhauling strategy is better to cope with the goals of the decision-maker cannot be gen-
eralized to all problems. For instance, the mixing strategy is shown to be a better strategy
in the work of Tarantilis et al. (2013), but it is not an efficient option in the work of Reil
et al. (2018).
The second aspect concerns the nomenclature of a VRPB with several scenario and physi-
cal characteristics. A widespread trend for any variant of the VRP is the inclusion of more
practical aspects in the problem, aiming to bring the problem closer to reality of compa-
nies. This has motivated the emergence of a new class of problems denominated Rich
VRP. Caceres-Cruz et al. (2014) proposed a classification for a Rich VRP in three levels,
according to the degree of realism associated to the model. The first level is attributed
to classic VRP variants that borrow aspects of other variants (e.g., a VRPB with TW and
HF). The second level is attributed to classical-advanced VRPs, such integrated problems
(e.g., inventory-routing problems (IRP)) or multi-objective models (e.g., minimizing rout-
ing costs, GHG emissions and accident rate). The third level is attributed to VRPs that
are able to provide more accurate data through the incorporation of uncertainty and/or dy-
namism, for instance. For third level Rich VRP, hybrid methods such as matheuristics are
usually applied. Hereupon, it can be recognized that the majority of VRPB papers falls into
the first and second categories of the Rich VRP classification.
Finally, the third aspect, which drives the main purpose of the present paper, concerns the
goals of VRPBs. From Table 2.2, it becomes evident that some effort has been carried
in order to incorporate objectives other than economic in VRPBs, particularly in the last
years. As mentioned throughout this paper, the main goal of the present review is to analyze
sustainable VRPBs. Thus, the first group of VRPB papers selected for the analysis is those
that include, apart from the economic concerns, environmental and/or social aspects in
the objective function, such as in Ubeda et al. (2011), Eguia et al. (2013), Chávez et al.
(2016) and Turkensteen and Hasle (2017). The next group of papers addressing VRPB with
sustainability concerns are those that also address collaboration, for which three works are
identified: Bailey et al. (2011), Pradenas et al. (2013) and Juan et al. (2014). The last group
includes the VRPB works that are studied in the context of reverse logistics, for which six
papers are identified: Privé et al. (2006), Gribkovskaia et al. (2008), Rahimi et al. (2016)
Hoff et al. (2009), Nagy et al. (2013) and Soleimani et al. (2018). Overall, 13 sustainable
VRPB works are analyzed in the next section, as summarized in Figure 2.5.
Figure 2.5 – Number of VRPB works covering different dimensions of sustainability.
2.5. Analysis of VRPBs with sustainability concerns

The VRPB works that focus on sustainable aspects, namely by considering environmental
and/or social objectives, collaboration or reverse logistics, are described and discussed in
this section. We designate the problems tackled in such works as VRPBs with sustainability
concerns.
2.5.1 VRPBs with sustainable objectives

For the majority of VRPBs described in literature, the formulation of the problem still
focus on minimizing costs (usually the distance). Accordingly, as GHG emissions and fuel
consumption are proportional to the distance traveled, minimizing the distance indirectly
reduce both emissions and fuel. Nevertheless, most of the studies and practical applications
of the VRPB focus on the costs savings by performing inbound and outbound trips together
and neglect the quantitative assessment of the impact of transportation on the environment
and society.
The first study to address the environmental impact of a VRPB is described in Ubeda et al.
(2011). The authors also investigate the impact of different transport strategies, namely
re-scheduling deliveries, backhauling and green optimization, on the economic and envi-
ronmental costs and on the operation profitability. The first strategy relies on formulating a
VRP minimizing the total distance, the second strategy involves formulating a VRPB min-
imizing also total distance and the third strategy involves formulating the VRP with mini-
mization of total CO2 emissions. In all three strategies, the emissions are estimated based
on the distance and distance-based emission factors. The Mole and Jameson’s method is
2.5. Analysis of VRPBs with sustainability concerns 29
used to solve the VRPs and the Nearest Neighbor Insertion algorithm is used to solve the
VRPB. From the three strategies studied, backhauling provides the highest reduction of
the total distance traveled: more than 15%. The results of this case study establish that
i) backhauling increases the truck’s fill-rate and, consequently, the vehicles efficiency, and
ii) increasing the backhauling rate leads to an increase in both economic and ecological
benefits.
Chávez et al. (2016) formulate a multi-depot VRPB with minimization of the travel dis-
tance, the travel time and the total energy consumption. The GHG emissions and energy
consumption are computed based on the Pollution-Routing Problem (PRP) introduced by
Bektaş and Laporte (2011). In the PRP, the energy consumption is as a function of distance,
load and vehicle speed. Vehicle speed, in turns, is determined according to a series of road
and vehicle factors, such as road angle and friction coefficient, among others. The authors
propose a Pareto ant colony optimization (PACO) algorithm and test the algorithm on 33
adapted instances of Salhi and Nagy (1999) in order to find ordered Pareto solutions. The
resulting Pareto fronts suggest that energy and time have low correlation between them,
whereas distance travelled affects both, time and energy consumption.
A method to assess the carbon emission effects when using a classic VRP, a VRPB and
a VRPMB is presented in Turkensteen and Hasle (2017). The carbon emissions are a
function of distance driven and average load factor. The results point out that, comparing
with a classic VRP, the total distance is reduced in about 10-20% for the VRPB and in
about 20-40% for the VRPMB. The average load in the VRPB is very similar to that of the
VRP but it fluctuates considerably in the VRPMB. Moreover, savings in carbon emissions
up to 25% and up to 42% can also be reached for the VRPB and the VRPMB, respectively.
Although the VRPMB can induce higher benefits (lower costs and emissions), the time
to rearrange the load in the vehicle during mixed visits are not considered in this study.
Nevertheless, this extra time may lead to a drastic increase in the total time required to
complete a route.
Pradenas et al. (2013) and Juan et al. (2014) studied a VRPB with environmental objectives
in the context of collaborative networks. These works are analyzed in subsection 2.5.2.
The first work to address social objectives in a VRPB is presented in Eguia et al. (2013).
In fact, the problem considers all dimensions of sustainability but translates all parameters
in costs. Thus, the problem aims to minimize internal and external costs. Internal costs
refer to the total costs with fuel, drivers, vehicles and tolls. External costs comprise both
environmental (climate change and air pollution) and social impacts (noise and accident
rate). The problem considers different type of fuels and is solved with a Clarke and Wright
savings heuristic extended by the authors to include the ability to perform with a hetero-
geneous fleet. With a case study, the authors show that including backhauling is always a
better strategy than perform individual inbound and outbound routes. They also conclude
that considering social and environmental impacts results in choosing the less pollutant ve-
hicles, if an heterogeneous fleet is considered in the routing problem. Up to the moment,
only Rahimi et al. (2016) has also covered the social dimension of a VRPB, in the context
of reverse logistics. This work is analyzed in subsection 2.5.3.
2.5.2 VRPBs in collaborative networks

Collaborative vehicle routing can be defined as all forms of cooperation between the partic-
ipants in the network that aim to improve the efficiency of transport operations (Gansterer
and Hartl, 2018). Typical participants in a transportation network are the shippers (owners
of the load to be carried) and the carriers (owners of the vehicles). If the shipper owns self-
support vehicles, shipper and carrier are the same entity. Otherwise, the shipper outsources
a carrier to perform the transport operations.
Bailey et al. (2011) investigates the impact of a collaborative VRPB, when a carrier seeks
for collaborative shipments to reduce the empty distances and maximize savings. The
collaboration can occur with a shipper which offers a pickup-delivery task close to the
backhaul routes of interest of the carrier. The collaboration can also occur with another
carrier that does not have sufficient capacity to fulfill all its tasks. The authors developed
two models and two solution methods to solve a real-world freight network problem. The
first model allows at most one collaborative shipping per truck and is solved with a greedy
heuristic. The second model allows multiples collaborative shipments in each route and
is solved using a tabu search algorithm. The collaboration provides cost savings between
13% and 28% in backhaul routes compared with a non-collaborative network.
Besides reducing costs, the collaborative VRPB can also bring benefits from an environ-
mental perspective, as demonstrated by Juan et al. (2014) for collaborative shippers and by
Pradenas et al. (2013) for collaborative carriers.
In Juan et al. (2014), each shipper has a private fleet of vehicles, a unique depot and a
unique set of customers to serve. The authors compare a non-cooperative scenario, where
each shipper serves only its customers, and a cooperative scenario where a vehicle of one
shipper can perform a pickup and delivery tasks for another shipper, after serving its own
customers. The problem is modelled as a multi-depot VRP and solved with an Iterated
Local Search algorithm. With a numerical experiment, they conclude that a reduction of
about 24% in GHG emissions can be achieved with a collaborative network. Furthermore,
they determined that cost savings are highly influenced by the geographical dispersion
of the customers in the network but, at least, a 5% cost reduction can be achieved with
collaboration.
Pradenas et al. (2013) investigated the effect of three different collaboration strategies on
the costs and CO2 emissions of a transportation system. The first strategy - total compe-
tition, considers that each company performs in own deliveries (i.e. there is no collabora-
tion). The second strategy - total cooperation, assume that information and all resources are
shared among companies. The third strategy - mixed system, assume that information may
be exchanged but not the resources. They also propose a method to determine the profits
sharing between participants in the collaborative network, based on the popular Shapley
value of game theory (Shapley, 1953). The authors conclude that a decrease in 2% in GHG
emissions results in a cost increase between 2% and 8%. Nevertheless, a cooperative net-
work is able to bring economic profits for all companies, approximately 30% cost reduction
in the total transportation costs.
2.5.3 VRPBs in reverse logistics

Reverse logistics concerns the backward flows of products (e.g., returned products, end-
of-life products, waste) from its last use location to a final destination (e.g., manufacturer,
recycling company, waste treatment location).
When different vehicles must be used to perform independently the delivery and the col-
lection routes, the VRPRL cannot be seen as a VRPB or any of its variants. This occur
when the material to collect is unfitted to the delivery vehicle, such as in the case of waste
management (Ramos et al., 2014). On the other hand, when the same vehicle is used for
the delivery and the collection tasks, the VRPRL is analogous to a VRPSPD, if the delivery
and pickup occurs simultaneously (e.g., Privé et al. (2006)), or to a VRPDDP, if the tasks
occur in separate routes (e.g., Nagy et al. (2013)). Although most of the literature reviewed
in this paper concerning these VRPB variants can be linked to reverse logistics, this section
is devoted to those papers that specifically draw managerial insights for the application of
reverse logistics in practice. Consequently, papers focused on the development of solutions
methods are not described.
Privé et al. (2006) study a reverse logistics problem arising in the drink industry, where
the vehicles perform the deliveries of soft-drinks from a distribution center to retailers and
the collections of containers with empty cans and bottles to be returned to the distribution
center. Each recyclable container has an associated revenue and the delivery and collection
operations occur simultaneously. Three construction heuristics are developed to solve the
problem and the solutions are compared in terms of total costs, total revenues, percentage
use of vehicle capacity and number of vehicles required. Compared with the current routing
plan of the studied company, the proposed model and solutions methods are able to reduce
the total distance up to 23%. Gribkovskaia et al. (2008) investigate a similar problem
in reverse logistics, but considering that a revenue can only be collected if the delivery
and pickup occurs simultaneously. Otherwise, if two visits are necessary, no revenue is
collected. Note that both of the works presented can also be analogous to the VRPSB,
considering that the collection of recyclable material/containers is optional, based on the
revenue generated.
Hoff et al. (2009) investigate lasso solutions for the routing problem, which can be defined
as intermediate solutions between the VRPSPD and the VRPDDP. More precisely, a lasso
solution creates a route where a vehicle firstly delivers to one (or few) customers, then
performs a simultaneous delivery and pickup at the remaining customers, and after that, it
performs the pickups at the firstly visited customers while returning to the depot. A pioneer
work concerning lasso solutions can be found in Gribkovskaia et al. (2001). A load control
parameter β is introduced to set the free space of a vehicle before starting to collect empty
bottles and tested within the range [0.0, 1.0]. The decision-maker sets this parameter but
Casco et al. (1988) suggests a free space of 20%. Thus, with β = 0.0, the problem becomes
analogous to a VRPSPD; β = 1.0, the problem becomes analogous to a VRPDDP. They
applied a tabu search metaheuristic to solve several generated instances and demonstrated
the significant impact of the load control parameter on the solution costs. Although they
found out that, theoretically, solutions for the VRPSPD are less costly, lasso solutions with
a low value of β are preferable from a practical perspective.
The VRPDDP is investigated and compared with the VRPSPD in the study of Nagy et al.
(2013). The authors concluded that serving the customers twice instead of simultaneously
can reduce costs and the number of vehicles required, and such benefits are more evident
when there are considerable differences between delivery and collection quantities. An-
other important observation is that splitting delivery and pickup seems more advantageous
for customers located near the depot or when these have a higher pickup or delivery de-
mand.
Although the reverse logistics is itself a sustainable strategy, because is concerned with
reducing waste, recycling and circular economy, the optimization of the routing problem
still targets the cost reduction, as the works presented above. Notwithstanding, the litera-
ture seems to move towards more broader issues, namely by including environmental (e.g.,
Soleimani et al. (2018)) or social concerns (e.g., Rahimi et al. (2016)). Soleimani et al.
(2018) describe the problem of a newspaper company that uses a fleet of vehicles to deliver
newspapers to newsstands, while picking up unsold ones to return back to the distribution
center. The newsstands can also order an amount of old newspapers to the distribution
center. The problem is modelled as a non-linear multi-objective model, which is further
linearized and validated with a fuzzy approach. The objectives of the problem include the
minimization of total routing costs and total energy consumed. The CO2 emissions are a
function of the rate of fuel and considered as a constraint of the problem. Compared with
a separate planning of routes to deliver newspapers and routes to collect unsold ones, the
backhauling strategy allows a reduction of both costs (10.51%) and environmental impacts
(12.5%). Rahimi et al. (2016) describe an inventory-routing problem (IRP) for the distribu-
tion of perishable products to a set of retailers and the collection of expired products from
those retailers. They formulate a bi-objective mathematical model considering both eco-
nomic and social concerns. The economic objective aims to maximize the expected profits
and is determined by the total revenues obtained with the delivery of products minus the
total routing and inventory costs. The social objective concerns the minimization of acci-
dent rate and quantity of expired products. The accident rate is modelled as a function of
vehicle speed and the quantity of expired products is modelled as a function of the expira-
tion date of each product. The control of vehicle noise and limitation on GHG emissions
are introduced in the model as constraints. The authors analyze the conflicting nature of
the objectives in the problem, concluding with their experiments that increasing the relative
importance of the social objective over the economic one (e.g., 27% increase), improves
the social issues in the inventory-routing plan in 23% and reduces the total profits by about
17%. This is because increasing the number of vehicles used at a lower speed, reduces the
accident rate but increases the total routing costs.
2.5.4 Final remarks on the VRPBs with sustainability concerns

In this section, the literature on the VRPBs with sustainability concerns is firstly cross-
checked with the survey of Green VRP (GVRP), carried out by Lin et al. (2014), and then
with the survey of collaborative vehicle routing, carried out by Gansterer and Hartl (2018).
The survey of Lin et al. (2014) proposes a classification for the GVRP into three differ-
ent categories: i) the Green-VRP, ii) the Pollution Routing Problem, and iii) the VRP in
Reverse Logistics.
The first category covers the routing problems concerned with the energy consumption of
vehicles, and the primary goal is the minimization of fuel consumption, translated in a cost
function. The main factors that influence fuel consumption are the load, the distance trav-
eled and the speed of the vehicle. This category also includes the alternative-fuel powered
vehicles (AFV) and, as such, the risk of running out of fuel and the possibility of recharging
AFV en route are additional factors that must also be considered in this type of problems.
From the analysis of the VRPBs with sustainable concerns, the work of Eguia et al. (2013)
can be placed in the category of Green-VRP since it considers the energy consumption of
a heterogeneous fleet of vehicles.
The second category in the survey of Lin et al. (2014) covers the routing problem concerned
with the minimization of pollution emissions, in particular, CO2 emissions. Although a de-
crease in fuel consumption has an intimate relationship with a decrease in CO2 emissions,
additional factors must be considered in order to measure accurately the emissions, such
as those related to the engine of the vehicle and traffic conditions. The objectives of such
problems can also include the minimization of noise and congestion, among others. From
the analysis of the VRPBs with sustainable concerns, most of the works described in Sec-
tions 2.5.1 and 2.5.2 can be placed in the category of Pollution-Routing Problems.
Finally, the last category in the survey of Lin et al. (2014) covers the routing problems
applied in the context of reverse logistics operations. These can be further divided in i)
Selective Pickups with Pricing, when only the profitable collection sites are visited, ii)
Waste Collection, when the problem applies to the refuse collection services and waste
recycling, iii) End-of-life Goods Collection, for the pickups of components that may be
re-manufactured, and iv) Simultaneous Distribution and Collection, if the problem is anal-
ogous to a VRPSDP. From the analysis of the VRPBs with sustainable concerns, all the
works described in Section 2.5.3 can be placed in the category of VRP in Reverse Logis-
tics.
The survey of Gansterer and Hartl (2018) proposes a classification of collaborative VRPs
into i) centralized collaborative planning, ii) decentralized planning without auctions, and
iii) auction-based decentralized planning.
The centralized planning considers the joint optimization of the routing problem of all
participants. It assumes that all participants share all their information and, as such, the
problem reflects full collaboration between participants. The works of Juan et al. (2014)
and Pradenas et al. (2013) can be placed in this category.
The decentralized planning considers that only part of their information is shared among
participants. This information asymmetry may also reflect a hierarchy between participants
in the collaborative network. The work of Bailey et al. (2011) can be placed in this category,
since the studied carrier collaborates with other participants for its backhaul routes, but
its predefined delivery routes are not allowed to change. In other words, the carrier is
only willing to display information about its potential backhaul routes but does not share
information nor requests of the delivery routes to other participants (its delivery customers
are not considered in the collaborative network).
The auction-based decentralized planning assumes the existence of a bidding system, where
the participants submit their requests to a common pool, which are grouped into bundles,
and then offered to participants. The participants place their bids for the bundles, which
are then allocated to the best bid. From the review carried in the present paper, no VRPB
work falls in this category.
2.6. Future directions for sustainable VRPB

When planning routes according to a VRPB, an integrated inbound-outbound logistics
problem is considered. In the same route, the vehicle successively unloads all the re-
quested orders at the customers (outbound logistics) and, in the return trip to the depot,
the vehicle visits suppliers and load raw-materials until it reaches full capacity. As a re-
sult, it is possible to achieve significant savings by reducing the number of empty trips
of vehicles, the number of required vehicles and the total travel distance. This integrated
inbound-outbound perspective also highlights the sustainability side of the VRPB. Possible
reductions in the number of vehicles and in the distance they travel lead to consume less
fuel, generate less GHG and air pollutants and reduce noise pollution. This confirms the
strong relation between the VRPB and sustainability.
When evaluating the environmental impact in the VRPB, only the CO2 emissions have
been usually considered. This is due to the correlation of the CO2 emissions and fuel
consumption, as modelled in the PRP (Bektaş and Laporte, 2011). However, the emis-
sions of other pollutants, such as nitrogen oxides (NOx ), may depend on other parameters,
such as combustion dynamics or type of emission control, among others (Demir et al.,
2014). For instance, Naderipour and Alinaghian (2016) used the MEET model to estimate
the emissions of CO, CO2 and NOx for an Open VRP. The MEET model was developed
under an European project with the objective of providing a common procedure to evalu-
ate the impacts of transportation in the environment, covering methods, emissions factors
and functions to estimate the emissions of air pollutants. Therefore, in order to develop
more sustainable models for the VRPB, a possible direction may be to include specialized
functions in the mathematical model that estimate the individual emissions of different pol-
lutants. In this way, it is possible to evaluate and compare the effects of transportation on
local emissions (air pollutants, e.g., NOx , CO) and on global emissions (GHG, e.g., CO2 ).
The social concerns addressed in the VRPB literature are accident rate and noise, which
have direct impact on the well-being of the drivers. An interesting social aspect that could
be investigated for the VRPB is the equity in the working hours among drivers. Considering
that the wage is proportional to the working hours, routes with different duration, may be
seen unfair and, consequently, demotivate drivers. This in turns, may increase the accident
rate. For instance, Ramos et al. (2014) analyze the impact of balancing working hours
among drivers for a multi-objective multi-depot periodic VRP. Their results demonstrate
that minimizing the maximum working hours among drivers leads to increase both, the
costs and the CO2 emissions. Nevertheless, the authors propose a compromise solution
between economic, social and environmental objectives, and proof in a case study that it
has potential to save distance and fuel consumption and, simultaneously, promote equity in
routes duration.
The use of AFV in the VRP has been introduced in literature by Erdoğan and Miller-
2.6. Future directions for sustainable VRPB 35
Hooks (2012), which investigates the possibility of refueling a set of vehicles while en
route in order to avoid running out of fuel. From the literature review carried out in the
present paper, and also as highlighted by the review of Koç and Laporte (2018), no study
has yet investigated the use of AFV for a VRPB or its variants. The closest work may
be the one in Eguia et al. (2013), which considers different types of vehicles fueled by
diesel. Therefore, an important direction for the future work of sustainable VRPBs could
include the formulation of the VRPB with AFV. Additional research could cover the impact
of using different types of fuel (Ashtineh and Pishvaee, 2019) or the limited capacity of
alternative fueling stations (Bruglieri et al., 2019), as also suggested by Lin et al. (2014).
Given the importance of backhauling in collaborative networks, another future direction
for the VRPB may be to investigate in more detail the aspects of collaborative VRPB. An
example could be related to modelling a collaborative VRPB for which the goal pass be-
yond cost reduction, such as fairness in profit sharing between collaborators or balanced
distribution of drivers’ working hours. Another relevant study could be to investigate the
collaboration between a shipper and a carrier and at what extent is the VRPB profitable for
the entire network when compared with two different VRPs - one for the delivery routes
and one for the pickup routes. Furthermore, the different strategies of clustered, mixed or
optional backhauls could also be evaluated and compared for the collaborative VRPB. A
comparison between centralized and decentralized approaches for the collaborative VRPB
is also worthy of investigation. In this case, an analysis of the value of information shar-
ing could determine different collaborative strategies to be applied in real applications of
VRPB and its variants. Finally, departing from the work of Pradenas et al. (2013), which
determines the coalition savings in terms of costs and GHG emissions, the research on
collaborative VRPBs could be extended to determine coalition savings based on additional
environmental and social indicators.
The literature review on this paper shows that reverse logistics problems can be modelled
as a VRPSPD or a VRPDDP if the same vehicle is used to deliver products and pickup
returned items. Nevertheless, it would be very interesting to study a more broader VRPB
combined with reverse logistics. Particularly, the problem would consider a set of linehaul
customers and two sets of backhaul customers - one set of delivery customers that have
items to be returned (reverse logistics) and one set of suppliers of the depot (VRPB). In
this way, all possible routes can be effectively planned together and the company can lever-
age from better efficiency of inbound and outbound transportation needs, particularly if the
vehicles are self-supported. As mentioned in the previous section, several routing problems
tackling reverse logistics cannot be modeled as a VRPB or one of its variants, such as in
waste management. That occurs because the waste to be collected is not compatible with
the delivery items, or because the vehicles used for the collections have requirements not
supported by typical delivery vehicles. One possible way to overcome this drawback could
be to investigate multi-compartment vehicles in order to avoid cross-contamination of dif-
ferent products. Addressing multi-compartment vehicles and two- and three-dimensional
loading constraints for the waste collection problem is also suggested as a future research
direction in the review of Lin et al. (2014).
A major evidence from this review is that uncertainty was never addressed in a VRPB.
Chardy and Klopfenstein (2012) studied the impact of uncertainty inclusion in a VRP and
concluded that, while using the average scenario, which is the most used scenario in deci-
sion making, more than 15% of the tasks may not be concluded as previously estimated,
while when considering uncertainty this value falls to a maximum of 2.5%. Therefore,
another direction to promote more sustainable VRPBs could be to study the impact of un-
certainty inclusion and the development of adequate models. In fact, we highlight that
considerable attention should be given to this topic, since there are two different types of
customers in the VRPB, each one with distinct concerns and uncertainties. For example,
while the demand and time windows are major concerns for the success of delivery routes,
the success of backhaul routes may depend on the product availability or product quality at
suppliers, which may result in unmet demand for the depot or poor estimates of revenues
collected with pickup loads. The lack of stochastic routing problems in the literature is
highlighted in all reviews (Koç and Laporte, 2018; Lin et al., 2014; Gansterer and Hartl,
2018) and pointed as a future research direction to address more realistic cases of applica-
tions.
Stochastic programming and robust optimization are two popular approaches to deal with
uncertainty in VRPs (Averbakh, 2001). The main difference between them is that the for-
mer incorporates the uncertain parameter in the VRP formulation described by its prob-
ability distribution, while the latter represents the uncertain parameters as a bounded un-
certainty set, taking as input the worst realization of the uncertainty. For instance, the
VRP with uncertain travel times is modelled through stochastic programming in Li et al.
(2010), through robust optimization in Lee et al. (2012) and through a combined robust-
scenario approach in Han et al. (2014). In addition, comparing the performance of differ-
ent modelling approaches could be another line of research for the VRPB, since stochastic
programming may produce less costly solutions (Adulyasak and Jaillet, 2016) but robust
optimization may to provide less computational effort (Chen et al., 2016).
An additional suggestion for the development of more reliable and accurate models for
the VRPB derives from the survey of Ritzinger et al. (2016), which focuses on dynamic
and stochastic VRPs. In the last years, operations research has witnessed an increased in-
terest in dynamic routing problems due to, on the one hand, the increasing importance of
e-commerce and online transactions, and on the other hand, due to the exponential tech-
nological development of telematics systems that allows obtaining real-time data. Besides,
these telematics systems enable the collection of a large amount of data that could be further
processed to get useful statistics about stochastic information. Therefore, in order to keep
the VRPB research updated, the use of dynamic approaches and systems to collect real-time
information should be incorporated in the face of the traditional VRPB approaches. This
would lead not only to enrich VRPB models but also to increase the potential of solutions
methods to solve the related problems.
The future directions for the VRPB research and application are summarized in Figure 2.6.
2.7. Conclusions
The present analysis of literature demonstrated that backhauling is becoming an ever more
interesting option for solving real-world transportation problems and that it has the poten-
2.7. Conclusions 37
Figure 2.6 – Framework of future directions for the VRPB research.
tial to provide significant reductions in the total routing costs and increase sustainability of
transportation. Although the VRPB is commonly modelled as a cost minimization prob-
lem, some literature already extents the problem to include environmental and social ob-
jectives. Environmental objectives often include the minimization of CO2 emissions and
energy consumption. Social objectives include minimization of accident rate and control
of noise pollution. The analysis also reveal that the VRPB can well suit routing problems
in the context of reverse logistics and collaborative networks.
As a promoter of green logistics, possible future directions for the VRPB are: i) formu-
lation of sustainable VRPB models considering multi-objective functions with economic,
environmental and social issues; ii) investigation of additional and broader aspects in col-
laborative VRPBs, such as contracts and profits sharing among collaborators; iii) explore
the potential of different backhauling strategies when possible, such as trade-offs between
precedence and mixed backhauls or also between optional and restrictive backhauls; iv)
develop integrated models combining VRPB and reverse logistics, covering deliveries and
pickups at customers and pickups at suppliers; v) develop efficient models to incorporate
uncertainty and dynamism in VRPB models and measure the robustness of solutions.
This paper complements the reviews of Koç and Laporte (2018) and Parragh et al. (2008a)
on the models, solution methods, and applications of the VRPB, but particularly it extends
the literature to broader contexts, such as collaboration and reverse logistics. One inno-
vative aspect of this paper is the focus on the sustainable perspective and impacts of the
VRPB and, accordingly, the proposal of a future research agenda driven by sustainability
concerns.
Acknowledgements
This work is financed by the ERDF – European Regional Development Fund through
the Operational Programme for Competitiveness and Internationalisation - COMPETE
2020 Programme and by National Funds through the Portuguese funding agency, FCT
- Fundação para a Ciência e a Tecnologia within project POCI-01-0145-FEDER-016733
(Easyflow).
Bibliography
Adulyasak, Y. and Jaillet, P. (2016). Models and Algorithms for Stochastic and Robust
Vehicle Routing with Deadlines. Transportation Science, 50(2):608–626.
Anbuudayasankar, S., Ganesh, K., Koh, S. L., and Ducq, Y. (2012). Modified savings
heuristics and genetic algorithm for bi-objective vehicle routing problem with forced
backhauls. Expert Systems with Applications, 39(3):2296 – 2305.
Angelelli, E. and Mansini, R. (2002). The vehicle routing problem with time windows
and simultaneous pick-up and delivery. In Klose, A., Speranza, M. G., and Van Wassen-
hove, L. N., editors, Quantitative Approaches to Distribution Logistics and Supply Chain
Management, pages 249–267, Berlin, Heidelberg. Springer Berlin Heidelberg.
Anily, S. (1996). The vehicle-routing problem with delivery and back-haul options. Naval
Research Logistics, 43:415–434.
Ashtineh, H. and Pishvaee, M. S. (2019). Alternative fuel vehicle-routing problem: A life

cycle analysis of transportation fuels. Journal of Cleaner Production, 219:166 – 182.
Averbakh, I. (2001). On the complexity of a class of combinatorial optimization problems

with uncertainty. Mathematical Programming, 90(2):263–272.
Bailey, E., Unnikrishnan, A., and Lin, D.-Y. (2011). Models for Minimizing Backhaul
Costs through Freight Collaboration. Transportation Research Record: Journal of the
Transportation Research Board, 2224(1):51–60.
vehicle routing problem. Journal of the Operational Research Society, 61(7):1072–1077.
Bektaş, T. and Laporte, G. (2011). The Pollution-Routing Problem. Transportation Re-

search Part B: Methodological, 45(8):1232–1250.
Belloso, J., Juan, A. A., and Faulin, J. (2017a). An iterative biased-randomized heuristic
for the fleet size and mix vehicle-routing problem with backhauls. International Trans-
actions in Operational Research, 26(1):289–301.
Bibliography 39
Belloso, J., Juan, A. A., Faulin, J., and Serrano, A. (2015). Using multi-start biased ran-
domization of heuristics to solve the vehicle routing problem with clustered backhauls.
Lecture Notes in Management Science, 7:15–20.
Belloso, J., Juan, A. A., Martinez, E., and Faulin, J. (2017b). A biased-randomized meta-
heuristic for the vehicle routing problem with clustered and mixed backhauls. Networks,
69(3):241–255.
Belmecheri, F., Prins, C., Yalaoui, F., and Amodeo, L. (2013). Particle swarm optimization
algorithm for a vehicle routing problem with heterogeneous fleet, mixed backhauls, and
time windows. Journal of Intelligent Manufacturing, 24(4):775–789.
Berghida, M. and Boukra, A. (2016). Quantum Inspired Algorithm for a VRP with Hetero-
geneous Fleet Mixed Backhauls and Time Windows. International Journal of Applied
Metaheuristic Computing, 7(4):18–38.
Bianchessi, N. and Righini, G. (2007). Heuristic algorithms for the vehicle routing problem
with simultaneous pick-up and delivery. Computers Operations Research, 34(2):578 –
594. Reverse Logistics.
Bortfeldt, A., Hahn, T., Männel, D., and Mönch, L. (2015). Hybrid algorithms for the
vehicle routing problem with clustered backhauls and 3D loading constraints. European
Braekers, K., Ramaekers, K., and Van Nieuwenhuyse, I. (2016). The vehicle routing prob-
lem: State of the art classification and review. Computers & Industrial Engineering,
99:300–313.
Brandão, J. (2006). A new tabu search algorithm for the vehicle routing problem with
backhauls. European Journal of Operational Research, 173(2):540–555.
Brandão, J. (2016). A deterministic iterated local search algorithm for the vehicle routing
problem with backhauls. TOP, 24(2):445–465.
Bruglieri, M., Mancini, S., and Pisacane, O. (2019). The green vehicle routing problem
with capacitated alternative fuel stations. Computers Operations Research, 112:104759.
Caceres-Cruz, J., Arias, P., Guimarans, D., Riera, D., and Juan, A. A. (2014). Rich Vehicle
Routing Problem. ACM Computing Surveys, 47(2):1–28.
Casco, D. O., Golden, B. L., and Wasil, E. A. (1988). Vehicle routing with backhauls:
Models, algorithms, and case studies. In Golden, B. and A.A. Assad, e., editors, Vehicle
Routing: Methods and Studies, volume 16, page 127–147. North-Holland, Amsterdam.
Catherine Early (2011). Delivering greener logistics.
Chardy, M. and Klopfenstein, O. (2012). Handling uncertainties in vehicle routing prob-

lems through data preprocessing. Transportation Research Part E: Logistics and Trans-
portation Review, 48(3):667–683.
Chávez, J. J. S., Escobar, J. W., and Echeverri, M. G. (2016). A multi-objective Pareto ant
colony algorithm for the Multi-Depot Vehicle Routing problem with Backhauls. Inter-
national Journal of Industrial Engineering Computations, 7(1):35–48.
Chen, J.-F. and Wu, T.-H. (2006). Vehicle routing problem with simultaneous deliveries
and pickups. Journal of the Operational Research Society, 57(5):579–587.
Chen, L., Gendreau, M., Hà, M. H., and Langevin, A. (2016). A robust optimization
approach for the road network daily maintenance routing problem with uncertain service
time. Transportation Research Part E: Logistics and Transportation Review, 85:40–51.
Cheung, R. K. and Hang, D. D. (2003). Multi-attribute label matching algorithms for ve-
hicle routing problems with time windows and backhauls. IIE Transactions, 35(3):191–
205.
Cho, Y.-J. and Wang, S.-D. (2005). a Threshold Accepting Meta-Heuristic for the Vehicle
Routing Problem With Backhauls and Time Windows. Journal of the Eastern Asia
Society for Transportation Studies, 6:3022–3037.
Clarke, G. and Wright, J. W. (1964). Scheduling of Vehicles from a Central Depot to a

Number of Delivery Points. Operations Research, 12(4):568–581.
Crispim, J. and Brandão, J. (2005). Metaheuristics applied to mixed and simultaneous

extensions of vehicle routing problems with backhauls. Journal of the Operational Re-
search Society, 56:1296–1302.
Cuervo, D. P., Goos, P., Sörensen, K., and Arráiz, E. (2014). An iterated local search algo-
rithm for the vehicle routing problem with backhauls. European Journal of Operational
Research, 237(2):454–464.
Davis, L. B., Sengul, I., Ivy, J. S., Brock, L. G., and Miles, L. (2014). Scheduling food bank
collections and deliveries to ensure food safety and improve access. Socio-Economic
Planning Sciences, 48(3):175–188.
Deif, I. and Bodin, L. (1984). Extension of the clarke and wright algorithm for solving the
vehicle routing problem with backhauling. In Kidder, A., editor, Proceedings of the Bab-
son College Conference on Software Uses in Transportation and Logistic Management,
pages 75–96. Babson Park, MA.
Dell’Amico, M., Righini, G., and Salani, M. (2006). A branch-and-price approach to the
vehicle routing problem with simultaneous distribution and collection. Transportation
Science, 40(2):235–247.
Demir, E., Bektaş, T., and Laporte, G. (2014). The bi-objective Pollution-Routing Problem.
Demir, E., Huang, Y., Scholts, S., and Van Woensel, T. (2015). A selected review on the
negative externalities of the freight transportation: Modeling and pricing. Transportation
Research Part E: Logistics and Transportation Review, 77:95–114.
Bibliography 41
Derigs, U. and Metz, A. (1992). A matching-based approach for solving a delivery/pick-

up vehicle routing problem with time constraints. Operations-Research-Spektrum,
14(2):91–106.
Dethloff, J. (2001). Vehicle routing and reverse logistics: The vehicle routing problem with
simultaneous delivery and pick-up. OR-Spektrum, 23(1):79–96.
Dethloff, J. (2002). Relation between vehicle routing problems: an insertion heuristic

for the vehicle routing problem with simultaneous delivery and pick-up applied to the
vehicle routing problem with backhauls. Journal of the Operational Research Society,
53(1):115–118.
Dominguez, O., Guimarans, D., Juan, A. A., and de la Nuez, I. (2016). A Biased-
Randomised Large Neighbourhood Search for the two-dimensional Vehicle Routing
Problem with Backhauls. European Journal of Operational Research, 255(2):442–462.
DS Smith (2017). Tesco Food Waste Case Study.
Dueck, G. and Scheuer, T. (1990). Threshold accepting: A general purpose optimization

algorithm appearing superior to simulated annealing. Journal of Computational Physics,
90(1):161–175.
Duhamel, C., Potvin, J.-Y., and Rousseau, J.-M. (1997). A tabu search heuristic for the
vehicle routing problem with backhauls and time windows. Transportation Science,
31(1):49–59.
Eguia, I., Racero, J., Molina, J. C., and Guerrero, F. (2013). Environmental Issues in Vehicle
Routing Problems, pages 215–241. Springer Berlin Heidelberg, Berlin, Heidelberg.
Eksioglu, B., Vural, A. V., and Reisman, A. (2009). The vehicle routing problem: A
taxonomic review. Computers & Industrial Engineering, 57(4):1472–1483.
Erdoğan, S. and Miller-Hooks, E. (2012). A Green Vehicle Routing Problem. Transporta-

tion Research Part E: Logistics and Transportation Review, 48(1):100–114.

198:353–362.
Fischetti, M. and Toth, P. (1992). An additive bounding procedure for the asymmetric
travelling salesman problem. Mathematical Programming, 53(1):173–197.
Fisher, M. L. and Jaikumar, R. (1981). A generalized assignment heuristic for vehicle

routing. Networks, 11(2):109–124.
Gajpal, Y. and Abad, P. (2009). Multi-ant colony system (MACS) for a vehicle routing
problem with backhauls. European Journal of Operational Research, 196(1):102–117.
Ganesh, K. and Narendran, T. (2007). CLOVES: A cluster-and-search heuristic to solve

the vehicle routing problem with delivery and pick-up. European Journal of Operational
Research, 178(3):699–717.

García-Nájera, A., Bullinaria, J. A., and Gutiérrez-Andrade, M. A. (2015). An evolutionary

approach for multi-objective vehicle routing problems with backhauls. Computers &
Industrial Engineering, 81:90–108.
Gélinas, S., Desrochers, M., Desrosiers, J., and Solomon, M. M. (1995). A new branching
strategy for time constrained routing problems with application to backhauling. Annals
of Operations Research, 61(1):91–109.
Ghaziri, H. and Osman, I. H. (2003). A neural network algorithm for the traveling salesman
problem with backhauls. Computers & Industrial Engineering, 44(2):267–281.
Ghaziri, H. and Osman, I. H. (2006). Self-organizing feature maps for the vehicle routing
problem with backhauls. Journal of Scheduling, 9(2):97–114.
Goetschalckx, M. and Jacobs-Blecha, C. (1989). The vehicle routing problem with back-
hauls. European Journal of Operational Research, 42(1):39–51.
Golden, B. L., Raghavan, S., and Wasil, E. A., editors (2008). The Vehicle Rout-
ing Problem: Latest Advances and New Challenges, volume 43 of Operations Re-
search/Computer Science Interfaces Series. Springer US.
Granada-Echeverri, M., Toro, E. M., and Santa, J. J. (2019). A mixed integer linear pro-
gramming formulation for the vehicle routing problem with backhauls. International
Journal of Industrial Engineering Computations, 10:295–308.
Gribkovskaia, I., Halskau, , and Myklebost, K. (2001). Models for pick-up and deliveries
from depots with lasso solutions. In Stefansson, G. and Tilanus, B., editors, Proceedings
of the 13th Annual Conference on Logistics research NOFOMA 2001, page 279–293.
Chalmers University of Technology, Göteborg.
Gribkovskaia, I., Laporte, G., and Shyshou, A. (2008). The single vehicle routing problem
with deliveries and selective pickups. Computers & Operations Research, 35(9):2908–
2924.
Gutiérrez-Jarpa, G., Desaulniers, G., Laporte, G., and Marianov, V. (2010). A branch-and-
price algorithm for the Vehicle Routing Problem with Deliveries, Selective Pickups and
Time Windows. European Journal of Operational Research, 206(2):341–349.
Bibliography 43
Han, J., Lee, C., and Park, S. (2014). A Robust Scenario Approach for the Vehicle Routing
Problem with Uncertain Travel Times. Transportation Science, 48(3):373–390.
Hertz, A. and Widmer, M. (2003). Guidelines for the use of meta-heuristics in combinato-
rial optimization. European Journal of Operational Research, 151(2):247–252.
Hoff, A., Gribkovskaia, I., Laporte, G., and Løkketangen, A. (2009). Lasso solution strate-
gies for the vehicle routing problem with pickups and deliveries. European Journal of
Operational Research, 192(3):755–766.
Juan, A. A., Faulin, J., Pérez-Bernabeu, E., and Jozefowiez, N. (2014). Horizontal Co-
operation in Vehicle Routing Problems with Backhauling and Environmental Criteria.
Procedia - Social and Behavioral Sciences, 111:1133–1141.
Kara, , Kara, B. Y., and Yetis, M. K. (2007). Energy Minimizing Vehicle Routing Prob-
lem. In Combinatorial Optimization and Applications, pages 62–71. Springer Berlin
Heidelberg, Berlin, Heidelberg.
Katoh, N. and Yano, T. (2006). An approximation algorithm for the pickup and delivery
vehicle routing problem on trees. Discrete Applied Mathematics, 154(16):2335 – 2349.
Discrete Algorithms and Optimization, in Honor of Professor Toshihide Ibaraki at His
Retirement from Kyoto University.
Kontoravdis, G. and Bard, J. F. (1995). A grasp for the vehicle routing problem with time
windows. ORSA Journal on Computing, 7(1):10–23.
Küçükoğlu, and Öztürk, N. (2014). A differential evolution approach for the vehicle
routing problem with backhauls and time windows. Journal of Advanced Transportation,
48(8):942–956.
Küçükoğlu, and Öztürk, N. (2015). An advanced hybrid meta-heuristic algorithm for the
vehicle routing problem with backhauls and time windows. Computers & Industrial
Engineering, 86:60–68.
Kumar, R., Unnikrishnan, A., and Waller, S. T. (2011). Capacitated-Vehicle Routing Prob-
lem with Backhauls on Trees. Transportation Research Record: Journal of the Trans-
portation Research Board, 2263(1):92–102.
Labbé, M., Laporte, G., and Mercure, H. (1991). Capacitated vehicle routing on trees.
Operations Research, 39(4):616–622.
Lai, M., Battarra, M., Di Francesco, M., and Zuddas, P. (2015). An adaptive guidance meta-
heuristic for the vehicle routing problem with splits and clustered backhauls. Journal of
the Operational Research Society, 66(7):1222–1235.
Lai, M., Crainic, T. G., Di Francesco, M., and Zuddas, P. (2013). An heuristic search for the
routing of heterogeneous trucks with single and double container loads. Transportation
Lee, C., Lee, K., and Park, S. (2012). Robust vehicle routing problem with deadlines
and travel time/demand uncertainty. Journal of the Operational Research Society,
63(9):1294–1306.
Li, X., Tian, P., and Leung, S. C. (2010). Vehicle routing problems with time windows
and stochastic travel and service times: Models and algorithm. International Journal of
Production Economics, 125(1):137–145.
Lin, C., Choy, K., Ho, G., Chung, S., and Lam, H. (2014). Survey of Green Vehicle Routing
Problem: Past and future trends. Expert Systems with Applications, 41(4):1118–1138.
Lin, S., Bard, J. F., Jarrah, A. I., Zhang, X., and Novoa, L. J. (2017). Route design for
last-in, first-out deliveries with backhauling. Transportation Research Part C: Emerging
Technologies, 76:90–117.
Liu, S.-C. and Chung, C.-H. (2009). A heuristic method for the vehicle routing problem
with backhauls and inventory. Journal of Intelligent Manufacturing, 20(1):29–42.
Lu, E. H. C. and Yang, Y. W. (2019). A hybrid route planning approach for logistics with
pickup and delivery. Expert Systems with Applications, 118:482–492.
Mancini, S. (2016). A real-life Multi Depot Multi Period Vehicle Routing Problem with
a Heterogeneous Fleet: Formulation and Adaptive Large Neighborhood Search based
Matheuristic. Transportation Research Part C: Emerging Technologies, 70:100–112.
Min, H. (1989). The multiple vehicle routing problem with simultaneous delivery and
pick-up points. Transportation Research Part A: General, 23(5):377 – 386.
Mingozzi, A., Giorgi, S., and Baldacci, R. (1999). An Exact Method for the Vehicle Rout-
ing Problem with Backhauls. Transportation Science, 33(3):315–329.
Montané, F. A. T. and Galvão, R. D. (2006). A tabu search algorithm for the vehicle
routing problem with simultaneous pick-up and delivery service. Computers Operations
Research, 33(3):595 – 619.
Mosheiov, G. (1994). The travelling salesman problem with pick-up and delivery. Euro-
pean Journal of Operational Research, 79(2):299 – 310.
Mosheiov, G. (1998). Vehicle routing with pick-up and delivery: tour-partitioning heuris-
tics. Computers Industrial Engineering, 34(3):669 – 684.
Naderipour, M. and Alinaghian, M. (2016). Measurement, evaluation and minimization of

co2, nox, and co emissions in the open time dependent vehicle routing problem. Mea-
surement, 90:443 – 452.
Bibliography 45
Nagy, G. and Salhi, S. (2005). Heuristic algorithms for single and multiple depot vehi-
cle routing problems with pickups and deliveries. European Journal of Operational
Research, 162(1):126 – 141. Logistics: From Theory to Application.
Nagy, G., Wassan, N. A., Speranza, M. G., and Archetti, C. (2013). The Vehicle Routing
Problem with Divisible Deliveries and Pickups. Transportation Science, 49(2):271–294.
Nguyen, P. K., Crainic, T. G., and Toulouse, M. (2016). Multi-trip pickup and deliv-
ery problem with time windows and synchronization. Annals of Operations Research,
253(2):899–934.
Nikolakopoulos, A. (2015). A Metaheuristic Reconstruction Algorithm for Solving Bi-

level Vehicle Routing Problems with Backhauls for Army Rapid Fielding. In Zeim-
pekis, V., Kaimakamis, G., and Daras, N. J., editors, Military Logistics. Operations
Research/Computer Science Interfaces Series, volume 56, pages 141–157. Springer In-
ternational Publishing.
Oberhofer, P. and Dieplinger, M. (2014). Sustainability in the Transport and Logistics

Sector: Lacking Environmental Measures. Business Strategy and the Environment,
23(4):236–253.
Oesterle, J. and Bauernhansl, T. (2016). Exact Method for the Vehicle Routing Problem
with Mixed Linehaul and Backhaul Customers, Heterogeneous Fleet, time Window and
Manufacturing Capacity. Procedia CIRP, 41:573–578.
Osman, I. H. and Wassan, N. A. (2002). A reactive tabu search meta-heuristic for the
vehicle routing problem with back-hauls. Journal of Scheduling, 5(4):263–285.
Paraphantakul, C., Miller-Hooks, E., and Opasanon, S. (2012). Scheduling Deliveries with
Backhauls in Thailand’s Cement Industry. Transportation Research Record: Journal of
the Transportation Research Board, 2269(1):73–82.
Parragh, S. N., Doerner, K. F., and Hartl, R. F. (2008a). A survey on pickup and deliv-
ery problems Part I: Transportation between customers and depot. Journal für Betrieb-
swirtschaft, 58(1):21–51.
Parragh, S. N., Doerner, K. F., and Hartl, R. F. (2008b). A survey on pickup and delivery
problems Part II: Transportation between pickup and delivery locations. Journal für
Betriebswirtschaft, 58:81–117.
Potvin, J.-Y., Duhamel, C., and Guertin, F. (1996). A genetic algorithm for vehicle routing
with backhauling. Applied Intelligence, 6(4):345–355.
Pradenas, L., Oportus, B., and Parada, V. (2013). Mitigation of greenhouse gas emis-
sions in vehicle routing problems with backhauling. Expert Systems with Applications,
40(8):2985–2991.
Privé, J., Renaud, J., Boctor, F., and Laporte, G. (2006). Solving a vehicle-routing prob-
lem arising in soft-drink distribution. Journal of the Operational Research Society,
57(9):1045–1052.
Rahimi, M., Baboli, A., and Rekik, Y. (2016). Sustainable Inventory Routing Problem for
Perishable Products by Considering Reverse Logistic. IFAC-PapersOnLine, 49(12):949–
954.
Ramos, T. R. P., Gomes, M. I., and Barbosa-Póvoa, A. P. (2014). Planning a sustainable re-
verse logistics system: Balancing costs with environmental and social concerns. Omega,
48:60–74.
Reil, S., Bortfeldt, A., and Mönch, L. (2018). Heuristics for vehicle routing problems with
backhauls, time windows, and 3D loading constraints. European Journal of Operational
Research, 266(3):877–894.
Reimann, M., Doerner, K., and Hartl, R. F. (2002). Insertion based ants for vehicle routing
problems with backhauls and time windows. In Dorigo, M., Di Caro, G., and Sam-
pels, M., editors, Ant Algorithms, pages 135–148, Berlin, Heidelberg. Springer Berlin
Heidelberg.
Reimann, M. and Ulrich, H. (2006). Comparing backhauling strategies in vehicle routing

using Ant Colony Optimization. Central European Journal of Operations Research,
14(2):105–123.
Ritzinger, U., Puchinger, J., and Hartl, R. F. (2016). A survey on dynamic and stochastic
vehicle routing problems. International Journal of Production Research, 54(1):215–231.
Ropke, S. and Pisinger, D. (2006). A unified heuristic for a large class of Vehicle Routing
Problems with Backhauls. European Journal of Operational Research, 171(3):750–775.
Salhi, S. and Nagy, G. (1999). A cluster insertion heuristic for single and multiple depot ve-
hicle routing problems with backhauling. Journal of the Operational Research Society,
50(10):1034–1042.
Salhi, S., Wassan, N., and Hajarat, M. (2013). The Fleet Size and Mix Vehicle Routing
Problem with Backhauls: Formulation and Set Partitioning-based Heuristics. Trans-
portation Research Part E: Logistics and Transportation Review, 56:22–35.
Saremi, A., Elmekkawy, T. Y., and Wang, G. G. (2007). Tuning the Parameters of a
Memetic Algorithm to Solve Vehicle Routing Problem with Backhauls Using Design
of Experiments. International Journal of Operational Research, 4(4):206–219.
Seuring, S., Müller, M., Westhaus, M., and Morana, R. (2005). Conducting a Literature
Review — The Example of Sustainability in Supply Chains. In Research Methodologies
in Supply Chain Management, pages 91–106. Physica-Verlag, Heidelberg.
Shapley, L. S. (1953). A value for n-person games. Annals of Mathematics Studies, 28:307–
318.
Bibliography 47
Sofge, D., Schultz, A., and De Jong, K. (2002). Evolutionary computational approaches to
solving the multiple traveling salesman problem using a neighborhood attractor schema.
In Cagnoni, S., Gottlieb, J., Hart, E., Middendorf, M., and Raidl, G. R., editors, Appli-
cations of Evolutionary Computing, volume 2279, pages 153–162, Berlin, Heidelberg.
Springer Berlin Heidelberg.
Soleimani, H., Chaharlang, Y., and Ghaderi, H. (2018). Collection and distribution of
returned-remanufactured products in a vehicle routing problem with pickup and delivery
considering sustainable and green criteria. Journal of Cleaner Production, 172:960–970.
Solomon, M. M. (1987). Algorithms for the Vehicle Routing and Scheduling Problems
with Time Window Constraints. Operations Research, 35(2):254–265.
Subramanian, A., Drummond, L., Bentes, C., Ochi, L., and Farias, R. (2010). A par-
allel heuristic for the vehicle routing problem with simultaneous pickup and delivery.
Computers Operations Research, 37(11):1899 – 1911. Metaheuristics for Logistics and
Vehicle Routing.
Subramanian, A., Uchoa, E., and Ochi, L. S. (2013). A hybrid algorithm for a class of
vehicle routing problems. Computers Operations Research, 40(10):2519 – 2531.
Subramanian, A., Uchoa, E., Pessoa, A. A., and Ochi, L. S. (2011). Branch-and-cut with
lazy separation for the vehicle routing problem with simultaneous pickup and delivery.
Operations Research Letters, 39(5):338 – 341.
Subramanian, A., Uchoa, E., Pessoa, A. A., and Ochi, L. S. (2012). Branch-cut-and-price
for the vehicle routing problem with simultaneous pickup and delivery. Optimization
Letters, 7(7):1569–1581.
Süral, H. and Bookbinder, J. H. (2003). The single-vehicle routing problem with unre-
stricted backhauls. Networks, 41(3):127–136.
Tarantilis, C. D., Anagnostopoulou, A. K., and Repoussis, P. P. (2013). Adaptive Path

Relinking for Vehicle Routing and Scheduling Problems with Product Returns. Trans-
portation Science, 47(3):356–379.
Tavakkoli-Moghaddam, R., Saremi, A., and Ziaee, M. (2006). A memetic algorithm for
a vehicle routing problem with backhauls. Applied Mathematics and Computation,
181(2):1049–1060.
Thangiah, S. R., Potvin, J.-Y., and Sun, T. (1996). Heuristic approaches to vehicle routing
with backhauls and time windows. Computers Operations Research, 23(11):1043 –
1057.
Toth, P. and Vigo, D. (1996). A heuristic algorithm for the vehicle routing problem with
backhauls. In Bianco, L. and Toth, P., editors, Advanced Methods in Transportation
Analysis, pages 585–608, Berlin, Heidelberg. Springer Berlin Heidelberg.
Toth, P. and Vigo, D. (1997). An Exact Algorithm for the Vehicle Routing Problem with
Backhauls. Transportation Science, 31(4):372–385.
Toth, P. and Vigo, D. (1999). A heuristic algorithm for the symmetric and asymmetric
vehicle routing problems with backhauls. European Journal of Operational Research,
113(3):528–543.
Toth, P. and Vigo, D. (2002a). 8. vrp with backhauls. In Toth, P. and Vigo, D., editors, The
Vehicle Routing Problem, volume 9, pages 195–224. SIAM, Philadelphia.
Toth, P. and Vigo, D. (2002b). The Vehicle Routing Problem. SIAM monographs on
discrete mathematics and applications. Philadelphia: Society for Industrial and Applied
Mathematics.
Toth, P. and Vigo, D., editors (2014). Vehicle Routing: Problems, Methods, and Appli-
cations, Second Edition. Number 18 in MOS-SIAM Series on Optimization. SIAM,
Philadelphia.
Turkensteen, M. and Hasle, G. (2017). Combining pickups and deliveries in vehicle routing
– An assessment of carbon emission effects. Transportation Research Part C: Emerging
Tütüncü, G. Y. (2010). An interactive GRAMPS algorithm for the heterogeneous fixed fleet
vehicle routing problem with and without backhauls. European Journal of Operational
Research, 201(2):593–600.
Tütüncü, G. Y., Carreto, C., and Baker, B. (2009). A visual interactive approach to classical
and mixed vehicle routing problems with backhauls. Omega, 37(1):138–154.
Ubeda, S., Arcelus, F., and Faulin, J. (2011). Green logistics at Eroski: A case study.
International Journal of Production Economics, 131(1):44–51.
Vidal, T., Crainic, T. G., Gendreau, M., and Prins, C. (2014). A unified solution framework
for multi-attribute vehicle routing problems. European Journal of Operational Research,
234(3):658–673.
Wade, A. and Salhi, S. (2002). An investigation into a new class of vehicle routing problem
with backhauls. Omega, 30(6):479–487.
Wade, A. and Salhi, S. (2004). An Ant System Algorithm for the Mixed Vehicle Routing
Problem with Backhauls, pages 699–719. Springer US, Boston, MA.
Wang, Z. and Wang, Z. (2009). A novel two-phase heuristic method for vehicle routing
problem with backhauls. Computers & Mathematics with Applications, 57(11-12):1923–
1928.
Wassan, N. (2007). Reactive tabu adaptive memory programming search for the vehi-
cle routing problem with backhauls. Journal of the Operational Research Society,
58(12):1630–1641.
Bibliography 49
Wassan, N., Wassan, N., Nagy, G., and Salhi, S. (2017). The Multiple Trip Vehicle Rout-
ing Problem with Backhauls: Formulation and a Two-Level Variable Neighbourhood
Search. Computers & Operations Research, 78:454–467.
Wassan, N. A., Nagy, G., and Ahmadi, S. (2008a). A heuristic method for the vehicle
routing problem with mixed deliveries and pickups. Journal of Scheduling, 11(2):149–
161.
Wassan, N. A., Wassan, A. H., and Nagy, G. (2008b). A reactive tabu search algorithm
for the vehicle routing problem with simultaneous pickups and deliveries. Journal of
Combinatorial Optimization, 15(4):368–386.
Wu, W., Tian, Y., and Jin, T. (2016). A label based ant colony algorithm for heterogeneous
vehicle routing with mixed backhaul. Applied Soft Computing, 47:224–234.
Yalcın, G. D. and Erginel, N. (2015). Fuzzy multi-objective programming algorithm for ve-
hicle routing problems with backhauls. Expert Systems with Applications, 42(13):5632–
5644.
Yano, C. A., Chan, T. J., Richter, L. K., Cutler, T., Murty, K. G., and McGettigan, D.
(1987). Vehicle routing at quality stores. Interfaces, 17(2):52–63.
Yu, M. and Qi, X. (2014). A vehicle routing problem with multiple overlapped batches.
Transportation Research Part E: Logistics and Transportation Review, 61:40–55.
Zachariadis, E. E. and Kiranoudis, C. T. (2012). An effective local search approach

for the Vehicle Routing Problem with Backhauls. Expert Systems with Applications,
39(3):3174–3184.
Zachariadis, E. E., Tarantilis, C. D., and Kiranoudis, C. T. (2009). A hybrid metaheuris-

tic algorithm for the vehicle routing problem with simultaneous delivery and pick-up
service. Expert Systems with Applications, 36(2, Part 1):1070 – 1081.
Zachariadis, E. E., Tarantilis, C. D., and Kiranoudis, C. T. (2010). An adaptive memory

methodology for the vehicle routing problem with simultaneous pick-ups and deliveries.
European Journal of Operational Research, 202(2):401 – 411.
Zachariadis, E. E., Tarantilis, C. D., and Kiranoudis, C. T. (2015). Vehicle routing strategies
for pick-up and delivery service under two dimensional loading constraints. Operational
Research, 17(1):115–143.
Zhong, Y. and Cole, M. H. (2005). A vehicle routing problem with backhauls and time
windows: a guided local search solution. Transportation Research Part E: Logistics and
Transportation Review, 41(2):131–144.
Zhu, Z., Chu, F., and Sun, L. (2010). The capacitated plant location problem with cus-
tomers and suppliers matching. Transportation Research Part E: Logistics and Trans-
Çatay, B. (2010). A new saving-based ant algorithm for the vehicle routing problem with
simultaneous pickup and delivery. Expert Systems with Applications, 37(10):6809 –
6817.
2.A. Classification of VRPB works 51
Appendix 2.A Classification of VRPB works

