Liner Shipping Network Design: Dissertation

Dissertation
Liner Shipping Network Design

Decision Support and Optimization Methods
for Competitive Networks
M.Sc. Stefan Guericke
Schriftliche Arbeit zur Erlangung des akademischen Grades

doctor rerum politicarum (dr. rer. pol.)
im Fach Wirtschaftsinformatik
eingereicht an der
Fakultät für Wirtschaftswissenschaften der
Universität Paderborn
Paderborn, Dezember 2014
Gutachter:
1. Prof. Dr. Leena Suhl
2. Prof. Dr. Wilhelm Dangelmaier
The future has many names:
For the weak, it means the unattainable.
For the fearful, it means the unknown.
For the courageous, it means opportunity.
Victor-Marie Hugo
Acknowledgements
This thesis is the result of the research I conducted as a member of the International
Graduate School of Dynamic Intelligent Systems and the Operations Research &
Decision Support Lab at the University of Paderborn. I would like to thank everybody
that supported me to complete this dissertation within three years.
Foremost, I would like to thank Prof. Dr. Leena Suhl for sparking my interest on
Operations Research during my studies at the University of Paderborn. Furthermore,
for providing me the opportunity to work in an enjoyable and productive atmosphere
at her chair. I would like to thank my supervisors Prof. Dr. Wilhelm Dangelmaier
and Prof. Dr. Gregor Engels for the guidance and the ensuring of a descent research
quality.
The colleagues at the DS&OR working group provided a challenging and moti-
vating research atmosphere. In particular, I want to thank Jun.-Prof. Dr. Kevin
Tierney for his critical feedback on the work, Christoph Weskamp for always having
time for mutual discussions, and my office colleagues Torben Schramme and Marius
Merschformann for the pleasant time.
I am deeply thankful for the opportunity of doing the PhD as an ORCONOMY
GmbH fellow. I would like to thank the managing directors Dr. Ingmar Steinzen,
Dr. Jens-Peter Kempkes and Dr. Stefan Bunte for their guidance and the critical
discussions. I would like to thank the whole ORCNOMY team for the support
during the last three years, and especially Dr. Stephanie Heller for the feedback on
the mathematical proof and sharpening my view on mathematical details.
I am indebted to the team of the International Graduate School of Dynamic
Intelligent Systems, its head apl. Prof. Dr. Eckhard Steffen, and his assistance Astrid
Canisius for the incredible organization, support and the excellent PhD program.
I would like to thank anonymous referees for the feedback, discussions and pro-
found insight into liner shipping network planning.
Credits go as well to all students that participated to the success of this disserta-
tion with critical questioning and inspirational ideas.
I am deeply indebted to my whole family for sharing their interest in this exciting
research through discussions and support. In particular, I want to thank my parents
who always believed in me, financially safeguarding me during my studies and always
supporting me during my endeavors.
Most important, I deeply thank Daniela to share worries, delights and dreams.
Stefan Guericke
Paderborn, December 2014
ii
Contents
1. Introduction 3
1.1. Research Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2. Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2. Liner Shipping Network Planning 9

2.1. Planning Process Overview . . . . . . . . . . . . . . . . . . . . . . . 9
2.2. Basic Liner Shipping Network Design Problem . . . . . . . . . . . . 10
2.3. Route Types and Network Structure . . . . . . . . . . . . . . . . . 15
2.4. Transportation of Containerized Cargo . . . . . . . . . . . . . . . . 16
2.4.1. Demand Structure . . . . . . . . . . . . . . . . . . . . . . . 16
2.4.2. Container Sizes and Types . . . . . . . . . . . . . . . . . . . 17
2.5. Timing Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.5.1. Port Calls and their Impact on the Vessels’ Speed . . . . . . 18
2.5.2. Transit Times in the Maritime Context . . . . . . . . . . . . 20
2.6. Cooperative Agreements . . . . . . . . . . . . . . . . . . . . . . . . 22
2.6.1. Liner Conferences, Consortium Agreements and Global Alliances 22
2.6.2. Slot Sharing Agreements . . . . . . . . . . . . . . . . . . . . 24
2.7. Liner Service Capacities . . . . . . . . . . . . . . . . . . . . . . . . 24
2.7.1. Container Vessel Capacities . . . . . . . . . . . . . . . . . . 25
2.7.2. Port Depth and Vessels’ Deadweight Scale . . . . . . . . . . 25
2.8. Empty Container Repositioning . . . . . . . . . . . . . . . . . . . . 27
2.9. Costs and Revenues . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.10. Bunker Cost Uncertainty in the Tactical Planning Horizon . . . . . 32
2.11. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3. State-of-the-Art and Research Opportunities 35

3.1. Selected Optimization Techniques . . . . . . . . . . . . . . . . . . . 35
3.1.1. Linear and Mixed Integer Programming . . . . . . . . . . . 35
3.1.2. Delayed Column Generation . . . . . . . . . . . . . . . . . . 37
3.1.3. Metaheuristics . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.1.4. Fitness Approximation in Metaheuristics . . . . . . . . . . . 41
3.2. Related Combinatorial Optimization Problems . . . . . . . . . . . . 43
3.2.1. Vehicle Routing Problems . . . . . . . . . . . . . . . . . . . 43
3.2.2. Pickup and Delivery Problems . . . . . . . . . . . . . . . . . 44
3.2.3. Min-Cost Flow Problems . . . . . . . . . . . . . . . . . . . . 46
iii
3.3. Liner Shipping Network Planning . . . . . . . . . . . . . . . . . . . 48
3.3.1. Liner Shipping Network Design Problem . . . . . . . . . . . 48
3.3.2. Cargo Allocation and Empty Container Repositioning Problems 52
3.3.3. Speed Optimization . . . . . . . . . . . . . . . . . . . . . . . 55
3.4. Research Gap and Opportunities . . . . . . . . . . . . . . . . . . . 58
3.5. Goals of this Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4. Evaluating Networks - The Integrated Cargo Allocation Problem 63

4.1. Distinguishing Port Calls in Liner Services . . . . . . . . . . . . . . 63
4.2. Common Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.3. Arc-flow Formulation for the Cargo Allocation Problem . . . . . . . 67
4.3.1. Mathematical Model . . . . . . . . . . . . . . . . . . . . . . 68
4.3.2. Bunker Cost Discretization . . . . . . . . . . . . . . . . . . . 71
4.4. Path-Flow Formulation for the Cargo Allocation Problem . . . . . . 73
4.4.1. Restricted Master Problem . . . . . . . . . . . . . . . . . . . 74
4.4.2. Network Structure for determining Container Paths . . . . . 77
4.5. Relaxing the Integrality Constraints for the Bunker Cost . . . . . . 82
4.6. Numerical Results for the Integrated Cargo Allocation Problem . . 87
4.6.1. Problem Instances . . . . . . . . . . . . . . . . . . . . . . . 87
4.6.2. Arc-Flow Formulation . . . . . . . . . . . . . . . . . . . . . 90
4.6.3. Path-Flow Formulation . . . . . . . . . . . . . . . . . . . . . 93
4.6.4. Comparison of Numerical Results . . . . . . . . . . . . . . . 95
4.6.5. Choosing an appropriate Approximation Level . . . . . . . . 97
4.7. Comparison and Interpretation of Results . . . . . . . . . . . . . . 103
5. Improving Networks - The Liner Shipping Network Design Problem 105

5.1. Mixed Integer Formulation . . . . . . . . . . . . . . . . . . . . . . . 105
5.1.1. Mathematical Model . . . . . . . . . . . . . . . . . . . . . . 106
5.1.2. Subtour Elimination Constraints . . . . . . . . . . . . . . . 117
5.1.3. Numerical Results for the Mixed Integer Program . . . . . . 119
5.2. Metaheuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
5.2.1. Decomposition Approach for the Metaheuristics . . . . . . . 127
5.2.2. Determine the Fitness of a Solution . . . . . . . . . . . . . . 128
5.2.3. Construction Heuristics . . . . . . . . . . . . . . . . . . . . . 135
5.2.4. Improvement Heuristics . . . . . . . . . . . . . . . . . . . . 137
5.2.5. Numerical Results . . . . . . . . . . . . . . . . . . . . . . . 143
5.3. Surrogate Extensions to Metaheuristics . . . . . . . . . . . . . . . . 148
5.3.1. Metaheuristic Metacontrol . . . . . . . . . . . . . . . . . . . 148
5.3.2. Numerical Results . . . . . . . . . . . . . . . . . . . . . . . 151
5.4. Interpretation of Results . . . . . . . . . . . . . . . . . . . . . . . . 161
5.5. Bunker Cost Uncertainty in the Tactical Planning Horizon . . . . . 162
5.6. Numerical Results from a Global Liner Carrier . . . . . . . . . . . . 167
iv
6. Integration into a Decision Support System 175
6.1. Decision Support System Components . . . . . . . . . . . . . . . . 175
6.2. Process Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
6.3. Client-Server Communication . . . . . . . . . . . . . . . . . . . . . 178
6.4. Graphical User Interface . . . . . . . . . . . . . . . . . . . . . . . . 181
7. Conclusion 187
7.1. Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
7.2. Critical Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
7.3. Future Research Opportunities . . . . . . . . . . . . . . . . . . . . . 190
Bibliography 191
List of Figures 211
List of Tables 217
List of Algorithms 219
A. Glossary 221
B. Transformation Algorithm for the Layered Network Structure 223
C. Extended Numerical Results for the Integrated Cargo Allocation Prob-

lem 225
C.1. Problem Instances . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
C.2. Cargo Allocation Approximations . . . . . . . . . . . . . . . . . . . 227
D. Extended Numerical Results for the Liner Shipping Network Design

Problem 229
D.1. Liner Shipping Network MIP Formulation . . . . . . . . . . . . . . 229
D.2. Parameter Tuning for the Evolutionary Algorithm . . . . . . . . . . 230
D.3. Accuracy of VNS Surrogate Evaluation . . . . . . . . . . . . . . . . 234
E. Calculating Waypoints and Sea Distances 237

1. Introduction
According to the World Trade Organization, the worldwide merchandise trade was
estimated worth an 18.3 trillion US$ in 20131 , which is a worldwide annual increase
of 5.3% compared to 2012. The largest proportion of worldwide trade is transported
by sea: 80% of the weight and 70% of the value (see UN Conference on Trade and
Development (2013) and (Schieck, 2008, p. 177)).
Usually, the shipping industry is distinguished into three different operation modes:
industrial, tramp and liner (see Christiansen et al. (2004)). In industrial shipping,
enterprises own a fleet or have long-term time charter contracts, which basically
make them the ship owner. The enterprises are thereby responsible to manage it
profitably and ship their cargo to minimal costs. In tramp shipping, the operator
owns or charters a fleet and serves available cargo with a basic contracted amount
and tries to maximize the profit by working on the spot market (Christiansen et al.
(2004)). Tramp shipping usually transports break bulk, such as coal or grain, and
liquids, such as liquefied natural gas or crude oil (Stopford, 2009, p. 55). The third
mode of operation is liner shipping and is the scope of this thesis.
Liner shipping has been an important operation mode that emerged at the end
of the 19th century with the introduction of steamships (Stopford, 2009, p. 28, 31).
Liner operators steam according to a published regular schedule, similar to a bus
line, independent of the utilization of the vessels (Christiansen et al. (2004), Schieck
(2008)). From a customer’s (shipper’s) point of view, the advantages are regular
transportation opportunities, relatively reliable sailing schedules and predictable
transit times (Brooks, 2000, p. 2). Today, liner shipping is mainly connected with
the transportation of general cargo in standardized containers, which accounted for
16% of the seaborne trade in tons and 50% of the value in 2013 (see UN Conference
on Trade and Development (2013)). The main advantage of the containerization of
commodities during the 1950s is the decreased port time due to increased intermodal
efficiency (Levinson (2006)).
Beside the twenty foot equivalent unit (TEU), many different specialized container
types exist that allow the transportation of break bulk and liquids as well. In general,
container vessels are not competitive compared to large scale break bulk vessels
for high volumes (Schieck, 2008, p. 209) and are rather used to transport small
volumes for specific end customers. Other commodities that are transported by liner
operations are vehicles (using roll-on/roll-off vessels) or project cargo, such as sailing
boats or other bulky commodities (Schieck, 2008, p. 182,185).
1
Press release from 2014-04-14 http://www.wto.org/english/news_e/pres14_e/pr721_e.htm
3
1. Introduction
Several mergers and acquisitions in the liner shipping industry resulted in a highly
concentrated market with a few, dominating carriers (see Stopford (2009)). The
largest 15 liner carriers operate about 79.3% of the worldwide capacity in TEU
(see Table 1.1)2 , that is either owned or chartered. The four largest carriers provide
about 50% of the worldwide capacity. All of the carriers listed in Table 1.1 operate
global networks and serve almost all areas in the world (Ihde, 2001, p. 136).
Capacity P
Rank Operator TEU % %
1 APM-Maersk 2,728,567 15.6% 15.6%
2 Mediterranean Shipping Company 2,481,967 14.2% 29.8%
3 CMA CGM Group 1,578,483 9.0% 38.8%
4 Evergreen Line 886,577 5.1% 43.8%
5 COSCO Container L. 786,585 4.5% 48.3%
6 Hapag-Lloyd 767,703 4.4% 52.7%
7 CSCL 635,625 3.6% 56.3%
8 Hanjin Shipping 593,739 3.4% 59.7%
9 APL 578,772 3.3% 63.0%
10 MOL 569,582 3.3% 66.3%
11 OOCL 500,500 2.9% 69.2%
12 Hamburg Süd Group 498,826 2.8% 72.0%
13 NYK Line 494,458 2.8% 74.8%
14 Yang Ming Marine Transport Corp, 402,786 2.3% 77.1%
15 Hyundai M.M. 377,319 2.2% 79.3%
Table 1.1.: Largest liner operators, July 2014. The percentage is based on the existing
worldwide cellular fleet of 4,992 ships with 17,785,515 TEU, (source:
http://www.alphaliner.com/top100).
The top 15 liner carriers concentrate on nine countries (see Table 1.2). Today,
the Chinese carriers COSCO, CSCL, Hanjin and Yang Ming provides the largest
capacity. The largest single carrier is Maersk Line in Denmark, followed by the
Mediterranean Shipping Company (MSC) in Switzerland and the French CMA
CGM Group. The German carriers Hapag-Lloyd and Hamburg Süd Group provide
about 7.2% of the worldwide capacity, followed by other Asian operators.
Capacity
Country TEU %
China 2,919,235 16.7%
Denmark 2,728,567 15.6%
Switzerland 2,481,967 14.2%
France 1,578,483 9.0%
Germany 1,266,529 7.2%
Japan 1,064,040 6.1%
Taiwan 886,577 5.1%
Singapore 578,772 3.3%
South Korea 377,319 2.2%
Table 1.2.: Operated capacity of the top 15 carriers by country, aggregated data
from Alphaliner’s top 100 ranking, July 2014, (source: http://www.
alphaliner.com/top100).
2
The capacities in Table 1.1 includes owned and chartered vessels
4
1.1. Research Goals
Currently, liner carriers are facing different challenges. The increasing competition
due to market regulations and increased vessel sizes lead to decreased freight rates
in the last years (Stopford (2009)). Additionally, bunker prices increased by more
than 200% in the last five years (see Section 5.5). As a result, liner carriers have very
low operating margins3 : On average, carriers had to deal with a negative margin of
-2.6% in the first quarter of 2014 (see Alphaliner Newsletter, Volume 22 (2014)).
These challenges let the liner carriers implement different strategies beside mergers
and acquisitions, such as reducing the vessel speed, further increase the vessel size
or cooperating with each other. These strategies result in frequent adjustments of
the operated network. This is a very complex task that requires many steps, such
as assessing the network’s feasibility and evaluating the network performance on
a monetary basis. The network must also be coordinated with other partners the
carrier is cooperating with.
1.1. Research Goals