52
Table 2.2 – Classification of VRPB works (1984-2004)
Solution methods Backhauls Scenario characteristics Physical characteristics Objective function Information
Reference E H MH Pr Mx PD LS SWT TW MP MD HF Ecn Env Scl Un
Deif and Bodin (1984)1,2,3 • • •
Yano et al. (1987)2,3 • • • •
Casco et al. (1988))1,2,3 • • •
Goetschalckx and Jacobs-Blecha (1989)1,2,3 • • •
Min (1989)1,2 • • •
Derigs and Metz (1992)1 • • • • • •
Gélinas et al. (1995)1 • • • •
Kontoravdis and Bard (1995)2 • • • •
Anily (1996)1,2 • • •
Potvin et al. (1996)1 • • • •
Thangiah et al. (1996)1,2,3 • • • •
Toth and Vigo (1996)1,2,3 • • •
Duhamel et al. (1997)1,2,3 • • • •
Toth and Vigo (1997)1,2,3 • • •
Mosheiov (1998)1,2
) • • •
Mingozzi et al. (1999)1,2,3 • • •
Salhi and Nagy (1999)1,2,3 • • • • •
Toth and Vigo (1999)1,2,3 • • •
Dethloff (2001)1,2 • • •
Angelelli and Mansini (2002)1,2 • • • •
Dethloff (2002)1 • • •
Osman and Wassan (2002)1,2,3 • • •
Wade and Salhi (2002)1,2,3 • • •
Cheung and Hang (2003)3 • • • • •
Ghaziri and Osman (2003) • • •
Chapter 2. Vehicle Routing Problem with Backhauls
Süral and Bookbinder (2003)1,3 • • •

Wade and Salhi (2004)1,2 • • •
Paper reviewed in: 1 Parragh et al. (2008a), 2 Toth and Vigo (2014), 3 Koç and Laporte (2018)
Legend: E - Exact, H - Heuristic, MH - Metaheuristic, Pr - Precedence, Mx - Mixed, PD - Pickup and delivery, LS - load split, SWT - Service and waiting time, TW - Time-windows, MP - multi period, MD -
Multi-depot, HF - Heterogeneous fleet, Ecn - Economic, Env - Environmental, Scl - Social, Un - Uncertainty
Table 2.2 (cont.) – Classification of VRPB works (2005-2009)
Cho and Wang (2005)3 • • • •
Crispim and Brandão (2005)1,2,3 • • • •
Nagy and Salhi (2005)1,2,3 • • • • •
Zhong and Cole (2005)1,2,3 • • • • •
Brandão (2006)1,2,3 • • •
Chen and Wu (2006)1,2 • • •
Dell’Amico et al. (2006)1,2 • • •
Ghaziri and Osman (2006)1,3 • • •
2.A. Classification of VRPB works
Katoh and Yano (2006)2 • • • •

Montané and Galvão (2006)1,2 • • •
Privé et al. (2006)2 • • • • • •
Reimann and Ulrich (2006)1,3 • • • • •
Ropke and Pisinger (2006)1,2,3 • • • • • • • •
Tavakkoli-Moghaddam et al. (2006)2,3 • • • •
Bianchessi and Righini (2007)1,2 • • •
Ganesh and Narendran (2007)1 • • •
Saremi et al. (2007) • • • •
Wassan (2007)2,3 • • •
Gribkovskaia et al. (2008)2 • • •
Wassan et al. (2008a)2 • • •
Wassan et al. (2008b)2,3 • • •
Gajpal and Abad (2009)2,3 • • •
Hoff et al. (2009)2,3 • • •
Liu and Chung (2009)3 • • •
53
Wang and Wang (2009)2,3 • • • • • •

Tütüncü et al. (2009)3 • • • •
Zachariadis et al. (2009)2 • • •
Legend: E - Exact, H - Heuristic, MH - Metaheuristic, Pr - Precedence, Mx - Mixed, PD - Pickup and delivery, LS - load split, SWT - Service and waiting time, TW - Time-windows, MP - multi period, MD
- Multi-depot, HF - Heterogeneous fleet, Ecn - Economic, Env - Environmental, Scl - Social, Un - Uncertainty
54
Çatay (2010)2 • • •
Gutiérrez-Jarpa et al. (2010) • • • • • •
Subramanian et al. (2010)2 • • •
Tütüncü (2010)2,3 • • • •
Zachariadis et al. (2010)2 • • •
Zhu et al. (2010) • • • •
Bailey et al. (2011) • • •
Kumar et al. (2011) • • •
Ubeda et al. (2011) • • • •
Anbuudayasankar et al. (2012)2 • •
Paraphantakul et al. (2012) • • • • • • •
Zachariadis and Kiranoudis (2012)2,3 • • •
Belmecheri et al. (2013)3 • • • • •
Eguia et al. (2013)3 • • • • • • •
Lai et al. (2013) • • • • •
Nagy et al. (2013)3 • • •
Pradenas et al. (2013)3 • • • • • •
Salhi et al. (2013)3 • • • •
Tarantilis et al. (2013)2 • • • • •
Cuervo et al. (2014)3 • • •
Davis et al. (2014) • • • • •
Chapter 2. Vehicle Routing Problem with Backhauls
Juan et al. (2014) • • • •

Küçükoğlu and Öztürk (2014) • • • • • •
Vidal et al. (2014)3 • • • • •
Yu and Qi (2014)3 • • • •
Legend: E - Exact, H - Heuristic, MH - Metaheuristic, Pr - Precedence, Mx - Mixed, PD - Pickup and delivery, LS - load split, SWT - Service and waiting time, TW - Time-windows, MP - multi period,
MD - Multi-depot, HF - Heterogeneous fleet, Ecn - Economic, Env - Environmental, Scl - Social, Un - Uncertainty
Belloso et al. (2015)3 • • •
Bortfeldt et al. (2015)3 • • •
García-Nájera et al. (2015)3 • • •
Küçükoğlu and Öztürk (2015) • • • • • •
Lai et al. (2015) • • • •
Nikolakopoulos (2015) • • • • •
Yalcın and Erginel (2015)3 • • •
Zachariadis et al. (2015) • • •
2.A. Classification of VRPB works
Berghida and Boukra (2016)3 • • • • •

Brandão (2016)3 • • •
Chávez et al. (2016)3 • • • • •
Dominguez et al. (2016)3 • • • •
Nguyen et al. (2016) • • • • • •
Oesterle and Bauernhansl (2016)3 • • • • •
Rahimi et al. (2016) • • • • • • •
Wu et al. (2016)3 • • • • •
Belloso et al. (2017a)3 • • • •
Belloso et al. (2017b)3 • • • •
Lin et al. (2017)3 • • • • •
Turkensteen and Hasle (2017) • • • •
Wassan et al. (2017)3 • • • •
Reil et al. (2018) • • • • •
Soleimani et al. (2018) • • • • • •
Granada-Echeverri et al. (2019) • • •
55
Lu and Yang (2019) • • • •

Legend: E - Exact, H - Heuristic, MH - Metaheuristic, Pr - Precedence, Mx - Mixed, PD - Pickup and delivery, LS - load split, SWT - Service and waiting time, TW - Time-windows, MP - multi period,
MD - Multi-depot, HF - Heterogeneous fleet, Ecn - Economic, Env - Environmental, Scl - Social, Un - Uncertainty
2.B. Benchmark instances 57
Appendix 2.B Benchmark instances
Table 2.6 – Main benchmark instances used in VRPB literature
Instance Variant References
Toth and Vigo (1996, 1997, 1999); Mingozzi et al. (1999); Wade
and Salhi (2002, 2004); Osman and Wassan (2002); Brandão
(2006, 2016); Tavakkoli-Moghaddam et al. (2006); Ropke and
Pisinger (2006); Ganesh and Narendran (2007); Wassan (2007);
Goetschalckx and
VRPB Wassan et al. (2008a, 2017); Gajpal and Abad (2009); Tütüncü
Jacobs-Blecha
et al. (2009); Anbuudayasankar et al. (2012); Zachariadis and Ki-
(1989)
ranoudis (2012); Cuervo et al. (2014); Vidal et al. (2014); Bel-
loso et al. (2015, 2017b); Bortfeldt et al. (2015); Yalcın and
Erginel (2015); García-Nájera et al. (2015); Nguyen et al. (2016);
Granada-Echeverri et al. (2019)
Potvin et al. (1996); Thangiah et al. (1996); Duhamel et al.

(1997); Zhong and Cole (2005); Reimann and Ulrich (2006);
Gélinas et al. (1995) VRPB Ropke and Pisinger (2006); Ganesh and Narendran (2007);
Pradenas et al. (2013); Tarantilis et al. (2013); Küçükoğlu and
Öztürk (2014, 2015); Vidal et al. (2014); Nikolakopoulos (2015);
Nguyen et al. (2016)
Toth and Vigo (1997, 1999); Mingozzi et al. (1999); Dethloff

Toth and Vigo VRPB (2001); Wade and Salhi (2002); Ghaziri and Osman (2006);
(1996) Ropke and Pisinger (2006); Ganesh and Narendran (2007); Was-
san (2007)
Thangiah et al. VRPB Ropke and Pisinger (2006); Tarantilis et al. (2013)
(1996)
Dell’Amico et al. (2006); Brandão (2006, 2016); Tavakkoli-

Moghaddam et al. (2006); Saremi et al. (2007); Wassan (2007);
Wassan et al. (2008a, 2017); Gajpal and Abad (2009); Tütüncü
Toth and Vigo VRPB et al. (2009); Tütüncü (2010); Anbuudayasankar et al. (2012);
(1997, 1999)
Salhi et al. (2013); Cuervo et al. (2014); Bortfeldt et al. (2015);
Dominguez et al. (2016); Nguyen et al. (2016); Belloso et al.
(2017a,b); Granada-Echeverri et al. (2019)
Kontoravdis and VRPMB Zhong and Cole (2005); Tarantilis et al. (2013)
Bard (1995)
Dethloff (2001, 2002); Nagy and Salhi (2005); Crispim and

Brandão (2005); Chen and Wu (2006); Montané and Galvão
Salhi and Nagy VRPMB/
(2006); Wassan et al. (2008a,b); Zachariadis et al. (2009, 2010);
(1999) VRPSDP
Çatay (2010); Subramanian et al. (2010, 2011, 2012, 2013); Nagy
et al. (2013); García-Nájera et al. (2015); Chávez et al. (2016)
Min (1989) VRPSDP Dethloff (2001); Montané and Galvão (2006); Çatay (2010)
VRPSDP Montané and Galvão (2006); Zachariadis et al. (2009, 2010);

Dethloff (2001)
Çatay (2010); Subramanian et al. (2010, 2011, 2012)
Montané and VRPSDP Zachariadis et al. (2009, 2010); Subramanian et al. (2010, 2011,
Galvão (2006) 2012, 2013)
2.C. Cases studies and applications of VRPB and variants 59
Appendix 2.C Cases studies and applications of VRPB and vari-

ants
Table 2.7 – Case studies and applications of VRPB
Reference Case study
A retail company delivers goods to stores and picks

Yano et al. (1987); Cheung and Hang (2003)
up goods from suppliers
A supermarket chain delivers goods to customers
Casco et al. (1988); Eguia et al. (2013)
and picks up goods from suppliers
A public library sends material (e.g., books) to
Min (1989)
branch libraries and picks up material from them
A beverage company sends full bottles to customers
Privé et al. (2006); Hoff et al. (2009)
and collects the empty ones
A food retailer sends the requests to customers and
Ubeda et al. (2011)
picks up goods from suppliers
A bank needs to collect and replenish automated
Anbuudayasankar et al. (2012)
teller machines (ATM)
A cement producer sends bagged cement to its cus-
Paraphantakul et al. (2012)
tomers and picks up lignite from mines
An express delivery company receives packages at
Yu and Qi (2014) the hub, where are sorted and send to transfer sta-
tions
A food bank collects donations from local sources
Davis et al. (2014)
and delivers food to charitable agencies
From a port, a carrier provides door-to-door freight
Lai et al. (2015) to import customers and collects containers at export
customers
A ceramics company sends requests to customers
Yalcın and Erginel (2015)
and picks up raw-materials at suppliers
A food industry delivers perishable goods to cus-
Oesterle and Bauernhansl (2016)
tomers and pickups raw-materials from suppliers
A mail distribution centre delivers packages to post
Wu et al. (2016)
offices and collects others
A company delivers industrial equipment to cus-
Dominguez et al. (2016)
tomers and collects items at suppliers
A food company sends perishable goods to cus-
Rahimi et al. (2016)
tomers and collects the expired ones
A grocery chain sends requests to customers and
Lin et al. (2017)
picks up goods from suppliers
A newspaper vendor delivers newspapers to cus-
Soleimani et al. (2018)
tomers and collects the unsold ones
Chapter 3
A Rich Vehicle Routing Problem

with Backhauls
Integrated planning of inbound and outbound logis-

tics with a Rich Vehicle Routing Problem with Back-
hauls
Alexandra Marques∗ · Ricardo Soares ∗† · Maria João Santos ∗† · Pedro Amorim ∗†
Published in Omega, 2019
Abstract This paper addresses the integration of the planning decisions concerning in-
bound logistics in an industrial setting (from the suppliers to the mill) and outbound lo-
gistics (from the mill to customers). The goal is to find the minimum cost routing plan,
which includes the cost-effective outbound and inbound daily routes (OIRs), consisting of
a sequence of deliveries of customer orders, pickup of a full truck-load at a supplier, and
its delivery to the mill. This study distinguishes between three planning strategies: oppor-
tunistic backhauling planning (OBP), integrated inbound and outbound planning (IIOP)
and decoupled planning (DIOP), the latter being the commonly used, particularly in the
case of the wood-based panel industry under study. From the point of view of process
integration, OBP can be considered as an intermediate stage from DIOP to IIOP. The prob-
lem is modelled as a Vehicle Routing Problem with Backhauls, enriched with case-specific
rules for visiting the backhaul, split deliveries to customers and the use of a heterogeneous
fleet. A new fix-and-optimise matheuristic is proposed for this problem, seeking to obtain
good quality solutions within a reasonable computational time. The results from its ap-
plication to the wood-based panel industry in Portugal show that IIOP can help to reduce
total costs in about 2.7%, when compared with DIOP, due to better use of the delivery
truck and a reduction of the number of dedicated inbound routes. Regarding OBP, fos-
tering the use of OIRs does not necessarily lead to better routing plans than DIOP, as it
depends upon a favourable geographical configuration of the set of customers to be visited
in a day, specifically, the relative distance between a linehaul that can be visited last in a
route, a neighboring backhaul, and a mill. The paper further provides valuable manage-
rial insights on how the routing plan is impacted by the values of business-related model
parameters which are set by the planner with some degree of uncertainty. Results suggest
∗ INESC TEC - Instituto de Engenharia de Sistemas e Computadores, Tecnologia e Ciência, Porto, Portugal
† Faculdade de Engenharia, Universidade do Porto, Porto, Portugal
62 Chapter 3. A Rich Vehicle Routing Problem with Backhauls
that increasing the maximum length of the route will likely have the largest impact in re-
ducing transportation costs. Moreover, increasing the value of a reward paid for visiting a
backhaul can foster the percentage of OIR in the optimal routing plan.
Keywords logistics planning · vehicle routing with backhauls · rich vehicle routing · forest
industry
3.1. Introduction
The optimisation of the logistics processes has a whopping effect on improving the cost-
efficiency of supply chains. Specifically, in forest-based supply chains, the inbound logis-
tics bringing the wood from the forest to the mill can represent up to 30% of the total costs
(Audy et al., 2010), while the outbound logistics bringing the wood-based products from
the mill to the consumers can be equally high.
Despite recent studies showing that integrated planning of supply chain operations can lead
to better results than decoupled planning (e.g., Amorim et al., 2012), inbound and outbound
logistics planning are still dealt separately in most forest industries, as well as in other sec-
tors. The complexity of the logistics operations, specificities of the transportation fleet and
customer service levels are frequent justifications for this fact. In the wood-panel based in-
dustry, outbound logistics planning establishes the minimum-cost daily routes, henceforth
called outbound routes (ORs), for delivering the ordered amounts of finished products to
customers. This process accrues from the mill’s production plan and impacts on the cus-
tomer order lead time. Inbound logistics establishes the inbound routes (IRs), usually of
a dedicated log-truck, consisting of a sequence of full truck-load trips between a wood
sourcing location and the mill. The process is affected by wood procurement planning,
ultimately impacting on the upstream forest harvest scheduling decisions. Similar trans-
portation planning settings appear in the retail industry. Namely, in cases in which the
retailer has the option to pick-up products at suppliers besides just simply distributing to
stores (Yano et al., 1987).
This paper studies the integration of inbound and outbound logistics in the context of the
wood-based panel industry. The case study is driven from a real-life industrial application
that operates on a multi-mill setting. The production strategy of the wood-based panels at
each mill is Make-to-Order. The finished products are shipped to the customers in the day
after its production. The stock of raw materials should be at least one week to overcome
fluctuations in wood supply. The outbound logistics are planned locally, in the transporta-
tion department of each mill, while the inbound logistics are planned centrally, considering
the bulk demand for all the mills. The goal here is to find daily minimum-cost outbound
and inbound routes (OIRs) where the vehicle departing from each mill firstly performs a
sequence of deliveries of the amounts ordered by the customers, and secondly, whenever
is cost-effective, picks up a full truck-load of raw materials at a nearby supplier, and de-
livers it at the closest company’s mill. OIRs allow better use of the delivery truck, when
compared with ORs and further avoid dedicated IRs. This is possible because the driver
can easily adapt the same truck that transported the wood boards with reinforcements in its
structure so it can transport a full truck-load of wood chips. For wood-based supply chains,
it is common that the inbound transport is carried in full truck-loads (e.g., Carlsson and
Rönnqvist, 2007; Hirsch, 2011; Derigs et al., 2012).
In this paper, the problem of finding OIRs is modelled as a Vehicle Routing Problem with
Backhauls (VRPB). The VRPB is a variant of the well-known Vehicle Routing Problem
(VRP) where the route visits several customers, in some performing deliveries (referred
as linehauls) and in others pickups (the backhauls), but all deliveries must be made before
any pickups (Goetschalckx and Jacobs-Blecha, 1989). In this study, we use the VRPB as
a mean to tackle Integrated Vehicle Routing Problems, as outlined by Bektaş et al. (2015),
since the routing decisions related with the process of inbound logistics and those of the
outbound logistics are dealt jointly. Moreover, there are essential business-related rules
arising from our application to the wood-based panel industry that determine route feasi-
bility, which are not yet fully covered in the VRPB literature and justify the formulation of
a new variant of a rich VRPB, in line with the taxonomy proposed by Lahyani et al. (2015).
The first set of business-related rules addressed in this study relate to the conditions deter-
mining the visit to a backhaul: i) the backhaul can only be visited after all deliveries are
performed, here called a precedence constraint, because the reinforcement of the truck for
transporting the wood chips can only occur after the last delivery of the wood-based panels;
ii) there is at most one backhaul visited per route because the amount picked up is always
a full truck-load since there are no wood availability constraints at suppliers; iii) if there is
a pickup at a backhaul it is mandatory that the same route includes its delivery at a mill.
This is another type of precedence constraint ensuring that a mill is visited after a backhaul.
However, operational practice indicates that the unloading mill may or may not be the mill
of origin, because the company owns several mills geographically dispersed, and the truck
can end the route in any of these mills, as long as the compatibility requirements between
the types of raw materials available at the backhaul and accepted at the mill are accounted
for; iv) a backhaul may or may not be visited, which is known in the literature as selective
backhauling; v) routes without a backhaul are also feasible, and in this case, the route ends
in the last linehaul visited, similarly to what occurs in an Open VRP (see Figure 3.1). There
are other studies on VRPB that work with precedence constraints and selectiveness. How-
ever, the possibility to optimise the decisions about visiting or not a backhaul and further
choosing the delivering mill in order to minimise total logistics costs are new and impor-
tant features of the problem under study. Another important case-specific rule determines
that each customer may be visited more than once by different vehicles, known in the VRP
literature as split deliveries. The bundle of panels to be delivered at the linehaul customer
is of variable size and weight. Therefore, several smaller bundles can be transported by
the same truck, but larger bundles may need multiple trucks serving the same customer.
Lastly, the available fleet is composed of trucks which are heterogeneous in terms of the
transportation capacity. The transport is entirely outsourced to third-party carriers and paid
based on a fixed daily use cost and a variable cost depending on the travelling distances of
the ‘for-hire’ vehicles. We further emphasize that these business rules are also applicable
in other industries besides the wood-panel one, such as in grocery retail.
The complexity of this real-world problem motivates a study about the main strengths and
shortcomings of different inbound and outbound planning strategies, with greater or fewer
degrees of integration. Furthermore, given the considerable size that these problems can
achieve, it becomes relevant to envisage a scalable solution method, able to cope with the
operational reality.
This research builds on a literature review on VRPB and other rich VRP variants with
similarities to our problem. The first contribution of this paper is to develop a mathemat-
ical formulation to address a rich VRP that is primarily used to solve different planning
strategies for obtaining OIRs. We apply it to a case study in the wood-based panel indus-
try in Portugal and draw conclusions by comparing the routing plans obtained with those
3.2. Critical review of the state of the art 65
2
𝑘
9
𝑘 1
8 𝑘
7
5 4
10
Depot (mill) Linehaul (customer) Backhaul (supplier)
Outbound Inbound Vehicle w/o load

(finished product) (raw material)
Figure 3.1 – Problem representation
alternative planning strategies. Another contribution is to provide valuable managerial in-

sights for planners about the impact of business-related model parameters over the optimal
routing plan. Another contribution is to adapt the fix-and-optimise matheuristic presented
by Sahling et al. (2009) for obtaining good quality solutions for larger instances of this
problem within a reasonable computational time.
The remainder of this paper is organised as follows. Section 3.2 provides a critical review
of the literature regarding integrated transportation planning with a particular connection
to the VRPB. This review covers extensions of VRPBs and solution methods developed to
solve both artificial and real instances, and allows us to place our work in context. Section
3.3 presents the mathematical formulation of the three logistics planning strategies investi-
gated in this work, namely the opportunistic backhauling, the integrated and the decoupled
inbound-outbound transportation planning. Section 3.4 describes the solution approach de-
veloped, which is based on a fix-and-optimise algorithm. Section 3.5 presents the compu-
tational experiments performed with close-to-reality instances from a wood-based industry
in Portugal. The routing plans obtained for the three planning strategies are compared, and
relevant managerial insights are envisaged. The main conclusions are presented in Section
3.6.
3.2. Critical review of the state of the art

In the literature on logistics and transportation, the term integrated planning is broadly used
to refer to situations where the routing decisions are tackled jointly with other decisions
(Speranza, 2018). In some situations, the integration is between transportation decisions
of different planning levels, for example, strategic decisions concerning the design of the
transportation network and the tactical decisions related with the routes and assignment of
the transport vehicles (e.g., Bouchard et al., 2017). In other situations, the integration is
between the routing decisions and the decisions concerning other processes of the supply
chain. The special issue by Bektaş et al. (2015) on the integrated VRP shows examples of
cases where vehicle routing is interlinked with decisions related to loading, production (or
inventory), location, and speed optimisation. As an example, production-routing problems
integrate production, products delivery (i.e., outbound logistics), and usually also inventory
decisions (e.g., Adulyasak et al., 2015). There are several examples in the forest literature
where wood transportation to the mill (i.e., inbound logistics) and the upstream process of
forest harvesting are planned jointly (e.g., Marques et al., 2014).
As indicated by Speranza (2018), a common feature of the studies on integrated transporta-
tion planning is that dealing with those decisions separately or hierarchically by solving the
problems independently, leads to a sub-optimal solution for the integrated problem. In fact,
integrated planning potentiates global efficiency gains, usually translated into cost savings.
As an example, Archetti and Speranza (2015) present significant savings of around 9.5% in
terms of total cost and 9.0% in terms of the number of vehicles used when using a heuris-
tic solution for an inventory-routing problem, in comparison with the solution obtained by
sequentially and optimally solving the inventory management and the routing problems.
The main particularity of our study, not yet fully covered in the literature, is that the inte-
gration is between two processes of the supply chain – inbound and outbound logistics –
wherein both processes the relevant decisions are related with the optimal vehicle routes.
In fact, in our problem, it is the same vehicle that may perform both processes. There are
significant differences in respect to the modelling approach because, in the other cases of
integrated VRPs, such as production-routing, there are at least two types of decision vari-
ables, one for each process, and the correspondent linking constraints. While in ours, there
are only the decision variables related to routing. The linkage between the two processes
accrues from the way the routes are built.
The problem class that mostly resembles our problem is the VRPB, firstly introduced by
Deif and Bodin (1984). Since then, there are several VRPB variants being studied in the
framework of practical applications, as shown in the recent review of Koç and Laporte
(2018). In general terms, the VRPB consists in finding the minimum cost routes, which
start and end at the depot and visit a set of customers partitioned into linehauls (customers
who require deliveries), and backhauls (customers who require pickups), all must be visited
contiguously (e.g., Wade and Salhi, 2002).
The VRPB is not usually considered as an example of integrated vehicle routing planning.
In fact, many of the industrial applications of the VRPB focus on the outbound logistics
process, for example, in retail companies (e.g., Goetschalckx and Jacobs-Blecha, 1989;
Eguia et al., 2013). In these cases, the route prioritises first all the products deliveries, and
only afterwards the pickups, in order to attain a high vehicle utilisation. The customers
are all of the same type (e.g., stores), but with different requirements (i.e., pickup or de-
livery) and the picked up material can be of a different type that cannot be mixed with the
delivered products, such as empty boxes, damaged products or post-consumption material
in reverse logistics. In other applications, such as the distribution of equipment to rentals
(e.g., Dominguez et al., 2016), or package delivery over a distribution network (e.g., Yu and
Qi, 2014), the inbound and outbound material is the same, and it is all planned together as
a unique logistic distribution process.
3.2. Critical review of the state of the art 67
Contrarily, we argue that our case study can be considered integrated transportation plan-
ning because the inbound and outbound logistics are two separate processes that nowadays
are planned independently, involving different types of customers – i.e., suppliers of raw
materials vs. consumers of finished products – sharing in common the depot/mill. Yano
et al. (1987) study a case resembling ours, in a retail chain with one centralized distribution
centre, 40 stores and nearby vendors, where the route includes the delivery of goods to
stores and the pickup of goods in nearby vendors. Planning includes dedicated routes for
the vendors whenever it is not cost-efficient to include them in the delivery routes. The
results of this work allowed savings in the order of a half-million dollars. With a similar
strategy, Paraphantakul et al. (2012) report a case-study in a cement industry, where cement
customers are linehaul customers, and lignite mines are backhaul customers. The problem
was solved using an ant colony optimisation method, and the company was able to save
about 12% in the average tour duration.
The literature review on VRPB reveals examples of mathematical models, exact and heuris-
tic methods for solving distinct problem variants. A general integer linear programming
formulation and set partitioning formulation for the VRPB are presented in Koç and La-
porte (2018). Among the most common extensions of VRPB found in the literature are the
incorporation of time windows (Ropke and Pisinger, 2006; Gutiérrez-Jarpa et al., 2010;
Küçükoğlu and Öztürk, 2013; Nguyen et al., 2016), multi-periods (Davis et al., 2014;
Nguyen et al., 2016), multi-depots (Chávez et al., 2015), heterogeneous fleet (Salhi et al.,
2013; Lai et al., 2013) and split deliveries (Gutiérrez-Jarpa et al., 2010; Lai et al., 2015;
Nguyen et al., 2016; Wassan et al., 2017). There are also variants on the nature of the
backhauling, such as the mixed VRPB that also allows deliveries to linehauls after pickups
in backhauls (e.g., Yazgitutuncu et al., 2009).
As the research on transportation planning advances more and more towards its practi-
cal application, several extensions of VRPs that consider real-life aspects of the logis-
tics problems have emerged in the literature. The VRPs that cover such aspects, namely
the integration of different logistics operations (e.g., inbound and outbound transport), the
consideration of uncertainty or dynamism, or the inclusion of real constraints (e.g., time
windows and multi-periodicity), fall into the vast class of Rich VRPs (Lahyani et al., 2015;
Caceres-Cruz et al., 2014). As our problem concerns a VRP with selective backhauls, het-
erogeneous fleet, and split deliveries, we can classify it as a rich VRPB. Table 3.1 presents
a description of other VRPBs found in the literature that relate to our work, including the
real-life aspects addressed in the problem and the respective types of solution methods used
to solve the VRPB.
From Table 3.1, it is possible to observe that metaheuristics are the most popular methods
used to solve VRPBs. This results from the fact that the VRPB is an NP-hard problem
and, as such, very few exact methods are known to be efficient for large scale problems.
Yano et al. (1987) describe the problem using a set-covering formulation and then solve it
using a procedure based on a Branch-and-Bound that starts from an initial solution obtained
with simple heuristics. Gutiérrez-Jarpa et al. (2009) introduce a Branch-and-Cut algorithm
to solve a VRPB with split deliveries and test it in new problem instances adapted from
the VRP instances with up to 100 customers, but only those instances with 50 customers or
less can be solved to optimality. Davis et al. (2014) use a commercial solver to find optimal
Table 3.1 – Characteristics of the Rich VRPB under study and related works in the literature
VRPB features Solution method

Reference
TW HF SD MD MP SB MB Exact Metaheuristic Matheuristic
Yano et al. (1987) • •
Ropke and Pisinger (2006) • •
Gribkovskaia et al. (2008) • •
Gutiérrez-Jarpa et al. (2009) • • • •
Paraphantakul et al. (2012) • • •
Küçükoğlu and Öztürk (2013) • • •
Salhi et al. (2013) • •
Lai et al. (2013) • • •
Davis et al. (2014) • •
Chávez et al. (2015) • •
Nguyen et al. (2016) • • • •
Oesterle and Bauernhansl (2016) • • • •
Wassan et al. (2017) • •
Our problem • • • •
Legend: TW (time-windows), HF (heterogeneous fleet), SD (split deliveries), MD (multi-depot), MP

(multi-periodic), SB (selective backhauls), MB (mixed backhauls)
transportation schedules that allow food banks to collect food donations from local sources
and to deliver food to charitable agencies, through food delivery points. The problem is
solved in two phases: first, the problem is formulated as a set-covering model to assign
charitable agencies to food delivery points, and then, the problem is formulated as a VRPB
enriched with constraints related to food safety, operator workday and collection frequency,
also using the optimal solution of the first phase as an input. Oesterle and Bauernhansl
(2016) also study a logistic problem of a food company but considering a mixed VRPB
with time windows, heterogeneous fleet, manufacturing capacity and driving time limits.
The problem is formulated as a mixed integer programming model and also solved with a
commercial solver in two phases. The first phase creates clusters of customers to visit, and
at the second phase, the routes in each cluster are optimised.
With respect to metaheuristics, both local search and population-based methods have proved
to be very efficient to deal with VRPB and its extensions. Examples of local search meta-
heuristics include tabu search (Gribkovskaia et al., 2008; Nguyen et al., 2016), adaptive
large neighborhood search (Ropke and Pisinger, 2006), and variable neighborhood search
(Wassan et al., 2017). Examples of population-based metaheuristics developed for the
VRPB include ant colony optimisation (Paraphantakul et al., 2012; Chávez et al., 2015)
and evolutionary algorithms (Küçükoğlu and Öztürk, 2013). Moreover, two-phase heuris-
tics are also investigated in the works of Salhi et al. (2013) and Lai et al. (2013).
Regarding matheuristic approaches, no references related to its adaptation to the VRPB
were found. However, the literature accounts for several matheuristic approaches for var-
ious solving VRP variants. For example, the fix-and-optimise approach was initially pro-
posed by Sahling et al. (2009) for a lot-sizing problem, but it has been gaining recent in-
3.3. Problem formulation 69
terest in the literature for solving several rich routing problems with real-life aspects (e.g.,
Neves-Moreira et al., 2019). This matheuristic consists in iteratively fixing different sets
of binary variables from a mathematical model, thus allowing a commercial solver to only
solve smaller parts of the global problem. Depending on the problem, the selection of the
variables to be fixed or released needs to be carefully designed. Most references frame this
approach in a variable neighbourhood decomposition search (Hansen et al., 2001), where
the number of variables to be released is progressively increased as a way to increase the
neighbourhood sizes being explored (e.g., Darvish et al., 2019; Soares et al., 2019). Other
research works use distinct heuristic concepts, such as tabu search (e.g., Rieck et al., 2014)
by using a tabu list for the variables being fixed.
Our work is distinct from the ones revisited in this section. It contributes to the literature
because it not only describes a new formulation for a rich VRPB that can be used to address
different transportation planning strategies but also investigates a fix-and-optimise method
to solve the problem, which was not yet addressed in VRPB literature.
3.3. Problem formulation

This section outlines the main planning strategies for the integration of inbound and out-
bound logistics processes, which will be addressed in this paper. For each one of these
planning strategies, mathematical formulations will be provided, which will be the basis
for the sections that follow.
3.3.1 Logistics planning strategies

The integration of inbound and outbound logistics by finding the optimal OIRs can be
staged in two distinct planning strategies, in opposition to a simpler strategy of decoupled
planning, similar to what is used today by the company:
• Opportunistic backhauling planning (OBP): In this strategy, the primary process to

be considered is the outbound logistics. The outbound transportation plan encom-
passes ORs and cost-effective OIRs, but another plan exists for IRs. There is an
underlying idea that OIRs can provide only a residual amount of the raw materials
demanded and IRs assure the vast majority of the demand.
• Integrated Inbound and Outbound Planning (IIOP): In this strategy, both processes
of inbound and outbound logistics are planned jointly. The transportation plan en-
compasses all types of routes – ORs, OIRs and IRs.
• Decoupled Inbound and Outbound Planning (DIOP): This strategy implies that both
processes of inbound and outbound logistics are planned independently. The out-
bound transportation plan (or delivery plan) encompasses the ORs, while the inbound
plan (or supply plan) encompasses IRs, there are no OIRs. In the current situation of
the case study, logistics planning occurs in separate company departments. IRs are
planned centrally and ORs are planned in a department at each mill.
From the point of view of process integration, OBP can be considered an “intermediate”
stage, from DIOP towards IIOP, as well as from the point of view of the level of organi-
sational changes needed for its adoption. In fact, OBP impacts mostly on the planners of
the outbound logistics in each mill and on the truck drivers while IIOP implies a major
restructuring from merging (and possibly centralising) the inbound and outbound logistics
planning departments. From a modelling point of view, the mathematical formulation for
OBP and IIOP are similar. For the purpose of simplification, this section focuses on OBP,
making the necessary adjustments to IIOP afterwards. The section ends with the descrip-
tion of DIOP.
3.3.2 Opportunistic backhauling planning (OBP)

OBP can be modelled as a rich, capacitated Vehicle Routing Problem with selective back-
hauls and split deliveries. Considering a set of mills M, a set of linehaul customers L
whose demand needs to be fulfilled, and a set of suppliers backhauls B with raw materials
available for the mills that may or may not be visited. The problem consists in finding the
optimal daily minimum-cost routes for a set of trucks K, starting at the mill, encompassing
one or many deliveries to linehauls, and including at the most one pickup of a full truck-
load of a given type of raw materials at a backhaul, which is selected based on the best fit
with one of the possible destination mills. The set of types of raw materials to be collected
at a backhaul is represented by set P. Hence, the problem components include:
• the fleet of |K| trucks, where each truck k ∈ K has a given capacity (Qk ) and can
perform both deliveries and pickups.
There is a fixed cost for the daily usage of a
vehicle f k and a variable cost ckij proportional to the travelled distances;
• the |M| mills owned by the company that are geographically dispersed. Each mill
m ∈ M receives wood chips and produces wood-based panels on a make-to-order
basis. The fleet is assigned to a specific mill or origin (or depot), from where the
routes start. According to operational practice, in case of a route with a backhaul,
the truck can unload the raw materials in any of the company’s mills, which may or
may not be the mill of origin. There is a minimum amount of raw materials to be
backhauled to all mills (β);
• the |L| linehaul customers that are characterized by a given demand of a finished
product, which must be fulfilled (ql ) at each linehaul l ∈ L. Split deliveries can occur,
meaning that each customer may be visited more than once (each visit consisting in
at least a ψ amount), but each truck may visit a customer at most once;
• the |B| backhaul suppliers that are also geographically dispersed. Also, according
to the operational practice, it is assumed that all have unlimited availability, hence
pickups correspond to full-truck loads. The type of raw materials that are available
may also vary amongst them;
• the |P| types of raw materials consisting of wood chips of variable size and moisture
content, sawdust and recycled wood. Some types of raw materials are more desirable
to the mills than others. There are also compatibility issues with respect to the types
of raw materials available and demanded at the different locations.
Contrarily to other VRPs found in the literature, the time window constraints related to
the earliest or latest time to arrive at each location are not of importance. However, the
maximum distance travelled in a route is limited by a parameter α. It is noteworthy that
the route length can be constrained in terms of travelling time, to account for driving time
regulations stating maximum driving or working times. However, in this case, the value of
the maximum distance travelled was set with the planner as an average of the actual routes
length, already implicitly considering all the necessary stops, hence simplifying problem
modelling. In summary, the characteristics of the feasible routes are: i) start at a home
depot with the truck loaded up to its maximum capacity, with the products ordered by
the linehaul customers; ii) perform a sequence of deliveries to the linehauls; iii) if it is
cost-effective and doable during the maximum route length, the vehicle travels empty to
a nearby backhaul supplier to pick up a full truck-load of raw materials to be delivered at
any of the company’s mill, where the route ends (specific to OIRs); and iv) if a backhaul
is not visited, the route is ended when the truck is empty after visiting the last linehaul of
the route (specific to ORs), as the company does not pay for trips where the truck does not
transport merchandise.
3.3.2.1 Modelling approach

The rVRPB under study is modelled as a graph G = (V, A) where V is the set of all vertices,
V = {0} ∪ L ∪ B ∪ M and A is the set of all possible arcs. We adopt a standard flow VRP
formulation with 3-index decision variables xikj equal to 1 if vehicle k ∈ K travels from cus-
tomer i ∈ V to j ∈ V and zero otherwise. Like in the standard VRPB formulation proposed
by Parragh et al. (2008), we distinguish the vertices in linehauls and backhauls, in order to
model the precedence constraints.
However, the typical VRPB constraints assuring that each vertex is visited exactly once do
not apply, due to the possibility of selective backhauls (i.e., backhauls may or may not be
visited) and the split deliveries at the linehauls (i.e., linehauls are visited more than once).
To avoid the complexity of a multi-depot and open VRP, we propose a 2-echelon backhauls
network, starting and ending at the same fictitious depot 0. In fact, when the route starts,
the fictitious depot corresponds to the mill of origin from where the customers’ orders will
be delivered. Since there is a fleet dedicated to each mill when the route starts, routing
planning for each mill can be done separately as a single depot. When the route ends,
the fictitious depot corresponds to a fictitious location whose distance from the last vertex
visited in the route is equal to zero. Hence, the 2-echelon backhauls network is composed
by the first echelon of backhauls corresponding to the suppliers and the second echelon of
backhauls corresponding to the mills to be supplied by the backhauled amounts. Additional
constraints are needed to assure that a mill can only be visited after a backhaul (see Figure
3.2).
The decisions whether a backhaul is visited in a route or not, and if so, to which mill
to go next, are based on a new parameter related with the reward paid for visiting that
backhaul and a mill next (δbm ). Like in previous studies of VRP with selective pickups (e.g.,
1st linehaul last linehaul 1st echelon 2nd echelon

customer customer backhauls backhauls
(suppliers) (mills)
1 1
2 2 1
… 9
9 3 3
2
4 4 10
3
Depot “Fictitious”
of origin depot of
5 5 destination
Outbound Inbound Vehicle w/o load Trip to fictitious

(finished product) (raw material) depot
Figure 3.2 – Network representation of the problem
Gribkovskaia et al., 2008) and other formulations of VRP with profits (e.g., Aras et al.,
2011), the reward is used to make an arc linehaul to backhaul more or less attractive. The
reward corresponds to a payment per each ton of raw materials picked-up in a backhaul and
delivered in a neighbouring mill. If the route ends after visiting the last linehaul, then there
is no positive reward associated with that route. Hence, the reward parameter is used in the
objective function, which trades-off between the sum of the travelling costs for visiting the
backhaul after the last linehaul and moving from there to a mill, and the reward gained for
visiting that backhaul. The reward parameter is also used to address compatibility issues
related to the type of raw material p to be transported from a given backhaul b to a given
mill m. In fact, if p is not available in b or not accepted in m then δbm = 0. On the contrary,
if there are several types of raw materials that can be transported from b to m, the value
of δbm corresponds to the value of the most profitable material because there are no other
aspects determining the choice between them. Consequently, the set P does not need to be
considered in this model. However, in other real-life applications where the availability at
the backhauls and or demand at the mills is limited and varies per type of product, the set
P should be properly incorporated in the model, leading to a four-index decision variable
x.
A new decision variable is needed to assure that, despite the possibility of splitting the
deliveries to a linehaul, each delivery cannot exceed the truck capacity and that the total
amount delivered in the several routes that visit it meets the expected demand. Previous
studies used continuous variables wki representing the quantity transported by vehicle k ∈ K
to/from customer i ∈ V for a similar purpose (e.g., Nikolakopoulos, 2014). However, in
the rVRPB under study, without time windows, these variables are insufficient for sub-tour
elimination. In this context, a new set of continuous variables ukij represent the load of ve-
hicle k ∈ K when traversing arc (i, j) ∈ A. Variables ukij are a natural adaptation of variables
ui (Bektaş et al., 2015; Toth and Vigo, 2014) to a multi-route and split delivery situation.
Additional constraints are needed to account for the routes with backhauls. In this case,
the truck-load is higher before visiting the first linehaul, then progressively decreases until
reaching zero after visiting the last linehaul. If a backhaul is visited, the pickup corresponds
to a full truck-load. As an example, for a given route k, encompassing {0, i, i0 , i00 , j, 0}, where
i, i0 , i00 ∈ L and j ∈ B, then the following rules apply: uk0i ≤ ukii0 ≤ uki0 i00 , uki00 j = 0, ukj0 0 = Qk .
Figure 3.3 exemplifies a feasible solution for the OBP starting in the node 9, in a network
composed by 5 linehauls (numbered 1 to 5), 3 backhauls (numbered 6 to 8) and 3 mills
(numbered 9 to 11). For simplification purposes, only the arcs used in the solution are
represented in Figure 3.3a. The demand (in ton) at the linehauls is q1 = 30, q2 = 20, q3 =
20, q4 = 20, q5 = 70. The reward for visiting a backhaul is 0.1e/ton in all cases. The avail-
able fleet is composed of 5 trucks, with capacity (in ton) Q1 = 40, Q2 = 30, Q3 = 30, Q4 =
40, Q5 = 40. The linear distances between vertices (di j ) are computed in reference to the
background grid with 1km by 1 km, for example, d13 = 2 km. The fixed cost for using a
vehicle is zero, and the variable cost is 1 e/km.
Vehicle 𝑘
3 7 11
Vehicle
load Vehicle capacity
𝑘 ,20 40
𝑘 , 20
2 1
𝑘 ,0
𝑘 ,40
𝑘 ,30 𝑘 , 40 𝑘 , 30 20
10 6
𝑘 , 40
𝑘 ,0
4 𝑘 , 20
9 𝑘 , 30 5
𝑘 , 40 𝑘 ,0
𝑘 , 40 8 𝑘 ,0
0
9 1 2 6 10 Arrival
at vertex
(a) feasible solution for vehicles (b) variation of the load of vehicle k1 along
k1 , k2 , k3 , k4 , k5 the route
Figure 3.3 – Example of a feasible solution for a rVRPB
The routing plan foresees the use of all five vehicles: k1 , k2 , k4 and k5 are OIRs while k3 is
an OR ending after visiting linehaul 3. There are split deliveries in linehauls 1 and 5. Total
costs are 29e and total revenues are 15e. The values of ukij for truck 1 are shown in Figure
3.3b.
This example is instrumental in showing the impact of the reward value over the final rout-
ing solution. In fact, the route visiting linehaul 4 will always visit backhaul
6, and then mill
10, because the extra cost for visiting this pair backhaul-mill is 1e d4,6 = 1e, d6,10 = 0 =⇒ ck4,10 = 1
and the minimum revenue is 3e (δbm = 0.1e/ton, min{Qk } = 30 ton =⇒ uk6,10 ≥ 30, ∀k ∈
K, δ6,10 = 0.1. Applying a similar logic, it is expected that the route visiting linehaul 3 will
visit backhaul 7 if δ7,11 ≥ 0.4, since the extra cost for visiting the backhaul and mill is 4e
and uk7,11 ≥ 30, ∀k ∈ K.
3.3.2.2 Mathematical formulation

For the sake of convenience, before presenting the mathematical formulation, we resume
the necessary decision variables, sets and parameters.
Decision variables:

1 if vehicle k travels from location i to j;

xikj 

0 otherwise.

ukij load of vehicle k ∈ K when traversing arc (i, j) ∈ A
Sets:
L set of linehauls (customers where finished products are delivered)

B set of backhauls (suppliers where raw materials can be picked up)
M set of mills (where raw materials are delivered if a backhaul is visited)
V set of vertices; V = {0} ∪ L ∪ B ∪ M
K set of vehicles
Parameters:
qi quantity to be delivered to customer i ∈ L (ton)

ckij cost of transportation with vehicle k ∈ K from i ∈ V to j ∈ V (e)
k
f fixed cost of using vehicle k ∈ K in a daily route (e)
Qk transportation capacity of vehicle k ∈ K (ton)
di j travelling distance from i ∈ V to j ∈ V (km)
α maximum distance travelled in a route (km)
β minimum amount of raw materials to be backhauled (ton)
δbm reward for picking up one unit of raw material at backhaul b ∈ B and delivering
it to mill m ∈ M (e)
ψ minimum amount of order delivered to a linehaul (ton)
Model [P0]
X X XXX XXX
min f k x0k j + ckij xikj − δi j ukij (3.1)
k∈K j∈V\{0} k∈K i∈V j∈V k∈K i∈B j∈M
subjected to:
X
xikj ≤ 1 ∀ j ∈ L ∪ B, ∀k ∈ K (3.2)
i∈V
XXX
xikj = 0 (3.3)
i∈B j∈B k∈K
X XX
xikj = 0 (3.4)
i∈V\B j∈M k∈K
X X X
xikj = 0 (3.5)
i∈M j∈V\{0} k∈K
X X X
xikj = 0 (3.6)
i∈B j∈L∪{0} k∈K
XX
x0k j = 0 (3.7)
j∈B k∈K
X X
x0k j = k
xi0 b ∀k ∈ K (3.8)
j∈L i∈L∪M
X X
xikj = xkji ∀ j ∈ V, ∀k ∈ K (3.9)
i∈V i∈V
ukij ≤ Qk xikj ∀(i, j) ∈ A, ∀k ∈ K (3.10)
X X X
ukij − ukji ≥ ψ xikj ∀ j ∈ L, ∀k ∈ K (3.11)
i∈V i∈V i∈V
X X X
ukij = 0 (3.12)
i∈L j∈B∪{0} k∈K
XX
ukij − ukji = q j ∀j ∈ L (3.13)
i∈V k∈K
XX
di j xkji ≤ α ∀k ∈ K (3.14)
i∈V j∈V
XXX
ukij ≥ β (3.15)
i∈B j∈M k∈K
xikj ∈ {0, 1}, ukij ≥ 0 ∀(i, j) ∈ A, ∀k ∈ K (3.16)
The objective function (3.1) minimizes the total costs, decomposed into fixed costs (pro-
portional to the number of vehicles used) and the variable costs (proportional to the total
travelled distance), decreased by the revenue obtained for visiting backhauls and mills in
the course of the OIR. Constraints (3.2) assure that any location can be visited at most once
by each truck. Regardless, each linehaul and backhaul can be visited by several routes.
Constraints (3.3)–(3.7) deal with route precedence rules, resulting from the specificities of
this rVRPB for the wood-based panel industry. Specifically, constraints (3.3) state that the
transport from a backhaul to another backhaul is not possible. Constraints (3.4) assure that
the mill can only be visited after a backhaul. Constraints (3.5) assure that after visiting
a mill, the only possibility is to go to the ending depot. Constraints (3.6) state that after
visiting a backhaul, the next visit cannot be to a linehaul nor to the depot. Constraints (3.7)
assure that the route cannot visit a backhaul after the depot. Constraints (3.8) and (3.9)
are the typical VRP flow conservation constraints, at the depot and at each vertex, respec-
tively. Constraints (3.10) are linking constraints, assuring that there is only a given amount
transported to/from the customer if the customer is visited. Constraints (3.11) to (3.13)
assure the elimination of sub-tours. Specifically, constraints (3.11) assure that the load of
trucks progressively decreases as it visits the linehauls, and the amount delivered should be
higher than a minimum amount. By considering the lower bound of the minimum amount,
the model avoids undesirable solutions where xikj = 1 and ukji − ukij = 0, which may occur for
example if a linehaul (i0 ) is visited in the course of a route from i to j, i.e., xiik 0 = xik0 j = 1

(instead of xikj = 1) but the amount delivered in i0 is zero uki0 i − ukii0 = 0 due to the fact that
the distance matrix does not obey to the triangular inequality (i.e., ∃ di j : di j > dii0 + di0 j ).
Constraints (3.12) state that the truck leaves empty after visiting the last linehaul and con-
straints (3.13) assure that the demand at the linehauls is completely fulfilled. Constraints
(3.13) together with constraints (3.2) account for the possibility of split deliveries at the
linehauls. Constraints (3.14) assure that the maximum allowable distance of the daily route
cannot be exceeded. It is noteworthy that if the maximum route length is constrained by
the time travelled, then, this would require another type of auxiliary variables to count the
route duration and consequent changes in these constraints, with similarities with other
VRPs with time windows (e.g., Toth and Vigo, 2014). Constraints (3.15) set a minimum
amount of raw materials to be backhauled to mills. Finally, constraints (3.16) determine
the domain of the decision variables.
3.3.2.3 Special situation in which the rVRPB is simplified to a rich capacitated VRP
A problem variant of the rVRPB consists in removing constraints (3.14) and (3.15). In
this situation, where there is no limitation to the route length and there is no minimum
backhauling amount, a backhaul will be visited whenever it is cost-effective, according to
the trade-off between the extra transportation cost (from travelling from the last linehaul,
to that backhaul and to its closest mill) and the revenue (associated with delivering the
load from the backhaul to the closest mill). From a modelling perspective, this means that,
knowing which is the last visited linehaul in a route, it is possible to compute beforehand
if and which backhaul and mill should be visited to minimize total costs. Consequently,
the mathematical model can be simplified to a Rich Capacitated VRP (rCVRP) with split
deliveries. This problem will only consist in sequencing the linehauls to be visited in each
route, thus determining which linehaul will be last in each route.
This adaptation relies on a data pre-processing procedure (described in Algorithm 1) which
consists in computing the minimum cost of having a given linehaul visited last in a vehicle
route. If the cost of visiting a backhaul at the end of the route is lower than finishing the
route at the depot (line 5), the cost associated with the arc heading to the depot is updated
to the summed costs of pickup at the backhaul, delivering to the mill and returning to
the depot, subtracted by the corresponding reward for performing the delivery to that mill
(line 6). All combinations of vehicles, linehauls, backhauls, and mills are tested in this
pre-processing stage, therefore ensuring that the vehicle arcs heading to the depot account
for the minimum possible cost, which either corresponds to performing backhauling at the
most advantageous locations or finishing its route after visiting the last linehaul. Finally,
the sets of backhauls and mills are removed from the problem.
Algorithm 1 Data pre-processing for adapting the rVRPB to a rCVRP

1: for each vehicle k in K do
2: for each linehaul customer j in L do
3: for each backhaul customer i in B do
4: for each mill customer m in M do
5: if ckji + ckim + ckm0 − δim · Qk < ckj0 then
6: ckj0 := ckji + ckim + ckm0 − δim · Qk
7: V := V \ (B ∪ M); B := ∅; M := ∅
Afterwards, the new model for the rCVRP can be built upon [P0] by changing the objective
function and removing constraints related with the sets of backhauls and mills, as shown in
model [P1].
Model [P1]
X X XXX
min f k x0k j + ckij xikj (3.1b)
k∈K j∈V\{0} k∈K i∈V j∈V
subjected to (3.2), (3.8)–(3.13) and (3.16) of model [P0]
3.3.3 Integrated Inbound and Outbound Planning (IIOP)

As stated before, the IIOP strategy consists in jointly planning all types of routes, includ-
ing OIRs, ORs only for delivery of finished products and IRs for pickup of raw materials.
Model [P2] for IIOP can be built upon adaptations of [P0], that account for the IRs, as
follows. Constraints (3.7) are removed to allow dedicated routes from the depot to a back-
haul. A new parameter δbm D represents the reward for picking up one unit of raw material
at backhaul b ∈ B and delivering it to mill m ∈ M (e) in the course of the dedicated route.
A new set of auxiliary continuous variables ykij is needed to represent the amount picked
up in b ∈ B and delivered in mill m ∈ M by vehicle k ∈ K in a direct route. The objective
function (3.1c) is adapted accordingly. A new set of constraints (3.17) defines variables ykij
and constraints (3.18) set its bounds. Considering an arc (i, j), i ∈ B, j ∈ M, with xikj = 1, if
k = 1, i ∈ B, then k is in a dedicated route, and according to the conjugation of constraints
x0i
k = 0, i ∈ B, then k is in an OIR, and yk = 0.
(3.17) and (3.18), ykij = ukij . If x0i ij
Model [P2]
X X XXX XXX XXX
min f k x0k j + ckij xikj − δiDj ykij − δi j (ukij − ykij )
k∈K j∈V\{0} k∈K i∈V j∈V k∈K i∈B j∈M k∈K i∈B j∈M
(3.1c)
subjected to (3.2)-(3.6), (3.8)–(3.16) of model [P0] and
ykij ≤ Qk xikj ∀i ∈ B, ∀ j ∈ M, ∀k ∈ K (3.17)

ykij ≤ ukij − (1 − x0i

k
)Qk ∀i ∈ B, ∀ j ∈ M, ∀k ∈ K (3.18)
ykij ≥ 0 ∀i ∈ B, ∀ j ∈ M, ∀k ∈ K (3.19)
3.3.4 Decoupled Inbound and Outbound Planning (DIOP)

As stated before, DIOP corresponds to the planning strategy currently used, where the ORs
and IRs are planned independently and there are no OIRs. For the outbound logistics plan-
ning, the optimal ORs can be obtained by solving model [P3] that is an adaptation of model
[P0], considering the nonexistence of backhauls and mills. For the inbound planning, the
optimal IRs can be obtained by solving a model [P4], also an adaptation of model [P0],
acknowledging only the routes from the depot/mill of origin to the backhauls.
Model [P3]
X X XXX
min f k x0k j + ckij xikj (3.1d)
k∈K j∈V\{0} k∈K i∈V j∈V
subjected to (3.2), (3.8)–(3.11), (3.13)–(3.14) and (3.16) of model [P0]
Model [P4]
X X XXX XXX
min f k x0k j + ckij xikj − δiDj ukij (3.1e)
k∈K j∈V\{0} k∈K i∈V j∈V k∈K i∈B j∈M
subjected to (3.2)–(3.6), (3.8)–(3.10) and (3.14)–(3.16) of model [P0]
3.4. Solution approach

3.4.1 Fix-and-optimise approach
As stated in the literature review, the complexity of the VRP problems in real-life appli-
cations justifies the use of matheuristics. In this study, all the different models presented
before are solved with a fix-and-optimise (F&O) approach in case of the large instances
(i.e., more than 30 customers). This solution method was firstly presented by Sahling et al.
(2009) for lot-sizing problems, but has been successively used for solving complex routing
problems with promising results (e.g., Neves-Moreira et al., 2016; Larrain et al., 2017).
The F&O matheuristic approach consists in iteratively solving several smaller mixed in-
teger programming (MIP) sub-problems of the original model. The design of each sub-
problem is problem-dependent and the obtained results highly depend on its adequate de-
sign. In this approach, we define a sub-problem as a set of decision variables to be either
released or fixed in the original MIP model. Fixing a variable consists in setting its lower
and upper bounds to the current solution value, thus precluding it from being changed in a
solver iteration. On the other hand, releasing a variable consists in restoring a fixed vari-
able to its original lower and upper bound values. For the problem at hand, two distinct
sub-problem types were conceived, named RouteRelease and LocationRelease.
3.4. Solution approach 79
The RouteRelease sub-problem releases all decision variables associated with a given set
of routes in the incumbent solution, based on proximity criteria of these routes. Route prox-
imity is defined by the centroids of each route, which are computed as the non-weighted
averages of the location coordinates that are visited. The outline of the RouteRelease
sub-problem construction procedure is illustrated in Algorithm 2. The procedure starts by
computing the centroid of each route in the incumbent solution (lines 5–7). For unused
vehicles, the route’s centroid is given by the depot’s coordinates. A pivot route is selected
at random from the incumbent solution (line 8), after which all other routes are ordered by
its centroid’s distance to the pivot route (line 9). The n routes with the lowest distance to
centroid of the pivot route are then released in the sub-problem (lines 10–15).
Algorithm 2 Route Release sub-problem construction

1: Input: vars (MIP model routing decision variables)
2: sol (incumbent solution)
3: n (number of routes to be released in the subproblem)
4: released_routes = ∅; centroid_list = ∅
5: for each route ∈ sol do
6: compute centroid of route
7: append centroid of route to centroid_list
8: rt ← random route; cnt ← centroid of rt
9: order centroid_list by descending order of their distance to cnt
10: released_routes ← n first routes ∈ centroid_list
11: for each var ∈ vars do
12: if var is associated with a vehicle ∈ released_routes then
13: release var
14: else
15: fix var
The LocationRelease sub-problem consists in releasing a given set of linehaul locations

based on its geographical proximity. The procedure is described in Algorithm 3, and it
starts by selecting a pivot linehaul (line 4), after which we retrieve all routes in the incum-
bent solution where the pivot linehaul is visited. Afterwards, we retrieve all the additional
linehauls that are visited in these routes (lines 7–8). Finally, the n closest linehauls to the
pivot linehaul that were previously selected are released (lines 11–18).
The overall structure of the matheuristic is shown in Algorithm 4.

The solution method requires an initial solution s0 with objective function f0 , which is
obtained through a greedy nearest neighbour heuristic (line 7): we select a random vehicle
and construct its route by visiting the nearest unsatisfied linehaul until vehicle capacity
is exhausted. The process is repeated until all linehaul demand is satisfied. No routes to
backhauls are considered in the constructive phase.
After obtaining an initial solution, the matheuristic is then started. To that effect, sub-
problem construction is initiated, whose size is controlled through the general principles of
Algorithm 3 Location Release sub problem construction

1: Input: vars (MIP model routing decision variables)
2: sol (incumbent solution)
3: n (number of locations to be released in the subproblem)
4: released_locations = ∅; candidates = ∅; loc ← random linehaul
5: for each route ∈ sol do
6: if route transverses loc then
7: for each linehaul ∈ route do
8: append linehaul to candidates
9: order candidates by descending order of their distance to loc
10: released_locations ← n first locations ∈ candidates
11: for each var ∈ vars do
12: if var is associated with a linehaul ∈ released_locations then
13: release var
14: else
15: if var is associated with a mill then
16: release var
17: else
18: fix var
a Variable Neighbourhood Decomposition Search (VNDS), similar to what is presented in

Hansen et al. (2001). Sub-problems are constructed in line 13, after which the MIP model
is fed the incumbent solution scur and the sub-problem is solved by a MIP solver (lines
14–15).
After each solver iteration, the obtained solution s solve is evaluated against the incum-
bent solution (lines 16–21). If the obtained solution did not yield an improvement of at
least imp (line 16), we consider this a non-improvement iteration and increment the non-
improvement counter. Nevertheless, we will accept the obtained solution even if it is not
significantly better than the previous one (line 21). After a given number of consecutive
non-improvements, the VNDS framework takes place either by increasing sub-problem
size or switching the sub-problem type, if the current sub-problem size has been maxed out
(lines 22–32). In the occurrence of a significant improvement of the problem’s objective
function, sub-problem type is re-set to RouteRelease and its initial size (line 17).
The matheuristic approach always initializes with the RouteRelease sub-problem type
and the LocationRelease sub-problem is used after a significant number of non-improvements
of the RouteRelease sub-problem. This algorithmic structure was conceived by bear-
ing in mind that RouteRelease would be used as a more disruptive sub-problem, which
would explore more disperse sections of the solution space, while the LocationRelease
sub-problem focuses more on intensification.
3.4.2 Data pre-processing

Pre-processing the instance data related to the network generation is a common procedure
in VRPs (e.g., Parragh et al., 2008; Soares et al., 2019) to simplify the mathematical for-
3.4. Solution approach 81
Algorithm 4 Matheuristic outline

1: Input: MIPmodel (mixed integer programming model)
2: P (list of possible sub-problems to be used)
3: Np (initial neighbourhood size for sub-problem p)
4: Ip,n (limit of consecutive non-improvement iterations for sub-problem p of
size n)
5: TLp,n (time limit for solver iterations of sub-problem p of size n)
6: imp (minimum solution improvement to reset the no-improvement counter i)
7: s0 , f0 = nearest_neighbour()
8: scur = s0 ; fcur = f0 ; i = 0
9: p = “RouteRelease”
10: while termination criteria not met do
11: n = np
12: while n ≤ Np do
13: construct sub-problem of type p with size n
14: feed MIPmodel with initial solution scur
15: s solve , f solve = MIPsolve (MIPmodel, TLp,n )
16: if f solve < fcur − imp then
17: scur = s solve ; fcur = f solve ; i = 0; p = “RouteRelease”
18: break
19: else
20: if f solve < fcur then
21: scur = s solve ; fcur = f solve
22: i=i+1
23: if i > Ip,n then
24: i=0
25: if n = Np then
26: if p = “LocationRelease” then
27: p = “RouteRelease”
28: else
29: p = “LocationRelease”
30: break
31: else
32: increase n
33: return scur , fcur
mulation and achieve better performance in the optimisation solver. The pre-processing
procedure used prior to solving the models is threefold. The sub-set of arcs to be consid-
ered is presented in Table 3.2.
Table 3.2 – Pre-processing the problem network
fictitious depot
Destination
or mill of origin
depot
Linehaul
Backhaul
Linehaul
Mill
Origin
0 depot
● ● (*)
or mill of origin
Linehaul 𝑙 ● (**) ●
Linehaul 𝑙 ≠ 𝑙 ● (**) ●
Backhaul 𝑏 (+)
Mill 𝑚 ● Legend:
(*) is generated in IIOP but not in OBP;
(**) for each linehaul, generate the arcs to all the backhauls that lead to a cost-
{0} fictitious depot effective solution (i.e. satisfy line 5 of Algorithm 1);
(+) for each backhaul, only generate the arc to the minimum cost mill.
First, we remove all the arcs that lead to an unfeasible route, i.e., arcs that violate the prece-
dence constraints (3.3) to (3.7). Second, we eliminate all arcs from linehauls to backhauls
where its visit is not economically worthwhile, according to the given reward for visiting a
backhaul. These arcs are only generated if they respect the condition exhibited in line 5 of
Algorithm 1.
Third, arcs from backhauls are only generated to its closest mill, as delivering merchandise
to more distant mills will only induce an increase of the problem’s objective function.
It should be noted that this data pre-processing procedure does not cut off optimal solutions
only if we do not impose a minimum inbound quantity to be collected from backhauls
via constraints (3.15). If this is not the case, this procedure may induce sub-optimality
or even turn the model infeasible because there are no cost-effective backhauls to visit.
Therefore, in these situations, a trade-off between optimality and simplicity must be taken
into account.
3.5. Computational experiments

The proposed approach was applied in a case study in a wood-based panel company in
Portugal. The mathematical model was implemented in Gurobi 7.5 commercial solver.
The solution method was developed in Python 3.6. The mathematical models were subject
to the data pre-processing procedure described earlier and used to compare the gains of
the IIOP strategy with the DIOP one, which is currently done by the company. A set of
experiments were also done to provide valuable managerial insights for planners. Lastly,
the performance of the proposed solution method was compared with a commercial MIP
solver for problem instances of increasing size, which were based on real routing plans
3.5. Computational experiments 83
executed by the company.
3.5.1 Case study

This study was motivated by a real-life application in the wood-based panel company,
firstly presented in Amorim et al. (2014). The focal company owns several mills, each
producing a specific portfolio of wood-based panels mainly for furniture, construction and
decoration. The case study is at one of the mills in Portugal that produces around 1.2
thousand tons of wood-based panels per day, in a make-to-order basis, and assures its
delivery to an average of 30 customers distributed over the entire Iberian Peninsula. The
average daily consumption of raw materials is 1,750 . The study uses real data regarding
the customers’ orders in two of the most representative operational days. There are 30
customers to be visited, whose ordered amounts are in average 35.5 ton/customer, varying
between 0.05 and 399 ton. The values of the model parameters are an approximation
of those provided by the planners. The distances between locations were computed by
resorting to the Google Maps routing engine.
Nowadays, the outbound routes are planned to start in the morning of the next day at the
opening hour of the mill of origin. It is assumed that all routes can start at the same time,
and there are no time windows conditioning the time of arrival to customers, suppliers or
mills. The responsible for outbound logistics determines the exact number of trucks needed
for the next day and groups the customers to be visited in each route according to empirical
rules that rely on the customers’ geographical location. Then, the routes are assigned to
the third-party logistics operators (3PL) with whom there are valid outsourced contracts.
The generic contractual conditions are a fixed cost of 70 per truck used and a variable cost
of 1.7 per km travelled. The fleet available at the mill of origin in each day encompassed
100 trucks, of which 20 trucks have capacity up to 10 tons, 40 trucks have a capacity of 20
tons, and 40 trucks have a capacity of 40 tons, summing up a total transportation capacity of
2,600 ton. Each vehicle must deliver at least 0.5 ton to each customer visited (i.e., ψ = 0.5),
except when its demand is lower than this parameter. All trucks are prepared to do IRs, if
needed.
Overall, the current logistics process results in a low rate of inbound-outbound flow in-
tegration, and the logistics planner has very little visibility about the arrival time to the
customers and the time and characteristics of the inbound loads.
The 3PL assigns a truck driver to each route. Then, the driver is responsible for establishing
the sequence for visiting all outbound customers, the path and schedules, which may or may
not be optimal. The decision of either to visit a supplier (backhaul) or not is often taken
by the driver, based on the extra cost for visiting a known supplier in the vicinity of the
last costumer (linehaul) visited in the route, in case it is doable within the route maximum
duration length set by the 3PL business rules and conditioned by transportation legislation.
There are 75 possible suppliers of wood-chips for IRs. The raw materials may be delivered
back to the mill of origin or alternatively to any of the other three mills owned by the
company in the Iberian Peninsula. Currently, there is no minimum amount of backhauling
required. According to the experience of the planners, the reward for each backhauling can
go up to 10 e/ton.
The computational results were obtained with two major groups of instances (A and B)
built upon the previous case study description, each one of them corresponding to a repre-
sentative operational day. Instances A differ from instances B with respect to the average
distance between the linehauls. The linehauls in instances A are more geographically dis-
persed, with an average distance between linehauls and depot of 461 km, while in instances
B it is 197 km. Baseline instances A30 and B30 correspond to the situation described, with
30 linehauls, a total demand of 1,853 ton, 75 backhauls, 4 mills and 100 available trucks.
Instances among the same instance group differ in the number of linehauls to visit (10, 30
or 50 linehauls) and the number of possible backhauls (0, 25, 50, 75 or 100 backhauls). The
instances were generated in a cumulative manner, i.e., the largest instances contain all loca-
tions considered in smaller instances. The selection of the locations to be included/removed
in the instances was performed randomly from the dataset of the case study.
3.5.2 Comparison among distinct planning strategies

Instances A10 and B10 were used in these experiments to compare and quantify the benefits
of adopting an OBP or IIOP strategy versus DIOP because it is possible to solve the model
quickly to optimality while larger instances require the proposed matheuristic whose gaps
to optimality could bias the results. Furthermore, the resulting routing plan can be easily
visualized.
To perform this comparison, two different reward values were considered (1e/ton and
7e/ton). In order to avoid results biased by different reward values for IRs and OIRs,
the backhauling reward was set regardless of the type of route (whether it was a dedicated
backhaul route or an opportunistic one) and is generically called reward instead of back-
hauling reward. The inbound quantity to be satisfied was set to 160 ton of raw materials,
which corresponds to approximately twice the outbound quantities of finished products in
these instances, taking into account the mills’ productive efficiency. The remaining pa-
rameters remained unchanged throughout the instances, with α = 1, 200 km and ψ = 0.5
ton.
For the IIOP strategy, model [P2] was solved to optimality and a given backhauling amount
was set. The DIOP models [P3] and [P4] were also solved to optimality with this same
backhauling amount to allow a fair comparison. In respect to the OBP strategy, the ra-
tionale to allow its comparison with the remaining strategies consisted in: (i) solving the
OBP model [P0] to optimality, replacing constraints (3.15) by a similar set of constraints
where a maximum (instead of minimum) backhauling amount of 160 ton is set; (ii) solving
model [P4] to obtain the IRs for the differential amount between 160 ton and the already
backhauled amount via OBP; (iii) computing the total costs for these two models.
The obtained results are presented in Table 3.3. In these instances, the matheuristic was not
required, since the computational time for proving optimality in the solver was very short
(less than 5 min on average). In these experiments, the number of binary decision variables
ranged from 10,000 to 17,000.
The analysis of these results shows that the logistics planning strategy leading to the lowest
cost is IIOP in all the experiments. In some cases, the strategy OBP performs better than
DIOP, as intuitively expected, but in others, it does not. This is because the OBP model is
Table 3.3 – Comparison between alternative Inbound and Outbound Planning strategies
Reward Planning Objective Costs (e) No. routes Backhauled No. trucks Runtime
Instance
(e/ton) strategy Function Total Fixed Transport Total OIR OR IR amount (ton) used (s)
Integrated 2,536 3,656 350 3,306 7 0 3 4 160 5 821
A10 Opportunistic 2,714 3,834 420 3,414 7 1 3 3 160 6 45

Decoupled 2,536 3,656 350 3,306 7 0 3 4 160 5 37
7.00 Integrated 768 1,888 280 1,608 4 3 0 1 160 4 36
B10 Opportunistic 771 1,891 280 1,611 4 4 0 0 160 4 32

Decoupled 896 2,016 350 1,666 7 0 3 4 160 5 517
Integrated 3,496 3,656 350 3,306 7 0 3 4 160 5 60
A10 Opportunistic 3,496 3,656 350 3,306 7 0 3 4 160 5 47

Decoupled 3,496 3,656 350 3,306 7 0 3 4 160 5 34
1.00 Integrated 1,802 1,962 350 1,612 5 4 1 0 160 5 21
B10 Opportunistic 1,811 1,971 350 1,621 6 1 2 3 160 5 332

Decoupled 1,856 2,016 350 1,666 7 0 3 4 160 5 243
myopic in the sense that it includes all OIRs that are cost-effective for a given backhauling
reward value, but does not trade-off between OIRs and IRs as it happens with the IIOP
model.
Instance B10 with a reward value of 7e/ton exemplifies a case where the IIOP strategy is
better than OBP and better than DIOP strategies. The total costs of the resulting logistics
plans are 1,888e, 1,891e, and 2,016e respectively. The optimal IIOP routing plan consists
of three OIRs (for trucks k1 , k2 and k5 ) and one IR (for truck k4 ) (Figure 3.4a). While, the
optimal plan for OBP encompasses three OIRs identical to the later plan, and one extra
OIR (k3 ) (Figure 3.4b). The OIR k3 is still cost-effective for that reward value, but it is
costlier than doing the alternative IR k4 as in the IIOP plan. No IRs are foreseen in the
OBP strategy because the routing plan obtained by solving model [P0] already fulfils the
whole demanded backhauled amount; therefore there is no stimulus for finding IRs with
model [P4] afterwards.
The decoupled planning strategy for instance B10 leads to a 6.8% increase of the total costs
when compared with the previous, due to the increase of the transportation costs and also
the use of five vehicles instead of four (Figure 3.4c). The overall routing plan encompasses
four IRs (obtained with model [P4]), plus three ORs (obtained with model [P3]). The IRs
are similar to the ones of vehicle k4 in the IIOP strategy but the ORs are not. This is because
the linehauls are re-distributed in the routes in a different way when the visit to backhauls
is not considered in the same model. For example, linehaul 7 was split deliveries according
to the IIOP and OBP plans due to its geographical proximity to the backhaul 13. This
no longer happens in the decoupled planning, and this linehaul is visited only once in the
course of a longer route that extends up to linehaul 10.
Conversely, instance A10 with a reward value of 7e/ton, exemplifies a case where the
performance of the IIOP strategy is the same as the DIOP strategy (3,656e), and the OBP
performs worse than the other planning strategies (3,834e, 4.9% worse). The optimal IIOP
routing solution consists of three ORs and four IRs (Figure 3.4e). In this setting, with the
linehauls more geographically dispersed and farther from the suppliers and neighbouring
mills, it is cheaper to visit several times supplier 16 in dedicated IRs than considering OIRs.
However, the solution of the OBP model, which is myopic with respect to this possibility,
encompasses one OIR that visits the cost-effective backhaul 22 (Figure 3.4f).
The analysis of these results also shows that the backhauling reward value has a significant
impact on the routing plans and can lead to different conclusions with respect to the com-
parison between the alternative planning strategies. For example, when the reward value is
lowered to 1e/ton, the results for instance B10 show that the visit to backhauls 15 and 16
are no longer cost-effective in the IIOP strategy. Hence, the routing plan consists of four
OIRs, all visiting backhaul 13 (Figure 3.4d) and 1 OR. The total costs are 3.8% higher than
in the experiment with a reward of 7e/ton, due to an increase in the total transportation
distance and in the use of five vehicles instead of four.
It is noteworthy that lowering the value of the reward for visiting the backhauls has a
negative effect on revenue and consequently, increasing the value of the objective function
(134% higher than with previous experiment with 7e). For this case, the IIOP strategy
still performs better than the OBP and DIOP. However, the total cost savings are reduced
to 2.7% and 2.2%, respectively. This is due to the fact that with a lower reward value, the
use of OIRs is less attractive, and the inbound demand must, therefore, be satisfied with
dedicated backhaul routes.
In instance A10, when the reward value is lowered to 1e per ton, there is no visit to a back-
haul that is cost-effective. Hence the optimal plan for the OBP strategy does not consider
any OIR, and it is identical to the IIOP and DIOP strategies described before.
These findings suggest that IIOP is the strategy that allows the optimisation of the com-
bination between backhauling and inbound routing, but under specific circumstances that
favour the supply of raw materials through cost-effective OIRs instead of direct IRs, OBP
can perform better than DIOP. As shown in these experiments, these circumstances depend
on the backhauling reward value for visiting a backhaul in an OIR and on the geographical
configuration of the logistics network of the planning day, especially the relative distance
between a linehaul that can be visited last in a route, and a neighbouring backhaul and mill.
As stated before, the opportunistic planning strategy can be considered an “intermediate”
stage from DIOP towards IIOP. The transition from DIOP to opportunistic planning is
smoother since it is restricted to organisational changes within the local outbound logistics
offices in each mill, while the IIOP also impacts in the central office currently responsible
for inbound logistics planning. During this intermediate stage using the OBP strategy, the
planners need to compare the optimal routing plan with the outcome of the DIOP strategy
in order to establish if backhauling is favourable for the set of customers visited in each
day.
3.5.3 Managerial insights

Focusing on OBP, which is the strategy likelier to be adopted by the planners in this case
study, additional experiments were designed for instances with 10 linehauls (A10, B10), to
provide managerial insights on how the values of key parameters of model [P0] set by the
planners with some degree of uncertainty, may actually impact on the routing plan. The
parameters under study are:
• backhauling reward (δ), i.e., incentive for picking up one unit of raw material in
a backhaul b ∈ B and delivering it to a mill m ∈ M. For simplification purposes,
50 km
14 14
6 6
13 12 13 12
5 15 5 15
8 8
11 11
10 10
4 9 4 9
7 7 𝑘
3 3
𝑘
linehauls (1-10)
2 1 2 1 𝑘
backhauls (13-16)
0 depot/mills (0, 11-12) 0 𝑘
Only neighbouring backhauls

16 to mills are represented 16 𝑘
(a) B10, integrated strategy (7e/ton) (b) B10, opportunistic strategy (7e/ton)
14 14
6 6
13 12 13 12
5 15 5 15
8 8
11 11
10 10
4 9 4 9
7 7
3 3
2 1 2 1
0 0
4×
16 16
(c) B10, decoupled strategies (1 and 7e/ton) (d) B10, integrated strategy (1e/ton)
50 km
14 14
6 6
13 12 13 12
5 15 5 15
11 11
4 4
linehauls (2-6, 17-21)
22 22
backhauls (13-16, 22-23) 𝑘
3 3
depot/mills (0, 11-12) 𝑘
2 Only neighbouring backhauls 2
to mills are represented 𝑘
0 0 𝑘
4× 3×
16 16 𝑘
20 20
17 17
18 23 18 23
19 21 19 21
(e) A10, integrated, decoupled (1 and 7e/ton) and (f) A10, opportunistic (7e/ton)
opportunistic strategies (1e/ton)
Figure 3.4 – Graphical representation of the planning strategies for instances A10 and B10
it is assumed that the reward is the same for all backhauls and mills that accept
compatible types of raw materials. Based on experts opinions, the reward can vary
in a range from 1 to 10 euros per ton;
• minimum backhauled amount (β), i.e., amount of raw materials to be backhauled in

OIR. If parameter β = 0 then no visits to backhauls are required; if β ≥ 1 ton, then, at
least, one backhaul must be included; and if β ≥ 41 ton, then, at least, two backhauls
must be included), since the maximum truck capacity is 40 ton;
• maximum length of the route (α), i.e., maximum distance travelled in a route. The
values tested are 1, 200 km (corresponding to the longest distance from the mill to a
linehaul in instance A10) and 1, 500 km;
• minimum delivery (ψ), i.e., the minimum allowable amount of order delivered to a
linehaul, conditioning the possibility of splitting the order of a linehaul into more
than one deliveries done by different trucks. The values tested were ψ = 0.5 ton,
meaning that many split deliveries are allowed, ψ = 5 ton and ψ = 10 ton (corre-
sponding to the capacity of the smallest truck), meaning that split deliveries are more
restricted.
Let us state the baseline conditions for this analysis δ = 6 e/ton for instance A10 and δ = 1
e/ton for instance B10, α = 1200 km for A10 and α = 400 km for B10, β = 0 and ψ = 0.5
ton. The results of the experiments presented in Table 3.4 and in the Appendix are the basis
for managerial insights that can be valuable for route planning.
Impact of the variation of the value of reward for visiting a backhaul (δ)
Results generically confirm a positive effect in the objective function of increasing the value
of δ, because more OIRs are performed, often with the same number of trucks. The total
transportation costs increase, due to the increase in the total distances travelled, but these
are compensated with a higher total reward collected. The first managerial insight that can
be formulated is that planners wishing to foster an increase of OIRs should set the reward
value at least equal to the extra travelling costs for visiting the most cost-effective backhaul
(i.e. the costs for travelling from the last linehaul to the backhaul and from there to the
closest mill).
For instance A10 the minimum δ should be 7e/ton. Below that value, there is no backhaul
that is cost-effective, hence, no OIRs are included in the optimal routing plan. The number
of trucks needed increases for four to five. Increasing δ to 8 e/ton improve the value of
the objective function but do not change the costs, because the number of trucks and the
routing plan remains the same. However, very high values of δ are not beneficial as it leads
to the use of a large truck fleet. Hence, the percentual increase of total costs is much higher
than the gains in the value of the objective function, and the resulting routing plan is hardly
adopted in practice. For example, a δ equal to 10e/ton in instance A10 leads to costs 243%
higher than in the baseline, corresponding to the highest number of 34 OIRs out of the 36
routes that compose the optimal routing plan.
Regarding instance B10, the linehauls are less geographically dispersed than in A10; thus,
a slight increase in the δ leads to significant changes in the number of OIRs and the im-
provement in the objective function value. In fact, the baseline experiment with a δ equal
to 1e/ton already leads to 3 OIRs and one for each of the trucks used. For δ equal or higher
than 5e/ton the routing plan changes drastically to 39 OIRs requiring 39 extra trucks.
Impact of the variation of the required backhauled amount (β)

Experiments suggest that increasing β has a negative impact on the value of the objective
function because it increases the transportation costs for the mandatory visit to backhaul.
However, in some instances, such as A10, it leads to an increase in OIRs, while in others,
such as B10, it leads to an increase of the number of IRs. A second managerial insight for
planners relates to the fact that the geographical dispersion between the linehauls, back-
hauls and mills is the determining factor for finding the optimal routing plan, as discussed
in Section 5.2. It is also noteworthy that, under some circumstances (e.g. for δ ≤ 2 and β > 0
for A10), the solution turns infeasible because the pre-processing algorithm guarantees that
only cost-efficient integrated routes can be created.
Impact of the variation of the delivery amount at a linehaul (ψ)

Experiments indicate that increasing ψ has a slightly negative effect on the value of the
objective function. Although this may imply the use of fewer vehicles, this also decreases
the possibility of creating integrated routes and, as such, the possibility of collecting a
higher total reward.
There is a complementary relation between the key parameters ψ and δ in fostering the
number of OIRs in the optimal routing plan. In practice, if ψ is low, means that more visits
to the linehauls are allowed, and so, there is more flexibility in the routing plan to include
OIR, especially if the reward for visiting a backhaul is high. In fact, the number of OIRs
is maximized (34 out of 35 routes) if ψ is very low (e.g., 0.5 ton) and δ is very high (e.g.,
10e). However, these high number of integrated routes (e.g. 34 out of 36 routes) can
hardly represent the common practice (Figure 3.5). Hence, another managerial insight for
planners relates to the importance of properly addressing the trade-off between the offered
reward and the maximum number of visits allowed to a linehaul, which is specific for each
case.
Impact of the variation of the maximum length of the route (α)

Experiments show that increasing α tends to improve the objective function, due to the
decrease in the number of required vehicles and the possibility to visit a larger number of
linehauls is the same route. However, without a direct impact on the number of OIR. As an
example for instance A10, increasing α from 1, 200 km to 1, 500 km, all other parameters
remaining the same as in the baseline scenario, lead to a decrease of 32% in the value of
the objective function, related with the use of 3 vehicles instead of 4. In instance B10, the
increase from 400 km to 800 km, leads to a decrease of 22% in the value of the objective
function due to longer routes, using the same fleet of 3 trucks.
In summary, results show that there are several trade-offs that need to be analysed by plan-
ners to balance the increase of OIRs and the increase in transportation costs. In particular,
results suggest that α is the parameter that impacts the most in improving the value of the
objective function and costs (improvements of 32% in instance A10 e 22% in B10, because
it enables to use fewer vehicles, and fewer distances travelled, however, do not necessarily
foster OIRs.
Moreover, the main parameter to be taken into account for planners willing to improve
OIRs is δ. As discussed before, OIRs tend to be included in the routing plan when the
reward value is above a threshold, corresponding to visiting the first cost-effective back-
haul. The value of this threshold depends on the geographical dispersion of nodes in the
transportation network and particularly the distance between the last visited linehaul, the
closest backhaul and its neighbouring mill.
35 5000
30
4000
25
Objective function
3000
Number of routes
20
15 2000
10
1000
5
0 0
0 2 5 6 7 8 10 0 2 5 6 7 8 10 0 2 5 6 7 8 10
Ψ = 0.5 Ψ=5 Ψ = 10
Reward 𝛿 for different Ψ
# OR # OIR Objective function
Figure 3.5 – Impact of the variation of the reward (δ) and minimum delivery (ψ)
(α = 1, 200, β = 0)
3.5.4 Performance of the solution approach

Despite the fact that the solver is able to obtain optimal solutions within a few minutes for
problem instances of 10 linehauls, this is not the case for larger instances. In these cases,
the use of the matheuristic is justified in order to obtain good quality solutions in a shorter
computational time. A set of computational experiments was envisaged to validate the pro-
posed solution approach. Instances of group A and B were solved using the standalone MIP
solver approach and the fix-and-optimise matheuristic. These experiments were performed
in an Intel Xeon E5-2450 @ 2.10GHz CPU with capacity for 16 simultaneous processing
threads.
Both approaches were run for 3,600s, with α = 1, 200, β = 0 and ψ = 0.5. The MIP solver
was executed once for each instance using Gurobi’s default parameters and the fix-and-
Table 3.4 – Summary of the experiments on the impact of the values of model parameters
Parameter Costs (e) Routes

Objective Runtime MIP
Inst.
α β ψ δ Function Total Fixed Transport Total OIR OR (s) Gap
1,200 0 0.5 6 4,703 4,703 280 4,423 4 0 4 19 1.8%
– – – 7 -1% +5% +25% +4% +1 +1 0 105 0.6%
– – – 8 -2% +5% +25% +4% +1 +1 0 112 0.3%
– – – 10 -46% +243% +800% +208% +32 +34 -2 272 1.9%
– 1 – – 0% +5% +25% +4% +1 +1 0 19 1.0%
– 41 – – +2% +12% +50% +10% +2 +2 0 190 0.0%
A10
1,500 – – – -32% -32% -25% -32% -1 0 -1 26 1.6%
– – 5 – +1% +6% +25% +4% +1 +1 0 28 2.0%
– – 5 10 -7% +35% +75% +33% +3 +5 -2 304 0.4%
– – 10 – +1% +6% +25% +4% +1 +1 0 130 1.9%
– – 10 10 -5% +21% +25% +21% +1 +3 -2 308 0.7%
400 0 0.5 1 2,002 2,002 210 1,792 3 3 0 16 0.0%
– – – 2 0% +4% +33% +1% +1 -2 +3 28 0.0%
– – – 5 -105% +284% +1,300% +165% +39 +36 +3 32 0.0%
– 1 – – +2% +4% +33% +1% +1 -2 +3 39 0.0%
– 41 – – +7% +11% +67% +4% +2 -1 +3 41 0.0%
B10 800 – – – -22% -20% 0% -22% 0 -2 +2 72 0.0%
– – 5 – 0% 0% 0% 0% 0 -3 +3 25 0.0%
– – 5 5 -23% +57% +267% +32% +8 +5 +3 20 0.0%
– – 10 – 0% 0% 0% 0% 0 -3 +3 29 0.0%
– – 10 5 -15% +25% +133% +13% +4 +1 +3 18 0.0%
Legend: α: maximum length of the route (km); β: minimum backhauled amount (ton); ψ: min-
imum delivery amount (ton); δ: backhauling reward (e/ton); Runtime: computational time after
which no better solution was obtained (seconds); MIP Gap: percentual difference obtained by
Gurobi between the upper and lower bound of the branch-and-bound method. All models were
run with a maximum time limit of 3,600s. The first row of results for each instance (highlighted
in bold) contains the baseline values, and the rows that follow exhibit either the absolute or the
percentual variation compared with the baseline values (except for Runtime and MIP Gap values).
optimise approach was run 10 times for each instance, using the parameters described in
Table 3.5.
Table 3.5 – Used parameters for the matheuristic approach
Parameters Value
Termination criteria Time limit 3,600s
No-improvement criterion Improvement between 100
consecutive iterations lower than
RouteRelease sub-problem Subproblem sizes 4, 6, 8, 16 routes
No-improvement limit to 2 iterations
change subproblem size
MIP solver iteration time limit Multiples of 15s
(according to sub-problem size)
LocationRelease sub-problem Subproblem sizes 2, 4, 6, 8 linehauls
No-improvement limit to 2 iterations
change subproblem size
MIP solver iteration time limit Multiples of 15s
(according to sub-problem size)
Table 3.6 summarizes the computational results of the MIP solver and matheuristic ap-
proaches for the 30 problem instances.
The results demonstrate that both the MIP solver and the matheuristic are adequate for
solving instances up to 10 linehauls (groups A10 and B10), as the solver is able to prove
optimality for most instances and the matheuristic easily reaches the same solution as the
MIP solver. For larger instances, the MIP solver yields optimality gaps up to 32% for
instances of groups A30, A50, B30 and B50. Specifically to instances of group A, it is
possible to observe the increase in the number of routes that perform backhauling as the
number of backhaul locations progressively increases. In instances of group B30, a single
OIR is used when backhauling is allowed, and in group B50 no opportunistic backhauling
is performed. However, the obtained solutions by the MIP solver when the number of
backhauls increases do not necessarily improve, contrarily to what would be expected.
Furthermore, for instances of group B50, the solver is unable to find a single feasible
solution within the 1-hour limit for 4 out of the 5 instances. This fact is probably due to the
increase in model size and complexity when more backhaul locations are being considered,
thus requiring more time for Gurobi to reach identical solutions when exploring the branch-
and-bound tree.
The matheuristic approach exhibits small standard deviation values for the 10 repetitions
performed for each instance, thus suggesting that the obtained results are robust. The
negative percentual difference values between the solver and the matheuristic suggest that
the matheuristic is able to converge correctly to better solutions, as opposed to the solver,
which exhibits very high optimality gaps. This negative percentual difference tends to be
increasingly more expressive with the increase in instance size. Furthermore, results also
suggest that the matheuristic also takes better advantage of the increase in the number of
backhaul locations, as the objective function values generally decrease when the number
of backhaul locations increases.
Table 3.6 – Computational results of the MIP and matheuristic approaches for 30 problem
instances
Problem instance MIP Solver Fix-and-optimise
Objective % diff.
P P MIP No. No. Function No. No.
Group |L| |B| |K| qi Qk OF Runtime Runtime
Gap routes OIRs routes OIRs
Standard
Average
Deviation
10 0 100 82 2,600 3,215 0.0% 50 3 0 3,215 0 31 3 0 0.0%
10 25 100 82 2,600 3,215 0.0% 205 3 0 3,215 0 37 3 0 0.0%
A10 10 50 100 82 2,600 3,215 0.0% 60 3 0 3,215 0 28 3 0 0.0%

10 75 100 82 2,600 3,215 0.0% 55 3 0 3,215 0 30 3 0 0.0%
10 100 100 82 2,600 3,215 0.0% 31 3 0 3,215 0 33 3 0 0.0%
30 0 100 1,853 2,600 13,745 18.4% 3,110 53 0 13,627 19 2,957 53 0 -0.9%
30 25 100 1,853 2,600 13,696 19.6% 3,444 54 26 13,404 20 2,885 53 25 -2.1%
A30 30 50 100 1,853 2,600 13,701 20.1% 2,758 54 32 13,374 21 2,879 53 31 -2.4%
30 75 100 1,853 2,600 13,833 20.9% 3,187 55 33 13,366 19 2,576 53 31 -3.4%
30 100 100 1,853 2,600 13,702 20.1% 1,877 56 34 13,369 20 2,731 53 31 -2.4%
50 0 100 2,061 2,600 20,986 29.9% 2,968 65 0 19,465 363 3,463 64 0 -7.2%
50 25 100 2,061 2,600 19,717 26.5% 1,902 64 30 19,019 194 3,527 63 26 -3.5%
A50 50 50 100 2,061 2,600 19,907 27.3% 3,469 65 37 19,079 313 3,397 64 31 -4.2%
50 75 100 2,061 2,600 20,829 30.5% 3,019 65 36 19,167 423 3,464 64 32 -8.0%
50 100 100 2,061 2,600 21,365 32.3% 2,653 66 33 19,116 327 3,513 65 36 -10.5%
10 0 100 82 2,600 1,575 1.9% 129 3 0 1,575 0 39 3 0 0.0%
10 25 100 82 2,600 1,575 2.7% 117 3 0 1,575 0 39 3 0 0.0%
B10 10 50 100 82 2,600 1,575 0.0% 53 3 0 1,575 0 39 3 0 0.0%

10 75 100 82 2,600 1,575 0.0% 51 3 0 1,575 0 44 3 0 0.0%
10 100 100 82 2,600 1,575 0.0% 51 3 0 1,575 0 61 3 0 0.0%
30 0 100 818 2,600 8,695 20.1% 3,032 23 0 8,210 9 2,267 22 0 -5.6%
30 25 100 818 2,600 8,626 19.4% 3,354 23 1 8,212 15 2,730 21 1 -4.8%
B30 30 50 100 818 2,600 9,162 30.5% 2,473 23 0 8,206 12 2,292 22 1 -10.4%
30 75 100 818 2,600 8,350 16.8% 3,262 21 1 8,206 10 2,876 21 1 -1.7%
30 100 100 818 2,600 9,003 32.1% 441 23 1 8,198 2 2,474 22 1 -8.9%
50 0 100 2,054 2,600 – – – – – 45,393 956 3,419 68 0 –
50 25 100 2,054 2,600 – – – – – 45,325 793 3,420 69 0 –
B50 50 50 100 2,054 2,600 – – – – – 45,759 526 3,388 64 0 –

50 75 100 2,054 2,600 46,991 13.8% 2,830 69 0 45,592 655 3,359 65 0 -3.0%
50 100 100 2,054 2,600 – – – – – 45,768 646 3,412 67 0 –
Average -3.0%
P
Legend: |L|: number of linehauls to be visited; |B|: number of possible backhauls; |K| total number of vehicles available; qi : total quantity to be delivered to linehauls
P
(ton); Qk : total vehicle transportation capacity (ton); OF: Final value of the objective function (e); Runtime: average computational time after which no better solution
was obtained (seconds); MIP Gap: Percentual difference obtained by Gurobi between the upper and lower bounds of the branch-and-bound method; No. routes: total number
of routes; No. OIR: number of routes that include visit to a backhaul; % diff.: percentual difference of the fix-and-optimise average objective function towards Gurobi’s
objective function (a negative difference favours the matheuristic). For the fix-and-optimise approach, the route indicators correspond to the repetition whose objective
function value was closest to the obtained average.
From these results, we can say that the proposed matheuristic approach is adequate for
solving the problem at hand. For instances of considerable size, the MIP solver starts
to struggle in finding feasible solutions in an acceptable time limit, and apparently also
has greater difficulties taking advantage of backhauling, while the matheuristic is able to
decrease the overall logistics costs with an increase in the number of backhaul locations,
therefore yielding more consistent results.
3.6. Conclusions and future work

Integrating planning processes requires a thorough assessment of both quantitative benefits
pertaining to the expected decrease in the related costs and qualitative impacts related to the
usual need of breaking functional silos. This work explored, mainly, the quantitative aspect
of integrating outbound and inbound logistics routes. We used as a background a case-
study from the wood-based panel industry, but the results and approaches developed are
generalisable for other settings in which this integration can be modelled as an rVRPB (e.g.,
grocery retail, cement distribution). Besides modelling three possible planning strategies
(i.e., OBP, IIOP, and DIOP), we have also developed a matheuristic to tackle real-world
instances of this problem.
Three key conclusions emerge from our computational study. Firstly, the intuitive idea that
intermediate levels of integration would always result in better planning outcomes was not
verified. DIOP outperforms OBP in certain geographical contexts where the distribution
network is more dispersed. In our studies, this happened in instance A10 when the average
distance between linehaul customers and the depot of origin is 197 km. In this case, it was
actually cheaper to assure the supply of raw material through dedicated inbound routes (i.e.
going to and from the nearest supplier) than including a visit to a supplier at the end of the
outbound route (i.e. after visiting all customers). The IIOP model does this trade-off, but
the OBP model is myopic to the possibility of doing direct inbound routes, hence, leading
to worse results than DIOP.
Secondly, we confirm that there are important parameters dealt by the planners with some
degree of uncertainty that actually can have a great influence on the total costs of the rout-
ing plan. This study analysed four of these parameters - backhauling reward, minimum
backhauled amount, maximum length of the route and minimum delivery amount allowed.
Results suggest that increasing the maximum length of the route leads to the largest impact
in the performance of the routing plan but including a quantitative reward for each supplier
visited will likelier increase the proportion of integrated inbound and outbound routes in
the overall routing plan. In fact, the total reward (e/ton) should be equal or higher than
the extra transportation costs for the most cost-effective supplier. Meaning that the extra
distance travelled empty from the last customer to the nearest supplier and then full from
there to the neighbouring mill is minimized.
Finally, the developed matheuristic proved to be a suitable approach to tackle this problem
and this fact reiterated the interest of fix-and-optimise to solve routing problems.
Future work could be devoted to merging the qualitative and quantitative assessments re-
lated to the integration of planning processes. In particular, the study of integrated inbound
Bibliography 95
and outbound routes is of interest due to its potential in improving the ever-relevant sus-
tainability dimension.
Acknowledgements
This work is financed by the ERDF - European Regional Development Fund through the
Operational Programme for Competitiveness and Internationalisation - COMPETE 2020
Programme and by National Funds through the Portuguese funding agency, FCT - Fun-
dação para a Ciência e Tecnologia within project POCI-01-0145-FEDER-016733.
Further funding was received from the Portugal 2020 project “DM4Manufacturing - Align-
ing manufacturing decision making with advanced manufacturing technologies”, POCI-01-
0145-FEDER-016418, financed by UE/FEDER through the program COMPETE2020.
The second author also acknowledges the support of the Ph.D. Grant SFRH/BD/136314/2018,
awarded by the Portuguese Science and Technology Foundation (FCT) and financed by the
European Social Fund (ESF) and by National Funds of the Portuguese Ministry of Science,
Technology and Higher Education (MCTES) through the Human Capital Operational Pro-
gramme (POCH).
We would also like to thank the wood-based products company for the cooperation and for
supplying all the data.
Bibliography
Adulyasak, Y., Cordeau, J.-F., and Jans, R. (2015). The production routing problem: A
review of formulations and solution algorithms. Computers & Operations Research,
55:141–152.
Amorim, P., Günther, H.-O., and Almada-Lobo, B. (2012). Multi-objective integrated

production and distribution planning of perishable products. International Journal of
Production Economics, 138(1):89–101.
Amorim, P., Marques, A., and Oliveira, B. (2014). Wood-based products distribution plan-
ning with wood returns in a iberian company. In MOSIM 2014, 10ème Conférence
Francophone de Modélisation, Optimisation et Simulation, Nancy, France.
Aras, N., Aksen, D., and Tekin, M. T. (2011). Selective multi-depot vehicle routing prob-
lem with pricing. Transportation Research Part C: Emerging Technologies, 19(5):866–
884.
Archetti, C. and Speranza, M. G. (2015). The inventory routing problem: the value of
integration. International Transactions in Operational Research, 23(3):393–407.
Audy, J.-F., Lehoux, N., DAmours, S., and Rönnqvist, M. (2010). A framework for an
Bektaş, T., Laporte, G., and Vigo, D. (2015). Integrated vehicle routing problems. Com-
puters & Operations Research, 55:126.
Bouchard, M., D’Amours, S., Rönnqvist, M., Azouzi, R., and Gunn, E. (2017). Integrated
optimization of strategic and tactical planning decisions in forestry. European Journal
of Operational Research, 259(3):1132–1143.
Caceres-Cruz, J., Arias, P., Guimarans, D., Riera, D., and Juan, A. A. (2014). Rich vehicle
routing problem. ACM Computing Surveys, 47(2):1–28.
Carlsson, D. and Rönnqvist, M. (2007). Backhauling in forest transportation: models,

methods, and practical usage. Canadian Journal of Forest Research, 37(12):2612–2623.
Chávez, J. J. S., Escobar, J. W., Echeverri, M. G., and Meneses, C. A. P. (2015). A meta-
heuristic ACO to solve the multi-depot vehicle routing problem with backhauls. Inter-
national Journal of Industrial Engineering and Management, 6(2):49–58.
Darvish, M., Archetti, C., Coelho, L. C., and Speranza, M. G. (2019). Flexible two-echelon
location routing problem. European Journal of Operational Research, 277(3):1124–
1136.
Davis, L. B., Sengul, I., Ivy, J. S., Brock, L. G., and Miles, L. (2014). Scheduling food bank
collections and deliveries to ensure food safety and improve access. Socio-Economic
Planning Sciences, 48(3):175–188.
Deif, I. and Bodin, L. D. (1984). Extension of the Clarke and Wright Algorithm for Solv-
ing the Vehicle Routing Problem with Backhauling. In Proceedings of the Babson Con-
ference on Software Uses in Transportation and Logistic Management, pages 75–96,
Babson Parl.
Derigs, U., Pullmann, M., Vogel, U., Oberscheider, M., Gronalt, M., and Hirsch, P. (2012).
Multilevel neighborhood search for solving full truckload routing problems arising in
timber transportation. Electronic Notes in Discrete Mathematics, 39:281–288.
Dominguez, O., Guimarans, D., Juan, A. A., and de la Nuez, I. (2016). A biased-
randomised large neighbourhood search for the two-dimensional vehicle routing prob-
lem with backhauls. European Journal of Operational Research, 255(2):442–462.
Eguia, I., Racero, J., Molina, J. C., and Guerrero, F. (2013). Environmental issues in vehicle
routing problems. In Sustainability Appraisal: Quantitative Methods and Mathematical
Techniques for Environmental Performance Evaluation, pages 215–241. Springer Berlin
Heidelberg.
hauls. European Journal of Operational Research, 42(1):39–51.
Gribkovskaia, I., Laporte, G., and Shyshou, A. (2008). The single vehicle routing problem
with deliveries and selective pickups. Computers & Operations Research, 35(9):2908–
2924.
Bibliography 97
Gutiérrez-Jarpa, G., Desaulniers, G., Laporte, G., and Marianov, V. (2010). A branch-and-
price algorithm for the vehicle routing problem with deliveries, selective pickups and
time windows. European Journal of Operational Research, 206(2):341–349.
Gutiérrez-Jarpa, G., Marianov, V., and Obreque, C. (2009). A single vehicle routing prob-
lem with fixed delivery and optional collections. IIE Transactions, 41(12):1067–1079.
Hansen, P., Mladenović, N., and Perez-Britos, D. (2001). Variable neighborhood decom-
position search. Journal of Heuristics, 7(4):335–350.
Hirsch, P. (2011). Minimizing empty truck loads in round timber transport with tabu search
strategies. International Journal of Information Systems and Supply Chain Management,
4(2):15–41.
Küçükoğlu, İ. and Öztürk, N. (2013). A differential evolution approach for the vehicle
routing problem with backhauls and time windows. Journal of Advanced Transportation,
48(8):942–956.
Lahyani, R., Khemakhem, M., and Semet, F. (2015). Rich vehicle routing problems: From
a taxonomy to a definition. European Journal of Operational Research, 241(1):1–14.
Lai, M., Battarra, M., Francesco, M. D., and Zuddas, P. (2015). An adaptive guidance meta-
heuristic for the vehicle routing problem with splits and clustered backhauls. Journal of
the Operational Research Society, 66(7):1222–1235.
Lai, M., Crainic, T. G., Francesco, M. D., and Zuddas, P. (2013). An heuristic search for the
routing of heterogeneous trucks with single and double container loads. Transportation
Larrain, H., Coelho, L. C., and Cataldo, A. (2017). A variable MIP neighborhood descent
algorithm for managing inventory and distribution of cash in automated teller machines.
Computers & Operations Research, 85:22–31.
Marques, A. S., Audy, J. F., D’Amours, S., and Rönnqvist, M. (2014). Tactical and op-
erational harvest planning. In The Management of Industrial Forest Plantations, pages
239–267. Springer Netherlands.
Neves-Moreira, F., Almada-Lobo, B., Cordeau, J.-F., Guimarães, L., and Jans, R. (2019).
Solving a large multi-product production-routing problem with delivery time windows.
Omega, 86:154–172.
Neves-Moreira, F., Amorim, P., Guimarães, L., and Almada-Lobo, B. (2016). A long-
haul freight transportation problem: Synchronizing resources to deliver requests passing
through multiple transshipment locations. European Journal of Operational Research,
248(2):487–506.
Nguyen, P. K., Crainic, T. G., and Toulouse, M. (2016). Multi-trip pickup and deliv-
ery problem with time windows and synchronization. Annals of Operations Research,
253(2):899–934.
Nikolakopoulos, A. (2014). A metaheuristic reconstruction algorithm for solving bi-level

vehicle routing problems with backhauls for army rapid fielding. In Operations Re-
search/Computer Science Interfaces Series, pages 141–157. Springer International Pub-
lishing.
Oesterle, J. and Bauernhansl, T. (2016). Exact method for the vehicle routing problem with
mixed linehaul and backhaul customers, heterogeneous fleet, time window and manufac-
turing capacity. Procedia CIRP, 41:573–578.
Paraphantakul, C., Miller-Hooks, E., and Opasanon, S. (2012). Scheduling deliveries with
backhauls in thailands cement industry. Transportation Research Record: Journal of the
Parragh, S. N., Doerner, K. F., and Hartl, R. F. (2008). A survey on pickup and delivery
problems. Journal für Betriebswirtschaft, 58(1):21–51.
Rieck, J., Ehrenberg, C., and Zimmermann, J. (2014). Many-to-many location-routing

with inter-hub transport and multi-commodity pickup-and-delivery. European Journal
Ropke, S. and Pisinger, D. (2006). A unified heuristic for a large class of vehicle routing
problems with backhauls. European Journal of Operational Research, 171(3):750–775.
Sahling, F., Buschkühl, L., Tempelmeier, H., and Helber, S. (2009). Solving a multi-level
capacitated lot sizing problem with multi-period setup carry-over via a fix-and-optimize
heuristic. Computers & Operations Research, 36(9):2546–2553.
Salhi, S., Wassan, N., and Hajarat, M. (2013). The fleet size and mix vehicle routing prob-
lem with backhauls: Formulation and set partitioning-based heuristics. Transportation
Soares, R., Marques, A., Amorim, P., and Rasinmäki, J. (2019). Multiple vehicle synchro-
nisation in a full truck-load pickup and delivery problem: A case-study in the biomass
supply chain. European Journal of Operational Research, 277(1):174–194.
Speranza, M. G. (2018). Trends in transportation and logistics. European Journal of

Operational Research, 264(3):830–836.
Toth, P. and Vigo, D. (2014). Vehicle Routing: Problems, Methods, and Applications,
Second Edition. SIAM - Society for Industrial and Applied Mathematics.
Wade, A. C. and Salhi, S. (2002). An investigation into a new class of vehicle routing
problem with backhauls. Omega, 30(6):479–487.
Bibliography 99
Wassan, N., Wassan, N., Nagy, G., and Salhi, S. (2017). The multiple trip vehicle routing
problem with backhauls: Formulation and a two-level variable neighbourhood search.
Yazgitutuncu, G., Carreto, C., and Baker, B. (2009). A visual interactive approach to
classical and mixed vehicle routing problems with backhauls. Omega, 37(1):138–154.
Yu, M. and Qi, X. (2014). A vehicle routing problem with multiple overlapped batches.
Transportation Research Part E: Logistics and Transportation Review, 61:40–55.
Appendix 3.A Supplementary material

3.A. Supplementary material 101
Table 3.7 – Experiments on the impact of the values of model parameters

Inst.
A10 1200 0 0.5 6 4703 4703 280 4423 4 0 4 19 1.8%
A10 1200 0 0.5 7 4665 4945 350 4595 5 1 4 105 0.6%
A10 1200 0 0.5 8 4625 4945 350 4595 5 1 4 112 0.3%
A10 1200 0 0.5 10 2523 16123 2520 13603 36 34 2 272 1.9%
A10 1200 0 5 6 4732 4972 350 4622 5 1 4 28 2.0%
A10 1200 0 5 7 4692 4972 350 4622 5 1 4 219 0.6%
A10 1200 0 5 8 4652 4972 350 4622 5 1 4 123 0.3%
A10 1200 0 5 10 4356 6356 490 5866 7 5 2 304 0.4%
A10 1200 0 10 6 4732 4972 350 4622 5 1 4 130 1.9%
A10 1200 0 10 7 4692 4972 350 4622 5 1 4 175 1.1%
A10 1200 0 10 8 4652 4972 350 4622 5 1 4 237 0.5%
A10 1200 0 10 10 4482 5682 350 5332 5 3 2 308 0.7%
A10 1200 1 0.5 6 4705 4945 350 4595 5 1 4 19 1.0%
A10 1200 1 0.5 7 4665 4945 350 4595 5 1 4 109 0.8%
A10 1200 1 0.5 8 4625 4945 350 4595 5 1 4 153 0.2%
A10 1200 1 0.5 10 2523 16123 2520 13603 36 34 2 2523 0.7%
A10 1200 1 5 6 4732 4972 350 4622 5 1 4 137 0.5%
A10 1200 1 5 7 4692 4972 350 4622 5 1 4 357 0.7%
A10 1200 1 5 8 4652 4972 350 4622 5 1 4 241 0.6%
A10 1200 1 5 10 4356 6356 490 5866 7 5 2 165 0.2%
A10 1200 1 10 6 4732 4972 350 4622 5 1 4 96 0.2%
A10 1200 1 10 7 4692 4972 350 4622 5 1 4 57 0.4%
A10 1200 1 10 8 4652 4972 350 4622 5 1 4 200 0.1%
A10 1200 1 10 10 4482 5682 350 5332 5 3 2 94 0.5%
A10 1200 41 0.5 6 4802 5282 420 4862 6 2 4 190 0.0%
A10 1200 41 0.5 7 4722 5282 420 4862 6 2 4 226 0.1%
A10 1200 41 0.5 8 4642 5282 420 4862 6 2 4 126 0.2%
A10 1200 41 0.5 10 2523 16123 2520 13603 36 34 2 185 0.8%
A10 1200 41 5 6 4828 5308 420 4888 6 2 4 234 0.5%
A10 1200 41 5 7 4748 5308 420 4888 6 2 4 155 0.5%
A10 1200 41 5 8 4668 5308 420 4888 6 2 4 180 0.6%
A10 1200 41 5 10 4356 6356 490 5866 7 5 2 107 0.6%
A10 1200 41 10 6 4830 5310 350 4960 5 2 3 166 0.3%
A10 1200 41 10 7 4750 5310 350 4960 5 2 3 57 0.4%
A10 1200 41 10 8 4670 5310 350 4960 5 2 3 89 0.7%
A10 1200 41 10 10 4482 5682 350 5332 5 3 2 67 0.1%
A10 1500 0 0.5 6 3215 3215 210 3005 3 0 3 26 1.6%
A10 1500 0 0.5 7 3177 3457 280 3177 4 1 3 357 1.0%
A10 1500 0 0.5 8 3137 3457 280 3177 4 1 3 123 2.3%
A10 1500 0 0.5 10 1034 14634 2450 12184 35 34 1 504 0.3%
Legend: α: maximum length of the route (km); β: minimum backhauled amount (ton); ψ: minimum delivery amount (ton); δ: backhauling
reward (e/ton); Runtime: computational time after which no better solution was obtained (seconds); MIP Gap: percentual difference
obtained by Gurobi between the upper and lower bound of the branch-and-bound method. All models were run with a maximum time limit
of 3,600s.
Table 3.7 (cont.) – Experiments on the impact of the values of model parameters

Inst.
B10 400 0 0.5 1 2002 2002 210 1792 3 0 3 16 0.0%
B10 400 0 5 1 2002 2002 210 1792 3 0 3 25 0.0%
B10 400 0 10 1 2002 2002 210 1792 3 0 3 29 0.0%
B10 400 0 0.5 2 2001 2081 280 1801 4 1 3 28 0.0%
B10 400 0 5 2 2001 2081 280 1801 4 1 3 21 0.0%
B10 400 0 10 2 2002 2002 210 1792 3 0 3 18 0.0%
B10 400 0 0.5 5 -108 7692 2940 4752 42 39 3 32 0.0%
B10 400 0 5 5 1534 3134 770 2364 11 8 3 20 0.0%
B10 400 0 10 5 1710 2510 490 2020 7 4 3 18 0.0%
B10 400 1 0.5 1 2041 2081 280 1801 4 1 3 39 0.0%
B10 400 1 5 1 2041 2081 280 1801 4 1 3 37 0.0%
B10 400 1 10 1 2059 2099 280 1819 4 1 3 21 0.0%
B10 400 1 0.5 2 2001 2081 280 1801 4 1 3 33 0.0%
B10 400 1 5 2 2001 2081 280 1801 4 1 3 36 0.0%
B10 400 1 10 2 2019 2099 280 1819 4 1 3 30 0.0%
B10 400 1 0.5 5 -108 7692 2940 4752 42 39 3 27 0.0%
B10 400 1 5 5 1534 3134 770 2364 11 8 3 19 0.0%
B10 400 1 10 5 1710 2510 490 2020 7 4 3 16 0.0%
B10 400 41 0.5 1 2133 2213 350 1863 5 2 3 41 0.0%
B10 400 41 5 1 2133 2213 350 1863 5 2 3 32 0.0%
B10 400 41 10 1 2133 2213 350 1863 5 2 3 20 0.0%
B10 400 41 0.5 2 2053 2213 350 1863 5 2 3 27 0.0%
B10 400 41 5 2 2053 2213 350 1863 5 2 3 20 0.0%
B10 400 41 10 2 2053 2213 350 1863 5 2 3 10 0.0%
B10 400 41 0.5 5 -108 7692 2940 4752 42 39 3 38 0.0%
B10 400 41 5 5 1534 3134 770 2364 11 8 3 40 0.0%
B10 400 41 10 5 1710 2510 490 2020 7 4 3 14 0.0%
B10 800 0 0.5 1 1566 1606 210 1396 3 1 2 72 0.0%
B10 800 0 0.5 2 1503 1743 210 1533 3 3 0 23 0.0%
B10 800 0 0.5 5 -778 7222 2800 4422 40 40 0 54 0.0%
Legend: α: maximum length of the route (km); β: minimum backhauled amount (ton); ψ: minimum delivery amount (ton); δ: backhauling
reward (e/ton); Runtime: computational time after which no better solution was obtained (seconds); MIP Gap: percentual difference
obtained by Gurobi between the upper and lower bound of the branch-and-bound method. All models were run with a maximum time limit
of 3,600s.
Chapter 4
A Robust Vehicle Routing Problem

with Backhauls
A robust optimization approach for the vehicle rout-

ing problem with selective backhauls
Maria João Santos ∗† · Eduardo Curcio∗ · Mauro Henrique Mulati‡ · Pedro Amorim∗†
· Flávio Keidi Miyazawa‡
Published in Transportation Research Part E: Logistics and Transportation Review, 2020
Abstract The Vehicle Routing Problem with Selective Backhauls (VRPSB) aims to min-
imize the total routing costs minus the total revenue collected at backhaul customers. We
explore a VRPSB under uncertain revenues. A deterministic VRPSB is formulated as a
mixed-integer programming problem and two robust counterparts are derived. A novel
method to estimate the probabilistic bounds of constraint violation is designed. A robust
metaheuristic is developed, requiring little time to obtain feasible solutions with average
gap of 1.26%. The robust approach studied demonstrates high potential to tackle the prob-
lem, requiring similar computing effort and maintaining the same tractability as the deter-
ministic modeling.
Keywords vehicle routing problem with selective backhauls · robust optimization · branch-
and-cut · adaptive large neighborhood search
4.1. Introduction
The Vehicle Routing Problem with Backhauls (VRPB) is an extension of the classic Vehi-
cle Routing Problem (VRP), where two different types of customer are considered in the
network: linehaul customers, those who receive goods from a depot (outbound logistics),
and backhaul customers, those who send goods back to the depot (inbound logistics). Inte-
grated inbound-outbound logistics planning can reduce the distance of empty trips, which
are responsible for significant transportation costs (Liu and Chung, 2008), thus increasing
‡ Instituto de Computação - Universidade Estadual de Campinas (Unicamp) and Departamento de Ciência
da Computação – Universidade Estadual do Centro-Oeste (Unicentro), Brazil

104 Chapter 4. A Robust Vehicle Routing Problem with Backhauls
the efficiency of both vehicles and routes, and reducing fuel consumption and emissions of
air pollutants (Pradenas et al., 2013).
One approach whereby the VRPB includes an optional selection of backhauls is the Vehicle
Routing Problem with Selective Backhauls (VRPSB) (Baldacci et al., 2010). In this type
of problems, the backhauls are selected according to the revenues they provide. The first
application of the VRPSB is described in Yano et al. (1987) to solve a real transportation
problem of a company operating in a retail chain. The company owned a fleet of vehicles
and had to plan delivery routes from the distribution center to some stores and pickup
routes from vendors to the distribution center. Since the private fleet could not cover all
the vendors, due to capacity limitations, the company outsourced carriers to pickup the
material from the remaining ones. Therefore, one of the decisions of the problem was to
select which backhauls to visit. By optimizing the VRPSB, the company was able to save
a considerable amount of money.
Most of the exact approaches used to solve VRP rely on cut and column generation algo-
rithms (Poggi and Uchoa, 2014; Baldacci et al., 2012). Therefore, these are also the major-
ity of exact solution methods used to solve both VRPB and VRPSB. Gutiérrez-Jarpa et al.
(2009) formulate a VRPSB as a mixed integer linear programming model and use a branch-
and-price algorithm to solve adapted instances from Solomon (1987). Only those instances
with 50 customers or less could be optimally solved. Very recently, Granada-Echeverri
et al. (2019) model a VRPB with a mixed integer linear programming formulation based
on characteristics of an Open VRP, allowing to obtain several new best known solutions
for the instances of Goetschalckx and Jacobs-Blecha (1989) with up to 113 customers and
of Toth and Vigo (1997) with up to 100 customers.
Due to high computational performance, heuristics are the most used algorithms to solve
the majority of VRPB and its variants. For the VRPSB, García-Nájera et al. (2015) pro-
poses a similarity-based selection multi-objective evolutionary algorithm (SSMOEA) to
solve a multi-objective problem considering minimization of number of vehicles, travel
distance and uncollected revenue. The metaheuristic was tested for adapted instances of
Goetschalckx and Jacobs-Blecha (1989), providing solutions for instances with up to 150
customers in the range of 1000 to 8000 seconds.
Notwithstanding the practical application of the VRPSB to real-world problems (e.g., par-
cel services), few works have been described in the literature that use this variant of the
VRP (Gutiérrez-Jarpa et al., 2009). This statement is also valid for other similar problems
such as the VRP with Deliveries and Selective Pickups (VRPDSP). The VRPDSP often
arises in reverse logistics, where a vehicle delivers a request to a customer and may collect
an amount of material in the same location, resulting in a profit (Süral and Bookbinder,
2003; Privé et al., 2006; Bruck et al., 2012; Coelho et al., 2012). Thus, in the VRPDSP,
customers requiring deliveries are also those requiring pickups.
In the VRPSB, selecting one backhaul rather than another depends highly on the revenue
it provides and its impact in mitigating routing costs. For instance, the revenue may be
related to the quality of products to pick up or the suppliers may have several products
available with different quality attributes (Zanjani et al., 2010; Alvarez and Vera, 2014;
Andersson et al., 2016). However, to the best of our knowledge, only Allahviranloo et al.
(2014) reported the development of a selective VRP model with uncertainty in revenues,
showing evidence of the advantages of addressing uncertainty in humanitarian logistics.