The research scope of this thesis is to evaluate Operations Research (OR) methods
on real-world strategic planning problems of liner shipping carriers, namely the
cargo allocation and liner shipping network design problem (LSNDP). The cargo
allocation problem determines the cargo routing and service speeds on a given liner
network to evaluate it, whereas the network design problem determines the port
sequences and capacities. In particular, this thesis has the following four goals:
1. Evaluate large scale real-world liner networks
2. Formalize practical requirements for the LSNDP and generate optimal solu-
tions
3. Optimize the LSNDP for real-world instances in reasonable time
4. Integrate the developed methods into a decision support system
The cargo allocation problem (CAP) allows for the evaluation of a given liner
shipping network. In this thesis, the practical requirements for the planning task are
analyzed and optimization methods to support manual planning are developed. We
refer to manual planning as the process to design liner services without automatically
improving the networks. The cargo allocation problem evaluates designed networks
on a monetary basis. An important requirement that has to be dealt with are
fast solution times in order to get an instantaneous feedback on the impact of
the modified network. To the best of the authors knowledge, integrating cargo
allocation with several practical aspects, such as speed optimization, vessel drafts,
3
Ratio of operating income divided by net operating revenues.
5
1. Introduction
empty container repositioning and partner networks has not been done in previous
publications.
Based on the CAP, the more complex LSNDP’s requirements are analyzed from a
practical point of view. Based on these requirements, the research goal is to evaluate
the potential of OR methods in the network design. The benefit for the network
planners are automated improvements of existing networks and the exploration
of non-obvious potential alternatives. The liner shipping network design problem
includes all aspects of the integrated cargo allocation problem and extends them
by transit time and embargo aspects. This has not been done to the full extent in
previous publications.
The integration of the methods into a prototypical planning decision support
system (DSS) increases the usability to support network planners and releases the
potential of the optimization methods. In the scope of this thesis, challenges such
as the graphical presentation are approached.
These goals are derived from the research gap based on the state-of-the-art and
are presented in more detail in Section 3.4.
1.2. Outline
This thesis comprises seven chapters and is structured as follows: Chapter 1 corre-
sponds to this introduction which outlines the liner shipping market and introduces
the research topic.
In Chapter 2, the liner network planning process is outlined and the liner shipping
network design components are presented.
Based on this problem definition, Chapter 3 reviews models and solution ap-
proaches for the liner shipping network design and interwoven planning problems.
It provides an overview of the state-of-the-art regarding the cargo allocation and
network design problem. Based on the research gap and the practical requirements,
the goals of this thesis are derived in more detail.
Chapter 4 presents a new approach to solve the integrated liner shipping cargo
allocation problem. In particular, the speed optimization and the empty container
repositioning is integrated into a single planning problem. In the scope of this thesis,
two different mathematical formulations and solution approaches are developed and
evaluated on the LINER-LIB benchmark instances4 .
In Chapter 5, different approaches for solving the liner shipping network design
problem are described. More specifically, a mixed integer model is presented that
can be used to determine optimal liner networks. Next, two heuristic methods are
presented that determine good networks in a reasonable amount of time. To further
speed up the heuristics, the effects of approximate evaluations (so called surrogates)
4
The LINER-LIB benchmark suite contains data on ports, demands and vessel types and was
published in Brouer et al. (2013).
6
1.2. Outline
on the heuristics are analyzed. A sensitivity analysis shows the effects of varied
bunker cost on the network design. The concept can be used to analyze under which
bunker price changes carriers should adjust their network to cope with the new
situation. Finally, the developed methods are applied on a real-world liner network
to show the practical applicability.
In Chapter 6, the developed decision support system is presented. An example
process integration, functional description and implementation details to increase
the usability and acceptance for network planners are presented.
Finally, Chapter 7 concludes this thesis with a summary, a critical assessment of
the goals and methods and an outlook on future research topics.
7
2. Liner Shipping Network Planning
The objective of the liner shipping network planning in the liner industry is man-
aging the network efficiently. The network design problem is a complex strategic
planning problem, interwoven with several subproblems. This chapter is organized as
follows: First an overview of other liner shipping planning problems is given. Then,
the basic liner shipping network design problem, including the network structure,
transhipment operations and capacities, is presented. In the remaining sections the
reader is introduced into relevant subproblems and extensions of the basic problem
that are required to solve practical instances.
2.1. Planning Process Overview

In the liner shipping industry, different planning problems arise that depend on the
considered planning horizon. Schmidt and Wilhelm (2000) distinguish the strate-
gic, tactical and operational planning horizon. Strategic decisions are made for a
relatively long time period of two to five years and determine the scope of the tac-
tical planning problems with a shorter horizon of 6 to 24 months (Kjeldsen (2011),
Schmidt and Wilhelm (2000)). Operational planning problems are typically solved
on a daily or weekly basis in the scope of the tactical planning decisions.
Figure 2.1 shows selected planning problems from the liner shipping industry
in the different planning horizons. The investment in new vessels and their type
are strategic decisions ((Kjeldsen, 2011, p. 130), Meng and Wang (2011b)) because
container ships can have a lifespan between 15 and 26 years ((Bendall and Stent,
1987, p. 339), World Shipping Council (2014)). Another long term decision of a liner
carrier is the relevant regional focus. Based on the existing fleet and the regions,
the network design problem faces the decisions which liner services can be offered
to the customers. This liner shipping network design problem is sometimes also
called the vessel routing problem (Kjeldsen (2011), Álvarez (2009)). Depending on
the publication, the planning horizon can vary from several years to several months
(Brouer et al. (2013), Reinhardt et al. (2007) and Reinhardt and Pisinger (2010)).
From the tactical point of view, a specific fleet must be deployed on the services (see
for example Álvarez (2009)). Another problem is fleet repositioning that optimizes
the transition between two liner networks (Tierney and Jensen (2012) and Tierney
et al. (2014)).
Operational planning problems are for example revenue management (Zurheide
and Fischer (2012)), environmental routing (Windeck (2013)), vessel schedule recov-
ery (Brouer et al. (2012), Meng et al. (2013)) and stowage planning (see for example
9
Strategic problems
Fleet size and Alliance Liner network

Region selection
mix selection design
Tactical problems
Fleet Buffer Fleet

deployment optimization repositioning
Operational problems
Speed Vessel schedule Bunkering Stowage Revenue Environmental

optimization recovery optimization planning management routing
Figure 2.1.: Selected planning problems arising in the (liner) shipping industry in
the context of service planning and operating.
Avriel et al. (1998), Imai et al. (2006), Delgado et al. (2012) and Pacino (2013)).
The scope of this thesis is the liner shipping network design problem in a strategic
and tactical planning horizon. The problem is defined in more detail in the remainder
of this chapter.
2.2. Basic Liner Shipping Network Design Problem

In this section, the basic network design problem is described in detail. The planning
task is to define one or more liner services that are offered to the customer. Stopford
extends the definition of a liner service (also called sling, string or just service) from
Fayle (2005) as:
”A liner service is a fleet of ships, with a common ownership or manage-

ment, which provide a fixed service, at regular intervals, between named
ports, and offer transport to any goods in the catchment area served by
those ports and ready for transit by their sailing dates. A fixed itinerary,
inclusion in a regular service, and the obligation to accept cargo from all
comers and to sail, whether filled or not, on the date fixed by a published
schedule are what distinguish the liner from the tramp.” (Stopford, 2009,
p. 512)
According to this definition, a liner service consists of a served port sequence (port
rotation) that performs a round trip, a (single) deployed vessel type and a number
10
of vessels. The connection between two ports is called a leg in shipping context.
The vessel type determines the volume available on a regular (often weekly) basis,
whereas the number of vessels are required to hold the fixed itinerary’s frequency
subject to the vessel type’s minimum and maximum speed. The importance of fixed
itineraries is also highlighted by (Brooks, 2000, p. 89) since the shippers and their
customers negotiate import and export contracts for a year or more. Thus, the
shippers plan with reliable services and do not appreciate carriers changing the
service characteristics too frequently (Brooks, 2000, p. 89). Today, the fleet of ships
on a single liner services can be operated by several carriers, for example in vessel
sharing agreements (see Section 2.6). Furthermore, carriers are not obligated to
accept any cargo, but instead are able to select the most profitable.
A liner shipping network contains several liner services that are connected at
different ports. The set of all ports is assumed to be given in the scope of this thesis.
Parameters involved in the port selection are, for example, the hinterland freight
rates (see Rodrigue (2013), (Levinson, 2006, p. 10 and p. 170)), port congestion
and charges (see Slack (1985)). The reader is referred to the work of Slack (1985),
Malchow and Kanafani (2001) and Lirn et al. (2004) for port evaluation strategies.
At almost all ports, transhipment operations can take place. Transhipment is
the transfer of goods from one transportation mode to another and/ or from one
liner service to another or from one ship to another. In the scope of this thesis,
transhipment is performed between two services at a specific port. An important
reason to perform transhipment is freight consolidation which increases the utiliza-
tion on legs that do not fully utilize a transport carrier (Ballou, 1999, p. 249). Ports
where large volumes of containers are consolidated are called transhipment hubs. For
further transhipment hub benefits, see (Mattfeld, 2006, p. 2). Usually, transhipment
operations take place in free-trade zones. Thus, goods can be stored in the zone
without being subject to import duties. Typically, ports allow the free storage of
transhipment containers for several days. For example, in the port of Hamburg, dry
transhipment containers are free of charge for six days (see EUROGATE Container
Terminal Hamburg GmbH (2014)). The main drawback of transhipment operations
are the increased transit times due to the storage of containers at the container
yards and the associated costs (see Section 2.5.2).
Transhipment imposes huge complexity when optimizing liner shipping networks.
In this paragraph some of the challenges involved with transhipment operations
are presented for the network given in Figure 2.2. The network consists of three
liner services, each with its own characteristics. Services typically follow different
trades, i.e. connections between markets (see Section 2.4). The network in Figure
2.2 serves trades between Northern Europe, South America and Asia. Tangier is a
transhipment hub where cargo is transhiped from on service to another, allowing high
utilization for the transport between all geographical regions. Each liner service calls
different ports to bundle regional demand. As can be seen in the example service,
Brazil has Salvador in the center of its coast line and several other ports in the
11
Figure 2.2.: Example liner network with three services.
South. Thus, a large quantity of Brazil’s demand is served by Salvador in the North.
Since Brazil exports a lot of fresh products, the port is called on the service’s North
bound to offer low transit times for the perishable goods for Northern Europe and
the Mediterranean region. The route from China to Europe calls the ports Shanghai,
Tangier and Rotterdam in each direction. On the west bound direction’s Tangier
call, Chinese cargo with the destination of the Mediterranean region is unloaded and
further transported by a feeder service. The cargo from the Mediterranean region
can be picked up in Tangier at the return journey back to China. Thereby, the
Asia service has enough capacity to transport cargo from Northern Europe to the
Mediterranean region. Before leaving Tangier to return to China, the Asia service
can pick up cargo from South America leading to higher utilization on the East
bound direction with laden containers (see Section 2.8 for trade imbalances). The
South America service is also used to transport cargo within Brazil and between
Brazil and Argentina. North bound of this service, a connection in Rio Grande and
Salvador is used to free capacity South bound and keep the duration between the
ports of origin and European destinations short.
Beside the port rotation, the used capacity is also fixed for a liner service. Capacity
is provided by the deployed vessel type. Stopford differs seven container ship classes
that are extended by (Konings, 2008, p. 115) by ultra large container vessel (ULCV)
(see (Stopford, 2009, p. 584)). These classes are given in Table 2.1. However, the
12
Vessel type Vessel type TEU capacity

Feeder 0-499
Feedermax 500-999
Handy 1000-1999
Sub-Panamax 2000-2999
Panamax 3000-3999
Post-Panamax 4000-5999
Very large box carrier (VLBC) 6000-12000
ULCV 1200-18000
Table 2.1.: Typical container vessel types (see (Stopford, 2009, p. 584) , (Konings,
2008, p. 115)).
first vessel with 18000 TEU maximum capacity was already deployed by Maersk
Line (see Maersk Line (2014)).
The vessel types in Table 2.1 can be distinguished in three groups: Feeder and
feedermax vessels with relatively low capacity are used on short-haul operations.
Short-haul operations are referred to as regional distribution operations for deep-
sea services. Handy and Sub-Panamax vessels can be used for short-haul feeder
operations but also longer distances where port draft constraints restrict the use of
larger vessels, for example in the North Sea (see (Stopford, 2009, p. 582)). Larger
ships are used on long-haul trades where they spend up to 80% of their round trip
time at sea. Each vessel type has its own characteristics such as speed, bunker
consumption or load depending draft. Details on the capacities and the stated TEU
are given in Section 2.7.1. Details on the vessels’ speeds is given in Section 2.5.1.
Typically a carrier only owns part of its operated fleet. Depending on the carrier’s
strategy, up to 100% of the used capacity can be chartered (see Alphaliner Newsletter,
Volume 1, 2011 (2014)). For purchased vessels, depreciation has to be paid and for
time chartered vessels a charter rate has to be paid to the shipowner. This rate
can be highly volatile, depending on the current market situation. In the scope of
this thesis, chartered and owned capacity is not distinguished and the time charter
rate is assumed to be constant. Successive planning steps that deploy either own
or chartered vessels to the liner services must ensure further constraints, such as
cabotage or manning, that are not in the scope of network design.
Brooks (2000) states that shippers rely on regular services. This regularity is
reflected by the liner services port call frequency. Consider the South America
service in example Figure 2.2 with seven deployed vessels using a weekly frequency.
This leads to a round trip time of 7 · 7 = 49 days for each vessel. This means that
each vessel has a maximum of 49 days to perform the journey. The round trips are
performed consecutively, removing any buffer time after a round trip. Delays are
thereby potentially propagated to the next round trip. For further details and an
example schedule, see Section 2.5.1.
The last basic aspect of a liner shipping network are the used seaways. Although
ships can use nearly every route between two ports, major shipping routes were
established in the last centuries. Figure 2.3 shows the routes and their usage. The
13
Main Maritime Shipping Routes
Oresund
Dover Strait
Bosporus Tsugaru
Gibraltar Dardanelles
Suez Canal
Strait of Hormuz Taiwan Strait
Yucatan Channel
Luzon Strait
Mona Passage
Windward Passage
Bab el-Mandab
Panama Canal
Strait of Malacca
Makassar
Sunda
Lombok Torres
Core Route
Secondary Route
Cape of Good Hope
Magellan Passage
Dr. Jean-Paul Rodrigue, Dept. of Global Studies & Geography, Hofstra University.
Figure 2.3.: Main global maritime shipping routes, Rodrigue (2013).
route selection and the exact distance between the ports rely on a large number
of factors, such as weather, security implications and costs and can differ on each
journey. For simplicity, it is assumed that the distance between two ports is constant
and externally given. To calculate the distances either distance providers such as
www.sea-distances.com (2014), www.sea-rates.com (2014) or the method presented
in Appendix E can be used.
Figure 2.3 shows the high volume trades between northern Europe and Asia
and the trade between the North America east coast and Asia. One can see that
carriers heavily rely on using canals for their transportation. Although the usage of
canals can cost several hundred thousand US$, the alternatives usually lead to much
larger distances and much higher bunker costs. The canals with the highest vessel
frequencies are today the Suez and the Panama canal. Using the Cape of Good Hope
instead of the Suez on a journey between Shanghai, China and Hamburg, Germany
results in an additional distance of about 4000 nautical miles (nm) and additional
11.9 days at 14 knots (kn). Assuming a bunker consumption of 43 tons per day with
cost per ton of 600 US$, this alternative journey would cost at least 1,500,000 US$,
without additional fixed cost such as manning, maintenance or additional vessels
to hold the weekly frequency. Thus, there is basically no alternative in using the
large international canals on direct connections. Since 2010, the Suez canal has
been able to deal with the deadweight of the largest currently deployed ULCVs (see
Suez Canal Authority (2014)). The Panama canal’s bottleneck are its locks that
are currently being expanded, which will allow container vessels of up to 13,000
TEU to pass the locks with the planned opening in 2015 (see (Canal de Panamá,
2014, p. 6)). This Atlantic-Pacific expansion enables routes from Northern Europe
14
2.3. Route Types and Network Structure
to Asia via America and could lead to the merging of existing services as well as
the implementation of new trade routes (see Rodrigue (2010)). For a detailed study
on the impact of the new Panama locks on the North America freight distribution
see Rodrigue (2010).
In the following sections, further details on the network design problem are given.
These include the liner service route types, transported cargo, timing aspects, ca-
pacities, costs and revenues.
2.3. Route Types and Network Structure