More surprisingly, a recent literature review on the VRPB research shows that no work has
yet investigated this problem under uncertainty, which is usually present in practice (Koç
and Laporte, 2018).
To deal with uncertainty in VRPs, two main approaches stand out in the literature: stochas-
tic programming and robust optimization (Averbakh, 2001; Bertsimas and Sim, 2003).
Stochastic programming is a well-known technique used when the probability distribu-
tion of the uncertain parameters are known, and the goal is to optimize the expected value
of a solution while maintaining its feasibility for the scenarios considered (Birge and Lou-
veaux, 2011). When no knowledge on the probability distribution of the uncertainty is
available, robust optimization emerges as a suitable alternative (Chardy and Klopfenstein,
2012; Grossmann et al., 2016). In this case, the goal is to obtain a solution that is robust
(i.e. feasible) against all possible uncertainty realizations. Robust Optimization (RO) is, in
fact, the most recent trend to deal with uncertainty in optimization problems (Grossmann
et al., 2016).
The present paper analyses the robustness of routing plans obtained under revenue un-
certainty for a robust VRPSB, where it is required to achieve a minimum revenue from
collected raw-materials. The interest on studying the uncertain revenues derive from two
main aspects: (i) a deviation in the quality of raw-materials can impact the efficiency of the
productive processes (at the depot) and, consequently, the production costs of a company,
and (ii) for large distances, and particularly for daily routes, the deviation in the expected
revenue may turn the predefined routes very costly for consecutive days, which may cause
significant financial damage to the company. Particularly for the wood-based supply chain,
the quality of raw-materials is becoming more and more important, since, in one hand,
it allows to differentiate suppliers (Andersson et al., 2016) and, on the other hand, it im-
pacts the efficiency of subsequent manufacturing processes (Zanjani et al., 2010; Alvarez
and Vera, 2014). In addition, although several characteristics of raw-materials can be eas-
ily obtained, such as diameter and length, others can only be roughly estimated, such as
moisture.
An interesting finding from the literature review on RO is that, in opposition to exact
methods, few metaheuristics have been proposed to solve the VRP under uncertainty. In
fact, Solano-Charris et al. (2015) argue that, up to the date of their work, only four meta-
heuristics were proposed in the literature: a particle swarm optimization (PSO) algorithm
(Moghaddam et al., 2012), a differential evolution algorithm (DEA) (Cao et al., 2014), and
two Ant Colony Optimization (ACO) algorithms (Toklu et al., 2013; Toklu et al., 2014).
More recently, some research has been driven by improving or adapting well-known meta-
heuristics to handle the robust VRP. Gounaris et al. (2016) present an Adaptive Memory
Programming (AMP) metaheuristic to solve the VRP under demand uncertainty, which
considers two different cases of polyhedral uncertainty sets - budget support and factor-
model support. The AMP metaheuristic is enhanced with a new mechanism to identify
and select promising solution components and a new augmented objective function. The
LANTIME metaheuristic, based on the Tabu Search (TS) algorithm, is adapted by Groß
et al. (2019) to solve a robust VRP with uncertain travel times, which are represented by
interval scenarios. The adapted version of the metaheuristic includes a new procedure to
evaluate the cost of a solution, which compares a candidate solution in a given scenario
and in worst-case scenario. Hu et al. (2018) study a robust VRP where the uncertain travel
time and uncertain demand are assumed to belong to a new route-dependent budgeted un-
certainty set. To solve the robust problem, an Adaptive Variable Neighborhood Search
(AVNS) is modified in order to include two new steps in the standard algorithm: an adap-
tive shaking and a local search. The adaptive shaking step uses two exchange operators
to explore neighborhoods, whereas the local search step uses a randomized variable neigh-
borhood descent (RVNS) heuristic to improve the solutions. An enhanced Large Neighbor-
hood Search (LNS) is proposed in Wu et al. (2017) and in Pelletier et al. (2019) to solve
large scale VRP. The former represents the uncertain travel time as discrete scenarios and
proposes a new robust criterion to evaluate the robustness of solutions. The LNS is applied
to explore the solution space and it is hybridized with a series of local moves to improve
the solutions. The latter investigates several uncertainty sets for the robust Electric-VRP
with uncertain energy consumption, namely box, polyhedral (budget and route-dependent
budget), and ellipsoidal sets. The LNS used to solve the problem is combined with a Set
Partitioning (SP) formulation in a two-phase method. First, the metaheuristic generates
a set of candidate routes and then a complete robust solution is achieved by solving the
SP problem. Furthermore, the LNS proposed by Pelletier et al. (2019) uses all the com-
ponents of the Adaptive LNS (ALNS), except the adaptive procedure on the basis that
previous results have shown that no solution improvement was obtained for the particular
problem. Nonetheless, the ALNS has been successfully applied for the robust VRP under
uncertainty in travel/service times. Braaten et al. (2017) uses the ALNS in a modular fash-
ion, allowing to investigate different versions of the metaheuristic, namely by excluding
a preprocessing phase, excluding the local search heuristics, and replacing a new accep-
tance rule. This new acceptance rule only accepts a new solution that is not worse than a
previous one, instead of measuring it against a given threshold, as in the original ALNS
proposed by Ropke and Pisinger (2006). Eshtehadi et al. (2020) assumes that the uncertain
service times belong to an interval, but the robust counterpart derived is based on the very
conservative worst-case approach. The ALNS used to solve the problem departs from the
enhanced method proposed by Demir et al. (2012) and introduces new components, such
as new removal operators. Finally, metaheuristics have also been enhanced to solve prob-
lems beyond the scope of robust VRP, such as the location-routing problem (LRP) with
uncertain spatial customer distribution, demand and service time (Schiffer and Walther,
2018), the inventory-routing problem (IRP) with uncertain demand (Fardi et al., 2019), the
intermodal freight transportation problem (IFTP) with uncertain transportation costs and
uncertain capacities of terminals (Abbassi et al., 2019), and the Capacitated Arc Routing
Problem (CARP) with uncertain demand (Tirkolaee et al., 2018).
We explore the aspects of RO and its respective budget of uncertainty, using exact and
metaheuristics solution methods. We define a function to assess the robustness of solutions
embedded in a state-of-art metaheuristic for a large spectrum of VRPB variants, namely
the ALNS. In addition, we compare a RO model with a chance-constrained (CC) model
and study four different methods to derive bounds for the probability of constraint viola-
tion. Moreover, a factor model support describing the uncertain revenues is proposed as an
alternative to the budget of uncertainty. The main contributions of this paper are presented
4.2. Problem description and formulations 107
as follows, by decreasing order of significance.
1. A novel method to estimate tighter bounds for the probability of constraint violation
for the robust optimization under polyhedral uncertainty sets;
2. Two different formulations for the VRPSB under revenue uncertainty, contributing
simultaneously for two gaps in literature (lack of works dealing with uncertainty in
VRPBs and uncertain revenues in VRPs);
3. A simple and efficient procedure to check the feasibility of robust solutions embed-
ded in the state-of-art metaheuristic ALNS.
The remainder of this paper is organized as follows. The deterministic formulation, as

well as the RO models for the VRPSB are detailed in Section 4.2. The methods to derive
probabilistic bounds and an alternative uncertainty support are also defined. In Section 4.3,
both solution methods used in this work are presented, which are respectively a Branch-
and-Cut (B&C) algorithm and an ALNS metaheuristic. The main results are presented
and discussed in Section 4.4, particularly regarding the robustness of the models under
uncertainty, the structure of the solutions and computational performance of both solution
methods. Finally, conclusions and insights of this work are stated in Section 4.5.
4.2. Problem description and formulations

In this section, a mixed integer programming (MIP) formulation for the deterministic
VRPSB is presented first. The deterministic model is then reformulated with the proper
robust counterpart in order to incorporate uncertainty in revenues. Afterwards, four differ-
ent methods of estimating probabilistic bounds are designed to estimate the probability of
constraint violation. Furthermore, a new support for the uncertainty set, namely the factor
model support, is presented and illustrated for the VRPSB under revenue uncertainty.
4.2.1 Deterministic VRPSB

The problem can be defined by an undirected graph G = (V, E) and a fleet of K vehi-
cles. Set L = {1, . . . , n} comprises the linehaul customers with demands qi > 0 for all i ∈ L.
Set B = {n + 1, . . . , n + m} comprises the backhaul customers with revenues pi > 0 for all
i ∈ B. The depot is partitioned into two vertices: an origin depot 0 and a destination de-
pot n+m+1. We anticipate that a split depot allows a better adaptation of the problem
to be solved in the B&C, as discussed in Section 4.3.1. Finally, the set of all vertices
in the network is V = {0, n+m+1} ∪ L ∪ B. The edges in E are given by {(0, j) : j ∈ L} ∪
{(i, j) : i ∈ L ∪ {n+m+1}, j ∈ L ∪ {n+m+1}, i < j}∪{(i, j) : i ∈ L, j ∈ B}∪{(i, n+m+1) : i ∈ B}
with costs ci j ≥ 0 for all edges (i, j) ∈ E. Note that, for the sake of a clean notation we are
using the notation (i, j) to represent an (undirected) edge given by {i, j}. Set K = {1, . . . , K}
represents the fleet of identical vehicles, each one with a capacity C > 0, all of them initially
located at the depot. A route is a sequence of vertices given by r = (i0 , i1 , i2 , . . . , i s , i s+1 ) with
i0 = 0 and i s+1 = n+m+1 and with {i1 , i2 , . . . , i s } ⊆ L ∪ B. The total cost of a route is given
by the sum of its costs minus its revenue. The objective is to find a set of at most |K|
routes (tours) minimizing the total cost in such a way that: (i) all linehaul customers are
visited, each one in exactly one route; (ii) all the linehaul customers of a route must be
visited before the backhaul customer (if any); (iii) there is no route with just one customer
if it is a backhaul; (iv) the sum of the demands of the linehaul customers in a route is at
most C; and (v) a minimum revenue obtained from the raw-materials picked up at back-
haul customers is attained. Four additional features of real world transportation problems
are assumed for the VRPSB: (vi) the revenue associated to each backhaul customer is a
function of external factors not related to route planning; (vii) each route visits at most
one backhaul customer; (viii) backhaul customers have enough raw-material available to
supply the depot with multiple vehicles; and (ix) backhaul customers can only be visited
after all deliveries are performed (precedence constraint). These assumptions derive from
the fact that many inbound routes are supported by full truckload vehicles that, in a single
visit to a backhaul customer, collect an amount of load that equal their capacity. Example
of such cases are found in the manufacturing industry (e.g., forests with large supply for
mills) and in retail (e.g., supplier inbound operations to the retailers’ distribution centers).
In particular, for wood-based supply chains the inbound transport is often carried in full
truckloads (e.g., Carlsson and Rönnqvist (2007); Hirsch (2011); Derigs et al. (2012)). It
is also worth mentioning that woods parks do not usually present capacity problems to
accommodate several types of raw-materials. Finally, the precedence constraint is an im-
portant aspect in practice because: (i) vehicles are often rear-loaded and this precedence
constraint avoids the rearrangement of the loads at linehaul customers (Dominguez et al.,
2016; Kumar et al., 2011) and (ii) linehaul customers have higher service priority than
backhaul customers (Ropke and Pisinger, 2006; Tütüncü et al., 2009) because they have
time-windows necessary to inspect the received loads.
To better illustrate the network of the VRPSB, an example of a feasible solution is given in
Figure 4.1. In this example, the network is composed of seven linehaul customers (white
circles), five backhaul customers (grey circles) and a depot (partitioned into two squares).
Four different routes are created. Two vehicles visit backhaul customers after visiting the
linehaul customers (dashed and double lines) and two other vehicles only visit linehaul
customers, returning then back empty to the depot (full and dotted lines).
The deterministic VRPSB is presented as a two-index vehicle flow formulation, using two
decision variables: xi j , that represents the number of times the edge (i, j) is used in a
tour, and κ, that denotes the number of distinct paths that begin at the origin depot 0 and
end at the destination depot n+m+1. The following notations are also used to formulate
the problem: δ(X, Y) = {(i, j) ∈ E : i ∈ X, j ∈ Y}, δ(S) = {(i, j) ∈ E : i ∈ S, j < S}, δ(X, i)
corresponds to δ(X, {i}), and δ(i) is equivalent to δ({i}). The deterministic VRPSB reads as
follows:
X X
min ci j xi j − p j x j,n+m+1 (4.1)
(i, j)∈E j∈B
Figure 4.1 – Example of a feasible solution for a VRPSB instance with V = {0, 13} ∪ L ∪ B,
where L = {1, 2, 3, 4, 5, 6, 7} and B = {8, 9, 10, 11, 12}. Omitting costs, demands and rev-
enues. The depot are the squares and K = {1, 2, 3, 4}. Each type of line (full, dotted, dashed
and double) represent the total route of each vehicle.
subjected to
X
xi j = κ, (4.2)
(i, j)∈δ(0)
X
xi j = κ, (4.3)
(i, j)∈δ(n+m+1)
X
xi j = 2 ∀i ∈ L, (4.4)
(i, j)∈δ(i)
X
xi j = 2xi,n+m+1 ∀i ∈ B, (4.5)
(i, j)∈δ(i)
X
xi j ≥ κ ∀S ⊆ V, S ∈ 0, S < n+m+1, (4.6)
(i, j)∈δ(S)
X
p j x j,n+m+1 ≥ β, (4.7)
j∈B
X
xi j ≥ 2r(S) ∀S ⊆ L, |S| ≥ 2, (4.8)
(i, j)∈δ(S)
xi j ∈ {0, 1, 2, . . . , K} ∀(i, j) ∈ δ(B, n+m+1), (4.9)

xi j ∈ {0, 1} ∀(i, j) ∈ E \ δ(B, n+m+1), (4.10)
κ ∈ {1, . . . , K}. (4.11)
Equation (4.1) is the objective function, which aims at minimizing the total costs, i.e., total
travel costs minus total revenues obtained from visiting backhaul customers. Equations
(4.2) and (4.3) impose that exactly κ paths are connected to the origin depot 0 and to the
destination depot n+m+1. Equation (4.4) ensures that each linehaul customer is visited
exactly once by a vehicle by forcing the connection of exactly two edges – one input edge
and one output edge. By the definition of the problem any route is allowed to visit a
backhaul customer at most once, and a backhaul can be visited by several distinct routes.
In this way, Equation (4.5) makes the number of routes that enters in each backhaul equals
the number of routes that connects that backhaul to the destination depot n+m+1, taking
advantage that there are no edges among the backhauls neither between them and the origin
depot. Equation (4.6) ensures the existence of κ routes that begin at the origin depot 0
and end at the destination depot n+m+1, accomplishing this by delimiting the number of
selected edges to the routes in each possible cut S separating 0 and n+m+1. The equations
(4.5) and (4.6) together guarantee that whether a backhaul is chosen, there is only one in
a route and it is connected to the destination depot, i.e., both equations are necessary to
assure the precedence constraints among the linehauls and backhauls. Furthermore, these
equations assure that there are no routes with two extreme points in 0 nor two extreme
points in n+m+1. We anticipate that these constraints can be relaxed if the problem allows
for multiple backhaul visits per route. For that case, we demonstrate the modelling changes
in 4.B.
Equation (4.7) ensures that a minimum amount of revenue β is achieved with a given rout-
ing plan. We emphasize that, in practice, the wood industry is concerned with the different
characteristics of raw-materials, which have great impact on the manufacturing processes.
The work of Andersson et al. (2016) distinguishes between "hard" constraints that reflect
the logs requirements in the wood industry, and the "soft" constraints that reflect desirable
logs properties. The authors also note that both type of constraints are often uncertain.
Equation (4.8) denotes, simultaneously, the well-known capacity inequalities (CI) and the
subtour elimination constraints (SEC) for the part of the routes made only by linehaul cus-
tomers. The value of r(S ) represents the minimum number of routes (vehicles) necessary
to attend the demand of the linehaul customers in S ⊆ L. The optimal value of r(S ) can
be calculated by the Bin Packing Problem (Martello and Toth, 1990), which is NP-hard. A
widely used lower bound to r(S ) is given by k(S ) = dq(S )/Ce, where q(S ) = i∈S qi . These
P
constraints work as follows: (i) if S has demand q(S ) > C, at least r(S ) routes must con-
nect S with V \ S , and, since there is a factor of 2 on the right hand side, this ensures that a
vehicle must enter and exit the cut S ; and (ii) given a route of vertices in S that is a subtour,
it is known that connections between S and V \ S are such that (i, j)∈δ(S ) xi j < 2r(S ), thus
P
eliminating route S . In this paper, we substitute r(S ) by k(S ) in Equation (4.8), giving rise
to the well known rounded capacity inequalities (RCIs). We emphasize that this substitu-
tion maintains the correctness of the formulation (Irnich et al., 2014). Equations (4.9) to
(4.11) define the domain of the decision variables.
Remark If β = 0 then the deterministic VRPSB can be converted into an asymmetric

capacitated vehicle routing problem (ACVRP). To do that, it is necessary to execute the
steps described in Algorithm 1.
Algorithm 1 Pseudo-code to eliminate backhauls from the deterministic VRPSB (β = 0)

1: for each j ∈ L do
2: Calculate the new traveling costs c0 j,n+m+1 :
3: c0 j,n+m+1 = min{min{c ji + ci,n+m+1 − pi }, c j,n+m+1 }
i∈B
4: Formulate the model (4.1)–(4.11) with B = ∅ and new traveling costs c0 j,n+m+1 ∀ j ∈ L
Algorithm 1 calculates new minimal costs for the arcs from the linehaul customers to the
depot and then removes the selective backhaul customers from the problem. The algorithm
can be performed in O(|L| · (|B| + 1)) and its output results in an ACVRP.
4.2.2 Robust VRPSB

Robust optimization is a fairly recent research field (Gorissen et al., 2015) that opened a
plethora of possibilities to deal with uncertainty in many optimization problems. It is a
suitable approach to address uncertainty when its probability distribution is not known, or
when infeasibility cannot be tolerated (Li et al., 2011; Gorissen et al., 2015).
Robust optimization assumes that uncertain parameters belong to a bounded uncertainty
set, and the main aim is to provide an optimal solution that is feasible for this entirely
uncertainty set (Ordóñez, 2010). Defining the uncertainty set is, therefore, a major decision
that will greatly influence the design of the model and the formulation of the problem. One
of the main advantages of robust optimization is precisely its computational tractability for
many types of uncertainty sets (Gorissen et al., 2015).
In this work, we consider that revenue in backhauls is uncertain. As mentioned in Section
4.1, this is a very common issue in wood-based supply chains, where a minimum amount
of revenue related to quantity and quality of raw materials has to be collected everyday
and it is not possible to know the quality of wood beforehand. Revenue uncertainty is only
considered in Equation (4.7), meaning that a minimum revenue requirement of β must be
satisfied considering the deviations controlled by the uncertainty set. This allows for a
trade-off between the nominal profit obtained in the objective function and the feasibility
of Equation (4.7). In addition, it will be possible to assign probabilistic bounds of violating
Equation (4.7), as it will be detailed in Section 4.2.3. Finally, it is noteworthy that this
approach is very intuitive, meaning that the decision maker aims to optimize the nominal
objective function, but at the same time must meet a minimum revenue requirement for the
all cases considered in the uncertainty set.
The first robust optimization approach used in this work follows the budget of uncertainty Γ
of Bertsimas and Sim (2004) to develop the robust version of the VRPSB under study. The
uncertain revenues are assumed to belong to a space represented by a polyhedral set U(Γ).
The choice of the polyhedral set comes with two explanations. Firstly, the resulting model
is more computationally tractable than others with non-linear sets, which is desirable for a
problem that is already NP-hard. Secondly, by considering a Γ polyhedral set, it is possible
to derive tight probabilistic bounds, as shall be further demonstrated in the next subsection.
The uncertain revenue p̃ j of each backhaul j ∈ B is modeled as a symmetric and indepen-
dent variable, bounded by the interval [ p̄ j − ṗ j , p̄ j + ṗ j ], where p̄ j is the nominal revenue
of backhaul j and ṗ j is the deviation in the revenue. It is thus possible to define a scale
deviation ε j = ( p̃ j − p̄ j )/ ṗ j varying between [−1, 1], such that the revenue of each backhaul
is given by:
p̃ j = p̄ j + ṗ j ε j . (4.12)
The uncertainty set U(Γ) is represented as follows:

X
U(Γ) = {ε : |ε j | ≤ 1 ∀ j ∈ B, |ε j | ≤ Γ}, (4.13)
j∈B
where ε ∈ R|B| is the vector containing the deviation ε j for each j ∈ B. As observed from
Equation (4.13), the budget of uncertainty Γ is introduced in order to control the number of
times the worst realizations of revenues are allowed to occur. In the VRPSB, the budget of
uncertainty falls in the range [0, |B|], but Γ = min{|B|, |K|} is already the worst scenario for
the problem, since the maximum number of deviations is limited by either the number of
backhauls or the number of vehicles (number of routes). Therefore, when Γ = 0, the model
becomes the deterministic version of the problem – no deviations are considered for any
revenue; when Γ = min{|B|, |K|}, the model becomes its worst-case deterministic version
– all revenues suffer deviation, which is equivalent to the Soyster (1973) approach. Thus,
Γ is used to adjust the robustness of the method against the level of conservatism of the
solution (Bertsimas and Sim, 2004), or in other words, Γ is used to control the size of the
uncertainty set.
Based on the uncertainty set U(Γ), the uncertain revenues in Equation (4.7) are reformu-
lated in the problem, as described by the following semi-infinite and non-linear constraint:
X
p̄ j x j,n+m+1 + ṗ j x j,n+m+1 ε j ≥ β ∀ε ∈ U(Γ). (4.14)
j∈B
Following the first theorem from Bertsimas and Sim (2004), the dual formulation is incor-
porated into the original model, thus obtaining its tractable robust counterpart, where λ and
µ j represent the dual variables. Therefore, the tractable robust counterpart derived from the
dual is as follows:
X X
min ci j xi j − p̄ j x j,n+m+1
(i, j)∈E j∈B
subjected to
(4.2) − (4.6)
(4.8) − (4.11)
X X
p̄ j x j,n+m+1 − λΓ − µ j ≥ β, (4.15)
j∈B j∈B
λ + µ j ≥ ṗ j x j,n+m+1 , ∀ j ∈ B, (4.16)
λ, µ j ≥ 0, ∀ j ∈ B. (4.17)
Considering the above reformulation, the robust VRPSB model presents 1 + |B| more vari-
ables than the deterministic version, as a result from including the dual variables λ and
µ j . In addition, the robust model also includes 2 · |B| + 1 more constraints, as a result of
including Constraints (4.15), (4.16) and (4.17).
Note that the uncertainty only appears in the constraints but not in the objective function.
This is aligned with basic assumptions in robust optimization, namely that the objective is
certain and the uncertain constrains are hard, in the sense that cannot be violated (Gorissen
et al., 2015). As such, the decision-maker is optimizing the expected revenue but concerned
that, in the worst-case, the minimum revenue required is met. In this work, optimizing
over the expected value instead of the worst-case value is valid because this operational
routing problem is performed every day and worst-cases are not usually catastrophic and
can be compensated by better scenarios in the next day. Still, we include the constraint
of minimum required revenue, because a minimum quality/revenue of raw-material in the
worst-case scenario may be desirable for the sustainability of the production process. Ac-
cording to Mulvey et al. (1995), a solution is said to be solution robust if it remains close
to the optimal for all uncertainty realizations, and a solution is said to be model robust if it
remains feasible for all uncertainty realizations. Therefore, we can contextualize our solu-
tions as model robust, because the main concern is on the feasibility and not on the quality
of a solution. A similar approach can be found in Alvarez and Vera (2014).
4.2.3 Estimates of probabilistic bounds

In robust optimization, it is not necessary to acknowledge the probability distribution of
the uncertainty sources. Nevertheless, it is necessary to consider how many revenues will
deviate from their nominal value, represented by Γ. The budget of uncertainty determines
the size of the uncertainty set and can be defined using historical values and the opinion
of experts. Γ can also be defined based on probabilistic bounds of constraint violation, as
shown in Bertsimas and Sim (2004).
In this section, we describe how Γ can be defined using different probabilistic bounds.
We first start with a case where no probability distribution is assumed for the uncertainty
sources. Then we describe a priori and a posteriori probabilistic bounds that can be ob-
tained assuming that the uncertainty sources follow an uniform distribution. We also de-
scribe a novel method that provides tighter probabilistic bounds for U(Γ) when random
parameters are independent, identically and uniformly distributed.
The choice for an uniform distribution comes from the fact that, since the random parame-
ters are not known from historical data, one could assume that all of them can occur with
the same probability. It is important to note that it is still a conservative assumption, given
that the extreme values have the same probability to occur than the mean value. How-
ever, this assumption allows to derive much tighter probabilistic bounds for the problem.
Obviously, if it is not possible to assume any distribution, standard probabilistic bounds
(Bertsimas and Sim, 2004; Guzman et al., 2016) can still be assumed.
4.2.3.1 Probabilistic bound based on Bertsimas and Sim (2004)

Based on Bertsimas and Sim (2004), the budget of uncertainty can be calculated according
to probabilistic bounds of constraint violation without assuming any distribution for the
uncertain parameters. Assuming that ε j ∈ J are independent, symmetrically distributed
have the expected value 0, the bound of violating constraint (4.15) can be calculated as
follows:
 
Γ
!
X  − 1
Pr   p̃ j x( j,n+m+1) < β ≈ 1 − φ √
 , (4.18)
j∈B |B|
where φ(θ) is a standard Gaussian cumulative distribution function. According to Equa-

tion (4.18), the probability of constraint violation is expected to decrease as the budget of
uncertainty increases. Therefore, this bound can be used to estimate the budget of uncer-
tainty based on confidence levels and avoid solving the problem several times to obtain an
adequate budget of uncertainty.
4.2.3.2 A priori probabilistic bound based on Guzman et al. (2017a) and Kang et al.
(2015)
If the probability distribution function of the uncertainty parameters is assumed to be
known, it is possible to derive tighter probabilistic bounds, as presented by Guzman et al.
(2017a) and Kang et al. (2015). Assuming that each uncertain parameter ε j is independent,
has the expected value equal to 0, and has a known cumulant generating function given by
Λ j (θ), the following bound holds:
    
X   
 X 
 
p̃ j x( j,n+m+1) < β ≤ exp min  + Λ  .
 

Pr  −θΓ (θ) (4.19)

j
θ>0 


j∈B

 j∈B


Here, we assume that each parameter ε j follows an uniform distribution supported in

θ
[−1, 1], which gives Λ j (θ) = ln( sinh
θ ). Finally, Γ in Equation (4.19) is obtained by means of
Algorithm 1 from Guzman et al. (2017a), given a desired probability of constraint violation.
4.2.3.3 A priori probabilistic bound based on the Irwin-Hall distribution

In this section, we derive a new probabilistic bound assuming that each uncertain parameter
ε j is independent and uniformly distributed in [−1, 1]. This bound differs from Equation
(4.19) since it does not rely on Markov’s inequality and moment generating functions. It
is inspired by the geometrical interpretation of the uncertainty set U(Γ) through the Irwin-
Hall distribution. Some previous works have used the Irwin-Hall distribution in robust
optimization. Chassein (2016) uses the Irwin-hall to derive the volume of the uncertainty
set. Li and Morales (2017) mention the Irwin-hall to calculate the probability of the total
realized uncertainty being above a specific deviation of a defined Γ, but does not provide
a general framework that assigns a Γ according to the probability of constraint violation.
Finally, Fliedner and Liesiö (2016) applies the Irwin-hall to define the probability of the
summation of absolute deviations supported in U[−1, 1] being above a specified Γ in the
uncertainty set, however it assumes absolute deviations and does not associate the prob-
ability with constraint violation bounds. Thereby, consider the following box uncertainty
set:
U HC = {ε ∈ R|B| : −1 ≤ ε j ≤ 1 ∀ j ∈ B}, (4.20)
where its feasible solution space is defined by a hypercube in [−1, 1]|B| . Assuming that
each variable ε j is uniformly distributed in [−1, 1], U HC will have infinite equiprobable
points in [−1, 1]|B| . Given that U(Γ) ⊆ U HC and since all points in U HC are equiprobable,
the percentage of realizations considered by U(Γ) can be calculated by its hypervolume
divided by the hypervolume of U HC :
Vol(U(Γ))
Pr scen (Γ) = , (4.21)
Vol(U HC )
P
which also can be interpreted as Pr |B| k=1 U[−1, 1] ≤ Γ , with U[−1, 1] being an uniform
distribution in [−1, 1]. Since Pr scen (Γ) is the percentage of realizations acknowledged by
the polyhedral set U(Γ), Constraint (4.14) is satisfied at least with probability Pr scen (Γ)
and therefore:
 
X 
Pr  p̃ j x( j,n+m+1) < β ≤ 1 − Pr scen (Γ), (4.22)
j∈B
when each p̃ j is uniformly distributed in [ p̄ j − ṗ j , p̄ j + ṗ j ].

To calculate Pr scen (Γ) it is necessary to resort to the Irwin-Hall distribution, which is the
distribution of the sum of n independent and identically distributed uniform random vari-
ables in [0, 1]. The cumulative distribution of the Irwin–Hall is defined as:
 n  t !
X  1 X n
F IH (t, n) = Pr  U[0, 1] ≤ t =
  (−1) j
(t − j)n , (4.23)
k=1
n! j=0
j
where n is the number of independent and identically P distributed uniform

random vari-
ables and t ∈ [0, n]. Since our aim is to calculate Pr |B|
k=1 U[−1, 1] ≤ Γ , it is necessary to
parametrize F IH (t) to acknowledge the sum of uniform distributions supported in [−1, 1]
instead of [0, 1] which can be easily done by considering n = |B| and t = Γ+|B|
2 :
 n 
X 
Pr  U[0, 1] ≤ t
 (4.24)
k=1
 |B| 
X Γ + |B| 
= Pr  U[0, 1] ≤ 
k=1
2 
 |B|  
X  
= Pr  2 · U[0, 1] − |B| ≤ Γ
  
k=1
 |B| 
X 
=  U[−1, 1] ≤ Γ .
k=1
Thence, Pr scen (Γ) can be calculated as follows:

t !
1X j n
Pr scen (Γ) = F IH (t, n) = (−1) (t − j)n , (4.25)
n! j=0 j
with t = Γ+|B|
2 and n = |B|. With that, it is easy to see that when Γ = 0, Pr scen (Γ) is 50%,
in the same way, when Γ = |B|, Pr scen (Γ) becomes 100%. For a geometrical derivation and
further details on the Irwin-Hall distribution we refer the readers to Marengo et al. (2017).
Considering that the maximum number of visited backhauls is limited by n∗ = min{|B|, |K|},
a tighter probabilistic bound can be achieved by assuming that the actual support of random
∗
deviations from visited backhauls is given by U[−1, 1]n . Hence, Pr scen (Γ) can be tighten
as follows:
Pr∗scen (Γ) = F IH (t∗ , n∗ ), (4.26)

Γ+n∗
with t∗ = 2 and n∗ = min{|B|, |K|}.
4.2.3.4 A posteriori probabilistic bound based on Guzman et al. (2017b)

All the above methods to obtain probabilistic bounds are referred to as a priori bounds,
which means that the probabilistic bound of violating a constraint is determined before
considering the solution of the optimization model. The probabilistic bound (GMF12)
in Guzman et al. (2017b) estimates the a posteriori probability of constraint violation.
Although this method is more challenging than the previous, it provides a tighter bound
because it allows to characterize one single solution to the problem (Guzman et al., 2017b),
rather than every feasible solution of the uncertainty set. Considering that each ε j ∀ j ∈ B is
independent and follows a continuous uniform distribution supported in [−1, 1], the exact
probability of constraint violation of a solution x∗ can be defined by the following equation:
 
X X 
Pr   p̄ j x( j,n+m+1) +
∗
ṗ j x j,n+m+1 ε j < β = FΘ (h(x∗ )),
∗
(4.27)
j∈B ∗ j∈B
where FΘ is the cumulative distribution function of j∈B ṗ j x∗j,n+m+1 ε j , the deterministic

P
part of the constraint is given by h(x∗ ) = β − j∈B∗ p̄ j x(∗j,n+m+1) , and B∗ = { j : ṗ j x∗j,n+m+1 >
P
0}. Because calculating FΘ is quite challenging, the brute-force enumeration algorithm
proposed by Guzman et al. (2017b) is used to calculate it.
4.2.4 Factor model support for uncertain revenues

An alternative set for the uncertain revenues can be described by a factor model support
(Gounaris et al., 2013), which enables to consider not only the deviation from the nominal
revenues, as in the budget of uncertainty presented in Equation (4.13), but also the different
factors that influence that deviation. Consider now the following set:
U 0 = { p̃ ∈ R+|B| : p̃ = p̄ + Φξ for some ξ ∈ Ξ}, (4.28)
where
Ξ = {ξ ∈ RF : ξ ∈ [−e, +e], eT ξ ∈ [−ΩF, +ΩF]}. (4.29)
The parameter F ∈ N represents the number of independent factors ξ1 , ..., ξF that affect the
revenue deviation of each backhaul customer. Matrix Φ ∈ R|B|×F represents the disturbance
of each factor on the revenue of each backhaul. Parameter Ω ∈ [0, 1] represents the percent-
age of surplus factors that can be above or below the point estimate 0, such that if Ω = 0,
the number of factors ξ f below and above the point estimate 0 ∈ RF are the same. Finally, e
is column vector of ones with dimension F. The resulting robust counterpart for the factor
model support is given by:
X X
(i, j)∈E j∈B
subjected to
(4.2) − (4.6)
(4.8) − (4.11)
F
(ν+f + ν−f ) − ΩF(ρ+ + ρ− ) ≥ β,
X X
p̄ j x j,n+m+1 − (4.30)
j∈B f =1
(ν+f − ν−f ) + (ρ+ − ρ− ) =

X
Φ j f x j,n+m+1 ∀ f = 1..F, (4.31)
j∈B
ν+f , ν−f , ρ+ , ρ− ∈ R+ ∀ f = 1..F, (4.32)
with ν+f , ν−f , ρ+ and ρ− being the dual variables from U 0.

4.2.4.1 Simple illustration

Although this type of support can be of interest in optimization problems, parameters F,
Ω and Φ must be assigned by the decision maker. As such, we exemplify how the un-
certain revenues can be described by a factor model support. For this example, consider
|B| = 10, and that three factors (F = 3) have influence on the revenue provided by raw-
materials of backhauls customers, namely moisture (Φ j1 ), heterogeneity (Φ j2 ) and ration
length/thickness (Φ j3 ):
 
 50 121 71 

 47 97 50 

 79 
 40 120
 55 
 60 116
 54 115 60 

Φ =   .
100 70 31 
 
 10 108 117
111 30 81 
 

 20 115 94 


47 80 126
Finally, by considering Ω = 0 we are assuming that the same number of factors deviate
below and above the point estimate 0. On the other hand, when Ω = 1, it means that
all factors deviate above (or below) the point estimate. Intermediate values of Ω indicate
that some factors will deviate while others remain steady. Figure 4.2 illustrates how the
parameter Ω influences the solution. For this example, we have used instance B3 from
Goetschalckx and Jacobs-Blecha (1989) and a minimum revenue required of β = 750.
Figure 4.2 – Influence of Ω in the objective function of a robust solution.
From Figure 4.2, it can be observed that, in this example, a value of Ω within the range
[0, 0.7] leads to the same solution value. Increasing Ω = to 0.8 leads to a subtle increase in
the objective function, whereas it is observed a drastic increase if Ω = 0.9. The growth of
4.3. Solution methods 119
this parameter is of utmost importance, since when it reaches Ω = 1.0 no feasible solution
is obtained. This extreme case can also be analogous to the worst-case solution, where all
uncertain realizations occur at the same time.
4.3. Solution methods

This section first introduces the exact method, a B&C algorithm, and then the ALNS meta-
heuristic developed for this problem. The B&C method is designed following the work
of Lysgaard et al. (2004) and is combined with CPLEX to solve the robust VRPSB. The
ALNS developed is based on the work of Ropke and Pisinger (2006) and includes addi-
tional procedures to incorporate the robust nature of the problem while solving the prob-
lem. We anticipate that both of the solution methods presented in this section are used to
solve small, medium and large instances adapted from literature, and afterwards compare
their computational performance. For illustrative instances, where the goal is to investigate
the impact of robustness parameters, probabilistic bounds and different uncertainty sets, a
simple CPLEX is used.
4.3.1 Branch-and-Cut
The B&C algorithm developed is partially based on the B&C for the CVRP presented
in Lysgaard et al. (2004), which is well known for its efficiency. Since that work, the
methods for the CVRP itself evolved and the state-of-the-art algorithms heavy rely on the
branch-and-cut-and-price framework (Poggi and Uchoa, 2014), which involves separation
algorithms to identify promising cuts and pricing algorithms to generate suitable columns
for the problem, as these models are based on set partitioning-like formulations (Toth and
Vigo, 2014). Our exact approach is based on B&C, which can be viewed as a component
to an eventual branch-and-cut-and-price based method.
The authors of Lysgaard et al. (2004) have made the source code of a software library
called CVRPSEP publicly available, which implements several cut separation routines for
the CVRP. In this paper, some of these separation routines are used in our implementation
of the B&C to solve the robust VRPSB. More precisely, our B&C algorithm makes use of
the RCI separation routine from the CVRPSEP, as well as it utilizes the max-flow routine,
from the same library, to separate the κ-cuts we used in our approach. For a comprehensive
material about the separation algorithms we refer to Toth and Vigo (2014). The B&C is
applied to the resolution of the Robust VRPSB formulation, which is presented in Section
4.2.2, given by the following equations: (4.1), (4.2) – (4.6), (4.8) – (4.11), and (4.15) –
(4.17).
The formulation presents suitable features for the B&C method we developed, namely: (i)
the undirected graph, (ii) the split depot, and (iii) the κ-cuts. Using an undirected graph to
represent the network in the VRPSB avoids symmetric solutions and keeps the number of
variables in the formulation smaller than an hypothetically directed graph. In the VRPSB, it
is mandatory to constrain a backhaul vertex to be, at most, on one extreme point of a route,
and therefore, the identification of the depot-connected edges of a given route is essential.
To accomplish that we splitted up the depot into the origin depot 0 and the destination depot
n+m+1, as already mentioned. Considering this transformation, the problem can be seen
as seeking paths that connects 0 to n+m+1, and, if a backhaul customer is present in a path,
it must be connected to the n+m+1. A linehaul customer can also be connected to n+m+1,
meaning that its route has no backhaul at all. On the other hand, we must ensure that there
are no routes where there are two extreme points in 0 nor two extreme points in n+m+1.
The κ-cuts in Equations (4.6) solve this problem, as those cuts restrict the solution to have
at least κ paths between 0 and n+m+1 in a solution.
As there is an exponential number of κ-cuts, we rely on the B&C capabilities for seeking
and adding violated cuts to the model. To separate these constraints, we use the imple-
mentation of the max-flow algorithm, presented in the CVRPSEP. If the max-flow value is
smaller than κ that cut is violated and added to the model. Note that Equations (4.2) and
(4.3) are specialized and stronger κ-cuts.
The RCIs, at Equation (4.8), ensures that the capacity limit of each vehicles is respected.
Each time the B&C has to call the RCI separation routine from the Lysgaard’s CVRPSEP,
it is necessary to make a conversion on the graph. This transformations occurs as follows:
given a candidate solution, the vertices 0, n+m+1 and the vertices of B are all shrunk into
a single supervertex that plays the role of the depot on the separation algorithm. If violated
cuts are found, and considering that the edges are not incident to the depot, the cuts are
added to the model and it is re-optimized.
The B&C algorithm adds the valid inequalities described above via the branch-and-cut
framework of CPLEX by using its callback functions. The separation step is composed by
four heuristics for the RCIs, accordingly Lysgaard et al. (2004), and a max-flow algorithm
for the κ-cuts. The branching strategy is kept as the default branching of CPLEX, which
typically branches on variables. An RCI is considered violated if the difference between
its left and right hand sides is greater than an RCI tolerance limit. In this work, this limit is
0.2. A κ-cut is considered violated if the max-flow value is smaller than κ.
Finally, the initial model given to the B&C algorithm is composed by the objective function
in Equation (4.1), the constraints in equations (4.2)–(4.5) and (4.15)–(4.16), the bounds in
equations (4.9)–(4.11) and (4.17), and two additional dedicated RCI cuts in the form of
Equation (4.8): one for the origin depot with S = {0}, and the other for the destination
depot with S = {n + m + 1}. As the model has an exponential number of κ-cuts in Equation
(4.6) and also an exponential number of RCI constraints in Equation (4.8), they are added
on demand in a branch and cut fashion.
4.3.2 Adaptive Large Neighborhood Search

The ALNS was firstly proposed by Ropke and Pisinger (2006) and since then, it has shown
to be effective for many VRPBs (Masson et al., 2013; Li et al., 2016b; Ghilas et al., 2016),
as well as other VRPs (Li et al., 2016a; Dayarian et al., 2016; Zajac, 2017).
Large neighborhoods can be explored by consecutively applying two heuristics: first, a
destroy operator, which partially destroys the current solution and then a repair operator
to make the solution feasible again, generating thus a new solution. The main features
of ALNS are using diverse destroy and repair operators to create different neighborhoods
(Pisinger and Ropke, 2010) and the adaptive search behavior as a consequence of the per-
formance of each pair of destroy-repair operators during the iterative process.
In this work, the general procedures of the ALNS, along with all related ALNS parameters,
are conducted as in Ropke and Pisinger (2006). The pseudo-code of ALNS is outlined in
Algorithm 2. First, an initial solution x is generated with a simple greedy heuristic and the
weights of each destroy and repair operators are stored in ρ− and ρ+ , starting at 1 for each
operator. Then, for each iteration, a pair of destroy-repair operators are applied based on
the their past performance (score). First, the destroy operator removes some vertices from
the solution (d(x)), and next, a repair operator inserts the removed vertices into different
positions (r(d(x))), creating a new solution xt . If the new solution xt satisfies the acceptance
criteria, it replaces x. If the new solution xt is better than the best solution found so far, it
replaces xb . This comparison is based on the solution cost, c. Finally, scores and weights
are dynamically adjusted as the search for a new solution progresses.
Algorithm 2 Pseudo-code of ALNS framework

1: Input: x, ρ− , ρ+
2: Output: xb
3: generate a solution x
4: set xb = x; ρ− = (1, . . . , 1); ρ+ = (1, . . . , 1)
5: while acceptance criteria is not met do
6: select destroy operators d ∈ Ω− and repair operators r ∈ Ω+ , using weights ρ− and
ρ +
7: xt = r(d(x))
8: if accepted (xt , x) then
9: x = xt
10: if c(xt ) ≤ c(xb ) then
11: xb = xt
12: update ρ− and ρ+
13: return xb
The destroy operators used are the Shaw Removal, the Random Removal and the Worst
Removal heuristics. The Shaw Removal selects the nodes to remove according to sim-
ilarity criteria, namely distance and quantity (Shaw, 1998). As the name suggests, the
Random Removal selects at random the nodes to remove from the solution. Finally, the
Worst Removal selects the nodes to remove according to their impact on the decrease of
the objective function. The objective is to remove those customers located in positions that
largely increase the total routing costs.
The repair operators used are the Greedy Insertion and the k-Regret Insertion. The former
heuristic is very simple, inserting the removed nodes into positions in the routes such that
the objective function increases the minimum possible. The latter applies the concept of
regret to insert first the customers for which the insertion in a later stage would be more
costly (with higher regret). The regret is determined as the sum of differences in the costs
as a result of inserting customer i in the first best route and in the kth best route. The values
for k = 2, 3, 4 were used.
To develop the robust ALNS, robustness was embedded into the algorithm during the cre-
ation of new neighborhoods. Thus, instead of creating a new solution with the destroy-
repair operators and afterwards evaluating its robustness, the robustness parameters are
already introduced, based on a polyhedral uncertainty set, during the repair procedure, as
detailed in Algorithms 3 and 4.
The adapted procedure starts by inserting all linehaul customers into a previously destroyed
solution d(x). After that, a solution x1 is achieved with all linehaul customers and with the
backhauls not removed by the destroy operator. The backhauls in x1 are ordered by decreas-
ing values of deviation in revenue multiplied by the number of times that the backhaul is
present in the solution. Thus, if two different backhauls are visited the same number of
times in the solution, the one with a higher deviation is the first to be addressed by the
budget of uncertainty. However, if a backhaul is visited twice and another one only once,
unless its deviation is very small, the former backhaul is the first to be addressed by the
budget of uncertainty.
At this stage, the total revenue minus total deviation (Revenue’) is determined using Al-
gorithm 4 and the feasibility of the solution is accessed by considering if this Revenue’ is
sufficient to achieve the minimum revenue required, β. If the Revenue’ is higher or equal
to β, then the expected revenue of the solution x1 is the total revenues collected with all
selected backhauls. Otherwise, the solution is infeasible and the expected revenue takes a
very large negative value. The total cost of x1 is then recalculated. The respective heuris-
tic in use proceeds by inserting potential backhauls in available routes and creates a new
solution x2 . The total expected revenue and total cost of x2 are determined as described
previously and compared with x1 . At the end, the solution with the lowest total cost is se-
lected as the neighbor solution r(d(x)) which will be used in the ALNS (line 5 of Algorithm
2). An example of the procedure to obtain such solutions is illustrated in Figure 4.3.
Algorithm 3 Pseudo-code to embed robustness into ALNS

1: Input: d(x)
2: Output: r(d(x))
3: B = vector of backhauls in solution d(x)
4: x1 = d(x)
5: Insert all linehaul customers in the solution with repair operator r ∈ Ω+
6: x1 = r(d(x))
7: Revenue (x1 ) = Algorithm 4 (x1 )
8: Total cost: c x1 = Total distance - Revenue (x1 )
9: Add a backhaul to B
10: x2 = r(d(x))
11: Revenue (x2 ) = Algorithm 4 (x2 )
12: Total cost: c x2 = Total distance - Revenue (x2 )
13: if c x2 < c x1 then
14: r(d(x)) = x2
15: else
16: r(d(x)) = x1
17: return r(d(x))
Algorithm 4 Function to assess robustness

1: Input: x1
2: Output: (x1 )
3: Ni = number of times backhaul i is present in the solution x1
4: B̄ = vector of backhauls in solution x1 ordered by decreasing ṗ · Ni
5: Revenue0 (x1 ) = 0
6: Revenue (x1 ) = 0
7: for i ∈ B̄ do
8: Revenue (x1 ) += Ni · p̄i
9: if Γ ≥ 1 then
10: Revenue0 (x1 ) +=
11: Ni · ( p̄i − ṗi )
12: Γ = Γ−1
13: else
14: Revenue0 (x1 ) += Ni · ( p̄i − ṗi · Γ)
15: Γ=0
16: remove backhaul i from B̄
17: if Revenue0 (x1 ) ≤ β then
18: Revenue (x1 ) += −999999
19: return Revenue (x1 )
Figure 4.3 – Example of procedure to embed robustness during the creation of a feasible
solution. Solution x1 is first evaluated in terms of total cost, for a given Γ value; then a
backhaul is inserted, producing solution x2 , and the total cost is again evaluated, for the
same value of Γ. If the total cost of x2 is lower than x1 , solution x1 is discarded and the
iterative process continues, otherwise, solution x2 is discarded and the former is kept.
4.4. Computational tests

The aim of the computational tests is two-fold: (i) evaluate the performance of models and
robustness of solutions, namely in terms of probabilistic bounds, feasibility and robustness;
and (ii) evaluate the performance of the B&C and the ALNS to solve the robust VRPSB.
The first aim is addressed using CPLEX and the experiments are performed on small in-
stances, due to their rapid convergence to the optimal solution. The second aim is obtained
by a series of experiments on large size instances, testing the two solution methods devel-
oped. An additional set of tests are presented in 4.B that compares the solution quality and
computational performance of two different cases - one allowing single backhaul visits per
route and one relaxing this constraint by allowing multiple visits.
The deterministic, RO and CC models for the first aim are implemented in OPL and solved
using IBM ILOG CPLEX Optimization Studio 12.6. The CC model developed to be com-
pared with the RO model is presented in the Appendix A. It is important to note that a
single-commodity formulation is used with OPL, instead of the two-index vehicle flow for-
mulation presented in Section 4.2, because it avoids the exponential constraints (4.8). In
fact, the resolution time is not a concern at this stage, since the main goal is to analyze
the structure of the solutions and their feasibility for different robustness parameters. The
ALNS algorithm is implemented in C++. The B&C is implemented in C++, using the
C callable API of CPLEX 12.7.1. The experiments of the ALNS were performed on a
computer equipped with the processor Intel Core i7 of 2.20GHz and with 16 GB of RAM,
and the operating system is the Windows 10. Besides, the B&C experiments were run on
a computer with processor Intel Xeon Gold 6142 of 2.60GHz with 64 on-line cores, 256
GB of RAM, and the Ubuntu 18.04.2 LTS operating system with the Linux kernel version
4.15.0-47-generic. Each B&C test was performed sequentially, although at most 32 of them
were in parallel at any time in the same machine. We add that, as the B&C lower bounds
are used to evaluate the quality of the ALNS solutions, we do not avoid using a powerful
available machine to perform its tests.
4.4.1 Problem instances

The instances used in the experiments are adapted from the standard VRPB instances in
Goetschalckx and Jacobs-Blecha (1989). The only new parameters added to these instances
are the revenues and deviations, which are randomly generated for each backhaul of each
instance as follows. From the original class of instances B of Goetschalckx and Jacobs-
Blecha (1989), we have used their quantities as proxies for the minimum and maximum
revenues. Then, 50 values are randomly simulated within this range, for each backhaul
i. The average value from this simulation corresponds to the revenue, p̄, and the standard
deviation corresponds to the deviation, ṗ. Furthermore, these values are simulated for the
10 backhaul customers from class B instances and then replicated for instances with higher
values, such that the revenue and corresponding deviation for backhaul customer 1 are the
same for backhaul customers 11, 21, and so on. These values are reported in Table 4.1.
4.4. Computational tests 125
Table 4.1 – Values of revenue p̄ and deviation ṗ for each backhaul customer used for
instances in the computational tests
Backhaul number p̄ ṗ
1, 11, 21, 31, 41, 51, 61, 71 438 242
2, 12, 22, 32, 42, 52, 62, 72 404 194
3, 13, 23, 33, 43, 53, 63, 73 454 239
4, 14, 24, 34, 44, 54, 64, 74 392 231
5, 15, 25, 35, 45, 55, 65, 75 464 229
6, 16, 26, 36, 46, 56, 66 380 201
7, 17, 27, 37, 47, 57, 67 384 235
8, 18, 28, 38, 48, 58, 68 355 222
9, 19, 29, 39, 49, 59, 69 459 229
10, 20, 30, 40, 50, 60, 70 415 253
4.4.2 Evaluation of the VRPSB under uncertainty

Throughout this section, several tests are conducted in order to analyze and evaluate the
robust optimization model (Section 4.2.2). First, the different bounds obtained by the dif-
ferent methods that estimate a posteriori the probability of constraint violation are charac-
terized. Next, the exact probability of constraint violation of these solutions is presented
and compared with the a posteriori probabilistic bounds. Then, the nominal and robust
solutions obtained for different values of β and Γ are analyzed and compared with the so-
lutions obtained with a chance-constrained model. The structure of nominal and robust
solutions are finally analyzed and discussed.
All the tests presented in the section are solved with CPLEX and tested with four instances
and four different values of minimum revenue required to cover a wide range of instance
characteristics. The standard instances tested are: A4 (20 linehaul customers, 5 backhauls
and 3 vehicles), B3 (20 linehaul customers, 10 backhauls and 3 vehicles), C4 (20 linehaul
customers, 20 backhauls and 4 vehicles) and F4 (30 linehaul customers, 30 backhauls and
4 vehicles). The values of minimum revenue required β are 0, 500, 750 and 1000.
4.4.2.1 Probabilistic bounds
In this section, the methods that estimate the probability of violating Constraint (4.7), based
on the values of the budget of uncertainty Γ, are analyzed and compared. The results are
presented in Figures 4.4a, 4.4b, 4.4c and 4.4d for |B| = 5, |B| = 10, |B| = 20 and |B| = 30,
respectively. It is important to note that the a posteriori method based on Guzman et al.
(2017b) is not analyzed in this section, since it is not influenced by the robust parameters.
From these results, it can be shown that, as expected, the probability of constraint vio-
lation decreases as the budget of uncertainty increases, for all the methods investigated.
Also, as more backhauls exist in an instance, the sooner the probabilistic bound reaches its
minimum, i.e. the sharper is the curve provided by the bound.
(a) Instance A (|B| = 5) (b) Instance B (|B| = 10)
(c) Instance C (|B| = 20) (d) Instance F (|B| = 30)
Figure 4.4 – Probability of constraint violation obtained with different methods.

The method based on Bertsimas and Sim (2004) is the most conservative among the four a
priori probabilistic bounds. In fact, when the instance only has 5 backhauls (Figure 4.4a),
the probability of constraint violation never reaches zero. When the number of backhauls
increases to 10 (Figure 4.4b), a probability of zero only occurs when Γ is maximum, i.e.
when Γ = |B|. In the other cases (Figures 4.4c and 4.4d), a probability of zero is achieved
but only latter than any of the other methods.
The method based on Guzman et al. (2017a) is very conservative for the lowest values of
Γ in each instance, but the probability of constraint violation decreases drastically with
increasing the degree of conservatism.
Finally, the methods based on the Irwin-Hall distribution are indisputably the ones that
provide the tighter bounds, showing always the lowest values for the probability of con-
straint violation for any value of Γ. In fact, by considering that the worst-case occurs when
Γmax = min{|B|, |K|} (Irwin-Hall*), it is possible to obtain an even tighter bound than when
Γmax = |B| (Irwin-Hall). It is important to note that when the function Irwin-Hall* is used,
the same probabilistic bounds are achieved for the cases illustrated in Figures 4.4a and
4.4b, since on both cases the number of vehicles is three, which is lower than the number
of backhauls. Similarly, the same bounds are achieved for the cases illustrated in Figures
4.4c and 4.4d, where the number of vehicles is four.
4.4.2.2 Comparison of a priori and a posteriori probabilistic bounds

In this section, the method that estimates a posteriori the probability of constraint violation
is compared with the best a priori method, as concluded in Section 4.4.2, i.e. the Irwin-
Hall* distribution method.
Table 4.2 and Table 4.3 report the different values of probability of constraint violation for
instances with 3 and 4 vehicles, respectively. Since the method of Irwin-Hall distribution
only considers the value of budget of uncertainty and number of backhauls, the correspond-
ing probability values are only presented once for each instance. On the other, the method
based on Guzman et al. (2017b) depends on the value of β and the specific backhaul cus-
tomers visited, as well as the respective number of visits, and for that the tables report the
corresponding different values for each instance and β.
The results from both Tables 4.2 and 4.3 allow to conclude that much tighter bounds can
be derived for any instance, for any β and for any Γ using the method that estimates prob-
abilistic bounds a posteriori. An interesting observation is that it seems that the method
based on Guzman et al. (2017b) estimates a lower probabilistic bound for solutions that
visit the same backhaul more than once. Taking for example instance C4 with β = 1000,
where it is shown that the probability of constraint violation when Γ = 2 is lower than when
Γ = 2.5, for in the former case, the backhaul customer 15 is visited twice while in the later
case, it is only visited once.
4.4.2.3 Performance of the models

In this section, the two different models under uncertainty, RO and CC, are analyzed and
compared in terms of their performance. The performance metric used is the Price of
Table 4.2 – Probability of constraint violation for instances with three vehicles.
Budget of uncertainty, Γ
Instance β Method 0 0.5 1 1.5 2 2.5 3
Irwin-Hall* 50.00% 31.77% 16.67% 7.03% 2.08% 0.26% 0.00%
0
Guzman post 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
A4 500 Guzman post 0.85% 0.85% 0.85% 0.85% 0.00% 0.00% 0.00%
750 Guzman post 26.17% 26.17% 0.69% 0.69% 0.25% 0.14% -
1000 Guzman post 14.41% 14.41% 8.97% - - - -
Irwin-Hall* 50.00% 31.77% 16.67% 7.03% 2.08% 0.26% 0.00%
0
Guzman post 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
B3 500 Guzman post 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
750 Guzman post 2.21% 2.21% 2.21% 1.42% 1.42% 0.20% -
1000 Guzman post 16.69% 16.69% 15.38% 6.40% - - -
(-) No feasible solution was found.
Table 4.3 – Probability of constraint violation for instances with four vehicles.
Budget of uncertainty, Γ
Instance β Method 0 0.5 1 1.5 2 2.5 3 3.5 4
Irwin-Hall* 50.00% 33.81% 20.05% 10.11% 4.17% 1.32% 0.26% 0.02% 0.00%
0
Guzman post 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
C4 500 Guzman post 1.07% 1.07% 1.07% 1.07% 0.00% 0.00% 0.00% 0.00% 0.00%
750 Guzman post 22.95% 22.95% 0.29% 0.29% 0.29% 0.12% 0.00% 0.00% 0.00%
1000 Guzman post 8.15% 8.15% 8.15% 6.40% 0.23% 0.34% 0.17% 0.01% -
Irwin-Hall* 50.00% 33.81% 20.05% 10.11% 4.17% 1.32% 0.26% 0.02% 0.00%
0
Guzman post 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
F4 500 Guzman post 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00% 0.00%
750 Guzman post 0.38% 0.38% 0.38% 0.38% 0.38% 0.12% 0.00% 0.00% 0.00%
1000 Guzman post 10.07% 10.07% 10.07% 6.26% 0.18% 0.18% 0.18% 0.02% -
Robustness (PoR), which measures the trade-off between constraint violation and the effect
on the nominal objective function (Bertsimas and Sim, 2004), and is determined as the
difference between robust and nominal solutions. Note that the metric PoR must only be
used when the objective is certain, otherwise it would be comparing two different scenarios
- the worst-case and the nominal case (Gorissen et al., 2015).
The results are reported for instances with three and four vehicles respectively in Tables
4.4 and 4.5. The first two columns identify respectively the instance and the model. The
third column presents the value of minimum revenue β and the fourth reports the value of
the nominal solution. The next columns present the value of the PoR obtained with each
model, for each level of probability of constraint violation. Note that the probabilities used
to test the CC model correspond to the probabilities obtained with Equation (4.26) from
the Irwin-Hall distribution method. Thus, a probability of 50.00% corresponds to a Γ = 0,
a probability of 0% corresponds to the maximum Γ applied, and so forth.
Table 4.4 – PoR obtained for RO and CC solutions of instances with three vehicles.
Nominal Probability of constraint violation

Instance Model β
solution 50.00% 31.77% 16.67% 7.03% 2.08% 0.26% 0.00%
0 136,068 0 0 0 0 0 0 0
500 136,581 0 0 0 0 3,193 3,193 3,193
RO 750 136,581 0 0 3,193 3,193 4,284 6,719 -
1000 139,774 0 0 1,091 - - - -
A4
0 136,068 0 0 0 0 0 0 0
500 136,581 0 0 0 0 3,193 3,193 3,193
CC 750 136,581 0 2,762 3,193 3,193 - - -
1000 139,774 0 0 - - - -
0 132,731 0 0 0 0 0 0 0
500 132,731 0 0 0 0 0 0 0
RO 750 132,731 0 0 0 770 770 1,623 -
1000 132,731 0 0 770 9,635 - - -
B3
0 132,731 0 0 0 0 0 0 0
500 132,731 0 0 0 0 0 0 0
CC 750 132,731 0 0 0 5,043 - - -
1000 132,731 0 0 - - - - -
The analysis of both Tables indicate that, firstly, as expected, the PoR increases while
increasing the budget of uncertainty (or decreasing the probability of constraint violation),
for a given level of β. When the PoR remains the same, it means that the solution structure
also remains the same, and so it does not impact the total expected revenue. These values
provide the minimum value of budget of uncertainty that defines the worst-case for a given
β, for each instance. Secondly, increasing the value of β tends to force the creation of a
new routing plan, probably due to the need of augmenting the visits to backhaul customers.
Taking for example, the results obtained with instance B3: when β = 500, the PoR = 0 and
remains the same for all levels of probability; when β = 750, the PoR changes when the
probability is 7.03%; and when β = 1000, the PoR changes sooner, when the probability is
16.67%. Nevertheless, these results are not always in accordance, as the tests with instance
Table 4.5 – PoR obtained for RO and CC solutions of instances with four vehicles.
Nominal Probability of constraint violation

Instance Model β
solution 50.00% 33.81% 20.05% 10.11% 4.17% 1.32% 0.26% 0.02% 0.00%
0 128,792 0 0 0 0 0 0 0 0 0
500 131,464 0 0 0 0 2,117 2,117 2,117 2,117 2,117
RO 750 131,464 0 0 2,117 2,117 2,117 5,371 11,245 11,245 11,245
1000 133,580 0 0 0 3,254 9,341 9,932 10,035 15,317 -
C4
0 128,792 0 0 0 0 0 0 0 0 0
500 131,464 0 0 0 0 369 2,116 2,116 2,116 2,116
CC 750 131,464 0 369 2,116 2,116 2,174 11,244 11,244 11,244 11,244
1000 133,580 0 0 9,128 9,128 11,777 - - - -
0 145,456 0 0 0 0 0 0 0 0 0
500 145,456 0 0 0 0 0 0 0 0 0
RO 750 145,456 0 0 0 0 0 1,861 9,254 9,254 9,254
1000 145,456 0 0 0 1,861 9,253 9,254 9,254 15,061 -
F4
0 145,456 0 0 0 0 0 0 0 0 0
500 145,456 0 0 0 0 0 0 0 0 0
CC 750 145,456 0 0 0 0 9,254 9,254 9,254 9,254 9,254
1000 145,456 0 0 9,254 9,469 19,444 - - - -
C4 reveals that when β = 750, the PoR changes sooner than when β = 1000, although the
PoR value is higher for the highest β.
From the comparison between the two models, it is possible to conclude that the CC model
is more conservative than the RO model, at least for the same probabilistic bounds obtained
with the method of Irwin-Hall distribution. The results with the RO model always provide
a PoR lower or, at most, equal to the CC model, for any of the instances tested, for a given
value of probability. In addition, a PoR > 0 seems to occur sooner for the CC model than
for the RO model. Finally, for all instances tested, the CC model provides more infeasible
solutions than the RO model.
4.4.2.4 Solution structure

For each value of β and Γ, the solution structure (routing plan) obtained with each instance
tested with the RO model is presented in Figure 4.5, specifying which backhaul customers
are visited in each routing plan. All solutions depicted in Figures 4.5a and 4.5b comprise a
plan with 3 routes, while all the solutions in Figures 4.5c and 4.5d comprise a plan with 4
routes.
As expected, increasing the minimum revenue required leads to increase the number of
backhaul customers visited, except in the case of instance B3, as shown in Figure 4.5b. Al-
though, in this case, the first routing plan obtained includes two visits to the same backhaul
customer (1), whereas all the other plans include distinct backhaul customers. It also seems
that, at least, one specific backhaul is always included in any route. For example, backhaul
5 is included in any routing plan of instance A4, backhaul 1 in instance B3, backhaul 15 in
instance C4 and backhaul 9 in instance F4.
With the exception of instance C4, at least one backhaul customer is visited when there is
no requirement of minimum revenue (β = 0). This means that in these cases, it is profitable
to visit a backhaul customer, even if there is no need to. Also, in instances A4 and C4, all
(a) Instance A4
(b) Instance B3
(c) Instance C4
(d) Instance F4
Figure 4.5 – Solution structures obtained by changing β and Γ.

routing plans obtained when there is a minimum revenue required, are different than the
cases when such requirement does not exist. In fact, it seems that both instances are very
sensitive to changes in any parameter, β or Γ. On the other hand, for instances B3 and F4,
the same routing plan is obtained whenever β = 0 or β = 500 or even for any value of β
with, at least, a Γ ≤ 0.5.
4.4.3 Evaluation of solution methods
To evaluate the computational performance of the B&C and the ALNS, both methods are
tested for instances with different vehicle capacities C, numbers of vehicles |K|, linehaul
customers |L| and backhaul customers |B|. These instances are randomly selected from the
universe of Goetschalckx and Jacobs-Blecha (1989), such that, at least, two instances from
each class (from A to N) are obtained (see Table 4.6). The instances are divided into 14
classes A–N. The classes G and L have, each one, 3 instances, and the remaining classes
have 2 instances each. The instances of a class are almost equal: they are different only in
respect to the vehicle capacity C and the number of vehicles |K| while their graphs, costs,
revenues, and deviations are identical. The number of backhaul customers |B|, in relation
to |L|, corresponds to: 25% in class A; 50% in B; 100% in C; about 25% in D; 50% in E;
100% in F; about 25% in G; about 50% in H; 100% in I; about 25% in J; about 50% in K;
100% in L; 25% in N; and 50% in M.
Each instance is tested for two different values of budget of uncertainty, Γ = 0 and Γ = 2,
and for two different values of minimum revenue required, β = 0 and β = 750. For the
B&C, the running time is limited to one hour for each test and is worth recovering that the
RCI tolerance is 0.2. For the ALNS, 30 replications of each test are performed.
Table 4.6 – Instance characteristics
Inst C |K| Inst C |K| Inst C |K| |L| |B|

A2 2550 5 A4 4050 3 20 5
B1 1600 7 B2 2600 5 20 10
C1 1800 7 C4 4150 4 20 20
D2 1700 11 D3 2750 7 30 8
E1 2650 7 E3 5225 4 30 15
F1 3000 6 F4 5500 4 30 30
G2 4300 6 G4 5300 5 G6 8000 4 45 12
H2 5100 5 H3 6100 4 45 23
I1 3000 10 I4 5700 6 45 45
J1 4400 10 J3 8200 6 75 19
K2 6000 8 K4 6200 7 75 38
L1 4400 10 L3 5000 9 L5 6000 8 75 75
M2 5200 10 M4 8000 7 100 25
N2 5700 10 N6 8500 8 100 50
As already mentioned, the VRPSB is NP-hard and, therefore it is not expected to achieve
optimality for solutions obtained with exact methods for large instances. However, the
advantage of the B&C developed is to obtain an adequate lower bound for each instance
and, consequently, measure the efficacy of the ALNS.
The results obtained from these tests are presented in Table 4.7. The first three columns
present respectively, the instance name, the budget of uncertainty Γ and the minimum rev-
enue required β. The next four columns refer to the main results obtained from the B&C,
ordered by the lower bound (LB), solution value (z), the gap of the B&C (GB&C ), and the
time of the tests in seconds (T(s)). The gap corresponds to the percentage difference be-
tween solution value and the respective lower bound. When the solution time of the B&C
reaches 3600s, it means the computing time has stopped at its limit.
The next six columns in Table 4.7 show the results from the ALNS, namely the worst, the
average, and the best solution values given by z, the standard deviation (sd(z)), the total
number of infeasible solutions (#i) obtained out of 30 runs, and the average solution time
in seconds (AT(s)). Finally, the last two columns present the most common performance
metrics used to compare the methods, namely the GSOL and GLB . The former metric is de-
termined by the percentage difference between the best solution value found by the ALNS
and the solution value obtained with the B&C, while the latter compares the ALNS solution
with the lower bound value from the B&C. In addition, the last line in the table presents
the average values of gap and solution time determined for both methods, as well as the
average standard deviation of solutions obtained with ALNS.
The B&C is able to provide optimal solutions for almost all tests with classes of instances
A to H (up to 45 linehaul customers), as determined by the values of GB&C . On the other
hand, the exact method is unable to find a feasible solution in 18 out of the 90 tests. An
interesting observation from the B&C results is related to the tests on the class of instances
I. These have the same number of linehaul customers as the class of instances H, but almost
the double number of backhauls. The difference between instances I1 and I4 are the number
of vehicles, respectively 10 and 6, and vehicle capacities, respectively 3000 and 5700. Yet,
the tests with instance I1 do not provide an optimal solution in one hour, while with instance
I4, the optimal solution is found in no more than one minute.
The class of instances J have almost the same total number of customers (94) of class I
(90), however, the number of linehauls of the J instances represents 167% of the |L| of the
I instances. From this point forward (classes J, K, L, M, and N), no optimal guaranteed
solutions are found and 18 of the 33 tests failed in returning a feasible solution. It is
also noteworthy that among L5 instances the better gap (2.60%) is presented by the robust
version with Γ = 2 and β = 750. Despite its usefulness in the actual solution of the problem,
all the B&C tests provide lower bounds that are useful to analyze the performance of the
ALNS.
The solutions obtained with ALNS can be considered reasonably stable, as indicated by
the average standard deviation of 1.5%. On the other hand, infeasible solutions are found
in seven instances, but only for the worst-cases, i.e. when Γ = 2.
Comparing solutions obtained with the B&C and with the ALNS, it can be observed that
the metaheuristic is able to reach 31 out of 51 of the known optimal solutions proven by the
B&C, given by a GLB of 0.00%. It can also be demonstrated that the highest GLB values
are obtained for the worst-cases. Furthermore, except for two cases (instance H2, Γ = 2,
Table 4.7 – Computational results from B&C and ALNS
B&C Results ALNS results

Inst Γ β LB z GB&C T(s) Worst z Avg z Best z sd(z) #i AT(s) GSOL GLB
A2 0 0 164,457 164,457 0.00% 0.5 164,457 164,457 164,457 0.0% - 0.7 0.00% 0.00%
0 750 164,457 164,457 0.00% 0.8 166,001 165,176 164,457 0.4% - 0.9 0.00% 0.00%
2 750 164,457 164,457 0.00% 0.3 168,480 165,909 164,457 0.5% - 1.0 0.00% 0.00%
A4 0 0 136,068 136,068 0.00% 0.2 139,157 137,181 136,068 1.1% - 0.5 0.00% 0.00%
0 750 136,581 136,581 0.00% 0.2 137,092 136,617 136,581 0.1% - 0.7 0.00% 0.00%
2 750 140,865 140,865 0.00% 0.2 150,866 144,846 140,865 2.2% 17 0.5 0.00% 0.00%
B1 0 0 209,034 209,034 0.00% 1.5 209,089 209,047 209,034 0.0% - 1.0 0.00% 0.00%
0 750 209,034 209,034 0.00% 1.0 209,634 209,222 209,034 0.1% - 1.5 0.00% 0.00%
2 750 209,034 209,034 0.00% 2.3 211,511 209,985 209,089 0.2% - 1.9 0.03% 0.03%
B2 0 0 162,983 162,983 0.00% 3.5 165,428 163,236 162,983 0.3% - 0.7 0.00% 0.00%
0 750 162,983 162,983 0.00% 1.5 168,811 163,343 162,983 0.6% - 1.1 0.00% 0.00%
2 750 163,431 163,431 0.00% 5.9 177,880 168,503 163,473 2.0% 2 1.0 0.03% 0.03%
C1 0 0 186,636 186,636 0.00% 2.4 191,913 187,691 186,636 0.9% - 2.2 0.00% 0.00%
0 750 187,655 187,655 0.00% 6.7 194,101 189,743 187,655 1.2% - 3.7 0.00% 0.00%
2 750 188,842 188,842 0.00% 24.3 200,405 192,907 188,842 1.9% - 4.2 0.00% 0.00%
C4 0 0 128,792 128,792 0.00% 0.2 130,760 129,953 128,792 0.7% - 1.6 0.00% 0.00%
0 750 131,464 131,464 0.00% 0.5 142,010 137,440 131,464 1.9% - 1.7 0.00% 0.00%
2 750 133,580 133,580 0.00% 0.2 147,191 141,817 135,682 2.6% 23 1.5 1.57% 1.57%
D2 0 0 296,077 297,853 0.60% 3600.0 317,855 300,514 297,853 1.8% - 2.3 0.00% 0.60%
0 750 297,853 297,853 0.00% 2724.5 315,653 300,486 297,853 1.5% - 4.6 0.00% 0.00%
2 750 295,983 297,853 0.63% 3600.0 318,202 301,785 297,853 1.8% - 5.5 0.00% 0.63%
D3 0 0 214,017 214,017 0.00% 292.8 215,464 214,395 214,017 0.2% - 1.7 0.00% 0.00%
0 750 214,017 214,017 0.00% 98.8 224,750 215,665 214,017 1.3% - 2.5 0.00% 0.00%
2 750 214,598 214,598 0.00% 294.0 227,802 219,152 214,598 1.5% - 2.8 0.00% 0.00%
E1 0 0 198,239 198,239 0.00% 45.5 204,070 199,989 199,005 0.9% - 1.4 0.39% 0.39%
0 750 198,239 198,239 0.00% 75.0 221,881 211,357 203,036 2.2% - 2.3 2.42% 2.42%
2 750 198,649 198,649 0.00% 41.7 230,836 221,851 217,390 2.0% - 3.4 9.43% 9.43%
E3 0 0 149,576 149,576 0.00% 2.3 161,201 152,294 149,576 3.1% - 2.3 0.00% 0.00%
0 750 151,519 151,519 0.00% 0.8 164,712 156,179 151,519 2.5% - 5.8 0.00% 0.00%
2 750 154,399 154,399 0.00% 2.8 179,894 162,333 154,399 3.7% - 2.8 0.00% 0.00%
F1 0 0 183,723 183,723 0.00% 18.5 193,113 186,508 183,723 1.7% - 4.2 0.00% 0.00%
0 750 183,723 183,723 0.00% 20.5 209,893 188,612 183,723 3.4% - 4.4 0.00% 0.00%
2 750 183,723 183,723 0.00% 23.8 209,893 190,643 183,723 3.8% - 4.6 0.00% 0.00%
F4 0 0 145,456 145,456 0.00% 0.1 149,676 149,041 145,456 0.9% - 2.3 0.00% 0.00%
0 750 145,456 145,456 0.00% 0.1 159,387 150,835 145,638 1.9% - 5.8 0.13% 0.13%
2 750 145,456 145,456 0.00% 0.1 156,699 150,631 147,027 1.8% 16 3.9 1.08% 1.08%
G2 0 0 225,110 225,110 0.00% 80.8 234,233 226,795 225,110 1.0% - 5.7 0.00% 0.00%
0 750 225,110 225,110 0.00% 114.8 234,333 227,184 225,663 0.9% - 5.6 0.25% 0.25%
2 750 225,444 225,444 0.00% 459.2 239,360 231,080 226,568 1.6% - 6.2 0.50% 0.50%
G4 0 0 202,551 202,551 0.00% 1.6 202,961 202,725 202,551 0.1% - 3.3 0.00% 0.00%
0 750 202,551 202,551 0.00% 4.3 217,448 204,516 202,551 1.4% - 3.7 0.00% 0.00%
2 750 203,320 203,320 0.00% 22.8 221,936 209,843 204,237 2.0% 2 4.1 0.45% 0.45%
G6 0 0 180,176 180,176 0.00% 64.6 189,408 183,619 181,322 1.0% - 2.6 0.64% 0.64%
0 750 180,630 180,630 0.00% 60.5 196,594 187,711 180,947 2.4% - 3.9 0.18% 0.18%
2 750 182,081 182,081 0.00% 71.4 207,143 198,803 191,603 2.7% 13 3.7 5.23% 5.23%
Table 4.7 (cont.) – Computational results from B&C and ALNS
B&C Results ALNS results

Inst Γ β LB z GB&C T(s) Worst z Avg z Best z sd(z) #i AT(s) GSOL GLB
H2 0 0 192,446 192,446 0.00% 149.9 200,616 196,179 192,904 1.1% - 4.8 0.24% 0.24%
0 750 193,092 193,092 0.00% 103.2 207,403 200,886 195,225 1.8% - 6.9 1.10% 1.10%
2 750 197,064 197,545 0.24% 3600.1 216,619 207,616 202,268 2.0% - 10.7 2.39% 2.64%
H3 0 0 182,499 182,499 0.00% 11.5 195,735 186,102 182,499 2.1% - 9.6 0.00% 0.00%
0 750 186,316 186,316 0.00% 26.7 203,299 192,831 187,053 1.9% - 10.9 0.40% 0.40%
2 750 188,727 188,727 0.00% 74.9 206,504 198,095 190,855 1.7% - 16.6 1.13% 1.13%
I1 0 0 244,435 261,403 6.94% 3600.1 267,657 259,439 256,536 1.2% - 10.4 -1.86% 4.95%
0 750 244,074 260,212 6.61% 3600.0 262,249 258,349 256,474 0.7% - 12.1 -1.44% 5.08%
2 750 244,741 262,372 7.20% 3600.0 265,021 259,621 256,717 0.9% - 24.5 -2.16% 4.89%
I4 0 0 191,397 191,397 0.00% 49.2 201,384 193,115 191,397 1.3% - 7.8 0.00% 0.00%
0 750 191,397 191,397 0.00% 60.0 202,672 193,951 191,462 1.8% - 7.4 0.03% 0.03%
2 750 191,397 191,397 0.00% 46.3 204,411 195,577 191,619 2.2% 1 8.8 0.12% 0.12%
J1 0 0 274,601 - - 3600.3 318,741 303,254 304,320 1.2% - 22.6 - 10.82%
0 750 273,684 - - 3600.1 317,682 303,092 303,092 1.4% - 28.1 - 10.75%
2 750 275,849 - - 3600.1 324,314 306,533 309,652 1.3% - 35.5 - 12.25%
J3 0 0 216,813 245,947 13.44% 3600.2 262,717 234,677 225,100 3.9% - 14.7 -8.48% 3.82%
0 750 218,177 232,069 6.37% 3600.3 257,619 238,005 226,555 3.7% - 17.5 -2.38% 3.84%
2 750 215,679 238,870 10.75% 3600.1 253,178 238,957 228,175 2.3% - 16.1 -4.48% 5.79%
K2 0 0 264,088 327,849 24.14% 3600.4 284,098 278,224 275,513 0.9% - 24.7 -15.96% 4.33%
0 750 267,466 281,093 5.10% 3600.1 288,174 279,027 275,745 1.1% - 37.7 -1.90% 3.10%
2 750 264,448 314,323 18.86% 3600.1 295,169 283,514 276,723 1.5% - 40.3 -11.96% 4.64%
K4 0 0 245,828 269,840 9.77% 3600.1 266,449 259,759 257,710 0.8% - 17.8 -4.50% 4.83%
0 750 246,304 266,482 8.19% 3600.2 271,202 262,880 258,558 1.4% - 24.6 -2.97% 4.98%
2 750 247,350 269,572 8.98% 3600.1 271,429 266,705 260,791 1.1% - 29.6 -3.26% 5.43%
L1 0 0 266,306 - - 3600.2 307,468 294,782 291,604 1.1% - 37.6 - 9.50%
0 750 268,661 - - 3600.1 306,696 294,407 291,850 0.9% - 42.4 - 8.63%
2 750 268,488 - - 3600.2 300,141 294,134 291,505 0.8% - 53.4 - 8.57%
L3 0 0 254,446 - - 3600.8 291,922 279,328 274,465 1.1% - 39.6 - 7.87%
0 750 253,382 - - 3600.4 282,923 278,122 273,860 0.8% - 39.0 - 8.08%
2 750 251,903 - - 3600.3 284,272 278,334 274,561 0.9% - 96.3 - 8.99%
L5 0 0 236,821 287,121 21.24% 3600.3 270,441 257,813 249,581 1.7% - 34.4 -13.07% 5.39%
0 750 235,730 268,902 14.07% 3600.2 271,250 257,389 249,666 2.4% - 40.9 -7.15% 5.91%
2 750 243,111 249,420 2.60% 3600.1 276,468 261,355 250,822 2.3% 2 37.5 0.56% 3.17%
M2 0 0 311,663 - - 3600.1 356,401 348,356 342,540 1.0% - 52.3 - 9.91%
0 750 305,740 - - 3600.1 365,427 349,324 341,859 1.8% - 61.9 - 11.81%
2 750 306,403 - - 3600.1 383,867 354,778 342,164 2.6% - 69.9 - 11.67%
M4 0 0 267,153 316,326 18.41% 3600.1 295,958 287,917 282,514 1.3% - 30.8 -10.69% 5.75%
0 750 268,303 312,923 16.63% 3600.2 294,224 287,284 282,018 1.0% - 68.7 -9.88% 5.11%
2 750 266,505 - - 3600.5 306,852 293,250 284,497 1.7% - 81.5 - 6.75%
N2 0 0 300,459 - - 3600.3 347,618 338,144 328,669 1.3% - 57.5 - 9.39%
0 750 298,437 - - 3600.1 348,852 338,435 329,922 1.4% - 63.3 - 10.55%
2 750 302,526 - - 3600.4 354,570 338,931 328,484 1.4% - 78.8 - 8.58%
N6 0 0 251,446 332,761 32.34% 3601.0 289,457 278,882 275,535 1.1% - 29.1 -17.20% 9.58%
0 750 252,833 - - 3600.1 289,093 279,720 275,533 1.2% - 36.9 - 8.98%
2 750 247,635 - - 3600.1 297,811 283,606 276,690 1.9% - 45.6 - 11.73%
Avg 3.24% 1616.7 1.5% 18.5 -1.26% 3.27%
β = 750 and instance L5, Γ = 2, β = 750) the GSOL is always non-positive for instances
where feasible solutions – with no guarantee of optimality – were found by the B&C,
which means that the ALNS can generate better solutions for these instances than the B&C,
at least for one hour of computing time.
With respect to the solution time, the ALNS can obtain solutions very fast. In fact, even for
the larger instances (L, M and N), a solution is found in less than 2 minutes, on average.
From these results, we can argue that the ALNS developed is a very efficient and an effec-
tive solution method for the VRPSB under uncertainty, thus generating very good quality
solutions in a very short time.
From Table 4.7, the performance of the robust optimization approach used in this work can
also be demonstrated. The results do not allow to infer on the solution time of the robust
optimization, since for some instances, the computing time of deterministic solutions are
inferior to the computing time of robust solutions (e.g., G2), but for other instances, the
opposite occur (e.g., I4). Nevertheless, it can be shown that, at least, the robust modelling
has no more significant intractability than the deterministic modelling. Based on these
results, it can be concluded that the robust optimization used in this work is efficient to
solve the VRPSB under revenue uncertainty.

This work presents the first study of a Vehicle Routing Problem with Backhauls (VRPB)
under uncertainty. More precisely, we consider the problem where the visits to backhauls
are optional, each visit has an associated revenue, which is uncertain, and a minimum rev-
enue is required to satisfy the depot demand. Following a robust optimization approach,
the revenues are considered to be bounded by a polyhedron set of uncertainty and a param-
eter, called budget of uncertainty, is applied to control the size of the uncertainty set. An
alternative robust approach based on a factor model support is designed and illustrated in a
simple example from a hypothetical case of a wood-based industry. A chance-constrained
model is also designed for the VRPB under uncertain revenues and compared with the
robust optimization model. Four different methods to estimate probabilistic bounds for
the minimum revenue constraint violation are defined, tested and compared. Finally, two
different solution methods, one exact and one metaheuristic, are developed to solve both
deterministic and robust problems.
Several insights could be retrieved from this study. First, nominal solutions are very sensi-
tive to uncertainty, since the structure of the solution changes quickly even for low values of
budget of uncertainty. Second, and as expected, the cost of robust solutions is higher than
the cost of nominal solutions. The solutions obtained with the chance-constrained model
are much more conservative than the solutions obtained with the robust optimization model.
In terms of probabilistic bounds, the method that estimates a posteriori the probability of
constraint violation of a solution is the most rigorous one, since it produces the exact value
for that probability. However, it requires the knowledge of the solution structure. A novel
method based on the Irwin-Hall distribution provides much tighter bounds than the other
a priori probabilistic methods, which allows to better characterize the bounds of robust
Bibliography 137
solutions based only on values of budget of uncertainty.

The B&C developed in this work finds optimal solutions in a reasonable time, but only
for small and medium instances. However, the computing time was limited to one hour
in this work. As such, increasing the time limit may result in a higher lower bound and,
consequently, in a considerably lower gap. On the other hand, the ALNS shows high
efficiency and efficacy to solve all instances, thus obtaining high quality solutions in a very
short time (less than two minutes). It is also shown that the robust approach used in this
work can create a tractable robust model, where solutions may be obtained with similar
computing effort as with the deterministic model.
Different directions for future work on this subject arise. Firstly, considering that very few
exact methods have been successfully applied in VRPBs, it may be interesting to evalu-
ate more promising solution methods, such as Branch-and-Price algorithms. Thus, better
lower bounds of the problem may be found, which improves the evaluation of the per-
formance of metaheuristics, such as the ALNS developed. Another future direction for
research concerns the modeling aspects of uncertainty sets. For instance, the uncertainty
set may be built using a known distribution or some statistics obtained from historical in-
formation, such as the distributionally robust optimization (Gabrel et al., 2014). As such,
the uncertainty set can be represented by tight bounds, which may produce better proba-
bilistic bounds for the robust problem and decrease the conservatism of robust solutions.
Finally, it would be interesting to apply the robust VRPSB in practice and gather insights
from the perspective of integrated logistics.
Acknowledgements
This work is financed by the ERDF – European Regional Development Fund through
(Easyflow). This research is also financed by Project "TEC4Growth - Pervasive Intelli-
gence, Enhancers and Proofs of Concept with Industrial Impact/NORTE-01-0145-FEDER-
000020" is financed by the North Portugal Regional Operational Programme (NORTE
2020), under the PORTUGAL 2020 Partnership Agreement, and through the European
Regional Development Fund (ERDF). Finally, this research has also been partially sup-
ported by CNPq procs. 425340/2016-3 and 314366/2018-0, FAPESP procs. 2015/11937-9,
2016/01860-1 and 2018/08879-5, and SETI-PR / Fundação Araucária / CAPES.
Bibliography
Abbassi, A., hilali Alaoui, A. E., and Boukachour, J. (2019). Robust optimisation of the
intermodal freight transport problem: Modeling and solving with an efficient hybrid
approach. Journal of Computational Science, 30:127 – 142.
Allahviranloo, M., Chow, J. Y., and Recker, W. W. (2014). Selective vehicle routing prob-
lems under uncertainty without recourse. Transportation Research Part E: Logistics and
Transportation Review, 62:68 – 88.
Alvarez, P. P. and Vera, J. R. (2014). Application of robust optimization to the sawmill

planning problem. Annals of Operations Research, 219(1):457–475.
Andersson, G., Flisberg, P., Nordström, M., Rönnqvist, M., and Wilhelmsson, L. (2016).
A model approach to include wood properties in log sorting and transportation planning.
INFOR: Information Systems and Operational Research, 54(3):282–303.
Averbakh, I. (2001). On the complexity of a class of combinatorial optimization problems

with uncertainty. Mathematical Programming, 90(2):263–272.
vehicle routing problem. The Journal of the Operational Research Society, 61(7):1072–
1077.
Baldacci, R., Mingozzi, A., and Roberti, R. (2012). Recent exact algorithms for solving the
vehicle routing problem under capacity and time window constraints. European Journal
Bertsimas, D. and Sim, M. (2003). Robust discrete optimization and network flows. Math-
ematical Programming, 98(1):49–71.
Bertsimas, D. and Sim, M. (2004). The price of robustness. Operations Research,

52(1):35–53.
Birge, J. R. and Louveaux, F. (2011). Basic Properties and Theory. In Introduction to

Stochastic Programming, pages 103–161. Springer, New York, NY.
Braaten, S., Gjønnes, O., Hvattum, L. M., and Tirado, G. (2017). Heuristics for the robust
vehicle routing problem with time windows. Expert Systems with Applications, 77:136
– 147.
Bruck, B. P., dos Santos, A. G., and Arroyo, J. E. C. (2012). Hybrid metaheuristic for
the single vehicle routing problem with deliveries and selective pickups. In 2012 IEEE
Congress on Evolutionary Computation. IEEE.
Cao, E., Lai, M., and Yang, H. (2014). Open vehicle routing problem with demand uncer-
tainty and its robust strategies. Expert Systems with Applications, 41(7):3569 – 3575.
Carlsson, D. and Rönnqvist, M. (2007). Backhauling in forest transportation: models,

methods, and practical usage. Canadian Journal of Forest Research, 37(12):2612–2623.
Chardy, M. and Klopfenstein, O. (2012). Handling uncertainties in vehicle routing prob-

lems through data preprocessing. Transportation Research Part E: Logistics and Trans-
Chassein, A. (2016). Approximation of ellipsoids using bounded uncertainty sets.

Bibliography 139
Coelho, I. M., Munhoz, P. L. A., Haddad, M. N., Souza, M. J. F., and Ochi, L. S. (2012).
A hybrid heuristic based on general variable neighborhood search for the single vehi-
cle routing problem with deliveries and selective pickups. Electronic Notes in Discrete
Mathematics, 39:99 – 106.
Dayarian, I., Crainic, T. G., Gendreau, M., and Rei, W. (2016). An adaptive large-
neighborhood search heuristic for a multi-period vehicle routing problem. Transporta-
tion Research Part E: Logistics and Transportation Review, 95:95–123.
Demir, E., Bektaş, T., and Laporte, G. (2012). An adaptive large neighborhood search
heuristic for the pollution-routing problem. European Journal of Operational Research,
223(2):346 – 359.
Derigs, U., Pullmann, M., Vogel, U., Oberscheider, M., Gronalt, M., and Hirsch, P. (2012).
Multilevel neighborhood search for solving full truckload routing problems arising in
timber transportation. Electronic Notes in Discrete Mathematics, 39:281 – 288. EURO
Mini Conference.
Dominguez, O., Guimarans, D., Juan, A. A., and de la Nuez, I. (2016). A Biased-
Randomised Large Neighbourhood Search for the two-dimensional Vehicle Routing
Problem with Backhauls. European Journal of Operational Research, 255(2):442–462.
Eshtehadi, R., Demir, E., and Huang, Y. (2020). Solving the vehicle routing problem
with multi-compartment vehicles for city logistics. Computers & Operations Research,
115:104859.
Fardi, K., Jafarzadeh_Ghoushchi, S., and Hafezalkotob, A. (2019). An extended robust

approach for a cooperative inventory routing problem. Expert Systems with Applications,
116:310 – 327.
Fliedner, T. and Liesiö, J. (2016). Adjustable robustness for multi-attribute project portfolio
selection. European Journal of Operational Research, 252(3):931 – 946.
Gabrel, V., Murat, C., and Thiele, A. (2014). Recent advances in robust optimization: An
overview. European Journal of Operational Research, 235(3):471 – 483.
García-Nájera, A., Bullinaria, J. A., and Gutiérrez-Andrade, M. A. (2015). An evolutionary

approach for multi-objective vehicle routing problems with backhauls. Computers &
Industrial Engineering, 81:90–108.
Ghilas, V., Demir, E., and van Woensel, T. (2016). An adaptive large neighborhood search
heuristic for the pickup and delivery problem with time windows and scheduled lines.
hauls. European Journal of Operational Research, 42(1):39 – 51.
Gorissen, B. L., Yanikoǧlu, I., and den Hertog, D. (2015). A practical guide to robust
optimization. Omega, 53:124–137.
Gounaris, C. E., Repoussis, P. P., Tarantilis, C. D., Wiesemann, W., and Floudas, C. A.
(2016). An adaptive memory programming framework for the robust capacitated vehicle
routing problem. Transportation Science, 50(4):1239–1260.
Gounaris, C. E., Wiesemann, W., and Floudas, C. A. (2013). The robust capacitated vehicle
routing problem under demand uncertainty. Operations Research, 61(3):677–693.
Granada-Echeverri, M., Toro, E. M., and Santa, J. J. (2019). A mixed integer linear pro-
gramming formulation for the vehicle routing problem with backhauls. International
Journal of Industrial Engineering Computations, 10:295–308.
Groß, P.-O., Ehmke, J. F., and Mattfeld, D. C. (2019). Cost-efficient and reliable city logis-
tics vehicle routing with satellite locations under travel time uncertainty. Transportation
Research Procedia, 37:83 – 90. 21st EURO Working Group on Transportation Meeting,
EWGT 2018, 17th – 19th September 2018, Braunschweig, Germany.
Gutiérrez-Jarpa, G., Marianov, V., and Obreque, C. (2009). A single vehicle routing prob-
lem with fixed delivery and optional collections. IIE Transactions, 41(12):1067–1079.
Guzman, Y. A., Matthews, L. R., and Floudas, C. A. (2016). New a priori and a poste-
riori probabilistic bounds for robust counterpart optimization: I. unknown probability
distributions. Computers & Chemical Engineering, 84:568 – 598.
Guzman, Y. A., Matthews, L. R., and Floudas, C. A. (2017a). New a priori and a posteriori
probabilistic bounds for robust counterpart optimization: Ii. a priori bounds for known
symmetric and asymmetric probability distributions. Computers & Chemical Engineer-
ing, 101:279–311.
Guzman, Y. A., Matthews, L. R., and Floudas, C. A. (2017b). New a priori and a poste-
riori probabilistic bounds for robust counterpart optimization: III. Exact and near-exact
a posteriori expressions for known probability distributions. Computers & Chemical
Engineering, 103:116–143.
Hirsch, P. (2011). Minimizing Empty Truck Loads in Round Timber Transport with Tabu
Search Strategies. International Journal of Information Systems and Supply Chain Man-
agement, 4(2):15–41.
Hu, C., Lu, J., Liu, X., and Zhang, G. (2018). Robust vehicle routing problem with hard
time windows under demand and travel time uncertainty. Computers & Operations Re-
search, 94:139 – 153.
Irnich, S., Toth, P., and Vigo, D. (2014). Chapter 1: The Family of Vehicle Routing Prob-
lems, pages 1–33.
Bibliography 141
Kang, S.-C., Brisimi, T. S., and Paschalidis, I. C. (2015). Distribution-dependent robust

linear optimization with applications to inventory control. Annals of operations research,
231(1):229–263.
perspectives. Computers & Operations Research, 91:79 – 91.
Kumar, R., Unnikrishnan, A., and Waller, S. T. (2011). Capacitated-Vehicle Routing Prob-
lem with Backhauls on Trees. Transportation Research Record: Journal of the Trans-
portation Research Board, 2263(1):92–102.
Li, B., Krushinsky, D., van Woensel, T., and Reijers, H. A. (2016a). An adaptive large
neighborhood search heuristic for the share-a-ride problem. Computers & Operations
Research, 66:170–180.
Li, D. and Morales, D. R. (2017). The robust uncapacitated lot sizing model with uncer-
tainty range.
Li, Y., Chen, H., and Prins, C. (2016b). Adaptive large neighborhood search for the pickup
and delivery problem with time windows, profits, and reserved requests. European Jour-
nal of Operational Research, 252(1):27–38.
Li, Z., Ding, R., and Floudas, C. A. (2011). A comparative theoretical and computa-
tional study on robust counterpart optimization: I. robust linear optimization and ro-
bust mixed integer linear optimization. Industrial & Engineering Chemistry Research,
50(18):10567–10603.
Liu, S.-C. and Chung, C.-H. (2008). A heuristic method for the vehicle routing problem
with backhauls and inventory. Journal of Intelligent Manufacturing, 20(1):29.
Lysgaard, J., Letchford, A. N., and Eglese, R. W. (2004). A new branch-and-cut algorithm
for the capacitated vehicle routing problem. Mathematical Programming, 100(2):423–
445.
Marengo, J. E., Farnsworth, D. L., and Stefanic, L. (2017). A Geometric Derivation of

the Irwin-Hall Distribution. International Journal of Mathematics and Mathematical
Sciences, 2017:1–6.
Martello, S. and Toth, P. (1990). Knapsack Problems: Algorithms and Computer Imple-
mentations. John Wiley & Sons, Inc., New York, NY, USA.
Masson, R., Lehuédé, F., and Péton, O. (2013). An adaptive large neighborhood search for
the pickup and delivery problem with transfers. Transportation Science, 47(3):344–355.
Moghaddam, B. F., Ruiz, R., and Sadjadi, S. J. (2012). Vehicle routing problem with
uncertain demands: An advanced particle swarm algorithm. Computers & Industrial
Engineering, 62(1):306 – 317.
Mulvey, J. M., Vanderbei, R. J., and Zenios, S. A. (1995). Robust optimization of large-
scale systems. Operations research, 43(2):264–281.
Ordóñez, F. (2010). Robust Vehicle Routing, chapter Chapter 7, pages 153–178.
Pelletier, S., Jabali, O., and Laporte, G. (2019). The electric vehicle routing problem
with energy consumption uncertainty. Transportation Research Part B: Methodological,
126:225 – 255.
Pisinger, D. and Ropke, S. (2010). Large neighborhood search. In Handbook of Meta-

heuristics, pages 399–419. Springer US.
Poggi, M. and Uchoa, E. (2014). Chapter 3: New Exact Algorithms for the Capacitated
Vehicle Routing Problem. In Vehicle Routing, pages 59–86. Society for Industrial and
Applied Mathematics, Philadelphia, PA.
40(8):2985 – 2991.
Privé, J., Renaud, J., Boctor, F., and Laporte, G. (2006). Solving a vehicle-routing prob-
lem arising in soft-drink distribution. Journal of the Operational Research Society,
57(9):1045–1052.
Ropke, S. and Pisinger, D. (2006). An adaptive large neighborhood search heuristic for the
pickup and delivery problem with time windows. Transportation Science, 40(4):455–
472.
Schiffer, M. and Walther, G. (2018). Strategic planning of electric logistics fleet networks:
A robust location-routing approach. Omega, 80:31 – 42.
Shaw, P. (1998). Using constraint programming and local search methods to solve vehicle
routing problems. In Principles and Practice of Constraint Programming — CP98,
pages 417–431. Springer Berlin Heidelberg.
Solano-Charris, E., Prins, C., and Santos, A. C. (2015). Local search based metaheuristics
for the robust vehicle routing problem with discrete scenarios. Applied Soft Computing,
32:518 – 531.
Solomon, M. M. (1987). Algorithms for the Vehicle Routing and Scheduling Problems
with Time Window Constraints. Operations Research, 35(2):254–265.
Soyster, A. L. (1973). Convex programming with set-inclusive constraints and applications

to inexact linear programming. Operations Research, 21(5):1154–1157.
Süral, H. and Bookbinder, J. H. (2003). The single-vehicle routing problem with unre-
stricted backhauls. Networks, 41(3):127–136.
Bibliography 143
Tirkolaee, E. B., Mahdavi, I., and Esfahani, M. M. S. (2018). A robust periodic capacitated
arc routing problem for urban waste collection considering drivers and crew’s working
time. Waste Management, 76:138 – 146.
Toklu, N. E., Gambardella, L. M., and Montemanni, R. (2014). A multiple ant colony
system for a vehicle routing problem with time windows and uncertain travel times.
Journal of Traffic and Logistics Engineering, 2:52–58.
Toklu, N. E., Montemanni, R., and Gambardella, L. M. (2013). An ant colony system
for the capacitated vehicle routing problem with uncertain travel costs. In 2013 IEEE
Symposium on Swarm Intelligence (SIS), pages 32–39.
Toth, P. and Vigo, D. (1997). An Exact Algorithm for the Vehicle Routing Problem with
Backhauls. Transportation Science, 31(4):372–385.
Toth, P. and Vigo, D. (2014). Vehicle Routing. Society for Industrial and Applied Mathe-
matics, Philadelphia, PA, 2 edition.
Tütüncü, G. Y., Carreto, C., and Baker, B. (2009). A visual interactive approach to classical
and mixed vehicle routing problems with backhauls. Omega, 37(1):138–154.
Wu, L., Hifi, M., and Bederina, H. (2017). A new robust criterion for the vehicle routing
problem with uncertain travel time. Computers & Industrial Engineering, 112:607 –
615.
Zajac, S. (2017). An adaptive large neighborhood search for the periodic vehicle routing
problem. In Lecture Notes in Computer Science, pages 34–48. Springer International
Publishing.
Zanjani, M. K., Ait-Kadi, D., and Nourelfath, M. (2010). Robust production planning in a
manufacturing environment with random yield: A case in sawmill production planning.
Appendix 4.A Chance-constrained VRPSB

In this section, a chance-constraint (CC) approach is proposed in order to benchmark it with
the robust optimization method. In a CC model, the constraint of minimum revenue must
be satisfied with a probability (1 − α), as described in Equation (4.33), where α represents
the probability of constraint violation. In this case, it is also assumed that the parameters
p̃ j are independent and follow an uniform distribution function in [ p̄ j − ṗ j , p̄ j + ṗ j ]:
 
X 
Pr  p̃ j x( j,n+m+1) ≥ β ≥ 1 − α. (4.33)
j∈B
To produce a tractable counterpart of the CC model, an approximate method is applied.

The idea of the method is to approximate the resulting distribution in an uniform distri-
bution supported in [ j∈B ( p̄ j − ṗ j )x( j,n+m+1) , j∈B ( p̄ j + ṗ j )x( j,n+m+1) ], since it is extremely
P P
hard to calculate the convolution sum of the non-identical uncertain parameters p̃ j and the
approximate uniform distribution can be used as an over-estimator of the actual resulting
distribution. The approximation is described as follows:
P
j∈B 2 ṗ j x j,n+m+1
X
p̄ j x j,n+m+1 − Υ (1 − α)−1
√ ≥ β, (4.34)
j∈B 12
where Υ−1 is the inverse of √a cumulative

√ P distribution function of the standardized uniform
distribution bounded
P
in [− 3. 3], j∈B p̄ j x j,n+m+1 is the mean of the resulting uniform
2 ṗ x
distribution and j∈B √j j,n+m+1 is its standard deviation. By replacing (4.7) with (4.34), the
12
resulting CC model reads as follows:
X X
(i, j)∈E j∈B
subjected to
(4.2) − (4.6)
(4.8) − (4.11)
(4.34)
Appendix 4.B Comparison of single and multiple backhaul vis-

its
In this section, we present and discuss the modelling and algorithmic aspects of allowing
multiples backhaul visits per route in the robust VRPSB.
The robust model presented in Section 4.2.2 allows for only one backhaul visit per route.
4.B. Comparison of single and multiple backhaul visits 145
Allowing multiple backhaul visits forces the model to include all potential arcs between
one linehaul and one backhaul customers and between two backhaul customers. Thus,
the set of edges must change to Ē, given by {(0, j) : j ∈ L} ∪ {(i, j) : i ∈ L ∪ {n+m+1}, j ∈
L ∪ {n+m+1}, i < j} ∪ {(i, j) : i ∈ L, j ∈ B} ∪ {(i, j) : i ∈ B, j ∈ B} ∪ {(i, n+m+1) : i ∈ B}.
Therefore, the model with multiple backhaul visits is formulated as follows:
X XX
min ci j xi j − p̄ j xi, j (4.35)
(i, j)∈E i∈V j∈B
subjected to
(4.2) − (4.4)
(4.6)
(4.9) − (4.11)
(4.17)
X
xi j = 2, ∀i ∈ B, (4.36)
(i, j)∈δ(i)
X
xi j ≥ 2r(S), ∀S ⊆ L ∪ B, |S| ≥ 2, (4.37)
(i, j)∈δ(S)
XX X
p̄ j xi, j − λΓ − µ j ≥ β, (4.38)
i∈V j∈B j∈B
X
λ+µj ≥ ṗ j xi, j , ∀ j ∈ B, (4.39)
i∈V
Note that, in the case of single backhaul visits, the worst-case is achieved when the budget
of uncertainty is the minimum number between the number of routes and the number of
backhaul customers, i.e., Γ = min{|B|, |K|}. In the case of multiple backhaul visits, how-
ever, the number of routes do not influence the robust parameter Γ, since any number of
backhauls can be visited by all routes. Thus the worst-case is achieved when Γ = |B|, which
leads to a higher number of possible scenarios.
To analyze the different solutions and the computational performance of both modelling
cases, tests are carried with four small instances, for different values of β and Γ. The tests
are performed with OPL (using a commodity flow formulation) and the time limit is set to
600 seconds. The results are reported in Table 4.9. The first three columns describe the
instance, the value of β and the value of Γ, respectively. The next four columns report the
objective function (z), the total distance (d), the MIP gap (Gap) and the time in seconds
(T (s)) for the case of single visits. The last four columns report the same indicators for the
case of multiple visits.
In terms of solution structure, the results show that allowing multiple backhaul visits always
lead to better or equal solutions than in the case of restricting the visits to one per route. For
example, with instances A4, the same value of total revenue can be achieved with shorter
distances, since several backhaul customers can be visited in the same route (e.g., instead
of visiting three backhauls in three routes, the same three backhauls are visited in just two
routes). With instances C4, a lower value of total revenue may be achieved but this still
compensates the lower total distance traveled. In fact, allowing multiple visits requires
only three routes to satisfy both the linehaul customers demand and the minimum required
revenue, whereas in the case of single visits, four routes must be created. This also occurs
for instance F4. Moreover, it is possible to observe that allowing multiple visits always
lead to feasible solutions, even for the highest values of Γ and β. For the case of single
visits, however, no feasible solutions are obtained for the worst-cases if β = 1000, which
can be seen for all the four instances tested.
In terms of computational performance, allowing multiple visits tends to lead to high com-
putational times and MIP gaps, because the number of variables increases drastically and
several combinations of two or more backhauls can be created for the same route. In
fact, whereas the previous model that tackles single visits considers only the backhaul arc
xi,n+m+1 (∀i ∈ B), the model that allows multiple visits considers also the arcs linking line-
haul to backhaul customers and between backhaul customers, i.e. x j,i (∀ j ∈ V/{0}, ∀i ∈ B).
4.B. Comparison of single and multiple backhaul visits 147
Table 4.9 – Comparison of single and multiple backhaul visits.
Single backhaul visits Multiple backhaul visits

Instance β Γ z d Gap T(s) z d Gap T(s)
0 0 136,068 136,532 0.01% 7 136,068 136,532 0.01% 20
500 0 136,581 137,449 0.01% 9 136,581 137,449 0.01% 25
500 1 136,581 137,449 0.01% 11 136,581 137,449 0.01% 30
500 2 139,774 141,106 0.01% 31 139,010 140,342 0.01% 53
500 3 139,774 141,106 0.01% 27 139,010 140,342 0.01% 49
500 4 139,774 141,106 0.00% 8 139,010 140,342 0.01% 31
A4 500 5 139,774 141,106 0.01% 11 139,010 140,342 0.01% 24
1000 0 139,774 141,106 0.01% 8 139,010 140,342 0.01% 43
1000 1 140,865 142,171 0.00% 4 139,131 140,437 0.01% 23
1000 2 – 143,294 145,064 0.01% 92
1000 3 – 145,928 148,128 0.01% 35
1000 4 – 145,928 148,128 0.01% 37
1000 5 – 145,928 148,128 0.01% 14
0 0 132,731 134,061 0.01% 3 132,731 134,061 0.01% 2
500 0 132,731 134,061 0.01% 4 132,731 134,061 0.00% 8
500 1 132,731 134,061 0.01% 4 132,731 134,061 0.01% 5
500 2 132,731 134,061 0.01% 2 132,731 134,061 0.01% 6
500 3 132,731 134,061 0.00% 2 132,731 134,061 0.01% 6
500 4 132,731 134,061 0.01% 2 132,731 134,061 0.00% 3
500 5 132,731 134,061 0.01% 3 132,731 134,061 0.01% 8
B3 500 10 132,731 134,061 0.00% 2 132,731 134,061 0.01% 2
1000 0 132,731 134,061 0.01% 1 132,731 134,061 0.00% 7
1000 1 133,501 134,748 0.01% 1 133,501 134,748 0.01% 6
1000 2 – 134,010 136,179 0.01% 5
1000 3 – 134,010 136,179 0.00% 6
1000 4 – 134,010 136,179 0.00% 4
1000 5 – 134,010 136,179 0.00% 5
1000 10 – 134,010 136,179 0.00% 1
(–) No feasible solution was found.
Table 4.9 (cont.) – Comparison of single and multiple backhaul visits.
Single backhaul visits Multiple backhaul visits

Instance β Γ z d Gap T(s) z d Gap T(s)
0 0 128,792 128,792 0.01% 22 128,792 128,792 0.01% 59
500 0 131,464 132,366 0.01% 11 130,108 130,987 3.06% 600
500 1 131,464 132,366 0.01% 54 130,108 130,987 0.01% 150
500 2 133,580 134,918 0.01% 378 130,619 131,882 3.41% 601
500 3 133,580 134,918 0.01% 137 130,619 131,882 2.15% 600
500 4 133,580 134,918 0.01% 126 130,619 131,882 0.01% 191
500 5 133,580 134,918 0.01% 157 130,619 131,882 2.13% 601
500 10 133,580 134,918 0.01% 36 130,619 131,882 0.01% 103
500 15 133,580 134,918 0.01% 16 130,619 131,882 0.01% 311
C4 500 20 133,580 134,918 0.01% 16 130,619 131,882 0.01% 76
1000 0 133,580 134,918 0.00% 3 130,619 131,882 0.01% 144
1000 1 133,580 134,918 0.01% 5 130,619 131,882 0.01% 235
1000 2 142,921 144,700 0.01% 52 131,721 133,474 0.01% 421
1000 3 143,615 145,345 0.01% 26 131,721 133,474 0.01% 462
1000 4 – 132,232 134,369 3.02% 601
1000 5 – 133,031 135,248 0.01% 364
1000 10 – 133,031 135,248 0.01% 248
1000 15 – 133,031 135,248 0.01% 289
1000 20 – 133,031 135,248 0.01% 211
0 0 145,456 146,749 0.01% 207 145,456 146,749 0.75% 600
500 0 145,456 146,749 0.01% 60 145,456 146,749 0.85% 601
500 1 145,456 146,749 0.01% 115 145,456 146,749 0.01% 542
500 2 145,456 146,749 0.01% 243 145,456 146,749 0.01% 312
500 3 145,456 146,749 0.01% 202 145,456 146,749 1.26% 601
500 4 145,456 146,749 0.00% 97 145,456 146,749 0.01% 290
500 5 145,456 146,749 0.00% 73 145,456 146,749 1.54% 601
500 10 145,456 146,749 0.00% 72 145,456 146,749 0.01% 318
500 15 145,456 146,749 0.00% 78 145,456 146,749 0.56% 600
500 20 145,456 146,749 0.00% 63 145,456 146,749 0.19% 600
500 25 145,456 146,749 0.00% 59 145,456 146,749 0.01% 175
F4 500 30 145,456 146,749 0.00% 62 145,456 146,749 0.01% 258
1000 0 145,456 146,749 0.00% 53 145,456 146,749 0.01% 204
1000 1 145,456 146,749 0.00% 54 145,456 146,749 0.01% 329
1000 2 154,710 156,383 4.09% 600 150,349 152,456 4.57% 601
1000 3 155,185 156,882 2.57% 600 157,043 158,774 7.96% 600
1000 4 – 153,262 155,641 6.02% 600
1000 5 – 152,521 155,043 5.47% 600
1000 10 – 150,599 153,121 4.12% 600
1000 15 – 148,503 150,688 1.46% 600
1000 20 – 149,272 151,457 2.60% 600
1000 25 – 148,728 150,918 1.26% 600
1000 30 – 148,162 150,347 1.54% 600
(–) No feasible solution was found.
Chapter 5
A Collaborative Vehicle Routing

Problem with Backhauls
A bilevel approach for the collaborative transporta-

tion planning problem
Maria João Santos ∗† · Eduardo Curcio† · Pedro Amorim∗† · Margarida Carvalho‡
· Alexandra Marques†
Submitted to International Journal of Production Economics, 2020 (under review)
Abstract The integration of the outbound and the inbound logistics of a company leads
to a large transportation network, allowing to detect backhauling opportunities to increase
the efficiency of the transportation. In collaborative networks, backhauling is used to find
profitable services in the return trip to the depot and to reduce empty running of vehicles.
This work investigates the vertical collaboration between a shipper and a carrier for the
planning of integrated inbound and outbound transportation. Based on the hierarchical na-
ture of the relation between the shipper and the carrier and their different goals, the problem
is formulated as a bilevel Vehicle Routing Problem with Selective Backhauls (VRPSB). At
the upper level, the shipper decides the minimum cost delivery routes and the set of in-
centives offered to the carrier to perform integrated routes. At the lower level, the carrier
decides which incentives are accepted and on which routes the backhaul customers are vis-
ited. We devise a mathematical programming formulation for the bilevel VRPSB, where
a routing and a pricing problems are optimized simultaneously, and propose an equivalent
reformulation to reduce the problem to a single-level VRPSB. The impact of collaboration
is evaluated against non-collaborative approaches and two different side payment schemes.
The results suggest that our bilevel approach leads to solutions with higher synergy val-
ues than the approaches with side payments. Its main limitation regards the complexity of
solving exactly the mathematical formulation for large instances.
Keywords bilevel optimization · vertical collaboration · vehicle routing problem with se-
lective backhauls · exact reformulation

‡ CIRRELT and Department of Computer Science and Operations Research, University of Montreal,
Canada
150 Chapter 5. A Collaborative Vehicle Routing Problem with Backhauls
5.1. Introduction
Among the different logistics operations, transportation comprises the major portion of the
total costs and it is strongly associated with a negative impact on the environment (Wang
et al., 2019). Thus, promoting sustainable initiatives for transportation is becoming a target
for many companies and supply chains. Reducing empty running is one of the most pop-
ular initiatives to increase the efficiency of vehicles, which impacts directly on reducing
costs, fuel consumption, and pollutant emissions (Evangelista et al., 2017). Traditionally,
the vehicles travel empty when returning to their original location, and this empty distance
may represent up to 25% of the total route distance (Juan et al., 2014; Turkensteen and
Hasle, 2017). An efficient way to reduce empty running is to provide pickup loads for
vehicles that would return empty to their depot. This is known as backhauling. For a com-
pany, backhauling allows to reduce the total costs of transportation by creating integrated
outbound-inbound routes instead of dedicated delivery and dedicated pickup routes. For
example, Sainsbury’s uses backhauling to create integrated routes such that, after deliv-
ery all requests at stores, the vehicles collect stock at warehouses in the return trip to the
distribution centre (Early, 2011). Differently, Tesco uses backhauling under the context of
reverse logistics. After delivering the requests to a store, a vehicle can collect, at that store,
returned products to be delivered at the distribution centre.
Backhauling is also widely applied in the context of collaborative transportation. For in-
stance, Nestlé and United Biscuits, competitors in the food market, have arranged a col-
laboration to improve their logistics operations. The companies, which share a common
depot, make their own delivery transport but the vehicles collect loads from each others
customers in the return trip. This backhauling strategy allowed to reduce empty running
from 22% to 13% in four years (Early, 2011). In the study of Juan et al. (2014) differ-
ent carriers collaborate with each other through backhauling by allowing each carrier to
service customers from other carrier’s depot. For the overall network, the collaboration
have provided average reductions of 16% on the total distance costs and of 24% on the
environmental costs. Both of the above examples refer to horizontal collaboration, where
the participants in the collaborative network are stakeholders at the same level in a supply
chain. Vertical collaboration, on the other hand, refers to the case where participants are
stakeholders at different levels in a supply chain, and usually involves an hierarchical re-
lation between them. An example of vertical collaboration between a set of retailers and
a service logistics provider (LSP) is investigated in Cruijssen et al. (2005). The retailers
need to serve all their customers, either using outsourcing or collaborating with the LSP.
The LSP assumes the leading of the collaboration, and offers to retailers reduced tariffs to
serve their stores. Each tariff represents the cost reduction that the LSP is able to offer to a
retailer, and it depends on the degree of synergy achieved with collaboration.
In the present work, we investigate a case of vertical collaboration between a shipper and
a carrier, where the shipper is the leading entity of the collaboration. The shipper aims
to promote the creation of integrated routes such that transportation costs are minimized,
whereas the carrier aims to maximize the revenues collected during its backhaul trips. The
shipper must then offer incentives to the carrier to motivate it to perform integrated routes.
Based on the conflicting and hierarchical nature of the objectives of the shipper and carrier,
5.2. Background literature 151
we propose a bilevel formulation for the collaborative transportation planning problem. A

bilevel optimization model is composed of two levels: the upper level describes the prob-
lem of a leader and the lower level describes the problem of a follower. The main character-
istic of the bilevel problem is that the lower level is part of the constraints of the upper level
problem. Thus, this represents a sequential game where first, in the upper level, the leader
(shipper) takes a decision, and afterwards, in the lower level, the follower (carrier) observes
the strategy of the leader and solves its optimization problem. The collaborative problem is
formulated as a Bilevel Vehicle Routing Problem with Selective Backhauls (VRPSB) and
solved by reducing it to an equivalent single-level mixed integer program. The properties
of the bilevel problem are analyzed, and the efficiency of the formulation is evaluated and
compared against traditional modelling approaches.
The main contributions of this work are relevant for both literature and practice. First, the
bilevel formulation proposed allows to explicitly model the interactions and the goals of
both participants in the network, as well as it ensures the individual rationality. Second,
the collaborative problem is formulated such that it can solve simultaneously the routing
and pricing problems, where routing decisions are taken jointly by both participants, and
pricing decisions are taken by the shipper when offering incentives to the carrier. Third, a
thorough analysis on the properties of the bilevel approach, and a comparison with other
alternative approaches (e.g., side payments), allow to gather several managerial insights on
the potential application of the bilevel approach to form prominent collaborative networks.
Fourth, the problem studied fits well real cases where several backhauling opportunities
may arise. Particularly in the forestry industry, the idea of motivating carriers to use their
empty vehicles to perform backhauling for shippers is widely applied (Marques et al., 2020;
Audy et al., 2012).
The remainder of the paper is structured as follows. A literature review on the collaborative
vehicle routing is presented in Section 5.2. Section 5.3 defines the bilevel VRPSB, describ-
ing the mathematical formulation and assumptions of the problem, as well as the properties
of the model. An exact single-level reformulation developed for solving the bilevel model
is proposed in Section 5.4. The computational experiments are presented and discussed
in Section 5.5, covering the generation of the data sets used in this work, the managerial
insights obtained with the different transportation planning strategies, and the evaluation of
the computational performance of models and solution method. Section 5.6 concludes this
paper, presenting the main insights and limitations of this work, as well as suggestions for
future work in the research field.
5.2. Background literature

The main motivation for a player to join a collaborative network is to reduce its total costs
(or increase its profits). Therefore, each player expects that its costs (profits) are lower
(higher) in collaboration than in the case where they perform individually (stand alone so-
lution). This is designated as the individual rationality. The difference between a solution
for the entire collaborative network and the stand alone solution is known as the coalition
gain (Cuervo et al., 2016), whereas the augmented percentage profit defines the synergy
value (Cruijssen et al., 2007).