In general, liner services are classified as deep-sea and short-sea (also feeder) services.
Deep-sea services serve Transatlantic, Transpacific and Asian trades. Feeder services
are usually connected to one or more hubs and pick up or deliver cargo within a
smaller geographical region.
Additionally, liner services can be distinguished by their type of port rotation,
namely pendulum, cyclic, butterfly and conveyor belt routes (see (Andersen, 2010,
p. 28) and (Plum et al., 2013b, p. 2)). Pendulum and cyclic routes are often referred
to as simple; butterfly and conveyor belt routes as non-simple routes.
(a) Pendulum. (b) Cyclic. (c) Butterfly. (d) Conveyor belt.
Figure 2.4.: Liner services with different route types.
Pendulum routes in Figure 2.4(a), from a literature point of view, alternate be-
tween two ports. Cyclic routes call more than two ports, without calling a port
twice per round trip (see Figure 2.4(b)). Butterfly routes in Figure 2.4(c) are cyclic
routes that call one port twice. Conveyor belt routes (or routes with multiple but-
terfly ports) in Figure 2.4(d) call more than two ports and visit more than one port
twice per round trip. The advantages of butterfly and conveyor belt routes are the
increased capacity on single legs, decreased vessel draft due to increased port calls
and potentially improved transit times (see (Plum et al., 2013b, p. 2 and 3)).
In practice, pendulum routes are more often characterized as routes that transport
cargo without transhipment and usually call more than two ports. Discussions with
liner operators indicated that they rarely, if ever, call the same port more than twice
per round trip. This thesis uses the terminology for route types found in literature.
15
It is assumed that all routes perform a round trip, i.e. that the start equals the end
port.
2.4. Transportation of Containerized Cargo

In this section, the structure of the underlying demand, called cargo flows in shipping,
is explained in more detail. Furthermore, the available container types and sizes are
presented.
2.4.1. Demand Structure

Before containerization started in the 1950s, the transported cargo shipped by
sea was classified into four segments: Bulk cargo, liquids, and refrigerated cargo.
These segments where mainly served by tramp shipping or general cargo liners (see
(Stopford, 2009, p. 29)). The post- World War II liners transported a large variety
of cargo on the same vessel. These vessels required intensive work by longshoreman
to load and unload a vessel. Between 60 - 70% of the costs to transport cargo by
sea were accounted to the dock operations (see (Levinson, 2006, 21)), and each port
call required several days or weeks. A solution that was introduced in 1956 is the
containerization of cargo. This also let the shipping system shift to specialization:
homogeneous bulk cargo are transported by bulk carriers or tankers and liquids
or automobiles by specialized ships. General cargo is seen as small either loose or
unitized cargo that was containerized in the 1950s and is transported by container-
ships in standardized containers today (see (Stopford, 2009, p. 36)). General cargo
can be low value commodities such as scrap metal, lumber or newsprint, medium
valued goods such as textile, fruits or vegetables and high value commodities such as
consumer electronics and completely or semi knocked down automobiles (see Klug
(2010)).
The specific commodity that is transported via containers from an origin to
a destination depends on each country’s import and export trades. For example,
Germany imports the commodity groups coffee, fish meal and fruits from Peru,
whereas electrical machinery, apparels and furniture are imported from China (see
UN (2014)). The mentioned commodity groups are often transported via containers
by sea and specialized break bulk ships (e.g. banana ships), although fruits or high
value commodities compete with air freight transportation in case of high employed
capital costs or perishable goods.
The majority of cargo transported by carriers is based on medium term contracts.
The Federal Maritime Commission (2001) reports that in 2001 about 80% of the
cargo was transported under service contracts, whereas Yi (2008) reports up to 95%,
depending on the leg and the products. The contracts are usually negotiated once
per year between the carrier and the shipper (customer). Concerning the quantities
specified in the service contracts, 60% had a minimum commitment of 100 TEUs or
16
2.4. Transportation of Containerized Cargo
less (see Yi (2008)). The committed range was between 1 TEU up to 68,000 TEUs
(see Federal Maritime Commission (2001)). Beside the committed freight, carriers
also transport spot cargo. Spot cargo is short-term cargo that is typically not known
for certain when the liner services are planned. Thus, forecasts can be included in
the cargo flows used to plan the services.
In the scope of this thesis, cargo flows are used on a port to port basis without the
knowledge of the specific commodity, transported in the container. This is common
practice in shipping since it allows the focus on moving goods in standardized units
efficiently (see Levinson (2006)).
2.4.2. Container Sizes and Types

For the transportation of commodities, several standardized container types and
sizes exist (see International Organization for Standardization (2014)). The possible
lengths are 6.06m for a twenty foot equivalent unit (TEU), 12.19m for a forty foot
equivalent unit (FEU) and 13.72m for a 45 foot container. Containers are either
2.59m (8.6 ft) or 2.90m (9.6 ft) high, the latter are called high cube containers. For
details on the history of container size standardization, see (Levinson, 2006, Chapter
7).
There are a variety of container types that can be used to transport different
commodities (see Table 2.2). Fresh or perishable goods require the use of ventilated
or refrigerated (for chilled or frozen products) containers. Each liner carrier can
provide more commodity specific container interiors, such as for hanging cargo.
Between 2002 and 2012, especially reefer containers had a high combined annual
growth rate (CAGR) of 3.6% (see Research (2013)) which was partly driven by the
exotic fruits’ CAGR of 9.1%. In 2004, the worldwide container stock was estimated
at 28.5 million TEU, whereof 18% were TEU, 75% FEU, 4% reefer containers and
3% others (such as open top) (see (Stopford, 2009, p. 510)). The revenue and weight
varies per container type and is considered in this work. Considering the weight is
important to assure the correct vessel draft (see Section 2.7).
Name Purpose
General purpose container For any general cargo
Open top container Over high cargo or crane loading from top side
Flat container Heavy or oversize cargo and project cargo
Ventilated container Air circulating container, e.g. for coffee and fruits
Refrigerated (reefer) container Cooling or refrigerating container for example for meat
Tank container Transportation of chemicals, alcohols, fruit juices etc.
Table 2.2.: Selection of frequently used container types.
Especially perishable commodities transported in ventilated and reefer containers

often require a specific transit limited time between the origin and destination.
17
These durations are described in the next section.
2.5. Timing Aspects

Liner services are designed with a fixed round trip time that introduces several
requirements on the port and sea duration. In this section, the effects on the network
planning and the transportation of cargo subject to transit times are presented.
2.5.1. Port Calls and their Impact on the Vessels’ Speed

Historically, the majority of the deep sea services have been offered on a weekly
basis (see for example Agarwal and Ergun (2008), Yan et al. (2009), Stopford (2009)
and Song and Dong (2013)). Although liner shippers sometimes offer bimonthly or
biweekly port departures (see Reinhardt and Pisinger (2010)), variable frequencies
are rarely considered in network design (see Brouer et al. (2013) for a weekly and
biweekly formulations).
Depending on the frequency, a number of vessels has to be deployed on the service
to respect the service frequency. This results in a maximum port call duration to
perform a round trip for each vessel.
Figure 2.5 shows a deployed service that have to be served weekly. The num-
ber of vessels per service must be within the range resulting from the minimum
and maximum speed and the service distance. The service’s total distance is 6876
nm. Assuming deploying a vessel type with a minimum speed of 13 Knots (kn)
and 24 kn maximum speed, this would lead to 6876nm/13kn = 22.03 days and
6876nm/24kn = 11.94 days respectively, per round trip. Thus, d22.03/7e = 4 ves-
sels using minimum or d11.94.03/7e = 2 vessels using maximum speed are required
to hold the weekly round trip time. The same capacity is offered per week either
way because weekly round trips are assumed. The exact number of vessels is deter-
mined by the port durations (see Figure 2.6). It shows a service calling Hamburg,
305nm
Rotterdam Hamburg
3412nm
3159nm
Alexandria
Figure 2.5.: Round trip of example service with approximate leg distances.
18
2.5. Timing Aspects
Port Berthing
Casting off
Port duration
Sea duration
Rotterdam
Alexandria
Hamburg
Round trip time 14 days Time
Figure 2.6.: Duration of example service with two vessels serving the ports on a
weekly basis.
Alexandria and Rotterdam with two vessels and a weekly frequency.

A vessel’s round trip time is divided into the duration at sea and the duration at
ports. Increasing the port duration (by more container moves) decreases the time
at sea because the service’s distance stays constant. Therefore the speed must be
increased and this can lead to a capacity shortage when the maximum speed is
reached. A trade-off between increased bunker cost and time charter cost occurs
when more vessels are deployed to decrease the speed. Decreasing the speed lower
than 20 kn is called slow steaming in the shipping industry (see for example Rodrigue
(2013)). However, slow or even super slow steaming strategies can have a negative
impact on the transit time (see (Notteboom and Vernimmen, 2009, p. 335) and
Section 2.5.2). In practice, the port duration is determined by the following port
processes:
1. Berthing
2. Loading/unloading preparation
3. Cargo loading/unloading
4. Bunkering (i.e., fueling the vessel)
19
5. Port buffer to compensate uncertainties (such as port congestion etc.)
6. Casting off
The processes berthing and casting off are performed by the port’s tug boats and
typically take between one and several hours (for example six to seven hours through
the Elbe river to reach some of the container terminals in the port of Hamburg,
Germany). In the scope of this thesis, all processes except loading and unloading
preparation and bunkering are considered. The duration of loading and unloading
cargo at the ports is determined by the moved cargo. Other process’ durations are
given by external parameters that can be used to respect further port processes.
There is ongoing work in the research community that optimizes the additional
buffer in each port (see for example the work of Wang and Meng (2011), Wang and
Meng (2012a) and Wang and Meng (2012d)).
By using the port processes presented above, a proforma schedule can be gener-
ated. Proforma schedules define the arrival and departure time at each port within
a service and provide information on the duration at sea. Successive planning meth-
ods can adjust the port call durations, the buffer time and the leg specific speed
to meet the service synchronization. Synchronization is an important task when
transhipping perishable goods, adjusting the services with partners and handling
tide and container terminal berth windows.
2.5.2. Transit Times in the Maritime Context

Transit times are a major criterion for the service quality of a liner carrier (see
Notteboom (2006), Panayides and Song (2013) and Wang and Meng (2014)) and
highly influence the structure of the overall network. In general, transit times define
the maximum duration between two ports. Transit time can be seen from different
perspectives: From an ocean carrier’s point of view that offers schedules on a regular
basis, transit time can be defined “as the number of sailing days on a port-to-port
basis” (Notteboom (2006)). From a broader logistics perspective, transit time is the
door-to-door duration in days, including intermodal transportation systems. Transit
times are especially important for perishable commodities but also for the customers’
demand planning that is based on determined replenishment lead times (see (Busch
and Dangelmaier, 2002, p. 198)). Three factors determine the transit time between
ports:
1. Port sequence
2. Vessel speed and the duration at sea
3. Port call durations
20
2.5. Timing Aspects
The port call sequencing is crucial for the transit time: Ports that are called last
within a region and first in another region provide the lowest transit times. This is
especially important on the intercontinental trade-bundling services (see Notteboom
(2006)). The speed of the vessels determine the duration to steam the leg between
two ports and play another important role. Beside the days at sea, also the port call
duration highly influences the transit time. Today, liner carriers make increasing
use of hub-and-spoke systems that connect major deep-sea services with short-sea
services (see (Hsu and Hsieh, 2005, p. 209) and (Imai et al., 2009, p. 756)). Because
transhipment operations at these ports also impact the transit time, the duration
of transhipment operations must be considered as well.
In the scope of this thesis, transit times are considered on a port-to-port basis
due to the unknown commodity groups that are transported. The transit times are
regarded as a strategic or tactical decision of the carrier to create competitive liner
networks.
1 day
Rotterdam Hamburg
5 days
5 days
34 days
Alexandria Shanghai
34 days
Figure 2.7.: Example transit time between Rotterdam and Shanghai using service
A (black) and B (gray).
To get an insight into how transit times are considered in this thesis, see Figure
2.7. For simplicity, assume a fixed port duration of 24 hours per port call. Service A
calls Rotterdam, Hamburg and Alexandria with two vessels whereas service B calls
Alexandria and Shanghai with 10 vessels. The leg durations already consider the
distance and required speed. The transit time between Rotterdam and Shanghai is
the path in the network with the minimum duration: From Rotterdam to Hamburg
and Alexandria on service A, transhipping to service B and then from Alexandria
to Shanghai. The duration is approximated as 1 · 4 days at the ports and 1 + 5 + 34
days at sea, which is a total of 44 days. In the scope of this thesis, a constant
transhipment duration is added to include the duration containers are stored in the
container yards during transhipment.
The example shows two important aspects: First, a badly created network can
highly decrease the transit time performance. This can be partly compensated for
21
by increasing the speed of the incident services, if possible. Otherwise, the network
structure must be changed to respect the transit times. In the example in Figure
2.7, service B could call Rotterdam in its port rotation to reduce the transit time
between Rotterdam and Shanghai.
Second, the transit time is only relevant if a network calls ports that are con-
strained by a maximum transit time. If either one port is not called, no link between
the two ports exist or no cargo is transported between two ports, transit times
become irrelevant for this connection. However, as soon as both ports are called, a
link exists and cargo is transported, transit times must be respected.
2.6. Cooperative Agreements

Operating services on a regular basis leads to relatively inflexible port calls. Seasonal
demands, volatile cash flow due to volatile freight rates and large costs associated
with new vessels force carriers to search for opportunities to manage these challenges.
One way to tackle the difficulties is to cooperate with other carriers and is described
in this section.
2.6.1. Liner Conferences, Consortium Agreements and Global Alliances

The liner conference system, developed in the mid-1870s, is the industry’s oldest
and, in the past, most important kind of cartel-type agreement (see (Stopford, 2009,
p. 556), (Basedow et al., 2012, p. 6)). The membership of a liner conference could
be either closed or open subject to government agency monitoring (see (Wong and
Bamford, 2012, p. 3)). The common objective of the conference members was to
“fix and agree on schedules, in order to rationalize the capacity and the frequency of
services offered to their customers, as well as the tariffs that are publicly available”
(Basedow et al., 2012, p. 6).
Starting from October 2008, all carriers operating in the EU are subject to regula-
tion 4056/86 that removes the block exemption for the liner shipping industry from
the 1958’s Treaty of Rome on antitrust. In short, carriers are not allowed to operate
in price or capacity fixing conferences when offering routes into or out of Europe.
Shipping conferences outside of Europe are not affected by the EU regulation but
other anti-trust laws, such as the American Ocean Shipping Reform Act that took
effect in May, 1999. Today, most of the main trade lanes are subject to anti-trust
laws (see for example Shipping Watch (2014)).
As a result of increasing anti-trust popularity in the 1990s , the carriers switched
to other forms or cooperation, such as consortia, alliances and mergers.
Consortia, or consortium agreements, offer joint services by two or more carriers
on the same shipping route (see EU Article 1 of the Commission Regulation 870/95).
On these services, cross-slot charters are used to reserve capacity for each member
of the consortium. Subject to the anti-trust law, each member sets pricing and
22
2.6. Cooperative Agreements
Figure 2.8.: East-West Carrier Alliances 2013, from Alphaliner Newsletter, Volume
26, 2013 (2014). The implementation of the P3 alliance was rejected by
Chinese regulators in July, 2014 (see for example List (2014)).
transportation conditions with shippers individually (see (Basedow et al., 2012, p.