Collaboration can be one of the three types: i) horizontal collaboration, if players are at
the same level of the network, ii) vertical, if players are at different levels of the network,
and iii) lateral, if there is a combination of both. A recent review on collaborative ve-
hicle routing concludes that horizontal collaboration is the most investigated type in the
literature (Gansterer and Hartl, 2018). Thus, collaboration can be achieved differently de-
pending on the business context. This work focus on type ii) hence, next, we review the
literature on different business models considering vertical collaboration, the use of coop-
erative game theory to incentive collaboration and the application of bilevel programming
to model leadership in collaborative situations.
A case of vertical collaboration is studied in Ergun et al. (2007), which considers a shipper
interested in identifying repeatable and continuous tours for carriers, in order to minimize
their re-positioning needs and, consequently, the routing costs. By combining inbound and
outbound routes, the shipper can negotiate discounts with the carrier and thus pay less for
the overall transportation. The problem is modeled as a time-constrained lane covering
problem and the results show savings between 5.5% and 13%. Bailey et al. (2011) study
the problem of a carrier seeking for collaborative shipments with potential partners in a
transportation network, receiving a revenue for each shipment. The collaboration can occur
with a shipper, which offers a pickup-delivery task close to the backhaul routes of the carrier
of interest, or with other carriers, who do not have sufficient capacity to fulfill all their tasks.
In a case-study, the authors demonstrate that the carrier can reach savings between 13%
and 28% compared with the stand alone solutions. A problem investigated in Xu et al.
(2017), that involves a manufacturer with a private fleet and outsourcing options, shows
that collaboration with a third party logistics (3PL) may allow to reduce the total costs in
10%. Cruijssen et al. (2005) propose an alternative to outsourcing, which they designate as
insinking. In opposition to outsourcing, which is decided by shippers, the insinking allows
a logistics service provider (LSP) to motivate shippers to be their customers. In this case,
the LSP selects the shippers it wishes to serve in order to build strong synergies. Based on
the synergy value, the carrier then determines customized tariffs to motivate the shippers to
collaborate.
Collaboration may be tackled with cooperative game theoretical tools. In this context, typ-
ically, the participants problems are aggregated into one large optimization problem, and
afterwards, the benefits (savings or profits) are determined and shared among the partici-
pants. This usually requires solving a pricing problem. Moreover, the collaboration should
attend specific criteria, such as individual rationality (i.e., each participant cannot perform
better individually than in collaboration). Several methods to allocate the profits from the
collaboration have been investigated in the literature. Among them, the Shapley value
(Shapley, 1953) is the most commonly used. This method distributes the profits among the
players, taking into account the contribution of each player to the overall coalition gain.
For example, Krajewska et al. (2008) applies the Shapley value to fairly allocate the prof-
its in a collaborative Pickup and Delivery Problem with Time Windows (PDPTW), while
Pradenas et al. (2013) use it in a collaborative Vehicle Routing Problem with Backhauls
and Time Windows (VRPBTW). Cruijssen et al. (2005) have also used the Shapley value
to determine customized tariffs to offer the participants in the collaborative Vehicle Rout-
5.2. Background literature 153
ing Problem with Time Windows (VRPTW). Another approach to solve the profit sharing
problem is to optimize the routes of a leading participant selfishly and then provide side
payments for each of the remaining participants (Özener et al., 2011). These side pay-
ments represent a part of the profits generated for a leading entity that should be sufficient
to compensate the losses of one or multiple participants. Dahl and Derigs (2011) inves-
tigated different compensation schemes to solve a collaborative problem between carriers
that are allowed to share orders. These compensation schemes distinguish between orders
executed by private fleet or by a partner fleet, and orders served by dedicated vehicles or
inserted in already existing routes. In Liu et al. (2010), the compensation scheme cover the
side payments received by a carrier if it executes orders from their partners and the penalty
costs if the carrier needs to outsource. Nevertheless, using side payments is not always a
guarantee that an efficient collaboration is created (Özener et al., 2011).
Recently, Defryn et al. (2019) have put in evidence that the traditional modelling of collab-
orative vehicle routing problems presents some fragility when it comes to consider different
goals of participants in the collaboration. Due to the capabilities of bilevel optimization to
explicitly consider the different optimization problems of players, it is expected that bilevel
models are able to overcome this difficulty. To the best of our knowledge, only Xu et al.
(2018) developed a bilevel formulation for a collaborative VRP. The bilevel formulation
considers a centralized logistics platform that allocates vehicles to customers orders (upper
level) and a set of vehicle owners that execute those orders (lower level). The upper level
problem aims to minimize the variable and the fixed routing costs, whereas the lower level
aims to minimize the total empty distances. This work highlights the benefits of balancing
the different objectives of different players in a transportation network, but the coalition
gain or profit sharing are not discussed.
Finally, it is worth mentioning that although bilevel optimization models are still emergent
in the field of collaborative vehicle routing, they are of very much use in pricing problems
(e.g., define the price of vaccines to sell in the market (Lunday and Robbins, 2019), prices
of shared transportation (Qiu and Huang, 2016), storage price for outbound containers in
dry ports (Qiu et al., 2015), setting revenue shares of retailers and prices of suppliers in
marketplaces (de Matta et al., 2017). Thus, considering that we propose to solve simulta-
neously a routing and a pricing problem, a bilevel optimization model seems well suited
for the purpose of this work.
Our work differs in several aspects from the above literature. First, we study a problem of
vertical collaboration, where the upper level problem belongs to the shipper and the lower
level problem belongs to the carrier, and where the goals of each player are different. Most
of the literature on vertical collaborative transportation considers a single goal of minimiz-
ing the costs or maximizing the savings for the entire network. Only Defryn et al. (2019)
have recently put in evidence that different players in collaboration can pursue different
goals, but the problem they study involves the collaboration between carriers, and thus it is
a case of horizontal collaboration. Second, we assume that the shipper is the leading entity
of the collaboration and the incentives offered are based on the response function of the car-
rier. In opposition, Cruijssen et al. (2005) considers that the carrier is the leading entity and
the tariffs offered to shippers are based on the Shapley value. Third, we aim to demonstrate
the advantage of using a bilevel formulation to handle the collaborative problem instead of
a traditional planning with side payments. We further develop an exact reformulation to

solve the problem up to optimality. The work of Xu et al. (2018) provides only one single
example of the capabilities of the bilevel formulation, and it describes a genetic algorithm
to solve the hierarchical problem. Finally, our study contributes to the scientific literature
with a thorough analysis on the benefits of the bilevel approach, as well as its limitations,
against traditional modelling strategies.
5.3. Problem description

This section describes in detail the problem investigated in this work. First, the collabo-
rative problem investigated in this work is presented, describing the perspectives of both
players - the shipper and the carrier, and how their problems relate to each other. Next,
the bilevel formulation for the collaborative VRPSB is presented, where the upper level is
the cost minimization problem of the shipper and the lower level is the profit maximiza-
tion problem of the carrier. The relevant properties of the bilevel VRPSB, which allow to
propose a single level formulation, are defined afterwards. The section concludes with an
illustrative example of a collaborative network formed by a shipper and a carrier.
5.3.1 The collaborative transportation planning problem

The transportation network is composed of a common depot, a set of linehaul customers
(customers of the shipper) and a set of backhaul customers (suppliers of the shipper). The
fleet of vehicles of the carrier are located at the common depot.
The shipper does not own a fleet of vehicles but, on the basis of a contract with a carrier,
sends regular shipments to meet the demand of all its customers (outbound routes). The
shipper plans the delivery routes such that (almost) all capacity of the carrier’s vehicles
are used when departing from the depot. The shipper has also requests to be picked up at
different suppliers, for which typically it is the supplier who sends a full truck load vehicle
to the depot of the shipper (inbound routes).
The shipper recognizes that integrating some inbound trips in the outbound routes of the
carrier, may lead to reduce its total routing costs. For the carrier, this strategy may also
bring benefits, because guaranteeing a full truck load in the return trip to the depot reduces
empty backhaul distances. Thus, to motivate the carrier to collaborate and perform an inte-
grated outbound-inbound route, the shipper must pay an additional incentive. However, the
carrier, which may serve other requests to other shippers, may not be willing to collaborate.
For example, the carrier can get a better incentive from another service or the distance to
perform any integrated outbound-inbound route exceeds the maximum distance allowed.
A main distinct feature of this problem is the way we consider the competition between
incentives. The incentives for backhauling offered by the shipper compete with the costs
of pure inbound routes (for the shipper) and the external incentives (for the carrier). The
competition with the former follows the rational principle that the cost of integrating a
backhaul customer in a delivery route must be lower than the cost of a pure inbound route to
visit this customer. The competition with the latter comes from the fact that, after deliveries,
a vehicle of the carrier has a remaining distance that can be used to provide external services
5.3. Problem description 155
while returning to the depot. We assume that the total remaining distance is used to perform
a single backhauling service, either for the shipper or for an external entity.
Another feature of our problem is that outbound and inbound routes are treated differently.
The first reason comes from the fact that outbound routes are the priority in the routing
problem, which is based on the typical assumptions of the VRPB, and its variants. The
second reason is that each delivery customer can only be visited once, whereas the backhaul
customers can be visited multiple times. Therefore, inserting the same backhaul customer
in different routes would lead to different routing costs. By using an incentive for each
backhaul customer, independently on the route it is inserted, allows to standardize the
cost of backhauling. Third, the costs of delivery routes are predefined in a contract with
the carrier, i.e. it will execute them no matter what. After deliveries, the carrier seeks
for backhauling services but there is no guarantee that it will find one. Thus, only after
deliveries, the competition between backhauling services arises, and the carrier will select
the most profitable one (hence, the need to motivate the carrier to perform a service not
covered by the contract for the deliveries).
In summary, the collaborative problem not only should solve a routing problem but also a
pricing problem, i.e. the shipper must define which incentives for the backhauling should
offer to the carrier, considering other competing incentives. The problem should also con-
sider the hierarchical nature of the shipper and the carrier, and the different goals that each
one pursuits. The goal of the shipper is to define the minimum cost routes and the mini-
mum incentives necessary to motivate the carrier to collaborate. The goal of the carrier is
to select the most profitable services to perform with its fleet of vehicles, until returning to
the depot.
5.3.2 Mathematical programming formulation

The following sets are used in the formulation. Set V = {0 ∪ L ∪ B} represents all nodes
in the network, where {0} is the depot, L = {1, ..., n} is the subset of linehaul customers
and B = {n + 1, ..., n + m} is the subset of backhaul customers. Set K = {1, ..., k} denotes the
delivery vehicles of the carrier. Each arc (i, j) in the network has an Euclidean distance
di j and an associated symmetric cost ci j , such that ci j = c ji and i , j. The unitary cost of
distance travelled for the shipper is cU L
i j and for the carrier is ci j . The cost of a dedicated
inbound vehicle is 2cU i0 , which pays the load and no-load distances between a backhaul
customer i and the depot {0}. Each linehaul customer i requires a given quantity qi to be
delivered and the depot requires a minimum amount of raw-materials Q0 to be collected
at backhaul customers. All vehicles have similar capacity C. The total distance travelled
by one delivery vehicle cannot exceed the maximum distance allowed Dmax . The expected
unitary profit per unit of distance of an external service outside the collaboration performed
by the carrier is given by φ.
The routing problem is modelled using a single commodity flow formulation, since only
one type of product can be carried on each arc, for delivery or for pickup load. The rout-
ing problem is modelled as a VRPSB with the exception that pure inbound routes are also
allowed. Allowing the creation of all type of routes (only outbound, only inbound and in-
tegrated outbound-inbound) brings more benefits to the optimization than forcing backhaul
customers to be visited in integrated routes (Marques et al., 2020).

The profit sharing problem is combined with the routing planning through the incentives for
backhauling offered by the shipper to the carrier, which leads to the Collaborative VRPSB.
Finally, the collaborative VRPSB can be formulated as a mixed-integer bilevel VRPSB,
where the upper level describes the problem of the shipper (Problem (5.1)-(5.9)) and the
lower level describes the problem of the carrier (Problem (5.14)-(5.18)).
The upper level decision variables are:

1, if vehicle k travels on arc (i, j)
xikj

:= 
0, otherwise ∀k ∈ K, ∀i, j ∈ V \ B
Zb := incentive offered to visit backhaul customer b, ∀b ∈ B

Ob := number of visits to backhaul customer b by dedicated inbound vehicles, ∀b ∈ B
yi j := load in a vehicle between customers i and j, ∀i, j ∈ V \ B.
The lower level decision variables are:


k 1, if vehicle k travels on arc (i, j)

x̂i j := 
0, otherwise ∀k ∈ K, ∀i ∈ L, ∀ j ∈ B
k
Zext := external incentive offered in route k, ∀k ∈ K.
5.3.2.1 Upper level problem

The objective function of the shipper is the minimization of the total cost of the routing plan
as in Equation (5.1) below. The total cost comprises three aspects: i) the cost associated to
the total distance travelled to visit all linehaul customers, ii) the total incentives payed to the
carrier to visit backhaul customers, and iii) the total cost of outsourcing dedicated inbound
vehicles. Note that the upper and lower problems interact through the two variables present
in the second term of the objective function, namely Zb and x̂ikj . Moreover, this term of
the function is nonlinear, but it can be linearized as we will see later (through constraints
(6.4.2)-(6.4.5)).
X XX XXX X
i j · xi j +
cU +
k k
min Zb · x̂ib Ob · 2cU
b0 (5.1)
i∈V\B j∈L k∈K i∈L b∈B k∈K b∈B
All routes start at the depot, as expressed by Constraints (5.2) and the flow constraints
in delivery routes, given by Constraints (5.3), guarantee the connectivity of the locations
visited in each route.
X
k
x0i ≤ 1, ∀k ∈ K (5.2)
i∈L
X X X
xikj = xkji + x̂kjb , ∀ j ∈ L, ∀k ∈ K (5.3)
i∈V\B i∈V\B b∈B
Equations (5.4) guarantee that all linehaul customers are visited exactly once by only one
vehicle.
XX
xikj = 1, ∀j ∈ L (5.4)
i∈V\B k∈K
The total load carried on each vehicle must decrease gradually as linehaul customers are
visited in each route, which is ensured by Equations (5.5). These constraints are also
subtour elimination constraints. Furthermore, Constraints (5.6) guarantees that the capacity
of each vehicle is never exceeded on each route.
X X
yi j = y ji + q j , ∀j ∈ L (5.5)
i∈V\B i∈V\B
yi j ≤ xikj · C, ∀i, j ∈ V \ B, ∀k ∈ K (5.6)
The demand of linehaul customers is fully satisfied with Equations (5.7) and the minimum
demand of the depot is satisfied with Constraints (5.8). It is assumed that, each time a ve-
hicle visits a backhaul customer (either on integrated or on pure inbound routes), it returns
full to the depot, so that the quantity delivered at the depot matches exactly the total capac-
ity of the vehicle. This rationale is based on the common practice of vehicles travelling in
full truck load, such as in the forestry industry (Marques et al., 2020).
X X
y0 j = qj (5.7)
j∈L j∈L
XXX Q0
k
x̂ib + Ob ≥ (5.8)
i∈L b∈B k∈K
C
The domain of the upper level variables is as follows:
Q0
xikj ∈ {0, 1}, Zb , yi j ≥ 0, Ob ∈ {0, ..., d e}, ∀i, j ∈ V \ B, k ∈ K, b ∈ B. (5.9)
C
To linearize the objective function of the upper level, we use the McCormick constraints
(McCormick, 1976). Thus, we introduce a new variable Akb = Zb · i∈L x̂ib k and derive the
P
following constraints. If i∈L x̂kib = 0, then the inequality (6.4.2) ensures that Akb is also zero
P
(Akb is higher than a negative number from inequality (6.4.4) and cannot be negative due
to the Equation (6.4.5)). On the other hand, if i∈L x̂kib = 1, the inequality (6.4.2) ensures
P
that Akb is lower than M (large number), and inequalities (6.4.3) and (6.4.4) guarantee that
Akb = Zb . Note that we can make M = max{2cU i0 }, ∀i ∈ B, which is the upper bound of each
inventive Zb .
X
Akb ≤ M · k
x̂ib , ∀b ∈ B, ∀k ∈ K (5.10)
i∈L
Akb ≤ Zb , ∀b ∈ B, ∀k ∈ K (5.11)
X
Akb ≥ Zb − (1 − k
x̂ib ) · M, ∀b ∈ B, ∀k ∈ K (5.12)
i∈L
Akb ≥ 0, ∀b ∈ B, ∀k ∈ K (5.13)
5.3.2.2 Lower level problem

The objective function of the carrier is described by Equation (5.14), which is the maxi-
mization of the total profits collected with all routes. The profit collected with integrated
routes is determined as the difference between the total incentives accepted and the total
travelling cost of including the backhaul customers in delivery routes. The profit collected
with the external services is equivalent to the total net external incentives accepted. Fi-
nally, the profit collected with a delivery route corresponds to the difference between the
total cost charged to the shipper and the total effective cost of the deliveries routes paid by
the carrier. Note that this term could be removed since it is constant (it does not include
decision variables of the lower level).
XXX XXX
max Zb · x̂kib − x̂kib · (cib
L
+ cb0
L
)+
i∈L b∈B k∈K i∈L b∈B k∈K
X XX X XX (5.14)
k
Zext · (1 − x̂ib ) +
k
(cU
ij − ciLj ) · xikj
k∈K i∈L b∈B i∈V\B j∈L k∈K
Constraints (5.15) forces the precedence constraint of a typical VRPSB, where backhaul
customers can only be linked to a last linehaul customer in a route. First, delivery routes are
the priority in the VRPSB, since these can only be performed by delivery vehicles. Second,
the load to be collected at any backhaul customer can fill all the capacity of a vehicle. Thus,
no mixing of delivery and pickup loads is possible.
x̂bk j = 0, ∀b ∈ B, j ∈ L, k ∈ K (5.15)
Constraints (5.16) enforce that the maximum distance Dmax is never exceeded in any route.
X XXX
di j · xikj + k
x̂ib · (dib + db0 ) ≤ Dmax , ∀k ∈ K (5.16)
i, j∈V\B i∈L b∈B k∈K
The external incentive, given by Equations (5.17), must be equal to the remaining distance
of a delivery route multiplied by the unitary profit φ. This is an important aspect considered
by the carrier, since the shortest are the delivery routes, the higher the distance remaining
to perform additional services.
XX
k
Zext = (Dmax − di j · xikj ) · φ, ∀k ∈ K (5.17)
i∈V\B j∈L
The domain of the lower level variables is as follows:

k
x̂ib ∈ {0, 1}, ∀i ∈ L, ∀b ∈ B, ∀k ∈ K (5.18)
The nonlinear terms in the objective function of the lower level are linearized through the
definition of a new variable Gk = Zext
k · (1 − P k
P
i∈L b∈B x̂ib ) and the respective McCormick
constraints (6.4.7)-(5.22). Note that M can be set to the upper bound of the maximum
external service, i.e. M = Dmax · φ.
XX
Gk ≤ M · (1 − k
x̂ib ), ∀k ∈ K (5.19)
i∈L b∈B
Gk ≤ Zext
k
, ∀k ∈ K (5.20)
XX
Gk ≥ Zext
k
− k
x̂ib · M, ∀k ∈ K (5.21)
i∈L b∈B
Gk ≥ 0, ∀k ∈ K (5.22)
5.3.3 Properties of the bilevel VRPSB

One particular characteristic of bilevel problems is its intrinsic hierarchical structure. The
upper level is the dominant player and the first to select an action. Afterwards, the lower
level observes the decisions of the upper level and optimizes its own objective function.
Each strategy selected by the lower level is called a rational response and the set of all
responses is known as the rational set (Colson et al., 2007; Safari et al., 2014; Sun et al.,
2008). Knowing the rational set, the upper level can anticipate the response of the lower
level and decide the final strategy that minimizes its costs.
The bilevel VRPSB states that, for each strategy of the shipper, i.e. for a fixed set of upper
decision variables, the carrier accepts or withdraws each incentive offered by the shipper,
in each route. If any incentive offered is not accepted in a route, the carrier performs
an external service. The profit for the carrier of an integrated route Pkb is determined as
the difference between the incentive Zb and the additional travelling cost to visit backhaul
customer b, as follows:
X
Pkb = Zb − k
x̂ib L
· (cib + cb0
L
), ∀k ∈ K, ∀b ∈ B. (5.23)
i∈L
The profit of an external service PkE equals the external incentive on that route, as follows:
PkE = Zext
k
, ∀k ∈ K. (5.24)
Given a fixed input of upper decision variables, the optimal response of the lower level
problem, F k , corresponds to the optimal solution of the carrier, as follows:
F k = max{PkE , max{Pkb }}, ∀k ∈ K. (5.25)

b∈B
When the rational response is not singleton, i.e. more than one response of the lower
level may be obtained for a single strategy of the upper level, two different approaches
can be applied - the optimistic and the pessimistic approaches (Colson et al., 2007; Sinha
et al., 2018). The optimistic approach assumes that the lower level will select the ratio-
nal response that is more favourable to the upper level. In opposition, the pessimistic
approach assumes that the lower level will select the least favourable response. The op-
timistic approach is the most investigated in literature and allows to slightly reduce the
non-cooperative nature of bilevel models (Kozanidis et al., 2013; Sinha et al., 2018). For
this reason, the bilevel VRPSB is build upon an optimistic approach, such that when
maxb∈B Pkb = PkE , the optimal response of the carrier in route k is to accept the incentive
offered by the shipper.
It is also worth mentioning that, although we use an optimistic approach, the problem of the
lower level considers also its most optimistic case. In fact, as the upper level does not know
exactly the problem of the lower level, it is assumed that the lower level can always achieve
the highest possible profits in each route, considering that all the remaining distance can be
used to provide external services. Thus, the incentives offered by the upper level are set to
cover the highest external incentive in each route.
5.3.4 Numerical example

In the following numerical example, we demonstrate the rationale upon the bilevel model
is build to model collaboration. Consider a shipper that needs to send requests to customers
1, 2, 3 and 4, and requires a full truck load from one of the backhaul customers 5 or 6. The
carrier has two vehicles available.
Figure 5.1 illustrates the bilevel solution and two non-collaborative solutions, for the pur-
pose. The non-collaborative solutions correspond to the separated planning (VRP, op-
timization of inbound and outbound routes independently) and the integrated planning
(VRPSB, all routes are optimized under the perspective of the shipper only). The non-
collaborative models are detailed in 5..1.
With a VRP (separated planning) (Figure 6.1c), the solution of the shipper includes the least
cost routes for deliveries and an inbound route with the backhaul nearest to the depot. The
carrier benefits with the external incentives from both delivery routes. This is designated
as the stand alone solution.
With a VRPSB (integrated planning) (Figure 6.1d), where the shipper controls the fleet of
vehicles, it would select backhaul customer 6 instead of backhaul customer 5, due to the
lower backhaul distance, i.e. d46 + d60 < d15 + d50 . However, the carrier would lose profits
in visiting a backhaul customer in a route that generates a higher external incentive.
Under a bilevel approach (Figure 6.1a), the shipper sets only an incentive for backhaul
5.4. Single-level reformulation 161
(a) VRP (b) VRPSB (c) Bilevel VRPSB
Figure 5.1 – Routing plans obtained with each planning approach. The square is the depot,
the white circles are linehaul customers and the grey circles are backhaul costumers. Blue
lines represent the (part of) the routes performed by the carrier to serve the requests of the
shipper. Red lines represent the part of the routes performed by the carrier to serve external
services. Black line represents the pure inbound route outsourced to a supplier.
customer 5. No incentive is provided for backhaul customer 6 because it would be higher

or, at least, equal to the cost of an inbound route. The carrier, in turns, would integrate
the backhaul visit in its longest route, where the backhaul incentive competes with a lower
external incentive, leaving the shortest route to guarantee a higher external incentive.
In this example, it was demonstrated how the bilevel VRPSB incorporates the rational
response of the carrier into the problem of the shipper, by guaranteeing that the profits
achieved with collaboration are not lower than its stand alone solution. In Section 5.5.3,
we compare the bilevel approach with a traditional VRPSB with side payments, where the
shipper plans all the delivery and integrated routes and decides on the side payment of each
integrated route.
5.4. Single-level reformulation

One classic way to handle mathematically a bilevel problem is to reformulate it into a
single-level optimization problem (Sinha et al., 2018) and then solve it with an exact
method to find the global optimum. The reformulation based on Karush-Kuhn-Tucker
(KKT) conditions is the most popular method (Zeng and An, 2014), where the lower level
problem is replaced by its corresponding KKT conditions, which include primal feasibility,
dual feasibility and complementary slackness conditions. A similar reformulation method
is based on duality theory, where the lower level is replaced by its primal and dual con-
straints, enforcing strong duality. The duality-based reformulation leads to a lower num-
ber of variables than the KKT conditions, rendering a better computational performance
(Garcia-Herreros et al., 2016). Nevertheless, both of these methods can only be of use in
practice when the lower level is linear with continuous variables, for which the optimal
solutions correspond to vertices of the feasible region.
Recently, solvers for mixed integer bilevel programming are available: MibS∗ and bilevel† ,
whose technical details are respectively discussed in Tahernejad et al. (2017) and Fischetti
et al. (2017). In both solvers, only the integer upper level variables can appear in the lower
level, since such assumption guarantees the existence of an optimal solution (Vicente et al.,
1996). The bilevel VRPSB presents both binary and continuous decision variables, and
the upper level continuous variable Zb appears in the objective function of the lower level.
Therefore, the general purpose techniques mentioned above are not suitable for the problem
at hand. In order to overcome the lack of solution approaches to our problem, we will
focus in its specific structure. In particular, we show how the rational set of the lower level
problem can be represented by a finite set of inequalities, allowing to reduce the bilevel
VRPSB to a single-level mixed integer linear programming (MILP) problem. In this way,
the obtained single level problem can be solved by off-the-shelf solvers.
With the purpose of transforming the lower level objective function in a set of disjunctive
constraints describing its rational set, we make use of Big-M constraints. This new set
of constraints is then incorporated into the problem of the upper level, which results in a
single level MILP VRPSB model. The detailed description of the reformulation technique
applied in this work is presented in the following steps.
Step 1. The integer variables of the lower level problem, x̂ib k , are replaced by the corre-
sponding flow variables of the upper level, xikj . By extending the entire set of locations to
{i, j} ∈ V = {0} ∪ L ∪ B, all constraints of the lower level can be re-written using only the
decision variables of the upper level, as follows.
xbk j = 0, ∀b ∈ B, j ∈ L, k ∈ K (5.15’)
X
di j · xikj ≤ Dmax , ∀k ∈ K (5.16’)
i, j∈V
XX
k
Zext = (Dmax − di j · xikj ) · φ, ∀k ∈ K. (5.17”)
i∈V j∈L
Step 2. The objective function of the lower level problem is to maximize the total profits of
the carrier. The profits obtained with delivery routes are fixed a priori by the upper level,
because these routes are not influenced by the decisions of the carrier. The decisions of the
carrier only influence which backhaul customers are visited, if any, and on which routes.
Therefore, the objective of the lower level can be translated into the maximization of profits
obtained by accepting the different incentives, namely those offered by the shipper (5.26)
and those offered by its competitors (5.27).
F k ≥ Pkb , ∀k ∈ K, ∀b ∈ B (5.26)
k
F ≥ PkE , ∀k ∈ K. (5.27)
∗ available at https://github.com/coin-or/MibS
† available at https://msinnl.github.io/pages/bilevel.html
5.4. Single-level reformulation 163
Step 3. To enforce that, in each route, the carrier decides in favour for the incentive that
maximizes its profits, a set of disjunctive constraints are established, using M as a large
number (e.g., M = Dmax · φ). Constraints (5.28) introduce a binary variable HEk that takes
k is higher than the profit of visiting any backhaul
the value of 1 if the external incentive Zext
b in route k (i.e. PkE > Pkb ) and 0 otherwise. Similarly, Constraints (5.29) introduce a binary
variable Hbk that takes the value of 1 if Pkb > PkE and Pkb > Pki , ∀i ∈ B, b , i, and 0 otherwise.
F k − PkE ≤ (1 − HEk ) · M, ∀k ∈ K (5.28)

k
F − Pkb ≤ (1 − Hbk ) · M, ∀k ∈ K, ∀b ∈ B. (5.29)
Step 4. If the external incentive is accepted for a given route (HEk = 1), all the arcs be-
tween backhauls and linehaul customers must be zero, which is enforced by Constraints
(5.30). Otherwise, the incentive offered for a backhaul b is preferred (Hbk = 1), for which
Constraints (5.31) ensure that, for route k the arcs containing the remaining backhauls are
all set to zero, since each vehicle visits only one backhaul. Constraints (5.32) guarantee
that if Hbk = 1, then, at least, one arc linking a linehaul to a backhaul customer must ex-
ist. Equation (5.33) ensures that only a single incentive may be accepted by the carrier in
each route, while Constraints (5.34) are used to guarantee that either the incentive to visit
a backhaul b is accepted or the external incentive is accepted.
(1 − HEk ) ≥ xib
k
, ∀k ∈ K, ∀i ∈ L, ∀b ∈ B (5.30)
(1 − Hbk ) ≥ xikj , ∀k ∈ K, ∀i ∈ L, ∀ j , b ∈ B (5.31)
X
k
xib ≥ Hbk , ∀k ∈ K, ∀b ∈ B (5.32)
X i∈L
Hbk + HEk = 1, ∀k ∈ K (5.33)
b∈B
X
Hbk ≤ 1, ∀k ∈ K. (5.34)
b∈B
Concluding the reformulation procedure, the resultant model is a single level VRPSB with
the following (MILP) formulation:
XXX XX X
i j · xi j +
cU Akb +
k
min Ob · 2cU
b0 (5.35)
i∈V j∈L k∈K b∈B k∈K b∈B
subjected to
(5.2) − (5.9)
(5.150 ), (5.16) and (5.1700 )
(5.23) − (5.24)
(5.26) − (5.34)
Theorem 1. Any optimal solution of Problem (5.35) is also optimal to the bilevel VRPSB.
Proof. Next, we will show the correctness of each step in the single level reformulation.
Start by noticing that if a leader’s optimal solution has xib k , x̂k , for some i, j and k, then
ib
k k
changing xib to x̂ib still results in an optimal solution for the leader: the modified leader’s
strategy is feasible, the follower’s feasible region does not change, and none of the objec-
tive functions, (5.1) and (5.14), changes. Consequently, we can restrict xikj to mimic the
follower’s reaction as done in Step 1.
Note that the follower’s problem can be decompose in |K| maximization problems, one for
each vehicle, since there is no lower level linking constraint with the different vehicles.
Hence, we can focus on each of these optimization problems, namely, on the profit F k
that can be obtained by each vehicle k ∈ K. Recalling that F k can be modeled accordingly
with Equation (5.25), Step 2 and Step 3 linearize it through a set of 4 constraints (Con-
straints (5.26) to (5.29)), and new binary variables, HEk and Hbk , are added to model the
type of incentive accepted (external or backhaul). With these newly introduced variables,
in Step 4, we can ensure that the xib k reflect the follower’s reaction.
In this way, we can conclude that any optimal solution of Problem (5.35) is also optimal to
the bilevel VRPSB.

The computational experiments performed in this section cover three main analysis. The
first set of experiments aims to evaluate the solutions obtained with the bilevel approach
for the collaborative transportation problem. Bilevel solutions are compared with stand
alone solutions and the synergy values of collaboration are determined. The second set of
experiments aims to analyse potential side payments schemes, and to compare them with
our bilevel approach. We argue that our bilevel approach, as it incorporates the rational
response of the carrier into the problem of the shipper, provides balanced solutions as
well as higher synergy than side payments schemes. The third set of experiments analyses
the scalability of our approach to solve the bilevel model (single level reformulation) in
comparison with solving non-collaborative models.
Each set of experiments reports the results for two different scenarios: one that allows
multiple visits to the same backhaul customer, and one that forbids more than one visit
to the same backhaul customer. The motivation behind is to analyze the efficiency of the
bilevel approach to deal with different contexts of backhauling. Most of the literature
considers that backhaul customers can only be visited once, but in practice, they may have
enough availability of raw-material, which allows multiple pickups (Marques et al., 2020;
Santos et al., 2020). The bilevel model presented in Section 5.3.2 allows multiple backhaul
visits. To forbid multiple visits to the same backhaul customer, it is sufficient to include
constraint (5.36) and change the dimension of the variable to Ob ∈ {0, 1}. Hence, Constraint
5.36 ensures that each backhaul customer can only be visited once by all type of vehicles.
XX
k
x̂ib + Ob ≤ 1, ∀b ∈ B (5.36)
i∈L k∈K
All data sets used in the computational experiments are collected from Augerat et al. (1995)
and adapted to the VRPSB, as described next. All models are coded in Python 3.6.3 and
solved with Gurobi, on a computer equipped with the processor Intel Core i7 of 2.20GHz
and 16 GB of RAM.
5.5.1 Data sets

The original data sets define the number of locations (depot and linehaul customers) and
respective coordinates, the demand of linehaul customers, the number of vehicles and re-
spective capacity.
The adapted data sets comprise the following modifications: i) the backhaul customers
locations correspond to the 5 to 17 last locations in the original data set, and ii) the number
of vehicles available is determined as the minimum number required to meet all the demand
P qi
of linehaul customers without exceeding vehicle capacity, i.e. |K| = i∈L C in each data
set. The adapted instances used in this work are reported in Table 5.1.
Table 5.1 – Adapted instances used in this work
Original Adapted |L| |B| |K| Dmax

P-n16-k8 A 10 5 6 150
P-n19-k2 B 13 5 2 150
P-n20-k2 C 14 5 2 150
P-n22-k2 D 16 5 2 150
P-n22-k8 E 16 5 6 150
B-n31-k5 F 25 5 4 250
B-n31-k5 G 20 10 3 250
B-n38-k6 H 25 12 4 250
B-n41-k6 I 25 15 4 250
B-n45-k6 J 27 17 4 250
B-n45-k6 K 39 5 6 250
B-n45-k6 L 35 9 5 250
B-n50-k6 M 44 5 6 250
B-n50-k6 N 40 9 6 250
The remaining parameters of the bilevel VRPSB are set as follows. The depot demand
is a multiple of the capacity of a vehicle, such that Q0 = [C, 2C, . . . , |K|C]. The maximum
distance allowed is 150 for instances with less than 20 nodes, and 250 for the remaining
ones. These limits are sufficient to create integrated inbound-outbound routes. The unitary
cost per unit of distance is set to 1.2 e for the shipper and 1.0 e for the carrier, following the
study in Yu and Dong (2013). The unitary profit obtained with an external service equals
the unitary profit obtained with a delivery route, such that φ = 0.2 e\unit of distance.
Throughout the experiments, 10 small instances with varying Q0 are used, since optimal
solutions are obtained in reasonable time. The motivation to use different Q0 values is
to determine if increasing the number of required visits to backhaul customers increases
the performance of the collaboration. The last set of experiments uses also medium size
instances. We anticipate that large instances are not tested, as the exact method developed
in this work is not efficient to solve them. However, the motivation of our work is not on
the efficiency of the solution method but rather on the efficiency of the bilevel formulation
to handle the collaborative problem.
5.5.2 Bilevel versus traditional planning

With the bilevel VRPSB model proposed in this work, we aim to demonstrate that it is
efficient to handle the collaborative problem, while incorporating the different goals of the
players, and solving the routing and the pricing problems simultaneously. The impact of
collaboration is determined by two measures. The first is the network costs (NC), which
is given by the difference between the total costs of the shipper and the total profits of the
carrier (Equation 5.37). The second is the synergy value (SV), which provides the percent-
age gain that a collaborative network can reach compared with the stand alone solution
(Equation 5.38).
NC = Costs − Pro f its (5.37)

NCVRP − NCcollab
SV = (5.38)
NCVRP
Tables 5.2 and 5.3 present a comparison between stand alone solutions obtained with the
VRP model and collaborative solutions obtained with the bilevel VRPSB model, for the
scenarios investigated. Both read as follows: total costs of the shipper (column "Costs"),
total profits of the carrier (column "Profits"), costs of outsourcing inbound vehicles to
suppliers (column "Out."), network costs (column "NC"), total incentives offered by the
shipper and accepted by the carrier (column "Incent."), and the synergy value of the col-
laborative solution (column "SV").
Table 5.2 shows that seven out of ten instances do not promote collaboration, i.e. the
optimal solution for the shipper is to outsource all necessary vehicles to the suppliers.
Despite disappointing, this outcome is reasonable since the same backhaul customer can
be visited as many times as the depot demand requires. Thus, when the cost of visiting the
backhaul customer closest to the depot with an inbound vehicle is relatively low, and no
backhauling incentive is lower, the shipper always tend to outsource all necessary vehicles
to that closest backhaul customer. Nonetheless, when collaboration occurs, synergy values
can reach about 11%.
From Table 5.3, it is possible to observe very different results compared with the previous
scenario of unlimited visits to backhaul customers. When the number of visits is limited to
one, the incentives proposed by the shipper compete with more diverse options than only
the nearest backhaul customer. This seems to increase heavily the potential for collabora-
tion, as shown in eight out of ten instances. Moreover, any solution of the scenario with
limited visits provides equal or higher synergy values than in the scenario with unlimited
backhaul visits.
We were expecting to see an increase in the synergy value with increasing Q0 , but such was
not verified for all instances. Therefore, we cannot conclude on the impact of increasing
Table 5.2 – Stand alone and collaborative solutions in a scenario with unlimited visits to
backhaul customers.
VRP (stand alone) solutions Bilevel VRPSB (collaborative) solutions

Inst. Q0 Costs Profits Outs. NC Costs Profits Incent. Outs. NC SV
A C 275 180 68 95 265 180 58 0 85 10.0%
2C 342 180 135 162 324 180 117 0 144 11.3%
3C 410 180 203 230 387 180 180 0 207 10.0%
4C 478 180 270 298 454 185 247 0 270 9.4%
5C 545 180 338 365 521 187 314 0 334 8.6%
B C 192 60 38 132 192 60 0 38 132 0.0%
2C 230 60 76 170 230 60 0 76 170 0.0%
C C 204 60 38 144 204 60 0 38 144 0.0%
2C 242 60 76 182 242 60 0 76 182 0.0%
D C 207 60 38 147 207 60 0 38 147 0.0%
2C 245 60 76 185 245 60 0 76 185 0.0%
E C 323 180 53 143 323 180 0 53 143 0.0%
2C 376 180 106 196 376 180 0 106 196 0.0%
3C 429 180 159 249 429 180 0 159 249 0.0%
4C 482 180 212 302 482 180 0 212 302 0.0%
5C 535 180 265 355 535 180 0 265 355 0.0%
F C 499 200 113 299 478 200 88 0 278 7.0%
2C 611 200 226 411 581 200 177 0 381 7.4%
3C 724 200 338 524 685 200 281 0 485 7.4%
4C 837 200 451 637 798 200 281 113 598 6.1%
G C 429 150 91 279 424 150 86 0 274 1.7%
2C 520 150 182 370 515 150 86 91 365 1.2%
3C 611 150 273 461 606 150 86 182 456 1.0%
H C 444 200 51 244 444 200 0 51 244 0.0%
2C 495 200 101 295 495 200 0 101 295 0.0%
3C 546 200 152 346 546 200 0 152 346 0.0%
4C 596 200 203 396 596 200 0 203 396 0.0%
I C 469 200 35 269 469 200 0 35 269 0.0%
2C 504 200 70 304 504 200 0 70 304 0.0%
3C 538 200 105 338 538 200 0 105 338 0.0%
4C 573 200 140 373 573 200 0 140 373 0.0%
J C 379 200 50 179 379 200 0 50 179 0.0%
2C 429 200 100 229 429 200 0 100 229 0.0%
3C 479 200 150 279 479 200 0 150 279 0.0%
4C 529 200 200 329 529 200 0 200 329 0.0%
Average 2.3%
Table 5.3 – Stand alone and collaborative solutions in a scenario with limited visits to
backhaul customers.
VRP (stand alone) solutions Bilevel VRPSB (collaborative) solutions

Inst. Q0 Costs Profits Outs. NC Costs Profits Incent. Outs. NC SV
A C 275 180 68 95 265 180 58 0 85 10.0%
2C 345 180 137 165 324 180 117 0 144 12.5%
3C 416 180 209 236 387 180 180 0 207 12.4%
4C 488 180 281 308 457 180 245 0 277 10.1%
5C 562 180 355 382 528 180 248 68 348 8.9%
B C 192 60 38 132 192 60 0 38 132 0.0%
2C 264 60 110 204 246 60 53 38 186 8.5%
C C 204 60 38 144 204 60 0 38 144 0.0%
2C 276 60 110 216 255 60 50 38 195 9.6%
D C 207 60 38 147 207 60 0 38 147 0.0%
2C 270 60 101 210 253 60 45 38 193 8.3%
E C 323 180 53 143 323 180 0 53 143 0.0%
2C 393 180 123 213 393 180 69 53 213 0.2%
3C 466 180 196 286 466 180 69 126 286 0.1%
4C 542 180 272 362 542 180 69 202 362 0.1%
5C 623 180 352 443 622 180 69 283 442 0.1%
F C 499 200 113 299 478 200 88 0 278 7.0%
2C 616 200 230 416 581 200 177 0 381 8.4%
3C 733 200 348 533 685 200 281 0 485 9.1%
4C 854 200 468 654 802 200 281 117 602 7.8%
G C 429 150 91 279 424 150 86 0 274 1.7%
2C 534 150 197 384 515 150 86 91 365 5.0%
3C 647 150 310 497 609 150 179 91 459 7.6%
H C 444 200 51 244 444 200 0 51 244 0.0%
2C 502 200 108 302 502 200 0 108 302 0.0%
3C 584 200 191 384 583 200 77 108 383 0.3%
4C 672 200 278 472 671 200 83 191 471 0.2%
I C 469 200 35 269 469 200 0 35 269 0.0%
2C 508 200 75 308 508 200 0 75 308 0.0%
3C 554 200 121 354 554 200 0 121 354 0.0%
4C 601 200 168 401 601 200 0 168 401 0.0%
J C 379 200 50 179 379 200 0 50 179 0.0%
2C 439 200 109 239 439 200 0 109 239 0.0%
3C 499 200 169 299 499 200 0 169 299 0.0%
4C 561 200 232 361 561 200 0 232 361 0.0%
Average 3.7%
the demand of the depot. Overall, the results show that any solution of the bilevel model
leads to costs for the shipper that are always equal or lower than the stand alone solution.
Similarly, any solution of the bilevel model leads to profits for the carrier that are always
equal or higher than the stand alone solution. Based on these results, we argue that our
bilevel approach is efficient to deal with the collaborative problem where different goals are
considered and decisions are taken hierarchically. We must also emphasise that, although
the solution of the bilevel VRPSB seems to only benefit the shipper, in fact the carrier also
gains using a bilevel approach, since the upper level considers the most optimistic case of
the lower level problem, as implicit by the bilevel formulation (demonstrated in Section
5.3.3).
5.5.3 Bilevel versus compensation schemes

One alternative to model a collaborative transportation problem involves a leading partic-
ipant optimizing selfishly the routing problem, and then compensating another participant
with a side payment, so that the latter does not lose with the collaboration (Özener et al.,
2011). Some works in the literature determine the side payments as a fixed value (e.g.,
Caballini et al. (2016), Defryn et al. (2016)), whereas other compute the side payment as a
value dependent on the distance (e.g., Liu et al. (2010), Dahl and Derigs (2011), Archetti
et al. (2016)).
In this work, we use the integrated problem (VRPSB) to model selfishly the routing prob-
lem, and we propose two different compensation schemes. In the first scheme, the side
payment corresponds to the difference in the profits of the carrier between the stand alone
solution (VRP) and the integrated solution (VRPSB). The first side payment is designated
as S P∆ and their values are presented in Table 5.4, for each VRPSB solution (before side
payment). The second scheme computes the side payment as a value proportional to the
backhaul distance. This is designated as S P(sp), where sp stands for the proportion used,
as it is computed as Equation (5.39). The first compensation scheme provides the side pay-
ment after the routing, whereas the second determines the side payment while solving the
routing problem.
In this section, we compare the performance of the collaboration of our bilevel approach
with compensation schemes. Tables 5.5 and 5.6 report the objective functions of the ship-
per and the carrier, for the different side payments, along with the synergy value of the
solutions. To facilitate the comparison, the synergy values of the bilevel solutions are also
reported.
S P(sp) = sp · (dlb + db0 ), ∀ sp = {0.50, 0.75}, l ∈ L, b ∈ B (5.39)
In the bilevel model we assume an optimistic approach such that when the profit of an
external incentive equals the profit of a given backhaul visit, the carrier performs an inte-
grated route. The same assumption applies to the case of using side payments. Therefore,
as expected, a side payment S P∆ would always motivate the carrier to collaborate because
the profits gained are the same as in the stand alone solution. However, for most of the
Table 5.4 – VRPSB solutions and respective side payments determined after routing, for
the two different scenarios
Unlimited visits Limited visits

Inst. Q0 Costs Profits S P∆ Costs Profits S P∆
A C 250 164 16 250 164 16
2C 293 149 31 293 149 31
3C 339 131 49 341 134 46
4C 386 115 65 395 117 63
5C 433 100 80 458 104 76
B C 184 46 14 184 46 14
2C 222 46 14 232 39 21
C C 200 47 13 200 47 13
2C 233 39 21 248 41 19
D C 204 49 11 204 49 11
2C 239 40 20 246 41 19
E C 314 160 20 314 160 20
2C 367 143 37 367 143 37
3C 420 143 37 438 146 34
4C 473 143 37 514 146 34
5C 526 143 37 594 146 34
F C 450 172 28 450 172 28
2C 526 145 55 526 145 55
3C 611 126 74 611 126 74
4C 723 121 80 726 121 79
G C 398 124 26 398 124 26
2C 464 97 53 464 97 53
3C 556 97 53 557 97 53
H C 436 167 33 436 167 33
2C 486 167 33 493 167 33
3C 537 167 33 552 168 32
4C 588 167 33 626 149 51
I C 469 200 0 469 200 0
2C 504 200 0 508 200 0
3C 538 200 0 553 170 30
4C 573 200 0 599 170 30
J C 361 163 37 361 163 37
2C 403 130 70 403 130 70
3C 453 130 70 455 131 69
4C 503 130 70 515 131 69
Table 5.5 – Comparison of synergy values from bilevel solutions and from the integrated
planning with different side payment schemes, for the case of unlimited visits to backhaul
customers.
S P∆ S P(0.50) S P(0.75) Bilevel
Inst. Q0 Costs Profits SV Costs Profits SV Costs Profits SV SV
A C 266 180 9.3% 267 181 9.3% 275 180 0.0% 10.0%
2C 324 180 11.3% 329 185 11.3% 342 180 0.0% 11.3%
3C 388 180 9.5% 394 186 9.5% 410 180 0.0% 10.0%
4C 451 180 9.0% 461 190 9.0% 478 180 0.0% 9.4%
5C 514 180 8.6% 528 194 8.6% 545 180 0.0% 8.6%
B C 198 60 -4.7% 192 60 0.0% 192 60 0.0% 0.0%
2C 236 60 -3.7% 230 60 0.0% 230 60 0.0% 0.0%
C C 213 60 -6.1% 204 60 0.0% 204 60 0.0% 0.0%
2C 254 60 -6.6% 242 60 0.0% 242 60 0.0% 0.0%
D C 215 60 -5.5% 207 60 0.0% 207 60 0.0% 0.0%
2C 259 60 -7.4% 245 60 0.0% 245 60 0.0% 0.0%
E C 334 180 -7.6% 323 180 0.0% 323 180 0.0% 0.0%
2C 404 180 -13.9% 376 180 0.0% 376 180 0.0% 0.0%
3C 457 180 -11.0% 429 180 0.0% 429 180 0.0% 0.0%
4C 510 180 -9.0% 482 180 0.0% 482 180 0.0% 0.0%
5C 563 180 -7.7% 535 180 0.0% 535 180 0.0% 0.0%
F C 478 200 7.0% 475 197 7.0% 487 210 7.0% 7.0%
2C 581 200 7.4% 576 195 7.4% 600 210 5.1% 7.4%
3C 685 200 7.4% 688 198 6.5% 713 210 4.0% 7.4%
4C 802 200 5.4% 801 198 5.4% 826 210 3.3% 6.1%
G C 424 150 1.7% 420 146 1.6% 429 150 0.0% 1.7%
2C 517 150 0.7% 511 146 1.2% 520 150 0.0% 1.2%
3C 608 150 0.5% 602 146 1.0% 611 150 0.0% 1.0%
H C 469 200 -10.0% 444 200 0.0% 444 200 0.0% 0.0%
2C 520 200 -8.3% 495 200 0.0% 495 200 0.0% 0.0%
3C 570 200 -7.1% 546 200 0.0% 546 200 0.0% 0.0%
4C 621 200 -6.2% 596 200 0.0% 596 200 0.0% 0.0%
I C 469 200 0.0% 469 200 0.0% 469 200 0.0% 0.0%
2C 504 200 0.0% 504 200 0.0% 504 200 0.0% 0.0%
3C 538 200 0.0% 538 200 0.0% 538 200 0.0% 0.0%
4C 573 200 0.0% 573 200 0.0% 573 200 0.0% 0.0%
J C 398 200 -10.2% 373 175 -10.2% 379 181 -10.2% 0.0%
2C 473 200 -18.8% 423 175 -8.0% 429 181 -8.0% 0.0%
3C 522 200 -15.5% 473 175 -6.6% 479 181 -6.6% 0.0%
4C 572 200 -13.1% 523 175 -5.6% 529 181 -5.6% 0.0%
Average -2.7% 1.4% -0.3% 2.3%
Table 5.6 – Comparison of synergy values from bilevel solutions and from the integrated
planning with different side payment schemes, for the case of limited visits to backhaul
customers.
S P∆ S P(0.50) S P(0.75) Bilevel
Inst. Q0 Costs Profits SV Costs Profits SV Costs Profits SV SV
A C 266 180 9.3% 267 181 9.3% 275 180 0.0% 10.0%
2C 324 180 12.5% 329 185 12.5% 345 180 0.0% 12.5%
3C 387 180 12.4% 396 189 12.4% 416 180 0.0% 12.4%
4C 458 180 9.7% 468 189 9.6% 488 180 0.0% 10.1%
5C 534 180 7.3% 542 189 7.8% 562 180 0.0% 8.9%
B C 198 60 -4.7% 192 60 0.0% 192 60 0.0% 0.0%
2C 253 60 5.0% 257 68 6.9% 264 60 0.0% 8.5%
C C 213 60 -6.1% 204 60 0.0% 204 60 0.0% 0.0%
2C 267 60 4.3% 270 66 5.9% 276 60 0.0% 9.6%
D C 215 60 -5.5% 207 60 0.0% 207 60 0.0% 0.0%
2C 265 60 2.2% 265 72 8.3% 270 60 0.0% 8.3%
E C 334 180 -7.6% 323 180 0.0% 323 180 0.0% 0.0%
2C 404 180 -5.0% 393 180 0.0% 393 180 0.0% 0.2%
3C 471 180 -1.8% 466 180 0.0% 466 180 0.0% 0.1%
4C 547 180 -1.4% 542 180 0.0% 542 180 0.0% 0.1%
5C 628 180 -1.2% 623 180 0.0% 623 180 0.0% 0.1%
F C 478 200 7.0% 475 197 7.0% 487 210 7.0% 7.0%
2C 581 200 8.4% 576 195 8.4% 602 221 8.4% 8.4%
3C 685 200 9.1% 693 195 6.6% 719 221 6.6% 9.1%
4C 805 200 7.4% 811 195 5.8% 836 221 5.8% 7.8%
G C 424 150 1.7% 420 146 1.6% 429 150 0.0% 1.7%
2C 517 150 4.4% 512 144 4.3% 526 161 5.0% 5.0%
3C 609 150 7.6% 607 145 7.2% 630 169 7.2% 7.6%
H C 469 200 -10.0% 444 200 0.0% 444 200 0.0% 0.0%
2C 526 200 -8.1% 502 200 0.0% 502 200 0.0% 0.0%
3C 583 200 0.3% 570 187 0.3% 580 196 0.3% 0.3%
4C 678 200 -1.3% 658 187 0.2% 667 196 0.2% 0.2%
I C 469 200 0.0% 469 200 0.0% 469 200 0.0% 0.0%
2C 508 200 0.0% 508 200 0.0% 508 200 0.0% 0.0%
3C 583 200 -8.1% 554 200 0.0% 554 200 0.0% 0.0%
4C 629 200 -6.9% 601 200 0.0% 601 200 0.0% 0.0%
J C 398 200 -10.2% 373 175 -10.2% 379 181 -10.2% 0.0%
2C 473 200 -14.1% 427 178 -4.5% 434 184 -4.8% 0.0%
3C 524 200 -8.6% 483 158 -8.6% 494 185 -3.4% 0.0%
4C 584 200 -6.3% 542 158 -6.3% 554 185 -2.2% 0.0%
Average 0.0% 2.4% 0.6% 3.7%
instances reported in Tables 5.5 and 5.6, the costs of the shipper are higher than its stand
alone solution. In these cases, the network costs surpass the stand alone solution, which
result in negative synergy values. These results also put in evidence the lack of efficiency
of determining the side payment only after solving the routing optimization problem.
With respect to the schemes where the side payments are determined while solving the
routing problem, the results show that when using the highest proportion (sp = 0.75) the
solutions tend to be non-collaborative (S V = 0.0%). On the other hand, using the lowest
proportion (sp = 0.50), many solutions do not attend the individual rationality of the carrier,
since the profits are lower than in the stand alone solution (e.g., instances F, G, J). In such
cases, collaboration would not take place. Nonetheless, Tables 5.5 and 5.6 also show that,
when collaboration occurs, several solutions provided by these side payments allow the
carrier to collect more profits than with the bilevel approach, but this comes at higher costs
for the shipper.
Generally, the results demonstrate that any solution of the bilevel approach has higher or
equal synergy value than any solution obtained through the compensation schemes ana-
lyzed. Based on these results, we argue that the bilevel model proposed can more ef-
fectively capture the interactions of the different players than the compensation schemes,
guaranteeing a more balanced solution with a higher synergy value.
Finally, as also observed from the results in the previous section, synergy values obtained
with any side payments are higher, on average, for a scenario with limited visits than a
scenario with unlimited visits to backhaul customers.
5.5.4 Computational limits of the collaborative VRPSB