7)).
“A global alliance is a sort of consortium whose geographic scope is not a single
trade, but is instead worldwide” ((Basedow et al., 2012, p. 7)) or on certain major
trades (see (Panayides and Wiedmer, 2011, p. 2)). Global alliances effects on hor-
izontal integration (which refers to vessel employment and utilization) have been
stated as a key change in the last 20 year in the liner shipping industry (see (Lun
et al., 2010, p. 63)). The main advantages of alliances are the increased geographical
scope, possibility to perform vessel planning and coordination on a global scale, risk
and investment sharing, economies of scale, entry in new markets, increasing service
frequencies and combining purchasing power (see (Midoro and Pitto, 2000, p. 33)).
Similar to consortium agreements, global alliances can contain specific collabora-
tive agreements such as vessel sharing or slot sharing agreements (explained in the
next section). In vessel sharing agreements partners cooperate by optimizing their
container-vessel assignment to improve the utilization on specific legs or synchroniz-
ing the vessel departures (see (Panayides and Wiedmer, 2011, p. 2)). In Figure 2.8,
global alliances such as the Grand Alliance and CKYH for the East-West trades
are shown.
Cooperation plays an important role when designing the networks for each carrier
individually. Cooperation is not only relevant for the main carriers but also for the
23
small ones (see (Lu et al., 2010, p. 325)). The following sections provide both the
legal and economic overview on how cooperation is performed and how it can be
considered in the liner shipping network design.
2.6.2. Slot Sharing Agreements

An important form of cooperation is a slot sharing agreement. Within a consortium
or strategic alliance, partners share free capacity to maximize their vessel utilization.
Some authors claim that operational partnerships will become even more important
(see (Sys, 2009, p. 265)). A carrier can sell slots on an operated service, purchase
slots from a partner or do both with a slot swap.
Service Operator Deployed Allocation Allocation Reefer Allo- Reefer Slot
Vessel @ 14 tons @ 14 tons cation per leg Sur- Rate
EB/SB WB/NB charges
Service 1 Carrier 1 4000 3000 3500 150 - -
Service 2 Carrier 2 2000 500 300 100 150 800
Service 3 Carrier 3 2000 400 - 20 150 300
Table 2.3.: Example slot sharing agreement between three carriers.
Table 2.3 shows an example agreement between three carriers within a given
trade. Carrier 1 operates a service with a 4000 TEU vessel type. In the East and
South bound direction of its service, 1000 slots are available can be sold out. Of
these, carrier 2 purchases 500 and carrier 3 400 slots. On the West and North bound
direction, carrier 1 has more own cargo to transport and thus decreases its sold
slots. Carriers 2 and 3 call ports that are not served by carrier 1 and thus slots
are purchased at a specific rate. Carrier 1 has committed by contracts to transport
cargo for its partners but benefits from its increased network. If carrier 1 decides
to change its network, at least the already called ports have to be called due to
the contracts. However, due to the already mentioned anti-trust laws, the carriers
are not allowed to exchange their specific cargo flow information. Instead, it is
assumed that a carrier that wants to optimize the services (here carrier 1) can only
approximate the partners’ cargo flows between the previously called ports. When
he chooses to optimize the liner services he tries to decrease its slot costs, leading
to an imbalance between the slot rates (slot price) and the actually implemented
slot costs on his own service.
Slot sharing can be agreed on any subset of the legs within a liner service. In Table
2.3, the East, West, South and North bound directions represent these subsets.
2.7. Liner Service Capacities

In this section the liner service capacities are described in detail. Because vessels (of
a specific type) are assigned to the liner services, the capacity is referred to as the
24
2.7. Liner Service Capacities
liner service capacity. However, the service and vessel type capacities result from
specific vessels whose tonnage is subject to several physical constraints that also
apply to a vessel type. The assignment of a vessel type and a number of vessels to
a service is called a vessel deployment.
2.7.1. Container Vessel Capacities

The amount of containers on a vessel is limited. Usually, a vessel’s capacity is given
by its TEU tonnage (see (Schönknecht, 2009, p. 5)). This capacity is limited by
the container size and type and assumes a net weight of 14 tons per container. In
practice, a TEU container weighs between 2250 kg (empty) and 24000 kg (fully
loaded). Similar ranges hold for forty foot and high cube containers. Stowing a
container vessel with forty foot containers reduces the transportable containers by
approximately 50%, because their length is double compared to a TEU. A vessel
might not only transport dry containers but also reefer containers. These containers
need electricity and thus special container plugs to cool or freeze cargo within the
container.
For example, the container vessel “CHARLOTTE MAERSK” is currently oper-
ated on Asia – Europe trades and has a total nominal container capacity of 10,457
TEU and 700 reefer plug slots. Furthermore it has a maximum deadweight of 105,000
tons. It can transport 6,600 containers that weighs 14 tons each. The large amount
of reefer container plugs can also be used to stow dry containers.
Summarizing, a container vessel has the three capacity types: Dry slots, reefer
container plugs and maximum deadweight. These capacities must already be con-
sidered in strategic and tactical planning methods because every type can limit the
amount of containers on a vessel.
From an operational point of view, the capacities must be considered in much
more detail regarding the container content, minimum stowage distances between
containers and the vessel stability. The reader is referred to the work of Avriel et al.
(1998) and Imai et al. (2006) for details on container vessel stowage planning.
2.7.2. Port Depth and Vessels’ Deadweight Scale

The ports called within the services often have a maximum depth allowing vessels
to enter the ports. The ports’ depths depend on the tide levels, the geographical
position (salt water percentage) and the infrastructure offered by the port at the
berths. The depths can vary between 8 meters in small feeder ports up to 18 meters
in deep sea ports such as the Jade Weser port in Wilhelmshaven or 17.5 meter in
Shanghai (see World Port Source (2014)). These allow large vessels to enter the ports
independent of the tide levels that can influence the ports’ depth by several meters.
Tide levels become important when scheduling the liner service and reserving berth
windows at the container terminals. In the scope of the network design, tide levels
25
Draft Deadweight
Feet Meter
14
TF
45
F T
13 S
Summer Draft: 13.4 m
W
Lightship Draft: 4 M
Figure 2.9.: Example deadweight scale for a Panamax vessel.
are assumed to be constant and a maximum depth is used for each port.
A vessel has a lightweight displacement (based on the weight completely unloaded,
considering only the hull, machinery and installed equipment). This displacement
results in the vessel’s lightship draft. In practice, the draft of a vessel is subject
to the deadweight of the vessel. Depending on the loaded cargo, the vessel’s draft
increases according to the vessel’s deadweight scale, shown in Figure 2.9.
According to the deadweight scale, the example vessel has a draft of 4 meters.
When the vessel deadweight is nearly 60000 tons deadweight (tdw), the vessel draft
increases to 13.4 meters. For a specific vessel operating on a liner service, the
maximum deadweight and draft are subject to the geographical region and the time
of year according to the plimsoll mark. In the summer region (S) the maximum
draft is 13.4 meter, leading to a maximum deadweight of approximately 60000 tdw.
The zones are defined according to waves and winds: In the summer region (S), not
more than 10% and 8 Beaufort (34 kn) wind, in the tropical region (T) not more
than 1% or 8 Beaufort winds are allowed and further constraints regarding storms
exist.
For the maximum vessel types’ draft in the scope of this thesis, the summer
load line is used as an upper bound because it is the most restrictive. Example
deadweight scales from an ocean carrier’s vessel types indicate that the scales can
be linearized with relatively small errors. To calculate the draft it is assumed that
the draft only depends on the loaded cargo and the whole deadweight can be used to
transport cargo. Since the specific ports for bunkering are unknown in the planning
horizon, no constraints on the bunker amount are imposed. For details on how the
displacement and the draft of a vessel is calculated, see (Barrass, 2006, p. 383) and
26
2.8. Empty Container Repositioning
Ayre (1885).
Depending on the draft, a vessel cannot call all ports fully loaded. This is a key
problem for the increasingly larger vessels with larger draft. Ports must be called in
such a way that the vessels’ draft fits the port’s depth (see (Baird, 1996, p. 148) and
(Midoro et al., 2005, p. 98)). This requires the ocean carriers to carefully design the
liner services in such a way that draft constrained ports within a service are called
after ports with a larger import ratio to unload enough cargo from the vessels to
pass the draft - depth constraint (see (Stopford, 2009, p. 354)).
2.8. Empty Container Repositioning

In Section 2.4.1, the trade between China and Germany is described. The exam-
ple indicates that trade volumes between countries often differ from each other,
leading to trade imbalances. This has been a major problem since the beginning of
container shipping (see Levinson (2006), Shintani et al. (2007)). Figure 2.10 shows
the calculation of trade imbalances on the main shipping trades according to Fan
et al. (2012). The figure shows a trade imbalance of about 40% between Asia and
the US and South America, meaning the cargo volume is 40% higher from Asia
to America compared to the trade from America to Asia. Similar, the trade from
Asia to Europe is about 60% higher than the eastbound trade in 2009. The exact
calculation of the trade imbalance between two countries or regions differs in the
scientific community (see Xu et al. (2010), Fan et al. (2012)).
Ocean carriers often own or lease a stock of containers to brand containers for
marketing purposes (see (Notteboom and Rodrigue, 2008, p. 168)) and container
pools or the gray box concept are still not accepted in practice (see (Notteboom
and Rodrigue, 2008, p. 168) and Notteboom and Rodrigue (2007)). From the liner
shipping point of view, the effect of trade imbalances are stocks of empty containers
at different ports. Figure 2.11 shows the overall empty container repositioning flow.
In the example, customer A wants to transport one container to customer B. He
rents a container from the carrier that transports the empty container from either
Figure 2.10.: Trade imbalances on the main trades, (Fan et al., 2012, p. 249).
27
Figure 2.11.: Empty container repositioning using depots (on the basis of (Lai et al.,
1995, p. 689)).
a hinterland or port depot to the customer. The customer stows the container with
goods and moves the container to the port of origin’s container terminal. There,
the container is transported by sea to the port of destination (transhipment ports
can be used in between as well). Trucks, trains or both transport the container to
the final customer where the cargo is unloaded. Afterwards, the empty container at
customer B must be transported to an empty container depot in a port or hinterland.
From a more operational point of view, the container can be transported directly
to customer C for a new export.
This example shows that a lot of aspects are involved in the empty container
repositioning. Additional aspects such as depot optimization, container leasing or
container safety stocks have to be considered in the operational planning process
as well. For planning models focusing on the empty container depot optimization,
see (Flapper et al., 2005, p. 65 ff), Boile et al. (2008) and Mittal (2008), for empty
container safety stock and flow optimization see Feng and Chang (2008) and Erera
et al. (2009).
The overall empty container repositioning costs are estimated at 27% of the total
world fleet running cost (Song et al., 2005, p. 15). Thus, it is important to include
the empty container repositioning in the network design. From a strategic point
of view, specific customers and hinterland depots are often unknown because they
are determined by the network design. The hinterland repositioning costs (such
as intermodal transportation cost and safety stocks) are considered by a constant
cost factor per transported cargo flow unit. The main challenges arising with empty
containers are the increased vessel utilization and the opportunity cost by unavailable
slots to transport cargo flows. In the scope of this thesis, each transported cargo
flow requires an empty container of the cargo flow’s container type and size at its
port of origin and provides an empty container at its port of destination. Thus, a
repositioning (or balancing) of the empty containers must be performed. Note that
empty containers mainly use the capacity type of dry slots, whereas the deadweight
utilization is relatively low.
28
2.9. Costs and Revenues

Operating a liner network results in different costs and revenues gained from the
cargo’s shipper. Stopford defines building-blocks of liner shipping costs (Stopford,
2009, p. 539 ff):
1. Ship costs per day
2. Port charges (excluding cargo handling)
3. Container and container handling costs
4. Administration costs
According to (Stopford, 2009, p. 539 ff), ship costs per day include the operating
costs (OPEX), capital costs for the time charter (or vessel depreciation) and bunker
cost. Port charges are both, per TEU and per call. Container costs and container
handling costs include leasing, maintenance, repair, terminal, refrigeration, tranship-
ment, hinterland transportation and cargo claim cost. Of these costs, transhipment
and hinterland transportation are the largest cost factors. Administration costs are
calculated as cost per employee involved in running the enterprise.
In Figure 2.12, the costs per service voyage for different vessel types are given.
The vessel cost ratio decreases with the size of the vessel, whereas the container
handling and onward hinterland transportation costs increase with the vessel size.
This can be explained by increased port handling and transhipment costs because
larger vessels generally lead to more transhipment operations due to reduced port
calls. The economy of scale through large vessels decreases the total voyage related
costs per TEU from 1,299 US$ to 1,023 US$ for a 4300 TEU ship down to 902 US$
per ship, which is a total decrease of 69.44%.
In the scope of the network design, the following costs were identified with network
planners of a global liner carrier:
1. Bunker cost per day at sea in the planning horizon
2. Vessel capital cost in the planning period
3. Port cost per call per vessel type
4. Container handling costs (transhipment costs, loading and unloading cost at

the port of origin and destination and equipment cost per container type
including hinterland repositioning, depreciation and leasing cost)
5. Slot purchasing cost at partner services
29
Figure 2.12.: Voyage costs per vessel type (see (Stopford, 2009, p. 542)).
In Section 2.8, empty container repositioning costs were described in detail. In

the strategic planning, hinterland and safety stock costs are not included. Rather,
for each served laden container a fixed equipment depreciation must be paid. Fur-
thermore, for each loading and unloading operation of an empty container at a port,
fixed cost must be paid.
In general, the largest cost components are, depending on the liner service, the
capital, bunker, transhipment and slot purchasing costs.
The capital cost can be either time chartering costs or costs for running own ves-
sels (beside others, this includes depreciation, insurance, maintenance and manning
costs), depending on whether the carrier has bought the vessel or charters the vessel
from the charter market. Note that both costs are unified for the remainder of this
thesis as charter cost. For example, the capital costs for a 11,000 TEU vessel is
about 46,301 US$ per day. This is calculated with a depreciation period of 20 years
with an interest rate of 8% p.a. (see (Stopford, 2009, p. 540)).
In the literature (see Álvarez (2009), Brouer et al. (2013)), the vessels’ bunker
cost are often calculated approximately based on the cubic bunkera consumption
function proposed by (Stopford, 2009, p. 234): F (S) = F ∗ SS∗ , where a depends
on the vessel’s engine and is about 2 for steam engines and 3 for diesel engines.
30
Figure 2.13.: Bunker consumption in tons per day for different vessel types, (Notte-
boom and Cariou (2009)).
Notteboom and Cariou (2009) also consider the age and the power of the vessel and
find the bunker consumption subject to the speed in knots presented in Figure 2.13.
Figure 2.13 shows that the consumption highly varies with the speed and differs
from the formula in Stopford (2009): a 5000-6000 TEU vessel’s bunker cost per
day, assuming bunker cost per ton of 600 US$, are 88,768 US$ at 21 kn using
the data given in Notteboom and Cariou (2009) and 79,048 US$ using Stopford’s
formula. Wang and Meng (2012c) analyze the cubic approximation in more detail
and conclude that the cubic approximation should only be used when no historical
data is available. In Figure 2.14 both, the approximation from Stopford (2009) and a
polynomial of order three is fitted to the existing historical data for different speeds
(also known as bunker profiles). Below 17 knots, the cubic approximation becomes
clearly worse than the polynomial approximation using a linear least squares fitting
(see for example Bevington and Robinson (1969)).
Notteboom and Vernimmen (2009) mention the large bunker cost impact on the
liner services and thus the liner network design problem. The (super) slow steaming
strategy became more relevant in the last years. This thesis uses polynomial bunker
consumption in the real-world test instances and cubic bunker consumption curves
in the artificial instances. Thus, the approximation errors in the real-world instances
can be reduced. Based on the bunker consumption, the bunker cost are calculated
by multiplying the consumption per day at sea in the planning horizon for each
vessel and the bunker costs per ton, currently about 658 US$ per metric ton in
Busan and 585 US$ in Rotterdam (see Bunker Index (2013)).
The costs introduced above decrease the profit of a carrier. Within the network
design problem in a strategic or tactical planning horizon, an ocean carrier seeks to
maximize its profit (see Álvarez (2009), Agarwal and Ergun (2008) and Brouer et al.
(2013)). A carrier gets revenue from the shipper for transporting cargo of a specific
type from its origin to its destination port using one or more liner services. The
31
Figure 2.14.: Example bunker profile for a 4300 TEU vessel. The y-axis is scaled
logarithmically to the basis of 10.
carrier specific revenue per cargo flow depends on the trade region, the deployed
capacity (supply), the container type and other additional offered services, such as
intermodal transportation. The cargo can be either contracted with shippers in a
medium or long term or obtained in a short term through the spot market. Ocean
carriers tend to prefer having contracts with a negotiated freight rate to have a
higher certainty in their planning processes. The revenue often include surcharges
such as the bunker adjustment factor (BAF), the peak-season surcharge or the
currency adjustment factor (CAF), allowing the carrier to compensate for some of
the uncertainties.
Using the revenue and costs, the earnings before interest and taxes (EBIT) of a
network can be calculated. In the scope of this thesis, EBIT is referred to as profit.
2.10. Bunker Cost Uncertainty in the Tactical Planning Horizon

In the strategic and tactical planning horizon, several uncertainties exist that can
influence the network structure, whereas cargo volumes and freight rates as the
largest uncertainties, see (Buxton and Stephenson, 2001, p. 298). Stopford (2009)
32
Bunker Index 2009 - 2013

1800
Average Global Bunker Price (US$/mt)
1600
1400
1200
1000
800
600
400
200
0
Mrz Mai Jul Sep Nov Jan Mrz Mai Jul Sep Nov Jan Mrz Mai Jul Sep Nov Jan Mrz Mai Jul Sep Nov Jan Mrz Mai Jul Sep Nov
2009 2010 2011 2012 2013

Time
Figure 2.15.: Bunker Index, aggregated data from Bunker Index (2014).
state the dominant bunker costs for liner shipping networks.

As indicated in the previous section, bunker costs are a large factor of the overall
costs and play an important role when designing networks. Figure 2.15 shows the
average prices over time and the deviation from the average in different ports. The
data is given by the Bunker Index (BIX) that collects and aggregates bunker prices.
Currently, the data set contains 157 port prices. The average global bunker price
per ton increased from 350 US$ per ton to 800 US$ per ton in the last four years,
giving it an increasingly important role when designing networks.
As a result, carriers implement slow or even super slow steaming strategies to
decrease bunker consumption. When liner services are implemented, a planning
problem is to decide where and how much to bunker on each port within a round
trip.
Figure 2.16 shows the average price changes on a daily basis. The price is highly
volatile and increased by more than 30% between 2010 and 2011. However, due
to computational challenges involved with the liner network design problem, this
thesis assumes a constant bunker price. To deal with the volatile bunker cost in
the industry, sensitivity analysis should be performed to identify the bunker price
impact on the network structure.
33
Monthly changes of the Average Global Bunker Price (in %)

20%
Change of the Average Global Bunker Price (in %)
15%
10%
5%
0%
Mrz Mai Jul Sep Nov Jan Mrz Mai Jul Sep Nov Jan Mrz Mai Jul Sep Nov Jan Mrz Mai Jul Sep Nov Jan Mrz Mai Jul Sep Nov
2009 2010 2011 2012 2013

-5%
-10%
Time
Figure 2.16.: Monthly BIX changes, data from Bunker Index (2014).
2.11. Summary
In this chapter, the liner shipping network design problem, its objective and its
practical constraints are introduced. The problem’s objective in the scope of this
thesis is to maximize the profit in the planning horizon. The revenue for transporting
containers should be maximized under consideration of bunker, handling, container,
port and charter cost. Furthermore, cost for using slots on partner services are
imposed. The objective is subject to the following requirements:
1. Ensure a valid container flow in the network
2. Perform a balancing of empty containers
3. Consider container vessel capacities
4. Respect the port depth with load dependent vessel drafts
5. Ensure transit times between ports
It has been shown that the network design in liner shipping is a complex planning
problem with many practical requirements. In the next chapter, the state-of-the-
art regarding Operations Research methods in liner shipping is analyzed and the
research goals are derived from the research gap.
34
3. State-of-the-Art and Research Opportunities
This chapter introduces selected optimization techniques for solving combinatorial

optimization problems. Related planning problems, such as the vehicle routing prob-
lem, the pickup and delivery problem and the min-cost flow problem, are presented.
Based on these fundamental planning problems, the state-of-the-art in strategic mar-
itime liner research is given. In particular, publications on the liner shipping network
design, the cargo allocation, empty container repositioning and speed optimization
are presented.
3.1. Selected Optimization Techniques
In this section, a selected overview of the state-of-the-art methods to solve combina-

torial optimization problems is given. The introduced methods are selected to cover
the solution approaches developed in this thesis. First, techniques to solve optimiza-
tion problems to optimality and afterwards heuristic approaches are presented.
3.1.1. Linear and Mixed Integer Programming
Linear Programming has been used since before World War II to solve linear opti-
mization problems (see (Suhl and Mellouli, 2013, p. 9)). A (compact) linear opti-
mization problem is defined as
zLP = max{cx : Ax ≤ b, x ∈ Rn+ } (3.1)
with an m × n matrix A, 1 × n vector c and m × 1 vector b (see (Nemhauser and

Wolsey, 1999, p. 27)). x is a vector of dimension n and also called a vector of
decision variables. The function cx that is either minimized or maximized is called
the objective function. The inequalities defined as Ax ≤ b are called constraints.
For a detailed introduction on linear programming, see (Suhl and Mellouli, 2013, p.
33ff).
Problem (3.1) can be rewritten in another notation that is more common (see
Problem (3.2)). The vector of objective coefficients c is represented by a parameter
cj for each variable xj of the optimization problem. The right hand side b is defined
as the parameter bi for each constraint. The optimization uses a parameter aij for
each row i and column j instead of the matrix a.
35
n
X
(P ) zLP = max cj x j (3.2)
j=1
s.t
X n
aij xj ≤ bi ∀i = 1, 2, · · · , m (3.3)
j=1
xj ≥ 0 ∀j = 1, 2, · · · , n (3.4)
Efficient solution approaches to solve large scale liner programs exist, such as the
simplex method or the interior point method (see (Suhl and Mellouli, 2013, p. 46),
(Chvátal, 1983, p. 13) for the simplex method, (Chvátal, 1983, p. 97) for the revised
simplex method, (Nemhauser and Wolsey, 1999, p. 37) for the dual simplex method
and (Chvátal, 1983, p. 443) for the interior point method).
Integer and mixed integer programming extend linear programming (3.1) with
integer variables. This introduces computational complexity for solving the problem.
The program (3.5) is called an integer program (IP) because it only consists of
integer variables.
(IP ) zIP = max{cy : Ay ≤ b, y ∈ Zn+ } (3.5)
In comparison, (3.6) defines a mixed integer program where y is a n-dimensional

vector of integer and x a p-dimensional vector of continuous variables.
(M IP ) zM IP = max{cx + hy : Ax + Gy ≤ b, x ∈ Rp+ , y ∈ Zn+ } (3.6)
Mixed integer programs (MIP) allow much broader model formulations because
logical relationships, fixed cost etc. can be expressed using the integer (or binary)
y variables. Note that a MIP also enforces a linear combination of variables in the
objective function as well as the constraints. When relaxing the integer requirement
of a MIP, the resulting model is called the LP relaxation.
A drawback of MIPs is that the efficient solution methods to solve LPs are not
applicable to solve MIPs. The activity of variables in the optimal solution(s) of the
simplex method can be fractional whereas they are forced to be integer in order to
be feasible for the MIP (see (Suhl and Mellouli, 2013, p. 10)). Therefore, different
general algorithms such as Branch & Bound (B&B) and Branch & Cut (B&C) were
developed. B&B basically decomposes the problem into smaller problems. Starting
with the root node, B&B branches (or divides) the original problem into smaller
problems by choosing appropriate values for currently fractional integer variables.
36
Each resulting problem is solved using the LP relaxation and again separated into
smaller problems and so on. This approach results in a complete enumeration of the
search space and is impractical for large scale problems. Bounding is used to prune
nodes in the search tree that cannot contain better solutions as the currently best
found solution (incumbent). For details on the bounding step of B&B see (Wolsey,
2000, p. 94). For further information on B&B, see (Nemhauser and Wolsey, 1999,
p. 355).
In most of the state-of-the-art MIP solvers such as Cplex (see IBM CPLEX
(2014)), XPress (see Frontline Solvers (2014)) or Gurobi (see Gurobi (2014)), a
B&C is implemented that additionally uses preprocessing methods and heuristics
to improve the runtime. B&B will be used as an acronym for using a MIP solver as
solution method throughout this chapter.
3.1.2. Delayed Column Generation

“Column generation (CG) is nowadays a prominent method to cope with a huge
number of variables” (Lübbecke and Desrosiers, 2005, p. 1007). Among the first
suggesting delayed column generation to solve large linear programs were Gilmore
and Gomory (1961).
In some problems, the number of possible columns (variables) in problem (3.1) is
too large to be completely enumerated. The problem containing all columns is also
called the master problem. In some cases, it can be reasonable to work on a subset
of columns, instead of the problem containing all columns. The master problem
becomes a restricted master problem (RMP) when only a subset of the decision
variables are used:
(RMP) zRMP = max{c̃x̃ : Ãx̃ ≤ b̃, x̃ ∈ X} (3.7)
The main idea is to successively solve the RMP, retrieve information from the
optimal solution and use it to extend the RMP by columns that potentially improve
the objective value, through a subproblem called the pricing problem (subproblem).
The information relevant to the pricing problem are the dual values of the RMP
constraints. These represent the impact of changes of the constraints’ right hand side
on the objective value. The overall column generation solution process is outlined
in Algorithm 1.
The RMP is initialized with a set of (artificial) columns X 0 that is successively
extended. After the RMP is solved, the (relevant) duals are stored for the pricing
problem (subproblem) PRICING(πi ) to generate new columns that can improve the
objective value. Each new column is added to the RMP if it has positive reduced
cost (for a maximization problems). According to the duality theory, this can but
need not improve the objective value of the RMP. Typically, the pricing problem
37
is an optimization problem itself, such as a shortest path problem.
Algorithm 1: Delayed Column Generation Procedure

1 Create initial set of columns X
0
2 Solve the RMP and store duals πi incident to row i

3 Solve the pricing problem P RICIN G(πi ). Terminate if no columns with
positive reduced costs found
4 Add columns to RM P and goto line 2
The efficiency of the column generation method relies on efficient solution methods
for the pricing problem and (relatively) small restricted master problems. If the
master problem is a linear program without integer variables, column generation
converges to the global optimality of the master problem. For a detailed introduction,
survey papers and implementation details on column generation see Chvátal (1983),
Lübbecke and Desrosiers (2005) and Desaulniers et al. (2005).
3.1.3. Metaheuristics
Metaheuristics aim to find approximate solutions in a reasonable time. They do
not guarantee optimality as in the previously described methods. “A metaheuristic
is formally defined as an iterative generation process which guides a subordinate
heuristic by combining intelligently different concepts for exploring and exploiting
the search space, learning strategies are used to structure information in order to
find efficiently near-optimal solution” (Osman and Laporte, 1996, p. 513 and 514).
Well known metaheuristics include simulated annealing (SA), genetic algorithms
(GA), evolutionary algorithms (EA) and, increasingly, variable neighborhood search
strategies (see Mladenović and Hansen (1997)). Genetic algorithms and variable
neighborhood search methods are presented in the following paragraphs.
Genetic Algorithm A genetic algorithm (GA) was first used by Holland (1975)
to solve combinatorial optimization problems. Genetic algorithms are based on a
set (population) of individuals that is successively improved. Darwins survival of
the fitness principle is used to optimize the whole population according to a fitness
function. Glover and Kochenberger (2003) formulate a general version of a genetic
algorithm as shown in Algorithm 2. The very basic version of a genetic algorithm
works as follows: Each individual is associated with chromosomes that represent their
properties (used to represent a solution for the problem at hand). At the beginning of
the algorithm, an initial population is created. Afterwards, the algorithm is executed
until the termination criteria, such as a maximum runtime or iterations without
a population improvement, is reached. In each iteration, a new individual (child,
offspring) is created from existing individuals (parents) using their chromosomes.
38
Algorithm 2: Genetic Algorithm (see (Gendreau and Potvin, 2010, p. 113))