The VRPB is strongly NP-hard since it generalizes the VRP (Toth and Vigo, 2002). Thus,
since the VRPSB generalizes the VRP (note that if there are no bakchauls, we have a simply
a VRP), VRPSB is also NP-hard. In addition, bilevel optimization problems are proven to
be strongly NP-hard and a mere assessment of the optimal solution is also NP-hard, even
for the simplest linear bilevel program (Jeroslow, 1985).
In this section, we aim to evaluate the practical difficulty of solving the collaborative trans-
portation planning formulations proposed. More precisely, we aim to empirically analyze
at what extent the collaborative formulation increases the practical difficulty of the prob-
lem.
Instances from F to N are tested with a computing time limit of 3600 seconds. Tables 5.7
and 5.8 provide the upper bound (UB), the computing time required to achieve the best
solution and the percentage gap obtained with each model, for both scenarios (limited and
unlimited backhaul visits). To avoid repetition, the UB is only displayed for instances not
already covered by Tables 5.2, 5.3 and 5.4.
The results show that the bilevel formulation is effectively the main reason behind the
complexity of solving the collaborative problem. The computing time to solve an instance
with the bilevel model tends to be higher than with the non-collaborative models, as well as
the percentage gap. Nevertheless, the complexity of the bilevel approach does not seem to
be much different than for the traditional VRPSB for some instances (e.g., instances L and
M). Moreover, the bilevel model solves simultaneously a routing and a pricing problem,
which is expected to be more computational challenging than solving a routing problem

only.
In general, the computing time to solve the bilevel problem tends to increase with increas-
ing size of the instance. It seems that the number of linehauls customers have more influ-
ence than the number of backhaul customers, but some instances present exceptions. For
example, instance N has more linehaul customers than instance L but it is solved in a much
shorter time. On the other hand, the exact method seems suitable to solve bilevel instances
with relatively high number of backhaul customers. These type of instances fit well real in-
dustries that have a wide range of backhauling opportunities, such as the forestry (Marques
et al., 2020).
Furthermore, it seems more challenging to solve the problems in a scenario with limited
visits to backhaul customers, than in a scenario with unlimited visits. On average, both
the computing time and the percentage gap are higher for the scenario with limit visits.
These results support that limiting the number of visits brings additional complexity the
bilevel model, since the incentives for backhauling also compete with diverse options for
pure inbound routes other than the least costly one.
Finally, we point out that the focus of this work is on the modelling aspects of the collab-
orative problem rather than on the solution methods. Using the properties of the bilevel
optimization, we have demonstrated an effective way to solve a hierarchical collaborative
problem. The rationale used to design the reformulation method could be applied to design
a metaheuristic, and thus guarantee higher efficiency when solving the problem.
5.6. Conclusions and future research

This work investigates an innovative formulation for a collaborative transportation planning
between a shipper and a carrier. The shipper offers incentives to the carrier in order to
create cost-effective integrated inbound-outbound routes. These incentives compete with
each other with other potential incentives offered to the carrier by external companies. The
problem of the shipper is a cost minimization VRPSB and the problem of the carrier is a
VRP with profits. Based on the hierarchical nature of the players and on the conflicting
objectives, the collaborative problem is formulated as a bilevel optimization problem. The
upper level describes the problem of the shipper and the lower level describes the problem
of the carrier. To solve the bilevel problem, we convert it in an equivalent single-level
mixed integer linear program by exploring problem-specific characteristics of the lower
level, and then, standard linearization techniques.
This work conducts an extensive analysis on the properties of the bilevel approach to
handle the collaborative problem. The bilevel model is compared with traditional non-
collaborative routing problems and with different side payment schemes, in order to assess
the impact of the collaboration and the approach applied. In addition, the impact of limit-
ing the number of backhaul visits is also evaluated. Finally, the computational limits of the
collaborative formulation is compared against traditional single-level routing problems.
The results of this work put in evidence the advantages of the bilevel approach to handle a
collaborative transportation planning, although the computational effort tend to be higher
5.6. Conclusions and future research 175
Table 5.7 – Computational performance of each model, considering unlimited backhaul

visits.
VRP VRPSB Bilevel
Inst. Q0 UB gap time UB gap time UB gap time
F C 0.0% 10 0.0% 25 0.0% 92
2C 0.0% 11 0.0% 19 0.0% 1579
3C 0.0% 10 0.0% 43 0.0% 145
4C 0.0% 9 0.0% 76 0.0% 118
G C 0.0% 14 0.0% 21 0.0% 24
2C 0.0% 16 0.0% 39 0.0% 73
3C 0.0% 6 0.0% 30 0.0% 61
H C 0.0% 5 0.0% 17 0.0% 25
2C 0.0% 5 0.0% 5 0.0% 27
3C 0.0% 5 0.0% 10 0.0% 24
4C 0.0% 6 0.0% 16 0.0% 27
I C 0.0% 13 0.0% 36 0.0% 35
2C 0.0% 24 0.0% 30 0.0% 70
3C 0.0% 18 0.0% 25 0.0% 57
4C 0.0% 15 0.0% 45 0.0% 45
J C 0.0% 10 0.0% 37 0.0% 377
2C 0.0% 11 0.0% 34 0.0% 98
3C 0.0% 15 0.0% 19 0.0% 131
4C 0.0% 14 0.0% 19 0.0% 457
K C 472 0.0% 287 455 0.0% 1072 472 0.5% 3600
2C 522 0.0% 289 499 0.0% 161 522 0.0% 650
3C 571 0.0% 162 549 0.0% 652 571 0.0% 2568
4C 621 0.0% 114 599 0.0% 223 621 0.0% 416
L C 452 0.0% 169 435 0.0% 1079 452 0.0% 1224
2C 502 0.0% 283 477 0.0% 1248 502 0.0% 1949
3C 552 0.0% 825 527 0.5% 3600 552 0.0% 1531
4C 601 0.0% 877 577 0.0% 1216 601 0.0% 796
M C 542 1.9% 3600 527 2.7% 3600 543 2.1% 3600
2C 591 1.0% 3600 567 1.1% 3600 595 2.5% 3600
3C 640 0.0% 3545 617 1.6% 3600 646 3.3% 3600
4C 692 1.8% 3600 666 0.0% 2215 702 7.3% 3600
N C 474 0.0% 22 460 0.0% 11 474 0.0% 444
2C 523 0.0% 22 500 0.0% 14 523 0.0% 985
3C 573 0.0% 22 548 0.0% 4 573 0.0% 206
4C 622 0.0% 21 597 0.0% 34 622 0.0% 167
Average 0.1% 504 0.2% 654 0.4% 926
Table 5.8 – Computational performance of each model, considering limited backhaul visits.
VRP VRPSB Bilevel
Inst. Q0 UB gap time UB gap time UB gap time
F C 0.0% 9 0.0% 17 0.0% 45
2C 0.0% 12 0.0% 15 0.0% 114
3C 0.0% 11 0.0% 68 0.0% 242
4C 0.0% 23 0.0% 22 0.0% 37
G C 0.0% 12 0.0% 5 0.0% 79
2C 0.0% 15 0.0% 32 0.0% 40
3C 0.0% 7 0.0% 10 0.0% 121
H C 0.0% 4 0.0% 7 0.0% 33
2C 0.0% 12 0.0% 21 0.0% 41
3C 0.0% 4 0.0% 11 0.0% 362
4C 0.0% 8 0.0% 17 0.0% 509
I C 0.0% 11 0.0% 28 0.0% 451
2C 0.0% 12 0.0% 5 0.0% 202
3C 0.0% 23 0.0% 12 0.0% 224
4C 0.0% 26 0.0% 17 0.0% 146
J C 0.0% 22 0.0% 16 0.0% 338
2C 0.0% 14 0.0% 13 0.0% 1026
3C 0.0% 34 0.0% 14 0.0% 587
4C 0.0% 29 0.0% 24 0.0% 173
K C 472 0.0% 290 455 0.0% 368 472 0.0% 873
2C 538 0.0% 162 503 0.0% 217 538 0.0% 1800
3C 649 0.0% 73 578 0.0% 1374 637 1.7% 3600
4C 773 0.0% 81 654 0.8% 3600 734 1.2% 3600
L C 452 0.0% 173 435 0.0% 633 452 4.8% 3600
2C 512 0.0% 338 477 0.5% 3600 512 1.3% 3600
3C 578 0.0% 422 532 0.0% 635 580 11.4% 3600
4C 661 0.0% 324 591 0.0% 695 658 0.6% 3600
M C 542 2.0% 3600 527 2.2% 3600 578 11.8% 3600
2C 612 0.0% 2160 567 0.0% 2308 623 4.4% 3600
3C 725 0.7% 3600 627 1.3% 3600 701 3.5% 3600
4C 843 1.6% 3600 690 0.0% 1014 789 3.3% 3600
N C 474 0.0% 22 460 0.0% 8.5 474 0.0% 1062
2C 529 0.0% 19 500 0.0% 10 529 0.0% 799
3C 600 0.0% 22 548 0.0% 9 600 0.0% 235
4C 690 0.0% 20 600 0.0% 7 677 0.0% 1909
Average 0.1% 434 0.1% 629 1.3% 1356
Bibliography 177
than than the traditional non-collaborative formulations. Therefore, one main direction for
future work includes the development of an efficient metaheuristic to solve the problem.
Such metaheuristic could encompass a genetic algorithm or another method of the family
of evolutionary algorithms, since these are the most used advanced methods to solve bilevel
optimization problems (Sinha et al., 2018). The bilevel model proposed in this work is
designed under an optimistic approach for both players. Because the objectives of each
player are different and they collaborate in a hierarchical structure, considering the most
optimistic case of the lower level can be seen as a robust strategy to achieve robust solutions
for the upper level. An interesting future line of research could encompass investigating
the collaborative problem under a pessimistic approach and compare it the optimistic.
Acknowledgements
This work was financed by the ERDF – European Regional Development Fund through
(Easyflow).
The fourth author wishes to thank the support of the Institut de Valorisation des Données
and Fonds de Recherche du Québec through the FRQ–IVADO Research Chair in Data
Science for Combinatorial Game Theory.
Bibliography
Archetti, C., Savelsbergh, M., and Speranza, M. G. (2016). The Vehicle Routing Problem
with Occasional Drivers. European Journal of Operational Research, 254(2):472–480.
Augerat, P., Belenguer, J. M., Benavent, E., Corberán, A., Naddef, D., and Rinaldi, G.
(1995). Computational results with a branch and cut code for the capacitated vehi-
cle routing problem. Technical report, Institut National Polytechnique, 38 - Grenoble
(France).
Bailey, E., Unnikrishnan, A., and Lin, D.-Y. (2011). Models for Minimizing Backhaul
Costs Through Freight Collaboration. Transportation Research Record: Journal of the
Caballini, C., Sacone, S., and Saeednia, M. (2016). Cooperation among truck carriers
in seaport containerized transportation. Transportation Research Part E: Logistics and
Transportation Review, 93:38 – 56.
Colson, B., Marcotte, P., and Savard, G. (2007). An overview of bilevel optimization.
Annals of Operations Research, 153(1):235–256.
Cruijssen, F., Borm, P., and Fleuren, H. (2005). Insinking: A methodology to exploit
synergy in transportation. SSRN Electronic Journal.
Cruijssen, F., Bräysy, O., Dullaert, W., Fleuren, H., and Salomon, M. (2007). Joint route
planning under varying market conditions. International Journal of Physical Distribu-
tion & Logistics Management, 37(4):287–304.
Cuervo, D. P., Vanovermeire, C., and Sörensen, K. (2016). Determining collaborative

profits in coalitions formed by two partners with varying characteristics. Transportation
Research Part C: Emerging Technologies, 70:171 – 184.
Dahl, S. and Derigs, U. (2011). Cooperative planning in express carrier networks — an

empirical study on the effectiveness of a real-time decision support system. Decision
Support Systems, 51(3):620 – 626.
de Matta, R., Lowe, T. J., and Zhang, D. (2017). Competition in the multi-sided platform
market channel. International Journal of Production Economics, 189:40 – 51.
Defryn, C., Sörensen, K., and Dullaert, W. (2019). Integrating partner objectives in hori-
zontal logistics optimisation models. Omega, 82:1–12.
Defryn, C., Sörensen, K., and Cornelissens, T. (2016). The selective vehicle routing
problem in a collaborative environment. European Journal of Operational Research,
250(2):400 – 411.
Early, C. (2011). Delivering greener logistics. Available at https://transform.iema.

net/article/delivering-greener-logistics.
Ergun, O., Kuyzu, G., and Savelsbergh, M. (2007). Reducing truckload transportation costs
through collaboration. Transportation Science, 41(2):206–221.

198:353–362.
Fischetti, M., Ljubić, I., Monaci, M., and Sinnl, M. (2017). A new general-purpose algo-
rithm for mixed-integer bilevel linear programs. Operations Research, 65(6):1615–1637.

Garcia-Herreros, P., Zhang, L., Misra, P., Arslan, E., Mehta, S., and Grossmann, I. E.
(2016). Mixed-integer bilevel optimization for capacity planning with rational markets.
Computers & Chemical Engineering, 86:33–47.
Jeroslow, R. G. (1985). The polynomial hierarchy and a simple model for competitive
analysis. Mathematical Programming, 32(2):146–164.
Bibliography 179
Juan, A. A., Faulin, J., Pérez-Bernabeu, E., and Jozefowiez, N. (2014). Horizontal Co-
operation in Vehicle Routing Problems with Backhauling and Environmental Criteria.
Procedia - Social and Behavioral Sciences, 111:1133–1141.
Kozanidis, G., Kostarelou, E., Andrianesis, P., and Liberopoulos, G. (2013). Mixed integer
parametric bilevel programming for optimal strategic bidding of energy producers in
day-ahead electricity markets with indivisibilities. Optimization, 62(8):1045–1068.
Krajewska, M. A., Kopfer, H., Laporte, G., Ropke, S., and Zaccour, G. (2008). Horizontal
cooperation among freight carriers: request allocation and profit sharing. Journal of the
Operational Research Society, 59(11):1483–1491.
Liu, R., Jiang, Z., Liu, X., and Chen, F. (2010). Task selection and routing problems in
collaborative truckload transportation. Transportation Research Part E: Logistics and
Transportation Review, 46(6):1071 – 1085.
Lunday, B. J. and Robbins, M. J. (2019). Collaboratively-developed vaccine pricing and

stable profit sharing mechanisms. Omega, 84:102 – 113.
Marques, A., Soares, R., Santos, M. J., and Amorim, P. (2020). Integrated planning of
inbound and outbound logistics with a rich vehicle routing problem with backhauls.
Omega, 92:102172.
McCormick, G. P. (1976). Computability of global solutions to factorable nonconvex

programs: Part i — convex underestimating problems. Mathematical Programming,
10(1):147–175.
40(8):2985–2991.
Qiu, X. and Huang, G. Q. (2016). Transportation service sharing and replenish-

ment/delivery scheduling in supply hub in industrial park (ship). International Journal
of Production Economics, 175:109 – 120.
Qiu, X., Lam, J. S. L., and Huang, G. Q. (2015). A bilevel storage pricing model for
outbound containers in a dry port system. Transportation Research Part E: Logistics
and Transportation Review, 73:65 – 83.
Safari, N., Zarghami, M., and Szidarovszky, F. (2014). Nash bargaining and
leader–follower models in water allocation: Application to the Zarrinehrud River basin,
Iran. Applied Mathematical Modelling, 38(7-8):1959–1968.
Santos, M. J., Curcio, E., Mulati, M. H., Amorim, P., and Miyazawa, F. K. (2020). A
robust optimization approach for the vehicle routing problem with selective backhauls.
Transportation Research Part E: Logistics and Transportation Review, 136:101888.
Shapley, L. (1953). A value for n-person games. Annals of Mathematics Studies, (28):307
– 317.
Sinha, A., Malo, P., and Deb, K. (2018). A Review on Bilevel Optimization: From Classi-
cal to Evolutionary Approaches and Applications. IEEE Transactions on Evolutionary
Computation, 22(2):276–295.
Sun, H., Gao, Z., and Wu, J. (2008). A bi-level programming model and solution algo-
rithm for the location of logistics distribution centers. Applied Mathematical Modelling,
32(4):610–616.
Tahernejad, S., Ralphs, T. K., and DeNegre, S. T. (2017). A branch-and-cut algorithm

for mixed integer bilevellinear optimization problems and its implementation. Techni-
cal report, COR@L Technical Report 16T-015-R3, Industrial and Systems Engineering,
Lehigh University.
Toth, P. and Vigo, D. (2002). 8. vrp with backhauls. In Toth, P. and Vigo, D., editors, The
Vehicle Routing Problem, volume 9, pages 195–224. SIAM, Philadelphia.
Turkensteen, M. and Hasle, G. (2017). Combining pickups and deliveries in vehicle routing
– An assessment of carbon emission effects. Transportation Research Part C: Emerging
Vicente, L., Savard, G., and Judice, J. (1996). Discrete linear bilevel programming prob-
lem. Journal of Optimization Theory and Applications, 89(3):597–614.
Wang, J., Lim, M. K., Tseng, M.-L., and Yang, Y. (2019). Promoting low carbon agenda
in the urban logistics network distribution system. Journal of Cleaner Production,
211:146–160.
Xu, S., Liu, Y., and Chen, M. (2017). Optimisation of partial collaborative transportation
scheduling in supply chain management with 3PL using ACO. Expert Systems with
Applications, 71:173–191.
Xu, X., Zheng, Y., and Yu, L. (2018). A bi-level optimization model of LRP in collaborative
logistics network considered backhaul no-load cost. Soft Computing, 22(16):5385–5393.
Yu, J. and Dong, Y. (2013). Maximizing profit for vehicle routing under time and weight
constraints. International Journal of Production Economics, 145(2):573 – 583.
Zeng, B. and An, Y. (2014). Solving Bilevel Mixed Integer Program by Reformulations
and Decomposition. Optimization On-line, pages 1–34.
Özener, O. , Ergun, , and Savelsbergh, M. (2011). Lane-exchange mechanisms for truck-

load carrier collaboration. Transportation Science, 45(1):1–17.
Bibliography 181
5..1 Traditional models

Traditionally, the transportation planning does not consider collaboration. In this section,
two traditional formulations are presented for the transportation planning, namely a sep-
arated model (which is described by a typical VRP) and an integrated model (which is
described by a VRPSB). Both models are simplifications of the bilevel VRPSB presented
in Section 5.3.2 that exclude the lower level problem from the formulation. Consequently,
the variable and all constraints related to the incentives are also excluded in the non-
collaborative models. On the other hand, the maximum distance allowed per route is taking
into account in both models, since the carrier would never accept to exceed this distance,
even for a single delivery route. In addition, the objective function of the carrier is treated
as an expression in both non-collaborative models. The mathematical formulation of the
traditional models are presented.
Separated model
The separated model describes the problem where inbound and outbound routes are planned
separately and integrated routes are not allowed. The problem is formulated as an Open
VRP and the objective function in (5.40) minimizes the total routing costs of delivery vehi-
cles, plus the constant expression given by the minimum number of vehicles necessary to
satisfy the depot demand. This states that the demand of the depot can only be satisfied by
dedicated inbound vehicles. The complete formulation of the separated model is presented
as follows:
XXX X
i j · xi j +
cU k
min Ob · 2cU
b0 (5.40)
i∈V j∈L k∈K b∈B
subjected to
(5.2) − (5.4)
(5.5) − (5.36)
(5.16)
xikj , Ob ∈ {0, 1}, yi j ≥ 0, ∀i, j ∈ V = {0} ∪ L, b ∈ B, k ∈ K, w ∈ Z+0 (5.41)
Integrated model
The integrated model describes the problem where inbound and outbound routes are planned
jointly by the shipper. This model considers that the shipper assumes control over all ve-
hicles of the carrier used in the network. As the lower level variables are not considered,
the shipper no more compete with others for backhaul routes. Instead, the shipper assumes
that the unitary cost to visit a backhaul customer is the same as visiting a linehaul customer.
Thus, the objective function in (5.42) is to minimize the total routing costs and outsourc-
ing of dedicated inbound vehicles. The constraints of the integrated model are the same
as those from the upper level problem with the additional backhaul customers constraints
of the lower level problem. Note that the variables of the lower level are substituted by
variables of the upper level in the backhaul customers constraints.
XXX X
i j · xi j +
cU k
min Ob · 2cU
b0 (5.42)
i∈V j∈L k∈K b∈B
subjected to
(5.2) − (5.36)
(5.15 ) and (5.160 )
0
xikj , Ob ∈ {0, 1}, yi j ≥ 0, ∀i, j ∈ V, b ∈ B,k ∈ K, w ∈ Z+0 (5.43)

Chapter 6
A Green Vehicle Routing Problem

with Backhauls
A green lateral collaborative transportation problem

under different collaboration strategies and profit al-
location methods
Maria João Santos∗† · Sara Martins∗† · Pedro Amorim∗† · Bernardo Almada-Lobo∗†
Submitted to Journal of Cleaner Production, 2020
Abstract Collaboration between companies in transportation problems seeks to reduce

empty running of vehicles and to increase the use of vehicles’ capacity. Motivated by a
case study in the food supply chain, this paper examines a lateral collaboration between a
leading retailer (LR), a third party logistics provider (3PL) and different producers. Three
collaborative strategies may be implemented simultaneously, namely pickup-delivery, col-
lection and cross-docking. The collaborative pickup-delivery allows an entity to serve cus-
tomers of another in the backhaul trips of the vehicles. The collaborative collection allows
loads to be picked up at the producers in the backhauling routes of the LR and the 3PL,
instead of the traditional outsourcing. The collaborative cross-docking allows the produc-
ers to cross-dock their cargo at the depot of another entity, which is then consolidated and
shipped with other loads, either in linehaul or backhaul routes. The collaborative problem
is formulated with three different objective functions: minimizing total operational costs,
minimizing total fuel consumption and minimizing operational and CO2 emissions costs.
The synergy value of collaborative solutions is assessed in terms of costs and environmen-
tal impact. Three proportional allocation methods from the literature are used to distribute
the collaborative gains among the entities, and their limitations and capabilities to attend
fairness criteria are analyzed. Collaboration is able to reduce the global fuel consumption
in 26% and the global operational costs in 28%, independently of the objective function
used to model the problem. The collaborative pickup-delivery strategy outperforms the
other two in the majority of instances under different objectives and parameter settings.
The collaborative collection is favoured when the ordering loads from producers increase.
The collaborative cross-docking tends to be implemented when the producers are located
close to the depot of the 3PL.

184 Chapter 6. A Green Vehicle Routing Problem with Backhauls
Keywords Collaboration · Backhauling · Cross-docking · Profit allocation

6.1. Introduction
Transportation is a crucial activity in any supply chain, responsible for a large portion
of the total logistics costs. In terms of environmental impact, transportation is responsi-
ble for about 14% of the total CO2 emissions worldwide, from which road transportation
alone represents about 75% (Liu et al., 2020). Thus, under the contemporary need of im-
proving supply chains sustainability, adopting green initiatives for transportation activities
presents opportunities to promote such goal. Among these green initiatives, collaboration is
a relatively wide-spread practice among companies (Evangelista, 2014; Ramanathan et al.,
2014). The goal of collaborative transportation is to identify hidden costs in logistics op-
erations that cannot be managed individually, but can be reduced collectively (Ergun et al.,
2007). Although the literature on collaborative vehicle routing has been increasing, there
are still some limitations that we aim to address in this work.
The first limitation concerns the type of collaboration usually investigated in the literature.
There are three types of collaboration: i) horizontal, when the participants are at the same
level of the supply chain (e.g., two suppliers), ii) vertical, when the participants are at dif-
ferent levels (e.g., a carrier and a supplier), and iii) lateral, when there is a combination of
both previous types (e.g., two suppliers and two carriers). However, within collaborative
vehicle routing problems, horizontal collaboration is the most investigated type in the lit-
erature (Gansterer and Hartl, 2018), whereas less research has been devoted to vertical and
lateral collaboration.
The second limitation regards the analysis of the impact of different collaborative strate-
gies. Strategies applied in the collaborative transportation planning may include exchang-
ing customers between companies (Dai and Chen, 2012; Fernández et al., 2016), sharing
vehicles’ capacity of different companies with a common depot (Cruijssen et al., 2007),
consolidating cargo of different companies in single delivery routes (Ergun et al., 2007), or
providing loads in backhaul routes to reduce empty trips (Juan et al., 2014). Moreover, the
majority of the literature on collaborative vehicle routing focus only on economic indica-
tors to assess the gains of collaboration. Nonetheless, collaborative transportation should
also be assessed in terms of environmental impact.
Another limitation concerns the simultaneous analysis of the value of collaboration in terms
of impact on the routing problem and impact on the allocation of the collaborative gains
between the participants in the transportation problem. According to Gansterer and Hartl
(2018), the majority of the literature on collaborative transportation focus either on solving
the transportation problem or on solving the allocation problem, both usually not both.
Therefore, this work tackles the three aforementioned limitations found in the literature
of collaborative vehicle routing. We address the first limitation by studying a case of lat-
eral collaboration between different participants in a transportation network, where they
may perform similar services (e.g., serving customers of each other), leading to horizontal
collaboration, but also complementary transportation activities (e.g., cross-docking), lead-
ing to vertical collaboration. Studying a lateral collaboration will allow us to investigate
the impact of different collaborative strategies that may be applied simultaneously, namely
pickup-delivery, collection and cross-docking. The collaborative pickup-delivery refers to a
strategy where customers of one participant are served by vehicles of the other participant.
The collaborative collection considers a strategy that allows vehicles of other participants
to pickup cargo at the producers, thus avoiding empty running of vehicles when returning
to the depot and, at the same time, reducing the need of outsourcing. Finally, the collab-
orative cross-docking refers to a strategy that allows loads from different locations to be
consolidated in an intermediary location before arriving to the final destination. The second
limitation is addressed by analyzing the collaborative model under three different objective
functions, covering economic and environmental issues, allowing to measure the impact
of collaboration under a sustainable context. The third limitation is addressed by solving
the integrated routing and the profit sharing problems, and analyzing the effect of different
collaborative solutions on the distributed profit shares determined by different allocation
methods.
This approach is validated and applied to an European leading retailer (LR) and a third-
party logistics provider (3PL). Several practical constraints are taken into account, such as
the interdependence between different strategies, which allows to gather relevant insights
to implement in real contexts. The case-study is motivated by the potential of reducing
logistics costs through collaboration, enabled by the transportation network. The LR and
3PL distribution centers are strategically located in distinct geographical areas that benefit
their operational activities. The LR has distribution centers located closer to dense areas of
stores to be supplied, while the 3PL is centrally positioned to serve different national and
international customers. With greater or lesser frequency, both serve common areas, being
the 3PL located in an area with more dispersed retail stores and closer to the LR suppliers.
The participants of such transportation networks are usually aware of the hidden potential
of collaboration to mitigate costs and CO2 emissions, and the main goal of this paper is
to provide managerial insights on that subject. It is worth mentioning that this case study
represents a common transportation problem setting in practice. Thus, the approach devel-
oped in this work can be easily extended and replicated for other supply chains. Also, the
use of consolidation centers is a relatively common strategy used by carriers and shippers
to reduce transportation costs, but it is rarely examined under the context of collabora-
tion. Therefore, this work provides insights into the impact of collaboration achieved by
combining strategies commonly used in collaborative networks, such as backhauling, and
others that are not usually accounted for, such as cross-docking.
The remainder of this paper is structured as follows. Section 6.2 presents the literature
review on collaborative vehicle routing problems relevant for this work, and frame the
position of our paper within current research. Section 6.3 provides a description of the col-
laborative problem and strategies investigated, the mathematical formulation of the collab-
orative vehicle routing problem and the allocation methods used for profit sharing. Section
6.4 reports the computational experiments on randomly generated and case-study instances,
discussing the impact of collaboration under economic and environmental contexts and the
performance of different allocation methods. Section 6.5 concludes the paper describing
the main findings and managerial insights of this work, and proposing future research as
well.
6.2. Related literature 187
6.2. Related literature

The collaborative transportation problem can be characterized according to the type of enti-
ties involved, or type of logistics services offered in the transportation network. Horizontal
collaboration refers to the case where the entities involved are at the same level of network,
offering the same type of service (e.g., carrier collaboration) (Audy et al., 2012). Vertical
collaboration refers to the case when the entities are at different levels of the supply chain
(e.g., supplier-retailer collaboration). Lateral collaboration, which combines horizontal and
vertical collaborations, exploits the benefits of both types of collaboration, contributing to
the overall management of transportation activities (Mason et al., 2007). The present work
describes a case of lateral collaboration, involving different entities (LR, 3PL and produc-
ers) which may execute similar transportation services, leading to horizontal collaboration,
as well as complementary services, leading to vertical collaboration.
It has been demonstrated that collaboration unveils hidden opportunities to reduce the lo-
gistics burden of transportation for individual companies. For example, when a carrier
(or shipper) needs to deliver to long distance customers, which are conveniently located
near to other carrier, the two entities may collaborate through customer sharing. In such
cases, the customer of the former carrier can be supplied by the other carrier, allowing to
reduce the total distances travelled by the vehicles of both entities (e.g., Fernández et al.
(2016)). Another opportunity for collaboration arises when the total loads to deliver to a
customer, or in a region, are relatively small compared to the full capacity of the vehicle
used to carry them. In such cases, the carrier can exchange these loads with other carrier
operating in the same region, or serving the same customer, increasing the average load
of the vehicles in use (e.g., Paul et al. (2019)). In practice, most of the vehicles return-
ing to the depot travel empty. Thus, backhauling emerges as a powerful opportunity for
collaboration, by allowing vehicles to perform additional services in these returning trips,
thus reducing the total empty running. The additional services may include pickup and
deliveries between different locations (e.g., Juan et al. (2014)) or one or more pickups to
supply the destination depot of the vehicles (e.g., Frisk et al. (2010)). Another possibility
of collaboration relies on the use of intermediary facilities between a pickup and a delivery
location, in order to reduce the total traveled distance and number of vehicles required. In
such cases, the loads of different suppliers that serve the same depot can be consolidated
in a intermediary location, which are then carried by a single vehicle to the depot (e.g.,
le Blanc et al. (2006); Neves-Moreira et al. (2016)). Although some of the above men-
tioned works tackle more than one collaborative strategy simultaneously, they only cover
one type of collaboration, specifically horizontal collaboration. In opposition, this paper
tackles different collaborative strategies that imply a combination of horizontal and verti-
cal collaboration. Particularly, we consider strategies that are based on capacity sharing
(which we designate as collaborative pickup-delivery), backhauling (which we designate
as collaborative collection) and use of intermediary facilities (which we designate as col-
laborative cross-docking).
The literature on collaborative vehicle routing shows two main streams of research - the col-
laborative vehicle routing problem and the profit sharing problem. Recent reviews on each
stream can be found in Gansterer and Hartl (2018) and Guajardo and Rönnqvist (2016),
respectively. The first stream focus on optimizing the global collaborative problem (e.g.,
minimizing the total cost in the collaborative network), developing efficient methods to
solve it and measuring the impact of the collaboration. Common metrics applied in the
collaborative transportation are the performance and the synergy value. The performance
provides the absolute gains (or savings) obtained from collaboration, determined as the
difference between the profit (or cost) of a non-collaborative solution and a collaborative
solution (Cuervo et al., 2016). The value of a collaborative solution is provided by the
sum of all individual profits (or costs) of each entity before collaboration, also known as
stand alone solution. The synergy value provides the percentage profit increase (or cost
reduction) created by the collaboration, compared to the stand alone solution (Cruijssen
et al., 2007). More recently, some works also assess the impact of collaboration in terms
of environmental impacts (e.g., Juan et al. (2014); Chabot et al. (2018)). In this work, we
analyse the synergy value of collaborative solutions obtained with a fuel consumption min-
imization model and with a model minimizing both operational and environmental costs,
which we designate as the holistic model. We further assess the environmental impact of
such solutions in comparison with the traditional economic models.
The second stream of the literature focus on developing allocation methods to distribute
the gains from collaboration among the participants, according to some fairness criteria, or
properties (Frisk et al., 2010). Common properties desirable in the profit sharing problem
are efficiency, individual rationality and group rationality (Dahlberg et al., 2019). The first
property ensures that all costs (profits) obtained with the collaboration are fully distributed
among the participants. The second property guarantees that each participant is not allo-
cated higher costs (lower profit) than its stand alone solution. The third property applies the
same principle as the previous property for a coalition (group of participants), i.e. the par-
ticipants in a coalition cannot be allocated a higher cost (lower profit) than the cost (profit)
of that coalition. A stable allocation is one that ensures the properties of efficiency and
group rationality. The Shapley value is one of the most frequently used methods for profit
sharing in the literature, which allocates to each participant the respective average marginal
cost that results from its participation in the coalition. However, this method requires solv-
ing all possible coalitions, and in some cases it cannot guarantee stability (Guajardo and
Rönnqvist, 2016). Another method widely used is the nucleolus, which defines an ex-
cess vector related to the degree of acceptance of an allocated cost for each coalition, and
the goal is the maximization of the minimum excess. The nucleolus method outperforms
several methods proposed in the literature, in particular regarding stability, but it is rela-
tively hard to compute, and usually requires an algorithm to solve a sequential set of linear
programs (Guajardo and Rönnqvist, 2016). The proportional methods are the simplest al-
location methods in the literature, because in opposition to the previous methods they do
not require solving an optimization problem, and depending on the rules applied, they may
provide stability. Thus, although having been subject to some criticism from the research
community, the proportional methods are easy to compute and easy to communicate (Gua-
jardo and Rönnqvist, 2016), can be more easily scalable for real life instances (Ozener,
2014), and therefore they are commonly implemented in practice (Liu, 2010). Proportional
methods can also apply several rules to allocate the costs, namely based on demand, on
distances or on stand alone costs (Guajardo and Rönnqvist, 2016). Currently, there is no
6.2. Related literature 189
general acceptance of one method over another in the research community, since an al-
location method hardly fulfils all the properties defined in the literature. For this reason,
there is a vast literature on tailored methods adapted to specific case-studies or applications
(Guajardo and Rönnqvist, 2016). In this work, we investigate three proportional allocation
methods following the work of Lunday and Robbins (2019), and evaluate their performance
in terms efficiency and individual rationality.
One of the pioneer works tackling both the vehicle routing and the profit sharing problems
is described in Krajewska et al. (2008) for carrier collaboration. The collaborative problem
is modelled as a multi-depot Pickup and Delivery Problem with Time Windows (PDPTW)
which merges the requests of all carriers. The problem is optimized for distance-based
cost minimization and the profit sharing is solved with the Shapley value. Another col-
laborative PDP is investigated in Dai and Chen (2012) where both customer sharing and
vehicle sharing can be adopted between carriers. Three different methods are studied for the
profit sharing method, and each one requires solving a linear program. The first allocation
method minimizes the difference between the value of the collaborative (global) solution
and the Shapley value of each carrier. The other two methods minimize the difference be-
tween the allocation ratios of any pair of carriers, following different proportional rules.
In Zibaei et al. (2016), the collaborative problem of customer exchange is formulated as a
multi-depot VRP and the profit sharing is solved with different allocation methods already
proposed in the literature. Collaboration between a supplier and its customers is examined
in Özener et al. (2013), which describes an Inventory Routing Problem (IRP) minimizing
the average transportation costs and proposes three allocation methods to solve the profit
sharing problem. The first method allocates to each customer a cost proportional to its
individual cost. The second allocates the cost on a per-route basis, considering that several
customers can be served in the same route, thus considering synergies between customers.
The third method allocates the costs to customers based on the solution of the dual problem,
and considers the synergy between customers and between routes. In a latter work, Ozener
(2014) formulates a collaborative VRP with minimization of travelling and CO2 emissions
costs. The authors apply similar allocation methods presented in prior work, but now dis-
tributing both the transportation costs and emissions costs among the customers. Another
work that considers environmental issues on both routing and profit sharing problems is
described in Pradenas et al. (2013). The authors formulate a collaborative Vehicle Routing
Problem with Backhauls and Time Windows (VRPBTW) with minimization of energy con-
sumption, and the total costs are distributed according to the Shapley value. Sanchez et al.
(2015) also apply the Shapley value but the collaborative model is formulated as a VRPTW
with constraints limiting the carbon footprint of transportation. Wang et al. (2018) study
the collaboration of multiple centers (distribution centers (DCs), logistics centers (LCs),
depots) in a two-echelon VRP, where the first echelon contains DCs and LCs, and the sec-
ond echelon contains the DCs and the customers. The problem is solved considering the
minimization of costs and CO2 emissions. Afterwards, the Minimum Cost-Remaining Sav-
ings (MCRS) method distributes the costs among the collaborators. This allocation method
is based on a bilevel methodology that first allocates a minimum benefit to all participants,
and only after it distributes the remaining amount.
From the above literature, the most similar work to ours is the one of Wang et al. (2018).
However, while Wang et al. (2018) model the collaborative problem with a bi-objective
function, our work includes two models, one with only economic goals and one with only
environmental goals. Therefore, we examine how different formulations of the collabora-
tive model impact on the synergy value. Another distinct feature of our work is that we
study three different collaborative strategies that rely on backhauling, capacity sharing and
cross-docking, whereas Wang et al. (2018) only focus on customer sharing considering that
the demand of the customers of one participant can be satisfied by any other participant.
Finally, we study simple proportional allocation methods which are easily adapted to real
instances. Differently, Wang et al. (2018) uses an allocation method that may outperform
the proportional methods from a theoretical point of view, but which are hard to implement
in practice due to the need of solving an optimization for all possible coalitions formed.
6.3. The collaborative transportation planning

In this section, we detail the problem, and the methodologies involved to solve the routing
and the profit sharing problems. First, a description of each collaborative strategy and
their requirements is presented. Then, the mathematical models used to formulate the
collaborative problems under different contexts (minimization of costs and environmental
impacts) are formally defined. Finally, the proportional allocation methods used to solve
the profit sharing problem are described.
6.3.1 Illustration of collaborative strategies

Traditionally, each participant in the transportation network is responsible for serving its
customers, optimizing their routes individually. Figure 6.1a illustrates the case of individ-
ual planning, which does not consider collaboration. The LR plans the delivery routes to
satisfy the demand of its stores and the 3PL plans the delivery routes to satisfy the demand
of its customers. After all deliveries are made, their dedicated vehicles return empty to
the respective depot. The producers supply the LR through outsourced vehicles. Because
the LR and the 3PL are located in different areas but operate in the same region, there are
hidden opportunities to increase the efficiency of the logistics operations through collabo-
ration. Thus, different participants may collaborate through the joint planning of transport
operations, implementing one or more collaborative strategies.
Figure 6.1b illustrates the case of collaborative pickup-delivery, which allows vehicles of
the 3PL to serve stores of the LR. In this case, after all deliveries made to the 3PL cus-
tomers, a 3PL vehicle may pickup a set of requests at the depot of the LR, delivering them
to stores in the way back to its depot. This collaborative strategy does not include the possi-
bility of vehicles of the RL serving customers of the 3PL, due to practical or market-related
reasons (e.g., retailers do not serve other retailers). The collaborative pickup-delivery is
particularly important when the stores are located along the backhaul trips of the 3PL, or
located near to his depot. Figure 6.1c illustrates the case of collaborative cross-docking,
which allows the producers to ship their loads to the depot of the 3PL, instead of sending
directly to the depot of the RL. These loads can be consolidated in the delivery routes of the
3PL, or can be collected latter by the vehicles of the RL. The collaborative cross-docking
6.3. The collaborative transportation planning 191
emerges as a solution to lower the transportation costs of the producers. In particular, if

several loads of different producers are consolidated, it might be possible to reach full-
truck loads. Figure 6.1d illustrates the case of collaborative collection, which allows the
dedicated vehicles of both, the RL and the 3PL, to pickup loads at the producers. This
strategy provides a cost reduction for the producers, and allows to reduce empty running of
dedicated vehicles.
(a) No collaboration
(b) Collaborative pickup-delivery
(c) Collaborative cross-docking (d) Collaborative collection

Legend: Circles represent stores, squares represent 3PL customers and diamonds represent producers. The
blacks rectangles are the depots of LR and 3PL. Blue clusters and orange clusters represent the set of locations
visited by 3PL and by LR, respectively. Blue lines and orange lines represent the routes executed by dedicated
vehicles of 3PL and vehicles of LR, respectively. Black lines represent routes executed by outsourced vehicles
from producers. Full and dotted lines represent loaded and empty vehicles, respectively.
6.3.2 Vehicle routing problem

The transport network is composed of a set of stores S = {1..., s}, a set of producers P =
{1..., p}, a set of 3PL customers R = {1..., r} and the depots of LR and 3PL represented by
α and β, respectively. The set of all locations is given by V = α ∪ β ∪ S ∪ P ∪ R. Each store
i ∈ S and 3PL customer i ∈ R demands quantity qi to be delivered, whereas each producer
j ∈ P has quantity q j available to be sent to the LR. The subsets K LR = {1, ..., k LR } and
K 3PL = {k LR + 1, ..., k LR + k3PL } represent the limited number of dedicated vehicles of the
LR and the 3PL, respectively. The vehicles are homogeneous with capacity C, and the set
of all vehicles is given by K = K LR ∪ K 3PL . Each arc (i, j) ∈ V is associated to an Euclidean
distance di j = d ji . The transportation cost with the dedicated fleet is given by ci j for the arc
(i, j), whereas the unitary cost of outsourced vehicles is given by o. The 3PL can cross-
dock loads from the producers with an associated cost cd. The decision variables of the
problem are as follows:

1, if vehicle k from LR travels on arc (i, j)
xikj

:= 
0, otherwise ∀i, j ∈ V, ∀k ∈ K LR

1, if vehicle k from 3PL travels on arc (i, j)
zkij

:= 
0, ∀i, j ∈ V, ∀k ∈ K 3PL
otherwise

1, if producer i sends the load to depot j
∀i ∈ P, ∀ j = {α, β}

wi j := 
0, otherwise

1, if cross-docked load from producer i is carried by vehicle k
tik

:= 
0, otherwise ∀i ∈ P, ∀k ∈ K
yi j := load carried by a vehicle on arc (i, j), ∀i, j ∈ V.
6.3.2.1 Objective functions

In this work, three different objective functions are evaluated: i) minimization of oper-
ation costs, ii) minimization of fuel consumption, and iii) simultaneous minimization of
operation and CO2 emission costs.
The first objective function OF(C) describes the minimization of the total sum of distance-
based and cross-docking costs of the LR, 3PL and producers, as follows:
X
OF(C) = C LR + C3PL + Ci (6.1)
i∈P
where
XXX
C LR = xikj ci j (6.1.1)
i∈V j∈V k∈k
XXX X
C3PL = zkij ci j + wiβ cdi (6.1.2)
i∈V j∈V k∈k i∈P
Ci = wi j qi di j o, ∀i ∈ P, ∀ j ∈ {α, β} (6.1.3)
The second objective function OF(E) describes the minimization of fuel consumption,
which is dependent on the distance travelled and the load carried by all vehicles. We use
the Fuel Consumption Rate (FCR) introduced by Xiao et al. (2012) to determine the fuel
consumption of each entity in the transportation network. The FCR is influenced by the
rate of consumption of the empty vehicle (ρ), the rate of consumption of the full loaded
vehicle (ρ∗ ), the load in the vehicle on the arc (yi j ) and the capacity of the vehicle (C). The
objective function in (6.2) minimizes the total fuel consumption of all vehicles used by the
LR, the 3PL and the producers, as follows:
XXX
OF(E) = j + FCRi j + FCRi j di j
FCRiLR 3PL P
(6.2)
i∈V j∈V k∈K
where
ρ∗ − ρ
j = ρxi, j,k +
FCRiLR yi j xikj ∀i ∈ V, ∀ j ∈ V, ∀k ∈ K LR (6.2.1)
C
ρ∗ − ρ
FCR3PL
ij = ρzi, j,k + yi j zkij ∀i ∈ V, ∀ j ∈ V, ∀k ∈ K 3PL (6.2.2)
C
ρ∗ − ρ
FCRiPj = ρwi j + wi j qi ∀i ∈ P, ∀ j ∈ {α, β} (6.2.3)
C
The third objective function OF(CE) in (6.3) minimizes the economic and environmental
costs of the network, given by the sum of the costs of travelling, cross-docking and CO2
emissions. The parameter e represents the unitary cost of CO2 emissions and the parameter
η is used to convert the fuel consumption into the amount of emissions.
OF(CE) = OF(C) + e · η OF(E) (6.3)
Note that the right hand side of equations (6.2.1) and (6.2.2) is non-linear. To linearize
them, we first define the auxiliary variables Akij = yi j xikj and Bkij = yi j zkij , and then apply a
Big-M reformulation with M as a large number, as follows.
Akij ≤ Mxikj , ∀i ∈ V, ∀ j ∈ V, ∀k ∈ K (6.4.1)

Akij ≤ yi j , ∀i ∈ V, ∀ j ∈ V, ∀k ∈ K (6.4.2)
Akij ≥ yi j − (1 − xikj )M, ∀i ∈ V, ∀ j ∈ V, ∀k ∈ K (6.4.3)
Akij ≥ 0, ∀i ∈ V, ∀ j ∈ V, ∀k ∈ K (6.4.4)
Bkij ≤ Mzkij , ∀i ∈ V, ∀ j ∈ V, ∀k ∈ K (6.4.5)
Bkij ≤ yi j , ∀i ∈ V, ∀ j ∈ V, ∀k ∈ K (6.4.6)
Bkij ≥ yi j − (1 − zkij )M, ∀i ∈ V, ∀ j ∈ V, ∀k ∈ K (6.4.7)
Bkij ≥ 0, ∀i ∈ V, ∀ j ∈ V, ∀k ∈ K (6.4.8)
6.3.2.2 Constraints
In this section, we first present the constraints associated with the stores, followed sequen-
tially by those associated with the 3PL customers, producers, and vehicles. The last set
of constraints describe precedence requirements in the transportation network, which are
typical constraints of the Vehicle Routing Problem with Backhauls (VRPB).
Constraints (6.5) impose that each store is visited exactly once. The flow conservation is
defined by constraints (6.6). Constraints (6.7) and (6.8) force the vehicles of the LR to
depart from its depot. Constraints (6.9) ensures that the total total demand of the stores
is satisfied, and constraints (6.10) impose that the load carried by each vehicle decreases
along the visits to the stores.
XX
xikj + zkij = 1, ∀j ∈ S (6.5)
i∈V k∈K
X X
xikj = xkjl , ∀ j ∈ V, ∀k ∈ K (6.6)
i∈V l∈V
X
k
xαi ≤ 1, ∀k ∈ K (6.7)
i∈S
X
k
xαi ≥ xkjl , ∀ j ∈ V, ∀l ∈ V, ∀k ∈ K (6.8)
i∈S
X X
yα j = qj (6.9)
j∈S j∈S
X X
yi j = y ji + q j , ∀j ∈ S (6.10)
i∈V i∈V
Constraints (6.11) and (6.12) impose that each 3PL customer is visited exactly once by
vehicles of the 3PL only. The flow conservation is ensured by Constraints (6.13). Con-
straints (6.14) and (6.15) force the vehicles of the 3PL to depart from its depot. Constraint
(6.16) ensures that vehicles departing from the depot of the 3PL carry all loads necessary
to attend the demand of its customers, as well as the cross-docked loads from producers.
Constraints (6.17) impose that the load carried by a vehicle decreases along the visits to
3PL customers.
XX
zkij = 1, ∀j ∈ R (6.11)
i∈V k∈K
XX
xikj = 0, ∀j ∈ R (6.12)
i∈V k∈K
X X
zkij = zkjl , ∀ j ∈ V, ∀k ∈ K (6.13)
i∈V l∈V
X
zkβ j ≤ 1, ∀k ∈ K (6.14)
j∈R
X
zkβi ≥ zkjl , ∀ j ∈ V, ∀l ∈ V, ∀k ∈ K (6.15)
i∈R
X X X XX
yβ j = qi + qi wiβ + k
qi xiβ , (6.16)
j∈V i∈R i∈P i∈P k∈K
X X
yi j ≥ y ji + q j , ∀j ∈ R (6.17)
i∈V i∈V
Each producer can either send loads to the depot of the LR or to the depot of the 3PL, or
wait for the backhauling of the LR or the 3PL, which is guaranteed by Constraints (6.18).
Constraint (6.19) guarantees that the total load arriving at the depot of the LR corresponds,
at least, to the total load sent directly by the producers and the total load collected in
backhaul routes. If a producer sends loads to the depot of the 3PL, either vehicles of
the LR pickup them in backhaul trips or the 3PL send them in the delivery routes, as
enforced by Constraints (6.20). However, if the 3PL performs pickups at producers, the
LR must collect these at the depot of the 3PL, which is enforced by Constraints (6.21).
Furthermore, Constraints (6.22) ensures that vehicles of the LR only visit the depot of 3PL
if there are cross-docked loads. Also, mixed visits between 3PL customers and the depot of
LR are only allowed if there are cross-docked loads, as guaranteed by Constraints (6.23).
Constraints (6.24) ensures that the load in the vehicle increases while collecting loads at
the producers. The pair of Constraints (6.25) and (6.26) ensure that cross-docked loads
are not split when consolidated in backhaul routes for the LR or in delivery routes to 3PL
customers.
XX
wiα + wiβ + xkji + zkji = 1, ∀i ∈ P (6.18)
j∈V k∈K
X X X
yiα + qi wiα ≥ qi , (6.19)
i∈V i∈P i∈P
 
X  X X 
wiβ ≤  xk + xβk j + zkβα + zkjα  , ∀i ∈ P (6.20)
 βα
k∈K j∈P j∈R
 
X X  X 
zkiβ ≤  xk + xβk j  , ∀i ∈ P (6.21)
 βα
k∈K k∈K j∈P
XX X
zkiβ + zkij + wiβ ≥ xβk j . ∀k ∈ K (6.22)
i∈P j∈P j∈V
X X
zkαi ≤ wiβ , ∀k ∈ K (6.23)
i∈R i∈P
X X XX
y ji = yi j + q j xikj + zkij , ∀j ∈ P (6.24)
i∈V i∈V i∈V k∈K
X
Akβα = tik qi , ∀k ∈ K (6.25)
i∈P
X XX X
Bkβ j = zkij qi + tik , ∀k ∈ K (6.26)
j∈V i∈R j∈V i∈P
Constraints (6.27) ensure that the same vehicle is either used by the LR or the 3PL, but not
both. Constraints (6.28) ensure that the capacity of the vehicles is never exceeded in any
arc of the routes.
X
k
xαi + zkφi ≤ 1, ∀k ∈ K (6.27)
i∈V
X
yi j ≤ xikj + zkij C, ∀i, j ∈ V (6.28)
k∈K
The precedence of deliveries before pickups is ensured by Constraints (6.29) for LR ve-
hicles, and by Constraints (6.30) for 3PL vehicles. Additional precedence constraints are
necessary for the vehicles of the 3PL to forbid mixed visits between stores, 3PL customers
and suppliers, ensured by Constraints (6.31) to (6.33).
xkji + xβi
k
= 0, ∀i ∈ S , j ∈ P, k ∈ K (6.29)
zkji + zkjα = 0, ∀i ∈ S , j ∈ P, k ∈ K (6.30)
zkiα + zkβi = 0, ∀i ∈ S , k ∈ K (6.31)
zkij + zkji = 0, ∀i ∈ S , j ∈ R, k ∈ K (6.32)
zkij = 0, ∀i ∈ P, j ∈ R, k ∈ K (6.33)
6.3.3 Profit sharing problem

After solving both the stand alone problem (Appendix 6.A) and the collaborative problem
(described in Section 6.3.2), the global gains provided by the collaboration must be shared
among the participants. Due to the simplicity of application and practical use in real con-
texts, we investigate proportional allocation methods to solve the profit sharing problem.
Three allocation methods are collected from Lunday and Robbins (2019). The first alloca-
tion method (M1) distributes the collaborative gains equally among the participants. The
second allocation method (M2) first applies the principle of M1, and then reimburses the
potential losses of participants. The third allocation method (M3) follows an inverse ra-
tionale of that of M2, first reimbursing the losses and only after distributing the remaining
gains equally.
The following numerical example demonstrates the principles of each proportional method.
Consider that 4 participants are enrolled in the collaboration - the LR, the 3PL and two
producers (P1 and P2). Table 6.1 presents the individual cost of each participant, s0 , in
the stand alone solution and its difference, ∆∗ , compared with the collaborative solution. It
also shows the allocated share, a, of each participant obtained with each allocation method,
along with the total cost after receiving the allocated gain, s1 . The total gains is given by
the sum of the positive values only, i.e. 40 + 8 + 14 = 62.
In this example, it is possible to observe that an equal distribution of gains is not a good
approach to all participants. The total cost of the 3PL would be higher than its stand
alone cost, which puts in evidence that the criteria of individual rationality is not met if the
allocation method M1 applies. Using the method M2, the gains allocated to the 3PL would
only be sufficient to cover the additional costs and, as such, this participant would not
Table 6.1 – Numerical example

M1 M2 M3
s0 ∆∗ a s1 a s1 a s1
LR 100 40 15.5 84.5 14.0 86.0 10.5 89.5
3PL 80 -20 15.5 84.5 20.0 80.0 30.5 69.5
P1 20 8 15.5 4.5 14.0 6.0 10.5 9.5
P2 14 14 15.5 -1.5 14.0 0.0 10.5 3.5
receive any benefit compared with the stand alone solution. For this reason, the 3PL would
not be interested in participating in the collaboration. The method M3 allocates the gains
among participants such that the individual rationality is met, ensuring that all participants
are allocated with a total cost lower than its stand alone cost. For these reasons, the method
M3 dominates the other two methods in terms of fairness, which is also supported by the
results of Lunday and Robbins (2019).