1 Choose an initial population of chromosomes while termination condition not
satisfied do
2 repeat
3 if crossover condition satisfied then
4 select parent chromosomes
5 choose crossover parameters
6 perform crossover
7 end
8 if mutation condition satisfied then
9 choose mutation points
10 perform mutation
11 end
12 evaluate fitness of offspring
13 until sufficient offspring created;
14 select new population
15 end
As in the evolution process described by Darwin, mutation can randomly change

chromosomes of an individual. This is repeated until a predetermined number of
offspring are created. Finally, individuals from the previous population and offspring
are selected for the next generation based on their fitness (objective function).
Several extensions to this classical approach exist, such as different mutation
or parent selection strategies. One particular extension is the use of local search
procedures on the individuals. In this case, the genetic algorithm is also called
a hybrid genetic algorithm (see for example Gonçalves et al. (2005)) or memetic
algorithm (see for example Moscato et al. (1989)).
For the theoretical background and extensions of the GA see Michalewicz and
Fogel (2004), Glover and Kochenberger (2003) and Holland (1975). In the remainder
of this thesis, we will speak of evolutionary algorithms (EA) instead of genetic
algorithms due to the missing chromosome structure (see Kruse et al. (2013)).
Variable Neighborhood Search Variable neighborhood search is a metaheuris-

tics proposed by Mladenović and Hansen (1997). “The basic idea is a systematic
change of neighborhood within a local search” ((Glover and Kochenberger, 2003, p.
145)). Mladenović and Hansen (1997) present different approaches for the variable
neighborhood concept: Variable neighborhood descent (VND) and basic variable
neighborhood search (VNS).
The VND process is shown in Algorithm 3. To summarize, the algorithm searches
all neighborhoods of the initial solution x and selects the best one. It is a de-
39
Algorithm 3: Variable neighborhood descent (see (Gendreau and Potvin, 2010,

p. 64))
0 0
1 Select set of neighborhood structures Nk for k = 1, · · · , kmax
2 Find initial solution x
3 repeat
4 k=1
5 repeat
6 Find best neighbor x0 ∈ Nk0 (x) of x
7 if f (x0 ) > f (x) then
8 x = x0
9 else
10 k =k+1
11 end
0
12 until k = kmax ;
13 until no objective improvement;
terministic local search that should find the local optima for solution x with re-
spect to all neighbors. Whereas most local search heuristics use one neighborhood
structure, VND and VNS uses kmax 0
different neighborhood structures, denoted
Nk , k = 1, · · · , kmax . Algorithm 3 starts with the first structure and selects the best
0 0
neighbor in this structure as long as a better solution x0 is still found. Otherwise,

the next neighborhood structure is chosen. Typically the changes applied to the
solution i increases with increasing k using a nested structure.
Algorithm 4 outlines the basic variable neighborhood search. A random solution
x is created from the kth neighborhood structure of the current best solution x.
0
Afterwards, a local search is performed on this random solution. The result x00
should be a local optimum of x0 in the current neighborhood structure. If the local
optimum x00 is better than the current best solution x, the solution is applied and the
first neighborhood structure is used again. Otherwise, the next higher neighborhood
structure is explored.
Similar to other metaheuristics, many extensions and problem specific implemen-
tations exist. Mladenović and Hansen (1997) also propose reduced VNS and variable
neighborhood decomposition search. Furthermore (Glover and Kochenberger, 2003,
p. 150) introduce several hybrid approaches and present applications on the trav-
eling salesman problem, the vehicle routing problem and others (see (Glover and
Kochenberger, 2003, p. 152ff), Hemmelmayr et al. (2009a) and Parragh et al. (2010)).
For further metaheuristics the reader is referred to Reeves (1993), Michalewicz
and Fogel (2004) and Gendreau and Potvin (2010).
40
Algorithm 4: Basic variable neighborhood search (see (Gendreau and Potvin,

2010, p. 64 and 67))
0 0
1 Select set of neighborhood structures Nk for k = 1, · · · , kmax
2 Find initial solution x
3 repeat
4 k = 1 repeat
5 x0 = Shake(Nk (x))
6 x00 = LocalSearch(x0 )
7 if f (x00 ) > f (x) then
8 x = x00
9 k=1
10 else
11 k =k+1
12 end
0
13 until k = kmax ;
14 until stopping condition met;
3.1.4. Fitness Approximation in Metaheuristics

Throughout this thesis it will be shown that the main challenge of solving the liner
shipping network design problem is to solve the time consuming cargo allocation
problem as part of the metaheuristics. To tackle this, an approximation method for
the real fitness is developed. A function that determines the fitness heuristically is
also known as a fitness approximation or a surrogate (see Jin et al. (2003), Goh and
Tenne (2010) and Jin (2011)). In general, surrogates are intensively used in problems
that use a time consuming fitness function. In particular, mechanical optimization
problems where a simulation is used to evaluate a solution are often solved with
surrogates (see (Davarynejad et al., 2012, p. 245)).
Jin (2005) provides a comprehensive survey of surrogates in evolutionary algo-
rithms. Fitness approximations have been mainly applied to:
• Structural optimization problems, such as aerodynamics or fluid dynamics
• Solution approaches for optimization problems that require the user to evaluate
solutions
• Problems in noisy environments
• Fitness landscape approximation problems
Booker et al. (1999) present a general framework for the use of surrogates in
optimization approaches and evaluate it in a helicopter rotor blade design problem.
41
Sano and Kita (2000) use a history of previous fitness evaluations and samples
for the evaluation of a new individual. They use the maximum likelihood method
to evaluate a new individual approximately within a genetic algorithm that is
applied to noisy non-linear functions. Davarynejad et al. (2012) develop a method
to adaptively change the fitness granularity. They keep a pool of individuals with
exactly computed fitness functions. If a new individual is similar to an individual of
that pool according to some features, the pool’s exact fitness is used instead of the
crude estimate. Otherwise, the individual is added to the pool. Beside others, they
apply a genetic algorithm to an airplane wing design and a piezoelectric actuator
design problem1 .
Takagi and Iba (2005) provide an overview of interactive evolutionary computa-
tion that integrates the user into the optimization process. They provide example
applications such as industrial design, musical melody production, hearing aid and
others. Another application of interactive optimization is the work of Chou et al.
(2012). They develop districting plans for the city of Philadelphia with a special
focus of the fairness of the solutions.
Fitness approximations are also used in noisy environments. Papapanagiotou et al.
(2013) solve the orienteering problem with stochastic travel and service times. Based
on a set of scored vertices, the orienteering problem’s goal is to determine a length
limited path (not necessarily consisting of all vertices) that maximizes a score value
(see Vansteenwegen et al. (2011)). Papapanagiotou et al. (2013) present a monte
carlo simulation based objective function approximation to handle the uncertainties
in the problem. Papapanagiotou et al. (2014) extend this method by a combina-
tion of a monte carlo simulation and an analytical approximation. Weyland et al.
(2013a) develop a metaheuristic framework for stochastic combinatorial optimiza-
tion problems based on graphics processing units (GPU). The class of stochastic
optimization problems have in common that they can be approximately evaluated
using monte carlo sampling. In the framework of Weyland et al. (2013a), this is
performed in parallel on the GPUs. They apply their framework on the probabilistic
traveling salesman problem with deadlines and show significant improvements by
efficiently utilizing the GPUs. Weyland et al. (2013b) extend the work of Weyland
et al. (2013a) by improving their GPU algorithms.
Ratle (1998) present a method to speed up evolutionary optimization methods
by approximating the fitness landscape to quickly evaluate some of the individuals.
They apply their methods to test cases introduced by Keane (1994).
Beside the four areas that Jin (2005) mention, fitness approximation methods are
also used in graph theory. For example, Kim and Moon (2014) solve the linear
1
A piezoelectric actuator is a type of motor that transforms electric signals to mechanical output
(such as motion, pressure etc.) and uses piezoelectric materials, which induce voltage under
elastic deformation.
42
3.2. Related Combinatorial Optimization Problems
ordering problem using a genetic algorithm. The exact fitness function is given by
the sum of the weights of the used edges. Kim and Moon (2014) propose an approx-
imation of the fitness by using only a subset of all edges’ weights. They gradually
increase the number of edges during the evolutionary process.
Several problems occur when approximating the real (optimal, exact) fitness,
whereas the largest is misguiding the metaheuristics to a false optimum according
to the surrogate (see Jin et al. (2000)). For a survey on the application of surrogates
in evolutionary algorithms (EA), not only restricted to fitness approximations but
also EA operators, see Jin (2011).
Although mainly used in EAs, fitness approximations are also applied to other
metaheuristics. Singh et al. (2010) use surrogates in simulated annealing to solve
multi-objective test problem. Audet et al. (2008) extend the variable neighborhood
search metaheuristic by surrogates and apply the method to an engineering problem
in the chemical industry.
For a comprehensive survey on fitness approximations see Jin (2005). For further
approaches to approximate the fitness function, such as artificial neuronal networks
or fitness estimations using clustering algorithms, see (Jourdan et al., 2006, p. 59ff).
Summing up, the surrogate applications focus mainly on complex engineering
problems. Only a few attempts have been made to apply the concept of fitness
approximations to discrete optimization problems. To the best of the author’s
knowledge, surrogates have not been used for solving vehicle routing or maritime
problems yet.

In this section, related combinatorial optimization problems to the liner shipping
network design problem are presented. They provide a basic understanding of the
models developed in this thesis.
3.2.1. Vehicle Routing Problems

Vehicle routing and pickup and delivery problems arose from the classical traveling
salesman problem (TSP) and its multiple vehicle extensions (see (Bektas, 2006,
S. 212 - 213)). The planning task is to create a distance minimal round trip for a
salesman who wants to visit n cities exactly once. Miller et al. (1960) formulated the
problem as a mixed integer linear optimization problem. For further formulations
and solution approaches for the TSP see Bellmore and Nemhauser (1968), Lawler
et al. (1985), Dorigo and Gambardella (1997) or Basu (2012).
The vehicle routing problem (VRP) and extends the TSP by vehicle capacities
and dates back to the publication of Dantzig and Ramser (1959). The following
vehicle flow formulation (3.8) - (3.14) is based on (Toth and Vigo, 2002, p. 12).
43
VRP
XX
min cij xij (3.8)
i∈V j∈V
s.t.
X
∀j ∈ V \{0} : xij = 1 (3.9)
i∈V
X
∀i ∈ V \{0} : xij = 1 (3.10)
j∈V
X
xi0 = K (3.11)
i∈V
X
x0j = K (3.12)
j∈V
XX
∀S ⊆ V \{0}, S 6= ∅ : xij ≥ r(S) (3.13)
/ j∈S
i∈S
∀i, j ∈ V : xij ∈ {0, 1} (3.14)
V is set of customers, K the number of vehicles. The formulation minimizes the

total distance traveled by the vehicles. Furthermore, the vehicles’ round trips start
and end in a depot (node 0). Constraints (3.13) encodes the subtour elimination
constraint and the capacities of each vehicle. This basic formulation, also denoted
as the capacitated vehicle routing problem, is extended in many different ways such
as multiple depots (see Renaud et al. (1996)), split deliveries (see Archetti et al.
(2006)), backhauls (see Goetschalckx and Jacobs-Blecha (1989) and Toth and Vigo
(1997)) and time windows (VRPTW, see (Toth and Vigo, 2002, p. 6) and Kallehauge
et al. (2005)). An important subclass of vehicle routing problems with pickup and
delivery operations is described in the following section.
3.2.2. Pickup and Delivery Problems

The large family of pickup and delivery problems (PDP) has been extensively studied
in literature. The main difference to vehicle routing problems is the transportation
of demand between two or more specific nodes. Typically, two dimensions are dis-
tinguished in PDPs: commodity number and either traveling salesman or vehicle
routing constraints (see Table (3.1)).
According to (Cordeau et al., 2004, p. 431), three types of commodities exist
in the literature: one-commodity (single-commodity), two-commodity and multi-
commodity (n-commodity) pickup and delivery problems. In single-commodity
44
PDPs a single type of goods is either picked up at or delivered to a node. In

two-commodity PDPs each node can be used to pickup or deliver commodities.
This is a variant of the VRP with backhauls, where all deliveries must be performed
before any pickup. In the multi-commodity (n-commodity) PDP, each commodity
has a single pickup and a single delivery node. All three types of PDPs can be either
connected with the traveling salesman problem (PDTSP) or the vehicle routing
problem (VRPPD), respectively. The difference is that PDTSP need not consider
capacities whereas VRPPD consider capacities on a per vehicle basis. Note that
nowadays several PDTSP formulations consider per arc capacities.
P
SP
R
Commodity type
V
T
One-Commodity PDP 1-PDTSP VRPPD
Two-Commodity PDP 2-PDTSP 2-VRPPD
Multi-Commodity PDP m-PDTSP m-VRPPD
Table 3.1.: Classification of pickup and delivery problems.
For the single-commodity PDTSP, also called 1-PDTSP, see Hernández-Pérez and
Salazar-González (2003) and Hernández-Pérez and Salazar-González (2004). For the
multi-commodity PDTSP (m-PDTSP), see Hernández-Pérez and Salazar-González
(2009) and Plum et al. (2012) which also use arc capacities. The n-commodity pickup
and delivery requests can be also found in passenger transportation where they are
called dial-a-ride problems (DARP), see for example Jaw et al. (1986). For the liner
shipping industry, the 2-PDTSP and m-PDTSP are of special interest. The 2-PDTSP
represent the empty container repositioning problem, where demand origins at one
or more nodes and must be delivered to other one or more nodes. The m-PDTSP
represents the integrated routing of laden containers because each cargo flow is
transported between exactly one origin and one destination node. The m-PDTSP is
described as a mixed integer programming formulation in the following paragraphs
according to Hernández-Pérez and Salazar-González (2009). The objective is to find
a cost minimal Hamiltonian path such that all the commodities are collected and
delivered while holding the capacity constraints of the vehicle. A Hamiltonian path
is a closed path in a graph that visits every node exactly once.
Let G = (V, A) be a complete directed graph with n nodes V = {0, 1, · · · , n, n+1}
and the arcs A ⊆ V × V . The depot is at nodes 0 and n + 1. Let K be the set
of commodities, each associated with a weight qk , an origin sk and destination dk .
The vehicle capacity per edge is Q. For any subset S ⊂ V , let δ + (S) = {(i, j) ∈ A :
i ∈ S, j 6= S} and δ − (S) = {(i, j) ∈ A : i 6= S, j ∈ S}. δ + (S) stores the arcs whose
origin i lies but the destination j lies not in the set S and vice versa for δ − (S). ca
are the cost or distance for arc a. xa is a 0/1 variable indicating whether arc a is
used on the Hamiltonian path. fak is the flow of commodity k on arc a.
45
X
min ca x a (3.15)
a∈A
s.t.
X
∀i ∈ V \{0} : xa = 1 (3.16)
a∈δ − ({i})
X
∀i ∈ V \{n + 1} : xa = 1 (3.17)
a∈δ + ({i})
X
∀S ⊂ V, 0 6= S : xa ≥ 1 (3.18)
a∈δ − (S)
X
∀S ⊂ V, n + 1 6= S : xa ≥ 1 (3.19)
a∈δ + (S)