The computational experiments performed in this work have two main focus. The first is
to analyse the performance of the collaboration under different objectives (economic and
environmental) for various scenarios. The second research focus is to compare different
proportional allocation methods to solve the profit sharing problem based on some fairness
criteria. The ultimate goal of the computational experiments is to characterize the potential
of collaboration between the different participants in the case-study, which will allow to
draw a set of managerial insights.
The models are coded in Python 3.6.3 and solved with Gurobi on a computer equipped with
the processor Intel Core i7 and 16GB of RAM. The computational experiments are applied
to randomly generated instances from Solomon (1987) and to a case study instance.
6.4.1 Instances
The randomly generated instances depart from the data sets of 25 locations in Solomon
(1987), respectively for a cluster (C), random (R) and random cluster (RC) network. For
each type of network, 10 instances are created with 15 random locations and 2 fixed loca-
tions. The former represents the total number of stores, 3PL customers and producers, and
the latter represents the depots of the LR and the 3PL. These instances are then divided in
two 5-instance data sets designated as s8-r5-p2 (8 stores, 5 3PL customers and 2 produc-
ers) and s10-r3-p2 (10 stores, 3 3PL customers and 2 producers), respectively. In total, 30
different random instances are created. The demand of each store and 3PL customer are
randomly generated from an uniform distribution [2,5]. The ordering load from the LR to
each producer is set at 5. The number of dedicated vehicles available is 5, with a capacity
of 15. In the case study instance, the network is built based on real locations and distances
taken from the real context of application, which is detailed further in this section.
The remaining inputs of the models are considered in both randomly generated and case
study instances. The economic parameters are taken from the real case study, as follows.
The travelling cost of a dedicated vehicle is 0.8e per unit of distance. The cost of an
outsourced vehicle is 0.09e per unit of distance and unit of load. The cross-docking cost is
1.3e per unit of load. The parameters necessary to determine the FCR are set according to
Cheng et al. (2018), namely ρ∗ = 0.296 and ρ = 0.39. The parameters associated with the
CO2 emissions follow the convention that one liter of fuel generates 2.32 kg of CO2 per
liter of fuel and the cost of CO2 emissions is 37.2e per ton (Bektaş and Laporte, 2011).
6.4.2 Analysis of collaborative solutions

Each instance was tested with the three different objective functions, for both the collab-
orative and the non-collaborative models. The average results obtained for data sets s8-
r5-p2 and s10-r3-p2 are presented in Tables 6.2 and 6.3, respectively. Each one reports
the average operational cost and average fuel consumption of the stand alone (s0 ) and the
collaborative (s1 ) global solutions, the standard deviations (sd) of such solutions and the
synergy values (sv) achieved in terms of operational costs and fuel consumption.
The results demonstrate that the synergy value is influenced by the type of network. In
particular, type RC leads to the highest synergy values both in terms of cost and fuel con-
sumption, while type R leads to the lowest. Furthermore, these results are consistent among
the different objective functions and across the different data sets. The leverage of RC net-
works comes from the fact that customers and producers are more distant to the depots
than in the other types of network, which leads to higher total costs and fuel consumption
in the stand alone solutions. However, in full collaboration, the cost and fuel reductions are
usually higher for the RC networks than networks C and R, and consequently, the synergy
values achieved are also higher. It is also expected that full collaboration leads to higher
synergy values than the cases of limited collaboration (e.g., when only two strategies can
be applied). Nonetheless, limiting the type of strategies that can be selected may reduce
the leverage of RC networks, as documented in Appendix 6.B.
The holistic model provides very similar solutions than those obtained with the objective
function OF(C), either for the stand alone and the collaborative solutions. This occurs
because the cost of CO2 emissions is much smaller than the operational costs. The results
show that using the objective OF(CE) allows to modify the routing plan in order to reduce
fuel consumption, only as long as the cost of the economic solution does not increase. Also,
as expected, solutions obtained with the environmental objective OF(E) tend to prioritize
fuel reduction in detriment of the costs. Nevertheless, the relative difference between costs
obtained with OF(E) and OF(C) are very small.
A closer look to the collaborative solutions reported in Tables 6.2 and 6.3 allows to identify
and compare the strategies that promote deeper collaboration, under different contexts.
Following this reasoning, Tables 6.4 and 6.5 present the number of times each collaborative
strategy was selected in each 5-instance data set in the three types of network, and for each
objective function.
The collaborative pickup-delivery, which allows the 3PL to visit RL customers, seems to
influence heavily the collaboration, as it is presented in almost all solutions independent on
the objective function and type of network considered. The collaborative collection, which
Table 6.2 – Average results obtained with data set s8-r5-p2

Network type C Network type R Network type RC
Cost sd Fuel sd Cost sd Fuel sd Cost sd Fuel sd
s0 243.6 26.8 108.0 11.1 320.2 28.3 141.1 11.0 401.1 13.6 178.2 5.9
OF(C) s1 214.7 20.8 93.4 6.8 293.8 21.2 129.2 10.2 354.1 21.5 151.7 12.4
sv 12% 3% 13% 4% 8% 2% 8% 2% 12% 3% 15% 5%
s0 244.3 27.4 107.0 11.3 321.1 28.2 138.1 10.9 401.6 13.0 174.4 6.5
OF(E) s1 214.9 20.7 93.3 6.8 298.0 21.7 125.7 8.6 357.9 25.7 150.7 11.5
sv 12% 3% 12% 4% 7% 2% 9% 2% 11% 5% 14% 4%
s0 243.6 26.8 107.1 11.2 320.2 28.3 138.3 10.9 401.1 13.6 174.9 5.7
OF(CE) s1 214.7 20.8 93.3 6.9 293.8 21.2 128.4 10.4 354.1 21.5 151.5 12.3
sv 12% 3% 13% 4% 8% 2% 7% 1% 12% 3% 13% 5%
Table 6.3 – Average results obtained with data set s10-r3-p2

Cost sd Fuel sd Cost sd Fuel sd Cost sd Fuel sd
s0 229.4 31.6 101.4 14.4 329.7 24.4 143.6 11.2 390.8 21.5 172.1 9.6
OF(C) s1 205.8 20.9 87.9 8.4 300.8 19.2 129.6 9.1 341.5 24.0 144.4 10.3
sv 10% 4% 13% 5% 9% 1% 10% 1% 13% 2% 16% 2%
s0 229.7 31.8 100.7 14.1 329.9 24.4 142.1 10.3 391.2 22.3 169.6 7.0
OF(E) s1 207.4 21.6 87.3 8.3 305.1 21.5 127.8 8.2 344.7 28.0 142.7 9.2
sv 9% 5% 13% 4% 7% 1% 10% 1% 12% 2% 16% 2%
s0 229.4 31.6 100.8 14.2 329.7 24.4 142.2 10.3 390.8 21.5 170.1 8.0
OF(CE) s1 205.8 20.9 87.8 8.4 300.8 19.2 129.4 9.1 341.5 24.0 144.3 10.3
sv 10% 4% 12% 5% 9% 1% 9% 1% 13% 2% 15% 3%
allows dedicated vehicles to perform pickups at producers, is presented mostly in the solu-
tions obtained with the fuel minimization model. These results indicate that collaborative
collection is effective in reducing fuel consumption, but possibly at a relatively high cost.
The collaborative cross-docking is never presented in any solution. A possible obstacle
to implement the collaborative cross-docking may be the high cost of cross-docking com-
pared with the travelling cost of an outsourced vehicle, for small loads, or the relatively
high distance between producers and 3PL depot.
The results in Tables 6.4 and 6.5 also show that the network type RC creates the solutions
with the higher total number of collaborative strategies, followed by the network type C and
then type R. Based on these results, we can argue that, in general, the more strategies are
implemented, the higher is the impact on collaboration, and thus the higher is the synergy
value. Note that in RC networks, the number of collaborative collections is higher than
in networks C and R. This shows that collaborative collections are leveraged by RC net-
works. However, if this strategy is not allowed, the advantage of RC networks is reduced
comparing to networks C and R, as reported in Appendix 6.B.
Table 6.4 – Number of collaborative strategies used in solutions from the data set s8-r5-p2
OF (C) (E) (CE) (C) (E) (CE) (C) (E) (CE) Total
Collection 4 4 4 1 6 1 6 8 6 40
Pickup-delivery 5 5 5 5 5 5 5 5 5 45
Cross-docking 0 0 0 0 0 0 0 0 0 0
Total 9 9 9 6 11 6 11 13 11
Table 6.5 – Number of collaborative strategies used in solutions from the data set s10-r3-p2
OF (C) (E) (CE) (C) (E) (CE) (C) (E) (CE) Total
Collection 7 9 7 4 8 4 8 10 8 65
Pickup-delivery 5 4 5 5 5 5 5 5 5 44
Cross-docking 0 0 0 0 0 0 0 0 0 0
Total 12 13 12 9 13 9 13 15 13
6.4.3 Sensitivity analysis

As evidenced by the previous results, the collaborative pickup-delivery is adopted in almost
all solutions, but the collaborative cross-docking was never selected. Thus, with the goal of
exploring situations that may benefit cross-docking and determine the respective synergy
value, we perform several sensitivity analysis to different parameters involved in the routing
problem. These parameters comprise the geographic dispersion of the producers around
the 3PL depot, the quantity ordered to producers, the cost of cross-docking and the cost of
outsourcing.
The differences between the base scenario and the scenarios used in the sensitivity analysis
are provided in Table 6.6. The first scenario considers that in each 5-instance data set the
locations of the different producers are close to the depot of the 3PL. The second scenario
considers that the LR orders to each producer a quantity that matches the capacity of the
dedicated vehicles. The third scenario assumes that the cost of cross-docking is halved.
The last scenario assumes that the outsourcing cost is 0.3 per unit of load and unit of
distance, which is a common price in practice charged by outsourced carriers to transport
small loads.
Table 6.6 – Parameter settings of each scenario
Base scenario Scenario 1 Scenario 2 Scenario 3 Scenario 4
Location of producer p ∈ P random fixed random random random
Ordering quantity (q p ) 5 5 15 5 5
Cost of cross-docking (cd) 1.3 1.3 1.3 0.65 1.3
Cost of outsourcing (o) 0.09 0.09 0.09 0.09 0.3
The results from the sensitivity analysis to each scenario are presented in Table 6.7, which
provides the number of collaborative strategies implemented and the synergy values in
terms of operational costs and fuel consumption. As the solutions obtained with the holis-
tic model are the same as those obtained with the economic model, we only present the
sensitivity analysis for the objectives OF(C) and OF(E). Furthermore, as similar findings
are obtained for both data sets, this section only presents the results for the data set s10-
r3-p2. Finally, the results obtained with the scenarios 3 and 4 are the same as the ones
obtained in the base scenario, for the respective objective function. As such, Table 6.7 only
reports evidences from scenarios 1 and 2.
If the economic model is used in scenario 1, the number of strategies selected, particularly
collaborative collections, is reduced comparing to the base scenario. On the other hand,
using the environmental model tends to promote both collaborative collections and cross-
docking, but leads to lower synergy values.
In scenario 2, there is a general trend to promote collaborative collections with all producers
if the economic model is used. Otherwise, the number and type of collaborative strategies
implemented is the same as in the base scenario. Nonetheless, in both cases, the synergy
values tend to increase, as a result of the larger reduction of costs and fuel consumption at
the producers.
In order to further explore potentially beneficial situations for collaborative cross-docking,
sensitivity analysis to combined scenarios are also performed, namely by combining sce-
nario 1 with one of the other scenarios. For this purpose, only the economic model is used,
since the solutions obtained with the environmental model, in terms of collaborative strate-
gies, are not affected by the cost parameters nor the ordering quantity, as demonstrated
previously. The results from the sensitivity analysis to combined scenarios is presented
in Table 6.8. It can be observed that collaborative cross-docking is heavily influenced by
the costs parameters when the producers are located close to the 3PL depot. When the
cost of cross-docking reduces to half (combined scenarios 1 + 3), the number of collab-
orative cross-docking strategies increases, independently on the number of collaborative
Table 6.7 – Sensitivity analysis to different scenarios and objective functions

Base scenario Scenario 1 Scenario 2
C R RC C R RC C R RC
Collection 7 4 8 0 2 1 10 10 10
Pickup-delivery 5 5 5 5 5 5 3 5 5
OF(C) Cross-docking 0 0 0 0 0 0 0 0 0
svcost 10% 9% 13% 6% 8% 8% 21% 17% 25%
sv f uel 13% 10% 16% 6% 8% 8% 12% 10% 17%
Collection 9 8 10 5 5 5 9 8 10
OF(E) Cross-docking 0 0 0 5 5 5 0 0 0
svcost 9% 7% 12% 3% 7% 6% 21% 16% 25%
sv f uel 13% 10% 16% 12% 12% 9% 12% 10% 16%
collections previously established in scenario 1, except for network type C. Increasing the
outsourcing cost (combined scenario 1 + 4) leads to the highest total number of collabora-
tive strategies adopted. In particular, cross-docking is used in the majority of the solutions
obtained for all types of network. Furthermore, the synergy values of both costs and fuel
consumption increase substantially comparing to all other scenarios.
Table 6.8 – Sensitivity analysis to combined scenarios

Scenarios 1 + 2 Scenarios 1 + 3 Scenario 1 + 4
Collection 10 9 7 0 4 1 5 5 6
Cross-docking 0 0 0 0 4 1 5 5 4
svcost 18% 16% 11% 6% 8% 8% 24% 20% 13%
sv f uel 5% 8% 8% 6% 12% 9% 13% 13% 10%
6.4.4 Performance of allocation methods

In this section, we analyze the impact of each allocation method on the cost reduction
(gains) of each participant in the collaboration. In particular, we evaluate the performance
of the different proportional allocation methods and compare the individual gains achieved
by all entities in different collaborative solutions.
Table 6.9 presents the results obtained with each allocation method for different solutions,
namely the total allocated cost of each participant after the profit distribution (s1 ) and the
percentage cost reduction (g) in comparison with its stand alone solution (s0 ). Five differ-
ent collaborative solutions obtained with data set s10-r3-p2 are selected for this analysis:
a) a solution with only collaborative pickup-delivery (from base scenario, network type
R), b) a solution with only collaborative collection (from base scenario, network type C),
c) a solution with both collaborative pickup-delivery and collection (from base scenario,
network type RC), d) a solution with both collaborative collection and cross-docking and
e) a solution with all three collaborative strategies (both from combined scenarios 1 + 4,
network type C).
Table 6.9 – Percentage gains of each entity obtained with each allocation method, for dif-
ferent collaborative solutions
M1 M2 M3
s0 s1 g (%) s1 g (%) s1 g (%)
a) Only LR 212 177 17% 180 15% 196 7%
pickup-delivery 3PL 97 101 -4% 97 0% 81 16%
b) Only LR 122 125 -3% 122 0% 119 2%
collection P1 8 2 70% 4 51% 5 34%
P2 9 3 63% 5 46% 6 31%
c) Pickup-delivery LR 214 190 11% 196 8% 201 6%
and collection 3PL 142 160 -13% 142 0% 128 10%
P1 15 -9 160% -3 120% 1 90%
P2 18 -6 134% 0 101% 4 76%
d) Collection and LR 124 126 -1% 124 0% 108 13%
cross-docking 3PL 68 50 26% 51 25% 52 24%
P1 50 26 48% 27 47% 34 32%
P2 55 30 45% 31 43% 38 30%
e) Pickup-delivery, LR 138 110 20% 111 20% 118 15%
collection and 3PL 90 92 -3% 90 0% 69 23%
cross-docking P1 50 22 56% 23 55% 30 41%
P2 55 26 52% 27 50% 34 38%
The results in Table 6.9 put in evidence that both allocation methods M1 and M2 are not
suitable to benefit all participants, for any of the solutions presented. Method M1 always
lead to losses for the 3PL, while method M2 only compensates these losses. Both methods
would encourage the 3PL to leave the collaboration. In opposition, as expected, method
M3 provides an allocated cost that fulfils the fairness properties of efficiency and individual
rationality.
Another evidence of these results in the base scenario is the large percentage gains of the
producers when compared with the RL and the 3PL. This occurs because the costs of the
producers, even all together, represent a small portion of the total costs of the network
(e.g., solutions b) and c)). However, as the costs of producers increase, the individual
percentage gains seem to be more balanced between the different entities participating in
the collaboration (e.g., solutions d) and e)).
6.4.5 Case-study
In this section, we solve and analyze the routing and the profit sharing problems for a case-
study instance. The instance presents a transportation network for a typical day of the week
for both LR and 3PL. The network is composed of 20 stores (numbered from 0 to 20), six
3PL customers (numbered from 21 to 26) and two producers (numbered 27 and 28). The
demand from stores range between 2 and 10 pallets and from 3PL customers range between
1 and 18. Each dedicated vehicle has a standard capacity of 33. Each producer has a load
of 10 pallets to send to the LR.
Table 6.10 presents the set of routes created with the stand alone and the collaborative
solutions, for different objective functions. The operational cost and fuel consumption of
each solution, as well as the synergy of collaboration, are presented in Table 6.11. One first
evidence of these results is that, independently of the objective function used, only three
vehicles are necessary to perform the transportation services in a collaborative network,
whereas the stand alone solutions always require five vehicles (three for the deliveries to
stores and the other two for the deliveries to 3PL customers). As expected, the routes
created with the holistic model share many similarities with the routes created with the
economic objective function. In both cases, the collaborative solution implements the three
collaborative strategies.
Table 6.10 also shows that the collaborative solution obtained with the environmental ob-
jective function does not implement collections, which contrasts with the results obtained
with the randomly generated instances. In the case study instance, the producers are lo-
cated much closer to the depot of the 3PL than the depot of the RL and the stores, similarly
to scenario 1. From the sensitivity analysis, it was determined that, in fact, solutions from
scenario 1 generate lower number of collaborative collections when compared with the
base scenario. However, these are compensated by an increase of cross-docking strategies,
which is not verified in the case study instance.
In addition, Table 6.11 shows that high synergy values can be obtained with collaboration.
In particular, collaboration can reduce the operational costs in 28% and the fuel consump-
tion in, at least, 26%, independently of the objective function used to model the routing
problem. Moreover, the results suggest that an economic objective function is sufficient to
assess the collaborative potential of the network, instead of the holistic model.
Table 6.10 – Routes created in stand alone and collaborative solutions for each objective
function
OF Stand alone solution (s0 ) Collaborative solution (s1 )
LR - 1 - 18 - 19 - 16 - LR LR - 6 - 13 - 11 - 4 - 3PL - 27 - LR
LR - 6 - 20 - 9 - 10 - 17 - 14 - 7 - 8 - 3 - 12 - LR
(C) LR - 13 - 11 - 4 - 5 - 15 - 2 - LR
3PL - 26 - 21 - 25 - 23 - 3PL 3PL - 26 - 21 - 23 - LR - 1 - 18 - 14 - 17 - 10 - 7 - 9 - 20 - 5 - 3PL
3PL - 24 - 22 - 3PL 3PL - 22 - 24 - 25 - LR - 16 - 12 - 19 - 3 - 8 - 2 - 15 - 3PL
LR - 2 - 15 - 13 - 5 - 4 - 11 - LR LR - 16 - 19 - 1 - 18 - LR
LR - 12 - 3 - 8 - 14 - 17 - 10 - 7 - 9 - 20 - 6 - LR
(E) LR - 16 - 19 - 1 - 18 - LR
3PL - 26 - 21 - 23 - 25 - 3PL 3PL - 26 - 21 - 23 - LR - 6 -2 - 13 - 5 - 4 - 11 - 3PL
3PL - 24 - 22 - 3PL 3PL - 22 - 24 - 25 - LR - 12 - 3 - 8 - 14 - 17 - 10 - 7 - 9 - 20 - 15 - 3PL
LR - 1 - 18 - 19 - 16 - LR LR - 16 - 19 - 3 - 8- 7 -2 - 15 - 3PL - 27 - LR
LR - 2 - 13 - 11 - 4 - 5 - 15 - LR
(CE) LR - 6 - 20 - 9 - 10 - 17 - 14 - 7 - 8 - 3 - 12 - LR
3PL - 26 - 21 - 23 - 25 - 3PL 3PL - 26 - 21 - 23 - LR - 1 - 18 - 14 - 17 - 10 - 7 - 9 - 20 - 5 - 3PL
3PL - 24 - 22 - 3PL 3PL - 22 - 24 - 25 - LR - 12 - 6 - 13 - 11 - 4 - 3PL
The results obtained for the profit sharing problem with the three allocation methods are
presented in Table 6.12. As the solution obtained with the environmental objective func-
tion only considers collaboration between the LR and the 3PL through pickup-delivery
strategies, the producers do not receive any profit. Thus, since the profits are only shared
between the two, the cost reduction is much more pronounced than in the case when all
Table 6.11 – Synergy values of collaborative solutions obtained with each objective func-
tion
(C) (E) (CE)
Cost Fuel Cost Fuel Cost Fuel
s0 1895 780 1900 760 1895 772
s1 1364 570 1367 562 1364 570
sv 28% 27% 28% 26% 28% 26%
entities participate in the collaboration. In fact, if the producers also participate in the col-
laboration, the allocated share of profits is reduced by half for the LR and the 3PL. This
may raise some concerns among the participants in collaboration. On the one hand, the LR
and the 3PL may not be willing to give up half of the profits, and on the other, the producers
are certainly open to collaborate because it can only reduce their costs.
Therefore, one possible way to overcome this limitation may be to define a cap for the share
profit of the producers based on the carried and consolidated load quantities. For example,
a new rule of the allocation method could ensure that a producer would not receive a profit
share higher than the share it would receive by providing full truck loads. In order to
explore this setting, we perform an additional set of experiments in the case study instance
considering that the loads at each supplier match the capacity of the dedicated vehicles (i.e.
qi∈P = C = 33). The results obtained with each allocation method are reported in Table 6.13.
It is possible to observe that producers providing full truck loads receive a profit share such
that it does not impact the profit shares of the LR and the 3PL, if the allocation method
M3 is applied. In fact, the cost reduction of the LR and the 3PL is kept at 33% and 35%
respectively, with or without collaborating producers, and the producers can reduce their
costs in 61% and 44% respectively if they enter the collaboration.
Based on these results, we can argue that limiting the profit share of participants that have
relatively low contribution to the total costs (and/or environmental impact) of the network,
such as the producers, may be a good rule to promote the willingness to enter the collabo-
ration. Thus, combining this rule with method M3 seems to be a suitable strategy to handle
the limitation of overcompensating some participants while attending desirable fairness
criteria.
Table 6.12 – Allocated costs and percentage gains obtained with each allocation method
M1 M2 M3
s0 s1 g s1 g s1 g
OF(C) LR 806 630 22% 630 22% 674 16%
3PL 764 762 0% 762 0% 631 17%
P1 136 -41 130% -41 130% 3 98%
P2 189 12 91% 12 93% 56 70%
OF(E) LR 811 448 45% 448 45% 545 33%
3PL 764 595 22% 595 22% 498 35%
Table 6.13 – Performance of the allocation methods for the case where producers provide
full truck loads
M1 M2 M3
s0 s1 g s1 g s1 g
LR 806 531 37% 531 37% 562 33%
3PL 764 605 23% 605 23% 512 35%
OF(C) P1 448 144 68% 144 68% 175 61%
P2 624 321 49% 321 49% 352 44%
LR 811 448 45% 448 45% 545 33%
OF(E) 3PL 764 595 22% 595 22% 498 35%

This work investigates a case of lateral collaboration for the transportation planning prob-
lem of a network involving a LR, a 3PL and several producers. Three different strategies
can be applied in the collaborative network, namely pickup-delivery, collection and cross-
docking. The collaborative vehicle routing problem is formulated based on the Vehicle
Routing Problem with Backhauls (VRPB) under three different objective functions: min-
imization of operational costs, minimization of fuel consumption, and minimization of
operational and CO2 emissions costs. Three different proportional methods are applied to
solve the profit sharing problem. Method M1 distributes the collaborative gains equally
among the participants. Method M2 first distributes the gains equally and then reimburses
participants for their possible losses. Method M3 first reimburses the participants with
losses and the remaining gains are then distributed equally among the participants.
The results of this work indicate that the collaborative pickup-delivery is the most im-
plemented strategy in collaboration and tends to be independent of the objective function
used and type of network. The collaborative cross-docking is favoured when the producers
are located closer to the depot of the 3PL than the depot of the LR, particularly when an
economic objective function drives the optimization problem. On the other hand, the col-
laborative collection is favoured when producers are dispersed in the network (in particular
when using an environmental objective function) and the ordering quantities are relatively
high. The type of network RC tends to lead to higher synergy values than network types C
and R because it is more prone to incorporate all types of collaborative strategies in a single
solution. However, the leverage of RC may be reduced if collaborative backhauling is not
allowed. Another insight of this work is that using a holistic model that minimizes both
the operational and CO2 emissions costs leads to similar solutions obtained with the eco-
nomic model. Thus, it is possible to examine both the economic and environmental impact
using only a traditional economic model, and thus avoiding the increased complexity of
modelling the fuel consumption of vehicles in the transportation planning problem. For the
case study instance, the synergy value is relatively high both in terms of costs and environ-
mental impact. In particular, the results show synergies over 26% for both objective terms.
Moreover, among the methods used in the profit sharing problem, only method M3 is able
to meet both fairness properties of efficiency and individual rationality. Nonetheless, it may
Bibliography 207
be necessary to adjust the allocated shares of participants that provide a low contribution
to the collaboration, such that participants with higher contributions do not perceive to be
more harmed if entering the grand coalition than a sub-coalition.
This work contributes to the literature by providing a study of the collaborative vehicle
routing under a sustainable context, and by exploring the impact of different collaborative
strategies on both the routing and the profit sharing problems. It also contributes to practice
since it provides valuable managerial insights based on the case study investigated and
sensitivity analysis performed. It would be worth to integrate the routing and the profit
sharing problems, generating solutions that comprise the routes to execute and the profit
allocation to each participants after profit distribution. However, this suggestion for future
work is challenging from a modelling and solution approach perspective because it adds
the complexity of the profit sharing problem to an already complex routing problem.
Acknowledgements
This work was conducted under the project “MobFood – Mobilizing scientific and techno-
logical knowledge in response to the challenges of the agri-food market” (POCI-01-0247-
FEDER-024524), by “MobFood” Consortium, and financed by European Regional Devel-
opment Fund (ERDF), through the Incentive System to Research and Technological de-
velopment, within the Portugal2020 Competitiveness and Internationalization Operational
Program.
Bibliography
Audy, J.-F., D’Amours, S., and Rönnqvist, M. (2012). An empirical study on coalition
formation and cost/savings allocation. International Journal of Production Economics,
136(1):13 – 27.
Bektaş, T. and Laporte, G. (2011). The pollution-routing problem. Transportation Re-

search Part B: Methodological, 45(8):1232 – 1250. Supply chain disruption and risk
management.
Chabot, T., Bouchard, F., Legault-Michaud, A., Renaud, J., and Coelho, L. C. (2018). Ser-
vice level, cost and environmental optimization of collaborative transportation. Trans-
portation Research Part E: Logistics and Transportation Review, 110:1 – 14.
Cheng, C., Qi, M., and Rousseau, L.-M. (2018). Fuel consumption optimization model for
the multi-period inventory routing problem. Transportation Research Record Journal of
the Transportation Research Board.
Cruijssen, F., Braysy, O., Dullaert, W., Fleuren, H., and Salomon, M. (2007). Joint route
planning under varying market conditions. International Journal of Physical Distribu-
tion and Logistics Management, 37(4):287–304. Appeared earlier as CentER DP 2004-
80 (rt).
Cuervo, D. P., Vanovermeire, C., and Sörensen, K. (2016). Determining collaborative

profits in coalitions formed by two partners with varying characteristics. Transportation
Research Part C: Emerging Technologies, 70:171 – 184.
Dahlberg, J., Engevall, S., Göthe-Lundgren, M., Jörnsten, K., and Rönnqvist, M. (2019).
Incitements for transportation collaboration by cost allocation. Central European Jour-
nal of Operations Research, 27:1009–1032.
Dai, B. and Chen, H. (2012). Profit allocation mechanisms for carrier collaboration in
pickup and delivery service. Computers Industrial Engineering, 62(2):633 – 643.
Ergun, , Kuyzu, G., and Savelsbergh, M. (2007). Shipper collaboration. Computers Op-
erations Research, 34(6):1551 – 1560. Part Special Issue: Odysseus 2003 Second Inter-
national Workshop on Freight Transportation Logistics.
Evangelista, P. (2014). Environmental sustainability practices in the transport and logistics

service industry: An exploratory case study investigation. Research in Transportation
Business Management, 12:63 – 72. Sustainable Freight Transport.
Fernández, E., Fontana, D., and Speranza, M. G. (2016). On the collaboration uncapaci-
tated arc routing problem. Computers Operations Research, 67:120 – 131.
Frisk, M., Göthe-Lundgren, M., Jörnsten, K., and Rönnqvist, M. (2010). Cost alloca-
tion in collaborative forest transportation. European Journal of Operational Research,
205(2):448 – 458.

Guajardo, M. and Rönnqvist, M. (2016). A review on cost allocation methods in collabora-

tive transportation. International Transactions in Operational Research, 23(3):371–392.
Juan, A. A., Faulin, J., Pérez-Bernabeu, E., and Jozefowiez, N. (2014). Horizontal cooper-
ation in vehicle routing problems with backhauling and environmental criteria. Procedia
- Social and Behavioral Sciences, 111:1133 – 1141. Transportation: Can we do more
with less resources? – 16th Meeting of the Euro Working Group on Transportation –
Porto 2013.
Krajewska, M. A., Kopfer, H., Laporte, G., Ropke, S., and Zaccour, G. (2008). Horizontal
cooperation among freight carriers: request allocation and profit sharing. Journal of the
Operational Research Society, 59(11):1483–1491.
le Blanc, H., Cruijssen, F., Fleuren, H., and de Koster, M. (2006). Factory gate pricing:
An analysis of the dutch retail distribution. European Journal of Operational Research,
174(3):1950 – 1967.
Liu, G., Hu, J., Yang, Y., Xia, S., and Lim, M. K. (2020). Vehicle routing problem in cold
chain logistics: A joint distribution model with carbon trading mechanisms. Resources,
Conservation and Recycling, 156:104715.
Bibliography 209
Liu, P. (2010). Allocating collaborative profit in less-than-truckload carrier alliance. Jour-

nal of Service Science and Management - J Serv Sci Manag, 03:143–149.
Lunday, B. J. and Robbins, M. J. (2019). Collaboratively-developed vaccine pricing and

stable profit sharing mechanisms. Omega, 84:102 – 113.
Mason, R., Boughton, R., and Lalwani, C. (2007). Combining vertical and horizontal
collaboration for transport optimization. Supply Chain Management: An International
Journal, 12.
Neves-Moreira, F., Amorim, P., Guimarães, L., and Almada-Lobo, B. (2016). A long-
haul freight transportation problem: Synchronizing resources to deliver requests passing
through multiple transshipment locations. European Journal of Operational Research,
248(2):487 – 506.
Ozener, O. (2014). Developing a collaborative planning framework for sustainable trans-

portation. Mathematical Problems in Engineering, 2014.
Paul, J., Agatz, N., Spliet, R., and Koster, R. D. (2019). Shared capacity routing problem
an omni-channel retail study. European Journal of Operational Research, 273(2):731 –
739.
40(8):2985 – 2991.
Ramanathan, U., Bentley, Y., and Pang, G. (2014). The role of collaboration in the uk
green supply chains: an exploratory study of the perspectives of suppliers, logistics and
retailers. Journal of Cleaner Production, 70:231 – 241.
Sanchez, M., Rojas, L., Deschamps, J.-C., and Parada, V. (2015). Reducing the carbon
footprint in a vehicle routing problem by pooling resources from different companies.
NETNOMICS: Economic Research and Electronic Networking, 17.
Solomon, M. M. (1987). Algorithms for the vehicle routing and scheduling problems with
time window constraints. Operations Research, 35:254–265.
Wang, Y., Zhang, S., Assogba, K., Fan, J., Xu, M., and Wang, Y. (2018). Economic
and environmental evaluations in the two-echelon collaborative multiple centers vehicle
routing optimization. Journal of Cleaner Production, 197:443 – 461.
Xiao, Y., Zhao, Q., Kaku, I., and Xu, Y. (2012). Development of a fuel consumption
optimization model for the capacitated vehicle routing problem. Computers Operations
Research, 39(7):1419 – 1431.
Zibaei, S., Hafezalkotob, A., and Ghashami, S. S. (2016). Cooperative vehicle routing
problem: an opportunity for cost saving. Journal of Industrial Engineering Interna-
tional, 12:271–286.
Özener, O. , Ergun, , and Savelsbergh, M. (2013). Allocating cost of service to customers

in inventory routing. Operations Research, 61(1):112–125.
6.A. Stand alone model 211
Appendix 6.A Stand alone model

The stand alone problem considers that each entity in the transportation network operates
individually. Thus, the LR can only serve the stores, and afterwards the dedicated vehi-
cles return to the depot empty. The 3PL can only serve its customers, and afterwards the
dedicated vehicles return to the depot empty. Each producer may only send the available
loads directly to the depot of the LR, through outsourced vehicles. Therefore, the stand
alone model can be extended from the collaborative presented in Section 6.3 by including
a set of constraints that forbid the vehicles to travel in specific arcs of the network. These
additional constraints are as follows. Constraints (6.34) and (6.35) forbid vehicles used by
the RL to visit 3PL customers and producers, respectively. Similarly, Constraints (6.36)
and (6.37) forbid vehicles used by the 3PL to visit stores and producers, respectively. Con-
straint (6.38) ensures that vehicles of the LR and the 3PL do not visit the depots of each
other, and Constraints (6.39) ensure that all loads available at the producers are sent directly
to the depot of the LR.
xikj = 0, ∀i ∈ V, ∀ j ∈ R (6.34)
xikj = 0, ∀i ∈ V, ∀ j ∈ P (6.35)
zkij = 0, ∀i ∈ V, ∀ j ∈ S (6.36)
zkij = 0, ∀i ∈ V, ∀ j ∈ P (6.37)
k
xαβ + xβα
k
+ zkαβ + zkβα =0 (6.38)
wkiβ = 0, ∀i ∈ P (6.39)
Appendix 6.B Impact of full collaboration versus limited col-

laboration
The problem presented in Section 6.3 represents the case of full collaboration. This sec-
tion provides a comparison between the full collaboration setting and settings of limited
collaboration.
The first limited setting covers the case where only collaborative pickup-delivery and col-
laborative cross-docking can be selected. Thus, it is sufficient to include in the original
model the two additional constraints (6.35) and (6.37), which forbid backhauling. The sec-
ond limited setting covers the case where only collaborative backhauling and collaborative
cross-docking can be selected. Thus, it is sufficient to include in the original model the
additional constraints (6.36), which forbid vehicles of the 3PL to serve customers of the
RL.
The average synergy values obtained for limited and full collaboration settings are pre-
sented in Tables 6.14 and 6.15. The results show the benefits of full collaboration com-
paring with both cases of limited collaboration, as the synergy values regarding costs and
fuel consumption are much higher in the former case for both instances. It is also demon-
strated that the type of network is susceptible to the type of collaborative strategies allowed.
Forbidding collaborative pickup-delivery reduces drastically the synergy value of collab-

oration, which is more pronounced for network type R. On the other hand, forbidding
collaborative collection, the highest reduction of the synergy value is shown for network
type RC. Moreover, even in the cases of limited collaboration, the cross-docking strategy
does not provide any benefit, and thus it is never selected.
Table 6.14 – Average synergy values obtained with different collaborative strategies, for
instance s8-r5-p2
Pickup + Cross-docking Collection + Cross-docking Full collaboration
svcost 11% 8% 7% 4% 1% 6% 12% 8% 12%
OF(C) sv f uel 10% 8% 7% 8% 3% 11% 13% 8% 15%
svcost 11% 8% 7% 4% 0% 6% 12% 7% 11%
OF(E) sv f uel 9% 7% 6% 8% 4% 10% 12% 9% 14%
svcost 11% 8% 7% 4% 1% 6% 12% 8% 12%
OF(CE) sv f uel 9% 7% 6% 8% 3% 10% 13% 7% 13%
Table 6.15 – Average synergy values obtained with different collaborative strategies, for
instance s10-r3-p2
Pickup + Cross-docking Collection + Cross-docking Full collaboration
svcost 7% 8% 7% 5% 1% 7% 10% 9% 13%
OF(C) sv f uel 6% 7% 7% 10% 3% 12% 13% 10% 16%
svcost 7% 8% 7% 5% 0% 7% 9% 7% 12%
OF(E) sv f uel 6% 7% 6% 10% 4% 11% 13% 10% 16%
svcost 7% 8% 7% 5% 1% 7% 10% 9% 13%
OF(CE) sv f uel 6% 7% 6% 10% 3% 11% 12% 9% 15%
Chapter 7
Conclusions and future work
This thesis approaches the integrated transportation planning and the opportunities to cope
with sustainable challenges. Throughout the thesis, the integrated transportation problem
is addressed as a Vehicle Routing Problem with Backhauls (VRPB), which consists on the
integrated planning of outbound (delivery) and inbound (pickup) routes with the goal of
reducing empty running of vehicles. Even though the VRPB can reduce considerably the
number of vehicles, empty and total distances in comparison with the decoupled transporta-
tion planning, the problem is recurrently driven by economic concerns only. Nonetheless,
the concept of the VRPB itself matches two of the main targets of sustainable transporta-
tion, namely the efficient use of vehicles and the minimization of empty trips. Moreover,
the fuel wasted on empty trips and the resulting consequences (e.g., GHG emissions) can
be avoided, which remarks the green nature of VRPB.
Therefore, the first step conducted in this research encompasses a review on the VRPB
literature, allowing to establish the connections between the VRPB and sustainable trans-
portation. Chapter 2 is dedicated to this research step, which identifies the main limitations
of the current VRPB literature and the challenges to promote sustainable versions of the
VRPB. Some of these challenges are the drivers for the remaining chapters. Chapter 3
focuses on the analysis of different backhauling opportunities to improve the integrated
planning in real life contexts, addressing thus the challenge of efficient formulations for
the Rich VRPB, demonstrating the advantages of integrated planning in comparison with
decoupled planning. Chapter 4 is devoted to the mathematical formulation of a robust
VRPB, addressing the lack of studies on the VRPB under uncertainty. Chapter 5 focus on
the efficient formulation of a collaborative VRPB, solving some limitations of the currently
used approaches in practice. Chapter 6 focus on the development of green VRPB models
and their impact on collaborative networks, addressing sustainable challenges identified not
only in the VRPB literature but also in the literature of collaborative VRP. In this chapter,
the main findings and contributions of this thesis are presented, followed by the answers to
the research questions and suggestions for future work.
7.1. Contributions
The first contribution of this thesis is advanced in chapter 2 with the analysis, classifica-
tion and discussion of the VRPB literature in the light of sustainable transportation. The
VRPB problems are firstly classified according to a common taxonomy of VRP provided
in the literature, which is extended to distinguish the objective function in economic, en-
vironmental and social. Afterwards, the context of the VRPB application is examined in
order to highlight the works addressing VRPB models with environmental and/or social
concerns, and VRPB models used in both collaborative vehicle routing and reverse logis-
tics. This literature review provides statistical evidences that, out of 107 reviewed papers,
eight address environmental or sustainable concerns, six examine the VRPB in the reverse
logistics, three study the collaborative VRPB, and none addresses uncertainty. On the other
hand, using the VRPB as an instrument to promote sustainable transportation seems to be
arising the interest of the research community and practitioners, as the results also shows
that most of the sustainable VRPB literature was published in the last decade. In face of
these findings, chapter 2 reports the current challenges and future research directions for
the VRPB research. From the aforementioned work, of which the PhD candidate was the
main author, the following research paper has resulted:
• M. J. Santos, P. Amorim, A. Marques, A. Carvalho, A. Póvoa. The vehicle routing

problem with backhauls towards a sustainability perspective - a review. TOP, 2019.
https://doi.org/10.1007/s11750-019-00534-0
The main contribution of chapter 3 is a new mathematical formulation of the integrated

inbound-outbound transportation planning adapted to model real transportation problems
of the wood-based industry. The formulation relies on a Rich VRPB with several oper-
ational constraints (heterogeneous fleet, multi-depots, split deliveries, selective backhaul
customers) aiming to minimize fixed and variable routing costs and maximize revenues.
The integrated planning is compared with an opportunistic strategy, where the planning
of outbound routes are prioritized over the inbound routes, and a decoupled strategy, that
plans the inbound and outbound routes separately. For all the case-study instances tested,
the integrated planning allowed to achieve the best results, with an average cost reduction
of 2.7%. The computational experiments further analyse the influence of diverse param-
eters and network configurations on the integrated planning, which allows to draw some
insights. The most relevant one derives from the relationship between network configura-
tion and revenues collected at backhaul customers - the less dispersed are the locations and
the higher are the revenues, the higher is the potential to create integrated routes. However,
this may lead to an increase in the number of vehicles to perform non-viable integrated
routes (i.e. more than necessary). Therefore, the trade-off between maximum number of
vehicles and revenues should be a properly managed in the integrated planning. In this
work, the main responsibilities of the PhD candidate were the literature review, conducting
the experiments on sensitivity analysis and discussion of the results. The work described
in chapter 3 has result in the following research paper:
• A. Marques, R. Soares, M. J. Santos, P. Amorim. Integrated planning of inbound and

outbound logistics with a Rich Vehicle Routing Problem with Backhauls. Omega,
2019. https://doi.org/10.1016/j.omega.2019.102172
The contributions of chapter 4 are two-fold. Firstly, it provides the first study of a VRPB
under uncertainty. The revenues collected at backhaul customers are considered uncertain
and a robust version of the problem is formulated following a well-known robust opti-
mization approach in the literature. The robust model considers that uncertain revenues
are represented by a polyhedral uncertainty set and uses a parameter to control the size of
7.1. Contributions 215
that set, known as budget of uncertainty. The robust model is compared with a chance-
constrained model and the results indicate that the former approach is less conservative.
The budget of uncertainty is set by the user and reflects in some degree its risk aversion.
Thus, it can be related to the probability of constraint violation. In this context, the second
contribution of this chapter is an efficient method to estimate the probability of constraint
violation. This new method is inspired by the geometrical derivation of the uncertainty set
through the Irwin-Hall distribution proposed in the work of Marengo et al. (2017). Further,
it is compared with two methods proposed in the literature, and the results clearly demon-
strate that the new method outperforms them by providing the tightest probability bounds
of constraint violation. The PhD candidate was the main author of the this work, leading
the different tasks with the exception of the Branch-and-Cut algorithm which was devel-
oped by the third author. The second author was actively enrolled on the development of
the Adaptive Large Neighborhood Search (ALNS) metaheuristic, and on the development
of the methods to estimate probabilistic bounds. The following research paper result from
this work:
• M. J. Santos, E. Curcio, M. H. Mulati, P. Amorim., F. K. Miyazawa. A robust opti-
mization approach for the vehicle routing problem with selective backhauls. Trans-
portation Research Part E: Logistics and Transportation Review, 2020. https://doi.
org/10.1016/j.tre.2020.101888
The main contribution of chapter 5 is a new mathematical formulation of the collaborative
VRPB which allows to solve simultaneously the routing problem and the profit sharing
problem. The collaborative problem considers the case where a carrier, initially hired to
perform deliveries at the shipper’s customers, may be motivated to perform backhaul routes
for the shipper under proper incentives. Chapter 5 addresses this collaborative problem,
which implies solving a VRPB and defining proper incentives to induce shipper-carrier
collaboration. Based on the hierarchical nature between shipper and carrier, the new math-
ematical formulation is based on bilevel optimization which allows to solve the problem of
the shipper (upper level) while anticipating the rational response of the carrier (lower level
problem). As a result, the carrier would only accept incentives that generate higher profits,
or at least equal, to its stand alone solution, which puts in evidence the ability of the bilevel
formulation to tackle the property of individual rationality. The computational experiments
conclude that solutions obtained with the bilevel formulation tend to generate higher syn-
ergy values than solutions obtained with traditional planning with side payments. The PhD
candidate was the main author of this chapter, and the second and third authors contributed
to the development of the mathematical formulation and the exact reformulation technique
to approach the problem solving. The work on this chapter result in the following research
paper:
• M. J. Santos, E. Curcio, M. Carvalho, P. Amorim., A. Marques. A bilevel approach
for the collaborative transportation planning problem. Submitted to International
Journal of Production Economics, 2020.
The main contribution of chapter 6 is a set of efficient modelling approaches to tackle the
collaborative VRPB with environmental concerns. The collaborative problem considers
a set of different participants (a retailer, a 3PL and several suppliers) that operate at the
same level of the transportation network but may also perform complementary transporta-
tion services (e.g., cross-docking). Moreover, two different backhauling strategies can be
adopted in collaboration, namely pickup and delivery services and collections to supply
a depot. Solving the routing problem considering diverse strategies at the same time is
motivated by the fact that it may unveil opportunities otherwise hidden. In opposition to
the hierarchical approach described in the previous chapter, the collaborative problem in
this chapter follows a joint planning approach. The collaborative problem is formulated by
merging the problems of all participants into one large mathematical optimization problem
with an overall goal for the entire network, and not for the participants individually. Also,
the routing and profit sharing problems are solved sequentially. In particular, the profit
sharing problem is addressed by simple proportional allocation methods, which are easy to
communicate and implement in practice. The computational experiments conducted in this
research indicate that tackling environmental concerns in the objective function may reduce
substantially both environmental costs and fuel consumption, with a very slight increase on
the operational costs, and can create reasonably high values of synergy. Nonetheless, with
a holistic function, the collaborative solutions tend to follow the solutions obtained with an
economic objective function and lead to similar synergy values obtained with the environ-
mental objective. The PhD candidate is the main author of this work, and the second author
has contributed to the problem formulation. The following working paper has resulted from
the work in this chapter:
• M. J. Santos, S. Martins, P. Amorim., B. Almada-Lobo. A green collaborative trans-

portation problem - analysis of different strategies and profit allocation methods.
Working paper, 2020.
7.2. Answers to the research questions

In this section, we provide the answers to each research question raised in this thesis.
Research Question 0
What is the role of the Vehicle Routing Problem with Backhauls in terms of sustainability?
The VRPB focuses the reduction of empty running of vehicles by complementing out-
bound routes with inbound routes, which leads to reducing the total distances travelled by
the vehicles. As transportation is responsible for a significant part of GHG emissions, re-
ducing total distances is expected to reduce fuel consumption and emission of pollutants.
Nonetheless, other parameters besides the distance have influence on the environmental
impact of transportation, such as the load carried, the vehicle speed and the characteristics
of the vehicle, fuel and road. Optimizing transportation can also reduce the traffic acci-
dent rate and noise, which may also be influenced by the previously mentioned parameters.
Although few works have considered the environmental and social impacts of the VRPB,
the literature shows clear evidence of the increasing interest by the research community on
such topics in the last decade.
7.2. Answers to the research questions 217
The sustainable impact of the VRPB can be effectively demonstrated in real cases, consid-
ering rich contexts, collaborative networks or reverse logistics. The literature shows that
the VRPB allows reductions up to 20% in the total distance and up to 25% in CO2 emis-
sions in comparison with the traditional VRP. These reductions can almost double if mixed
deliveries and pickups are allowed. However, a mixed strategy should always be validated
in real cases in order to account for the effort required to rearranged the load in the vehicle.
The VRPB can be related to collaborative networks, as the main goal of collaboration is to
increase the efficiency of vehicles use and reduce empty trips, and as such, they are driven
by similar goals of the VRPB. The collaborative VRPB is demonstrated to be a powerful
strategy for both carriers and shippers, allowing savings of up to 24% in GHG emissions
and up to 30% in the total costs. The VRPB can also be linked to reverse logistics as this
problem considers the backward flows of products (e.g., returned products). In reverse
logistics, the linehaul and backhaul customers are the same, which implies that the clas-
sic VRPB is hardly applied. However, if the same vehicle is used to deliver and collect
products, other VRPB variants can be effectively applied. Case studies show that VRPB in
reverse logistics can reduce up to 23% in total distances, up to 12.5% in the environmental
impact and up to 23% in the social impact. Nevertheless, this requires analysis of trade-offs
between the three dimensions of sustainability.
Research Question 1
How can transportation with backhauling be enriched for real world contexts?
In order to investigate the real potential of the integrated inbound-outbound transportation
planning, the VRPB should be the closest as possible as real operations, and it should be
compared with the traditional decoupled planning. To bring the model closer to the real
problem, the VRPB can be enriched with several business-related rules. In the case study
investigated for the wood-based supply chain, the business-related rules include the com-
patibility between type of vehicles and loading requirements, the interdependence between
delivery and pickup operations, the selection of suppliers based on revenues, the possibility
to perform only deliveries, only pickups or both, and the possibility of splitting deliveries
to the same customer. Furthermore, other practical aspects such as heterogeneous fleet and
multi-depots are also included in this Rich VRPB.
However, as more practical constraints are considered in the VRB model, the harder is to
solve exactly the problem. To overcome this difficulty, we first reduce the complexity of
the model by formulating it as a two-echelon VRPB. The first echelon includes the routing
decisions between linehaul and backhaul customers, and as such it also includes the se-
lection of suppliers. The second echelon includes the routing decisions between backhaul
customers and last destinations, i.e. which mills are supplied by which backhaul customers.
Another suggestion to avoid the difficulty of solving exactly the problem is to solve it with a
matheuristic, because it allows to reduce the computational effort of exact methods, while
providing solutions as closest as possible to their optimum. Nevertheless, solutions ob-
tained with the matheuristic should always be compared with the lower bounds obtained
with exact methods, in order to measure the effectiveness of the approximate algorithm.
Research Question 2
How to address uncertainty in the Vehicle Routing Problem with Backhauls through robust
optimization?
Robust optimization is an emergent technique to deal with uncertainty in routing problems
when the probability distribution of the uncertainty is not known in advance. The literature
shows that there are three main alternatives to represent the uncertainty space, namely
the box, the ellipsoidal and the polyhedral uncertainty sets. The box uncertainty set only
allows to represent the worst-case values of the uncertainty, and as such it is considered a
very conservative approach. The ellipsoidal allows to reduce this conservatism but it also
reduces the tractability of the model as it originates quadratic functions in the constraints
that incorporate the uncertain parameters. Thus, the polyhedral uncertainty set provides a
suitable trade-off between conservatism of solutions and model tractability.
In this thesis, we demonstrate two alternatives to model the VRPB with uncertain revenues
using polyhedral uncertainty sets from the literature, namely the factor model support and
the budget of uncertainty. In particular, the budget of uncertainty approach leads to solu-
tions much less conservative than a chance-constrained programming model, and allows
to obtain different solutions for different types of decision-makers, i.e. different degrees of
risk aversion. Thus, it enables to investigate the trade-off between robustness of solutions
and risk aversion of the decision-maker. Nevertheless, the polyhedral uncertainty set re-
quires mild assumptions on the distributions and relationships of the uncertain parameters.
We execute the experiments considering that each uncertain parameter is independent on
the others and that their values rely on an uniform distribution interval, since this is a good
approximation when there is lack of knowledge on the distribution profile of the uncer-
tainty. Moreover, some statistical evidences from past information can be used to build
less conservative uncertainty sets and, as such, provide less conservative robust solutions.
Research Question 3
How to efficiently model and solve a collaborative Vehicle Routing Problem with Back-
hauls?
Bilevel optimization is a well-known strategy to model a collaborative game between two
(or more) entities that share common aspects of an optimization problem. In the context
of collaborative vehicle routing, a bilevel optimization model can effectively represent a
game between a shipper and a carrier, which have different objective functions and whose
relationship is based on a hierarchical structure.
In this thesis, we demonstrate how the bilevel optimization can leverage the collaboration
between a shipper and a carrier, which occurs mainly through backhauling. The problem
of the shipper, which aims to minimize total costs, can be described by the upper level
problem, while the problem of the carrier, which aims to maximize total profits, can be
described by the lower level problem. Furthermore, the profit sharing problem can also be
incorporated in the bilevel optimization model, which ensures the individual rationality of
both entities.
Although the literature provides some exact solution methods that are able to solve bilevel
optimization problems, in general these cannot be applied to our case, due to the existence
7.3. Future work 219
of continuous and integer variables in both upper and lower level problems. Thus, we de-
velop a reformulation technique that allows to solve exactly the problem, while considering
the individual rationality in the profit sharing problem. Furthermore, we show that a bilevel
approach is more effective than side payments approaches, which are common strategies
applied in practice.
Research Question 4
How to address the challenges of a practical sustainable collaborative Vehicle Routing
Problem with Backhauls?
In practice, collaborative networks involve several entities that may perform both similar as
well as complementary transportation services. However, the relationship or dependencies
between these services is either neglected in collaborative models, or they rely on a set of
assumptions that aim to reduce the complexity of the models.
In this thesis, we demonstrate how to mathematically formulate a collaborative model con-
sidering the dependencies of three transportation services, namely cross-docking, pickup-
delivery and backhauling, and show on which situations these services may conduct to
collaborative gains for the entire transportation network. We further show that indepen-
dently on the objective function used, either pure economic, pure environmental or a mixed
economic-environmental, significant savings in both costs and environmental impact can
be achieved. This means that modelling the collaborative problem using a traditional cost
minimization function is suitable to promote environmental sustainability of transportation.
Nevertheless, determining the collaborative gains for the entire network does not necessar-
ily implies that all participants would profit from collaboration. In fact, it is often the case
that one or more participants may increase their total costs in comparison with their indi-
vidual costs when collaboration does not occur. Hence, distributing the collaborative gains
among the participants as fairest as possible is an important requirement to implement the
collaboration in practice. The literature shows that there are plenty of allocation methods
used to solve the profit sharing problem, but each one should be carefully considered to
handle case-specific problems, since there is no "one size-fit all" method. Furthermore, to
be implemented in practice, the allocation method should provide simplicity and should
be easily understood by all the decision-makers. In this context, we show that using pro-
portional rules to distribute the gains among participant is a promising approach to solve
efficiently the profit sharing problem in practice.
7.3. Future work

The research conducted in this thesis allow to answer all the research questions previously
established, and which were mainly motivated by the findings on chapter 2. Nonetheless,
several opportunities and literature gaps related to the VRPB research still remain, and
deserve further attention by the research community.
Chapter 2 shows that there is room to explore further the potential of the VRPB to handle
sustainability concerns. For example, social concerns are hardly considered in the VRPB
literature and only cover accident rate and noise. However, it has been recognized that
the equity of working hours among drivers has a major impact on their performance and
motivation. Thus, an opportunity for research could be exploring the impact of this so-
cial concern in integrated routing plans in comparison with traditional decouple planning.
Also, the environmental aspects addressed in the VRPB literature only covers CO2 emis-
sions, which have a global impact on the environment, and can be easily converted based
on fuel consumption. However, the emission of other pollutants, such as NOx, is influ-
enced by other vehicle parameters and impact locally on the environment. Thus, another
opportunity for research entails an investigation of the impact of integrated planning con-
sidering both local and global emissions. Another interesting line of research could cover
the investigation of the logistics requirements and constraints of the VRPB in practice and
respective challenges. For example, it is often the case that delivery and pickup loads can-
not be transported at the same time due to compatibility issues, but this challenge may
be overcome with multi-compartment vehicles. Also, the use of Alternative Fuel Vehi-
cles (AFV) has been increasing due to their efficiency on reducing pollution, but so far no
VRPB study had considered this type of vehicles, which opens up another opportunity for
research. Another challenge not yet addressed in the VRPB literature relates with the use
of dynamic approaches to model and solve this problem.
In chapter 3, a matheuristic was developed to solve the Rich VRPB. This solution method
is based on a fix-and-optimize approach which iteratively solves smaller mixed-integer pro-
gramming (MIP) sub-problems of the original model, and where each sub-problem consists
of a set of decision variables to fix or release. Compared with a common MIP solver, the
matheuristic developed is able to provide higher solution quality and higher computing ef-
ficiency, but it still consumes a relatively high amount of time to provide a good solution
in large size instances. Nonetheless, matheuristics have been successfully applied in other
optimization problems, and their performance are highly dependent on the problem design.
Thus, another opportunity for future work is the improvement of the matheuristic proposed
in chapter 3, which may include redesign of the MIP sub-problems or testing different
heuristics.
The main motivation of chapter 4 was the lack of research on VRPB under uncertainty.
The approach used to model the robust VRPB was the well-known approach of budget of
uncertainty, and it has succeed as demonstrated by the computational results. However,
as the robust optimization is an emergent field of research, investigating new robust ap-
proaches and uncertainty sets is seen as a promising opportunity for future VRPB research.
For instance, the uncertainty set of revenues may be built using a different distribution or
leveraging some statistics obtained from historical information, such as the distribution-
ally robust optimization. The uncertainty set would then present tighter bounds, leading to
better estimates of the probability of constraint violation and also less conservative robust
solutions.
In chapter 5, the bilevel VRPB model proposed to study the vertical collaboration prob-
lem is shown to succeed for solving both the routing and the profit sharing problems. The
bilevel mathematical program is reduced to a single-level through an exact reformulation
technique, and a common solver is then used to solve the problem. This strategy how-
ever requires too much effort for solving large size instances. As evolutionary algorithms
are the most advanced solution methods applied in the bilevel optimization literature, a
7.3. Future work 221
possible direction for future VRPB research may entail the investigation of these type of
metaheuristics to solve the bilevel VRPB. Another opportunity for research involves the
analysis of different concepts that builds the bilevel model. In chapter 5, the bilevel VRPB
considers the most opportunistic case of the upper level, in face of the most optimistic sce-
nario of the lower level. Thus, the model can be designed using other rationales (e.g., less
opportunistic-pessimist scenario, less opportunistic-optimistic scenario) and the collabora-
tive solutions can be compared, which may allow to measure in some degree the robustness
of the different approaches.
Finally, in chapter 6, both the joint routing planning and the profit sharing problems are
investigated and solved for the lateral collaborative problem. Although this work fills some
of the gaps found in the related literature, a major challenge still remains - solving the
routing and the profit sharing problems simultaneously. Chapter 5 presents a formulation
that can handle this challenge, but such formulation is not suited for the problem carried
in chapter 6, as there are no hierarchical decisions to be made nor conflicting objectives
among the participants. Therefore, an opportunity for future research, not only concerning
the VRPB literature but extended also to the collaborative VRP literature, encompasses
the development of efficient formulations that allow to merge the routing and the profit
sharing problems into an integrated optimization problem, or the development of efficient
algorithms that can solve both problems simultaneously.

Porto: Towards Solving A Robust and Sustainable Vehicle Routing Problem With Backhauls

Uploaded by

Copyright:

Available Formats

Porto: Towards Solving A Robust and Sustainable Vehicle Routing Problem With Backhauls

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Porto: Towards Solving A Robust and Sustainable Vehicle Routing Problem With Backhauls

Uploaded by

Copyright:

Available Formats

Maria João Martins dos Santos

Towards solving a robust and sustainable

Submitted to Faculdade de Engenharia da Universidade do Porto in partial fulfillment of

Department of Industrial Engineering and Management

1 Motivation and overview 1

2 Vehicle Routing Problem with Backhauls 11

3 A Rich Vehicle Routing Problem with Backhauls 61

3.5.4 Performance of the solution approach . . . . . . . . . . . . . . . . 90

4 A Robust Vehicle Routing Problem with Backhauls 103

5 A Collaborative Vehicle Routing Problem with Backhauls 149

6 A Green Vehicle Routing Problem with Backhauls 183

7 Conclusions and future work 213

Motivation and overview

vertical collaboration, and involves decisions on similar and complementary transportation

1.1. Research objectives and methodology

Figure 1.1 – Research questions framework.

1.2. Thesis structure and synopsis

Dantzig, G. B. and Ramser, J. H. (1959). The truck dispatching problem. Management

Evangelista, P., Colicchia, C., and Creazza, A. (2017). Is environmental sustainability a

Gansterer, M. and Hartl, R. F. (2018). Collaborative vehicle routing: A survey. European

Guajardo, M. and Rönnqvist, M. (2016). A review on cost allocation methods in collabora-

Vehicle Routing Problem with

The vehicle routing problem with backhauls towards

Published in TOP, 2019

Keywords vehicle routing problem · backhauling · sustainability · collaboration · reverse

- Provide insights on how to leverage the sustainable potential of the VRPB.

2.2. The VRPB and its variants

1. Each vehicle travels exactly one route;

2. Each vehicle starts and ends at the depot;

4. Each customer (either linehaul or backhaul) is visited exactly once;

Descriptive analysis A preliminary analysis of all papers was undertaken, according to

- Measuring the externalities of transportation (impact of transport operations on the

- Collaboration is a sustainable strategy currently used by many companies in the

2.4. Literature overview

2.4.1 Descriptive analysis

Figure 2.1 – Distribution of published papers per year.

Figure 2.2 – Distribution of published papers per journal.

Figure 2.3 – Top 10 most cited VRPB papers.

2.4.2 Classification of VRPB works

2.4.2.1 Exact methods

(a) Solution method (b) Type of variant

(c) Scenario characteristics (d) Physical characteristics

(e) Objective function (f) Quality of information

Figure 2.4 – Number of VRPB papers classified according to each category.

Population-based metaheuristics Among population-based metaheuristics, evolution-

2.4.3 Final remarks on the VRPB

Figure 2.5 – Number of VRPB works covering different dimensions of sustainability.

2.5. Analysis of VRPBs with sustainability concerns

2.5.1 VRPBs with sustainable objectives

2.5.2 VRPBs in collaborative networks

2.5.3 VRPBs in reverse logistics

2.5.4 Final remarks on the VRPBs with sustainability concerns

2.6. Future directions for sustainable VRPB

Figure 2.6 – Framework of future directions for the VRPB research.

Ashtineh, H. and Pishvaee, M. S. (2019). Alternative fuel vehicle-routing problem: A life

Averbakh, I. (2001). On the complexity of a class of combinatorial optimization problems

Bektaş, T. and Laporte, G. (2011). The Pollution-Routing Problem. Transportation Re-

Catherine Early (2011). Delivering greener logistics.

Chardy, M. and Klopfenstein, O. (2012). Handling uncertainties in vehicle routing prob-