X X qk , if i = sk

∀i ∈ V, k ∈ K : k
fa − fa = −qk , if i = dk
k
(3.20)
0, else

a∈δ + ({i}) a∈δ − ({i}) 
X
∀a ∈ A : fak ≤ Qxa (3.21)
k∈K
∀a ∈ A : xa ∈ {0, 1} (3.22)
∀a ∈ A, k ∈ K : 0 ≤ fak (3.23)
The constraints (3.16) - (3.19) assure a simple cycle starting and ending at the
depot 0 = n + 1. Constraints (3.20) ensure the pickup qk ≥ 0 and delivery qk ≤ 0 for
all commodities. The capacity of the vehicle on all arcs a is defined in constraints
(3.21). Constraints (3.22) define the integrality of the xa variables and (3.23) the
nonnegativity of the fak variables. For further information and subtour elimination
constraints see Hernández-Pérez and Salazar-González (2009).
There exist several extensions for the PDTSP, such as the work of Şahin et al.
(2012) that extend the basic problem by multiple vehicles and split loads, tranship-
ment extension such as described in Shang and Cuff (1996) and time windows such
as in Mitrovic-Minic and Laporte (2006).
3.2.3. Min-Cost Flow Problems

A generalization of the classic transportation problem introduced by Orden (1956)
is the class of min-cost flow problems (MCFP). Problem (3.24) - (3.26) defines the
MCFP based on (Suhl and Mellouli, 2013, p. 192). The network is a directed graph
G = (V, E) with vertices N and edges E. The cost of transporting one unit on edge
46
(i, j) is cij and bi is the demand at vertex i. bi can be positive, meaning it offers
supply, or negative when is has a demand. A transhipment vertex is indicated by
bi = 0.
X
min cij xij (3.24)
(i,j)∈E
s.t.
X X
∀i ∈ V : xij − xji = bi (3.25)
(i,j)∈E (j,i)∈E
∀(i, j) ∈ E : lij ≤ xij ≤ uij (3.26)
The objective (3.24) is to minimize the transportation costs. Constraints (3.25)

balance the flow in the network and (3.26) imposes lower and upper bound on the
flow variables. This basic formulation uses a single commodity. A simple extension
leads to the class of multi-commodity flow problems, formulated in (3.27) - (3.29),
see for example (Fleischer and Skutella, 2006, p. 36). The main differences are the
flow variables per commodity k ∈ K. This leads to commodity specific arc cost cijk
and demands/supplies bik per node.
X
min cijk xijk (3.27)
(i,j)∈E,k∈K
s.t.
X X
∀i ∈ V, k ∈ K : xijk − xjik = bik (3.28)
(i,j)∈E (j,i)∈E
∀(i, j) ∈ E, k ∈ K : lijk ≤ xijk ≤ uijk (3.29)
In this formulation, each flow variable x is extended by a arc-commodity pair xijk .

The objective again is to minimize the commodity specific arc costs and balance
the supply and demand.
To summarize, several related planning problems are presented in this section.

Although these provide a basis, the problems arising in the maritime industry require
major extensions that lead to a relatively new maritime research stream. The specific
differences are outlined in Section 3.4.
47
3.3. Liner Shipping Network Planning

In this section, maritime liner shipping planning problems are presented. In partic-
ular, work in the field of liner shipping network design (LSNDP), cargo allocation
(CAP) and empty container repositioning (ECRP) as well as speed optimization
(SO) is analyzed regarding the cargo allocation and network design requirements
presented in Chapter 2. Afterwards, the research gap is presented and the goals of
this thesis derived.
One of the earliest publications specific to the maritime industry is the work of
Appelgren (1969) and Appelgren (1971) to solve the ship scheduling problem for
the tramp industry. He is followed by many publications in industrial, tramp and
liner shipping. The reader is referred to the following comprehensive survey papers
on routing and scheduling in the maritime industry: Ronen (1986), Ronen (1993),
Christiansen et al. (2004), Kjeldsen (2011), Christiansen et al. (2012), Tran and
Haasis (2013) and Meng et al. (2013).
3.3.1. Liner Shipping Network Design Problem

The liner shipping network design problem solves the complex planning task to create
liner services and deploy capacity to the services. Depending on the publication,
the capacity is either regarded as specific vessels or the vessel type. The LSNDP is
also called vessel deployment or vessel routing problem in the liner shipping context.
Most of the publications incorporate routing of laden containers into the network
design to calculate the transhipment costs. In Section 3.3.2, publications relevant
to this subproblem are discussed. Table 3.2 gives a comprehensive overview on
the LSNDP publications and their features relevant to this thesis. Some of the
works solve the similar hub-and-spoke network design problem where hub ports
or complete liner services are predetermined. Due to the similarity they are also
considered in Table 3.2.
The first column specifies the objective function of the publication that can be
either cost minimization or profit maximization. Cost minimization usually requires
the transport of all defined cargo flows. This can be unsuitable from a strategic
or tactical point of view since it can lead to low utilized services to fulfill the
cargo transportation. The second column specifies whether the repositioning of
empty containers is integrated into the network design decisions. The third column
specifies how the capacity constraints are integrated into the models, either per
vessel, per vessel type or per leg. The fourth column defines whether transhipment
and transhipment costs are considered. The next column defines whether the models
determine the routes or are externally given. The sixth column marks literature
that considers transit times in the network design. Columns seven to nine indicate
whether the speed is variable and if it is subject to a port duration based on
the moved containers and pilotage. Column ten and eleven indicate if the vessel
48
Largest instance’s ports/cargo flow count

Partner networks (SC=Slot Chartering)
Routes predetermined/variable (P,V)
Vessel speed fixed/variable (F,V)

Empty Container Repositioning
Cargo transhipment (-,T,TC)
Route types (C, B, CB)

Transit times per port
Objective (max/min)
Capacity (1,V,VT)
Deadweight scales
Solution method
Publication
Pape (1980) min - - - P - all - - - H 19/-
Rana and Vickson (1988) max - 1 - V - C F - - L,B 20/-
Rana and Vickson (1991) max - V - V - C F - - L,SG 20/-
Cho and Perakis (1996) max - V - P - C F - - BB -
Powell and Perkins (1997) min - V - P - C F - - BB -
Fagerholt (2004b) min - V - P - C F - - H 40/-
Reinhardt et al. (2007) min - V TC V - C,B F - - BC 5/5
Shintani et al. (2007) max X VT TC V - C V - - GA 20/-
Agarwal and Ergun (2008) max - VT T V - C F - - CG,B 20/114
Álvarez (2009) max - VT TC V - C F - - TS 120/14k
Fagerholt et al. (2009a) min - V - V - C V X - H -
Chen and Zeng (2010) max X V - V - C F - - GA 10/90
Agarwal and Ergun (2010) max - VT T V - C F - X CG,B 10/27
Reinhardt and Pisinger (2010) min - V TC V - C,B F - - BC 15/9
Andersen (2010) min - V TC V - C F - - LNS 16/325
Gelareh and Pisinger (2011) max - VT TC V - o F - - B -
Meng and Wang (2011a) min X VT TC P - C F - - BB 46/600
Wang and Meng (2012b) min - VT TC P - C V - - BB 46/652
Brouer and Desaulniers (2012) max - VT TC V - C,B F - - H -
Kjeldsen (2012) min - V TC V - C V - - CG 25/50
Plum et al. (2012) min - VT - V - C F - - CG 16/16
Brouer et al. (2013) max - VT TC V - C,B V - - TS,CG 110/4k
Plum et al. (2013b) max - VT TC V - all F - - CG 20/37
Gelareh et al. (2013) max - VT TC V - o F - - B,H 50/-
Mulder and Dekker (2013) max - VT TC V - C F - - GA 58/-
Wang (2013) min X VT TC P - C F - SC BB 20/-
Polat (2013) min - VT - V - C F - - VNS 26/-
Song and Dong (2013) min X VT o V - all V - - H 8/-
Plum et al. (2013a) min - VT - V - C F - - BP 25/-
Guericke and Suhl (2013) max X VT TC V x all V X SC GA 38/1.8k
Wang and Meng (2014) max - VT TC V o C F - - CG 12/-
Table 3.2.: Literature on the liner shipping network design problem

(X) feature considered
(-) feature not considered
(o) feature partly considered
Capacity: 1=Single vessel, V=Vessels, VT=Vessel types
Transhipment: T=Transhipment possible, TC=Transhipment with cost
Route types: C=Pendulum and cyclic, B=Butterfly, all=Conveyor belt
Solution methods: L=Lagrangian Relaxation, CG=Column Generation,
BB=Branch & Bound, BC=Branch & Cut, BP = Branch & Price, B = Ben-
ders decomposition, SG=Subgradient method, TS=Tabu Search, GA=Genetic
Algorithm, LNS = Large Neighborhood Search, H = Heuristic
49
type’s deadweight scale and partner networks are considered. The last two columns
give information on the solution approach and the solved instance sizes (where
applicable). Note that some authors do not provide information on the number of
ports or cargo flows that lead to the main complexity in the LSNDP. Unconsidered
or unmentioned features in a publication are indicated with a “-”.
In the remainder of this section, publications on the liner shipping network design
are described based on Table 3.2 in detail.
Publications without Transhipment and Single Service Planning

Pape (1980) is an early work on the fleet deployment problem. He assigns specific
vessels to predetermined port sequences. The cargo flows are approximated by fre-
quency the vessels call the ports. Port durations only depend on the size of the
vessel. The publications of Rana and Vickson (1988) and Rana and Vickson (1991)
are the first that solve the liner shipping network design problem. They formulate
a non-linear model and solve it with a lagrangian relaxation using the subgradient
method. Cho and Perakis (1996), Powell and Perkins (1997) and Fagerholt (2004a)
extend the model by further aspects. The main downside of the publications are
the lack of cargo transhipment operations that are nowadays heavily used in liner
shipping (see (Álvarez, 2009, p. 188), (Kjeldsen, 2011, p. 143)). These models are
not applicable to most of today’s carriers. Chen and Zeng (2010) and Song and
Dong (2013) focus on empty container repositioning and thereby do not include
container transhipment in their work. Plum et al. (2012) propose a column gener-
ation approach for the multi-commodity one-to-one pickup and delivery problem
with maximum path durations but do not include transhipment between paths.
Polat (2013) solves the feeder service network design problem but does not consider
transhipment costs because the cargo has to be transported directly. Plum et al.
(2013a) formulate an exact model to solve the single liner shipping service design
problem without transhipments between services. Gelareh et al. (2013) solve the
single string planning problem on a real-world instance using a heuristic. They also
do not consider transhipment for the existing services.
Publications without Transhipment Costs

Agarwal and Ergun (2008) introduce the ship scheduling and network design prob-
lem in the liner shipping context. They are the first who include transhipment in
the network design problem but do not associate any costs with these operations.
They consider weekly round trips in their model formulation and propose three
different solution approaches: A greedy heuristic, a column generation method and
a two-phase benders decomposition based algorithm. The methods are evaluated
on randomly generated instances. Agarwal and Ergun (2010) extend the model of
Agarwal and Ergun (2008) to support the joint optimization of networks of different
50
carriers. They incorporate a game theory approach with the model developed in
Agarwal and Ergun (2008) to find an optimal solution for several carriers.
Publications with Transhipment and associated Costs

The first publication that includes transhipment costs in the network design problem
is the working paper of Reinhardt et al. (2007) and the publication Reinhardt
and Pisinger (2010). Beside simple routes they include butterfly routes in their
model and solve it using a customized Branch and Cut method. They model inter-
service transhipment ports by counting the service legs using integer variables
and pay attention for the center (hub) port. They solve instances up to 15 ports
and nine demands to optimality. Álvarez (2009) also includes transhipment costs
and suggests a solution approach using a tabu search and a multi-commodity flow
model to evaluate a solution. Álvarez solves large scale instances up to 120 ports
and, although not mentioned, probably thousands of cargo flows. He shows that a
combination of metaheuristics joined with a multi-commodity flow problem (MCFP)
can work for large instances. He models the MCFP as an arc flow formulation without
empty container repositioning, deadweight scales, transit times or partners. The
transhipment constraints from Álvarez (2009) are also used in the cargo allocation
problem developed in this thesis. The work of Álvarez (2009) provides the basic
solution approach for the work of Brouer et al. (2013).
In 2011 and 2012 several authors tackle subproblems of the liner shipping network
design problem. Gelareh and Pisinger (2011) solve the fleet deployment, network
design and hub-location problem using a decomposition method. They allow tran-
shipment at hub ports but do not require the feeder networks to operate on round
trips or allow transhipment between these services. Meng and Wang (2011a) in-
clude the empty container repositioning problem in the network design problem
but require a manager to predetermine profitable services. Wang and Meng (2012b)
solve the liner ship fleet deployment problem and include transhipment operations.
Again, they assume externally given liner services. Brouer and Desaulniers (2012)
introduce a mathheuristic for the LSNDP, combining greedy approaches with integer
programming. They report promising preliminary tests but do not report numerical
results. Repositioning of empty containers, variable speed and transit times are not
included in the work of Brouer and Desaulniers (2012).
In 2013, several publications are released supporting not only pendulum, circle
and butterfly but also conveyor belt routes (see Plum et al. (2013b), Song and
Dong (2013) and Guericke and Suhl (2013)). Plum et al. (2013b) present a mixed
integer program for the liner shipping network design problem with transhipment
and conveyor belt routes. They introduce several additional constraints to tighten
the LP relaxation of their model. Thereby, they are able to evaluate their model on
the Baltic and WAF instances of the LINER-LIB 2012 (see Brouer et al. (2013)).
Song and Dong (2013) introduce a three stage optimization method to solve the
51
service route design problem. Repositioning of empty containers and conveyor belt
route types of a single long-haul service is included in their model. To the best of the
author’s knowledge, Guericke and Suhl (2013) are the first that consider variable
speeds, transit times and empty container repositioning in the LSNDP. They present
an evolutionary algorithm approach where the fitness function was evaluated using
a multi-commodity flow problem with an arc flow formulation based on Álvarez
(2009).
Further, more recent approaches to solve the network design problem are Mulder
and Dekker (2013) and Brouer et al. (2013). Mulder and Dekker (2013) propose
a genetic algorithm and use an arc flow formulation similar to Álvarez (2009) to
evaluate the fitness of the individuals. They introduce port clusters to reduce the
size of the cargo allocation problem from Álvarez (2009) to speed up the evaluation
of networks. These clusters are converted to feeder services during their solution
approach. Mulder and Dekker (2013) do not consider empty container repositioning,
transit times and variable speeds. Brouer et al. (2013) introduce the LINER-LIB
2012, which contains benchmark instances for the LSNDP. It contains seven instances
based on Maersk Line’s network and publicly available data. These instances also
include data for future work, such as transit times per cargo flow. The LINER-LIB
2012 is also used in the scope of this thesis. Brouer et al. (2013) extend the solution
approach from Álvarez (2009) and develop a column generation based method.
They do not consider transit times and empty container repositioning. The port
duration is assumed to be constant. They conclude their work with the importance
of including transit time and slow steaming strategies in the LSNDP.
Several authors include the transit times in their work. The network design
problem is extended by deadlines (LSNDPD) by Wang and Meng (2014). They
formulate the LSNDPD as a non-linear non-convex program and propose a column
generation based heuristic to solve the problem. The method is applied to a liner
shipping network with 12 ports and 73 cargo flows on the Asia-Europe trade. The
main drawback is that Wang and Meng (2014) do not allow transhipment between
the liner services.
3.3.2. Cargo Allocation and Empty Container Repositioning Problems

Liner shipping network design problems that consider cargo transhipment must
deal with the container routing because it determines a large proportion of the
operating costs. The routing of containers must be considered in strategic, tactical
and operational planning problems, each with specific requirements. In this section,
the cargo allocation (cargo routing) problem (CAP) is reviewed from a strategic
point of view. The repositioning of empty containers can be incorporated with the
CAP problem because the balancing is interconnected with serving the cargo flows
(see Chapter 2).
Therefore, this chapter not only reviews CAP publications, but also specific net-
52
Cyclic/Butterfly/Conveyor belt (C, B, CB)
Largest instance’s port/cargo flow number

including transhipment cost(-,T,TC)
Vessel speed fixed/variable (F,V)

Capacity constraints (VT,RG)
Empty container repositioning
Port duration/ pilotage

Transit times per port
Objective (max/min)
Cargo transhipment
Deadweight scales
Solution method
Route types
Publication
Álvarez (2009) max - VT TC - C F - - TS 120/-
Reinhardt and Pisinger (2010) min - VT TC - C,B F - - BC 15/9
Kjeldsen (2012) min - VT TC - C V - - CG 25/50
Mulder and Dekker (2013) max - VT TC - C F - - GA 58/-
Crainic et al. (1993) min X - - - C F - - BB -
Shen and Khoong (1995) min X VT - - C F - - BB -
Song and Carter (2009) min X VT T - C F - - BB -
Di Francesco et al. (2009) min X VT TC - C F - - BB -
Song and Dong (2011) min X VT - - C F - - Sim 6/-
Epstein et al. (2012) max X VT T - C F - - BB -
Brouer et al. (2011) max X VT TC - C F - - CG 234/16.3k
Wang et al. (2013b) max X VT - X C V X - BB 8/16
Wang et al. (2013c) min - - TC X C,B F - - BB 166/-
Wang (2014) min - VT TC - C F - - BB 40/458
Mhaky and Lee (2014) max - VT TC X CB F - - CG 33/300
Guericke and Tierney (2014) max X RG TC - CB V X - BB 39/365
Table 3.3.: Literature on the CAP within the LSNDP and the ECRP
Capacity: VT=Vessel types, RG=Resource groups
Solution methods: CG=Column Generation, BB=Branch & Bound,
BC=Branch & Cut, TS=Tabu Search, GA=Genetic Algorithm,
Sim=Simulation
work design problems and empty container repositioning problems that route laden
containers as well. Table 3.3 provides an overview on these publications. For presen-
tational purposes, first network design problems that incorporate cargo allocation
are presented. Afterwards, empty container repositioning problems (ECRP) and
specific cargo allocation problems are introduced. For each publication, features
relevant to the cargo allocation described in Chapter 2 are given.
53
Cargo Allocation within Liner Network Design

Álvarez (2009), Reinhardt and Pisinger (2010), Kjeldsen (2012) and Mulder and
Dekker (2013) provide multi-commodity flow formulations within the network design
to determine the transhipment costs. The publications have in common that they
use decision variables within the flow balance constraints to determine the loaded
and unloaded cargo at each port in each service. This modeling approach is also
used in the arc flow formulation in Chapter 4. None of these publications integrate
empty container repositioning or variable speeds.
Empty Container Repositioning

Crainic et al. (1993) formulate an operational empty container repositioning model
and analyze the container inventories under stochastic demand and supply by using
stochastic programming models. They do not consider the routing of laden contain-
ers but rather a fixed demand at each port. A decision support system to solve
the multiperiod empty container repositioning problem is introduced by Shen and
Khoong (1995). They include leasing-in and leasing-off decisions in their model.
Song and Carter (2009) analyze the scale of empty containers of three major ship-
ping routes and assess strategies to handle empty containers. They conclude that a
strategy that coordinates the empty containers at routes and share empty containers
among carriers reduced the repositioning costs by 12 – 18% in 2003 to 2007. They
show that up to 28% of the container movements were the movements of empty con-
tainers and indicate a trend that this figure might increase in future. Di Francesco
et al. (2009) use a multi-scenario approach to support the uncertain decision mak-
ing of empty container repositioning. They solve their formulation using system
dynamics. The ECRP with flexible destination ports is studied by Song and Dong
(2011). In their formulation, laden containers were routed and empty containers can
be spontaneously unloaded when needed. Epstein et al. (2012) develop a decision
support system for the South American carrier CSAV. They use a multi-commodity,
multi-period flow model and an inventory model to formulate this problem. They
incorporate safety stock levels with the empty container problem but do not consider
variable vessel speeds.
For further operational empty container repositioning problems the reader is re-
ferred to the work of Shintani et al. (2007), Saeidi et al. (2013) and Liu et al.
(2010).
Plain Cargo Allocation

In this paragraph cargo allocation problems that explicitly model container move-
ment is presented. Brouer et al. (2011) introduce the cargo allocation problem with
empty container repositioning to evaluate networks in a strategic planning horizon.
54
They used time periods for the repositioning and suggest a column generation ap-
proach to solve this problem. With the help of a relatively small restricted master
problem and shortest path pricing problems they solve large scale instances with up
to 9 time periods, 234 ports and more than 16,000 cargo flows in less than one hour.
Their solution method is promising for solving large scale instances of the cargo allo-
cation problem in this thesis because time periods are not considered in the network
design context. This has the ability to decrease the runtime. The work of Wang
et al. (2013b) schedules container ships under consideration of transit time sensitive
cargo flows and variable speeds. In their work, the demand quantity depends on
the transit times offered by the carrier. However, they do not include container
transhipments and empty container repositioning in their non-linear non-convex
formulation.
Wang et al. (2013c) uses a link-based multi-commodity flow formulation to solve
the container routing problem for one cargo flow under consideration of transit
times and transhipment operations. The vessels’ speed is assumed to be constant
and no capacity constraint is imposed to the legs. They solved their integer pro-
gramming model on 166 world wide ports using IBM CPLEX (2014). Wang (2014)
proposes a novel hybrid-link-based container routing model and considers butterfly
routes in mid-sized networks. They prove a unimodularity property of their integer
programming model and could therefore solve large instances.
Mhaky and Lee (2014) include transit times in their cargo allocation problem.
They define different service levels for their network that result in average speeds
for the network. On each network, the cargo is routed with the speed optimization
performed in a preprocessing step. They solve their cargo allocation problem using a
column generation approach. No empty container repositioning or speed adjustment
per leg is considered in their work.
Guericke and Tierney (2014) propose a cargo allocation model with speed op-
timization, empty container repositioning and cargo flow depended transit times.
Their model is solved with Gurobi and indicates large runtimes, especially for the
medium sized LINER-LIB networks. They do not consider deadweight scales or
partner networks in the model.
3.3.3. Speed Optimization

Bunker (fuel) costs are a huge factor when moving vessels. A lot of research has been
published on optimizing vessels’ speed, both for the tramp and liner shipping industry.
Table 3.4 gives an overview of literature on speed optimization problems. Note that
these publications are related to both liner and tramp shipping. Speed optimization
models from tramp shipping often incorporate time windows which make them also
applicable to liner shipping by fixing the end of the voyage according to the round
trip time. Table 3.4 shows the publications’ features: bunker consumption function,
port dependent bunker prices, uncertainty in bunker prices and port durations, the
55
stochastic/cargo dependent (D,S,C)

(cbc, cubic see Ronen (1982)/)
Bunker consumption function
Largest instance’s port count

Port dependent bunker costs
Port duration deterministic/

Uncertain bunker prices
(quad, quadratic/other)
Solution method
Publication
Ronen (1982) cbc - - D - -
Fagerholt et al. (2009b) cbc X - D NLP 16
Norstad et al. (2010) cbc - - D Alg -
Gatica and Miranda (2011) cbc X - D BB -
Meyer et al. (2012) other X - D Form -
Kim and Kim (2012) cbc X - - L,H 28
Qi and Song (2012) cbc - - S H 8
Yao et al. (2012) cbc X - D BB 15
Wang and Meng (2012c) cbc - - C BB 87
Vilhelmsen et al. (2013) other X - D BB 38
Hvattum et al. (2013) quad - - D Alg -
Norlund and Gribkovskaia (2013) cbc - - D Alg 10
Norstad et al. (2013) other X - D BB -
Kim (2013) cbc X - - L,H 14
Sheng et al. (2013) cbc X X D BB 15
Table 3.4.: Literature on the speed optimization

Solution methods: BB=Branch & Bound, L=Lagrangian relax-
ation approach, H=Heuristic, NLP=Nonlinear programming solver,
Form=Analytical formula, Alg=Exact algorithm
solution method and the instance size. Because speed optimization models typically
do not incorporate cargo routing decisions, the cargo flows are omitted in Table
3.4. As can be seen in Table 3.4, the instances deal with smaller instances since it
can be reasonable to decompose the problem per liner service. According to the
schedule, the vessel can be jointly loaded, unloaded and bunkered and thus can be
seen independent from other services.
One of the first publications in the area of speed optimization is Ronen (1982)
who provide a cubic consumption formula of the vessel’s speed. The optimal speed is
calculated algebraically. Although several authors (see Wang and Meng (2012c) and
Psaraftis and Kontovas (2013)) claim that the cubic consumption function leads to
bad approximations for low speeds and several vessel types, the cubic approximation
is often used in literature (see Table 3.4). Until the 2009, the attention on speed
optimization was relatively low. This might be related to the relatively low bunker
cost as indicated in Figure 2.15 in Section 2.10.
56
From an operational point of view, speed optimization must determine the speed
of the vessels on each voyage leg and the ports where to bunker. Section 2.10
shows that the bunker price highly depends on the port. Therefore, most of the
publications have port dependent bunker prices (see for example Fagerholt et al.
(2009b), Yao et al. (2012), Kim and Kim (2012) and Kim (2013)). A non-linear
speed optimization model considering time windows is presented by Fagerholt et al.
(2009b). They optimize the vessels’ speed on a per leg basis using an exact solution
strategy. Yao et al. (2012) propose an integrated model that determines the speed
per leg and the bunkering amount at each port of a liner service’s round trip. They
assume a deterministic port duration independent from the actual routed cargo.
Their non-linear model’s discretization using a piecewise linear function is solved
using IBM CPLEX (2014). Similar work is done by Kim and Kim (2012). Beside
the speed and bunkering port, they consider greenhouse gas emissions because of
the imposed carbon taxes. They develop a heuristic to solve a practical case study
on one transpacific liner service. Kim (2013) develops a lagrangian heuristic for the
problem introduced by Kim and Kim (2012).
An integrated ship routing and scheduling with speed optimization for the tramp
shipping industry is introduced by Norstad et al. (2010). The consideration of leg
dependent cubic speeds leads to a non-linear convex optimization problem that is
solved using a multi-start local search heuristic. A similar work is published by Gatica
and Miranda (2011). They discretize the pickup and delivery time windows to enable
practical constraints. Vilhelmsen et al. (2013) propose a column generation method
to solve the tramp shipping routing and scheduling with bunker optimization. The
application of the introduced tramp shipping routing models to the liner industry is
limited because tramp ships do not operate on services. A supply vessel optimization
model is presented in Norlund and Gribkovskaia (2013). Similar to tramp vessels
they do not operate on round trip services.
Meyer et al. (2012) propose an analytical operational speed optimization model
considering much more vessel specific parameters, such as fuel oil and lubricating
oil consumption as well as waiting time at the ports. They conclude that the
optimal vessel speed mainly depends on the freight rates and bunker prices, and
that slow steaming is a very good economical vessel operating mode. Additionally,
they conclude that the often applied cubic bunker consumption function is not
appropriate to reflect real-world consumption curves.
A large challenge in liner shipping are uncertain port durations (see Notteboom
(2006)). Therefore, Qi and Song (2012) include uncertain port times when optimizing
vessel schedules. They solve this problem using a simulation-based method. They
do not consider the routing of laden or empty containers. Recently, not only deter-
ministic bunker prices, but also uncertain prices have been considered. Sheng et al.
(2013) consider uncertainty in bunker prices and developed a multistage stochastic
program in a rolling horizon to determine the port where to bunker. They apply
their method to different liner services. An exact algorithm to determine the optimal
57
speed is given in Hvattum et al. (2013). They assume given time windows for the
fixed port calls but do not consider port durations by cargo flow routing.
Additional survey papers on tactical and operational speed optimization models are
presented in Wang et al. (2013a) and Psaraftis and Kontovas (2013).
3.4. Research Gap and Opportunities

In this section the state-of-the-art regarding the cargo allocation and the liner
shipping network design problem is summarized and the research gap is derived
according to the problem definition in Chapter 2. Based on the research gap, the
goals of this thesis are defined.
In the previous sections, related planning problems for the requirements specified
in Chapter 2 are presented. Although vehicle routing and pickup and delivery
problems share common properties with the liner shipping network design problem,
such as round trips, capacities and origin-destination demands, the following main
differences exist (for details on these aspects, see Chapter 2):
1. A liner service can start at every node (no fixed depot exists).
2. A port must not be served but can be served several times by one or more
vessels.
3. The vessels’ speed is a fundamental aspect in liner shipping because the cubic
fuel consumption function imposes large costs.
4. Container transhipment between vessels is common practice and can be done

at every port (more or less efficiently).
5. Empty containers must be repositioned in the network.
6. No nightly buffer between each round trip exists.
Due to these differences, the existing planning problems cannot be easily adapted
to the LSNDP. Thus, researchers developed specialized solution methods. The state-
of-the-art analysis has shown that lots of quantitative work has been performed to
design and optimize single liner shipping services and whole networks. Notteboom
states the key decisions for liner network planners as: Service frequency, fleet size,
vessel size mix and number of port calls (see (Notteboom, 2006, p. 20)). These key
decisions are already broadly supported in quantitative models in the literature.
The recent results of Plum et al. (2013b) show that small instances can be solved
close to optimality, whereas middle-sized instances with more than 100 ports can
be solved efficiently using heuristics (e.g. Álvarez (2009) and Brouer et al. (2013)).
58
3.5. Goals of this Thesis
Beside these fundamentals aspects, Chapter 2 introduces further practical require-

ments for liner shipping networks such as empty container repositioning, transit
times, slow steaming strategies and partner integration. The importance of these
aspects is also highlighted by Brouer et al. (2011), Notteboom (2006), Gelareh et al.
(2010), Brouer et al. (2013) and Wang and Meng (2014). They occur on different
planning levels in liner shipping because they are connected with the cargo alloca-
tion in the network. Section 3.3 shows that a research gap still exists between the
cargo allocation problem found in literature and the real-world requirements for this
problem. More specifically, integrating the aspects of empty containers repositioning,
speed optimization, deadweight scales, transhipment operations and liner carrier
cooperation into the cargo allocation problem is still an open research topic.
The more difficult LSNDP not only relies on the cargo allocation but also deter-
mines the underlying liner shipping network structure. The state-of-the-art analysis
in the previous sections shows that some of the aspects (such as transit times or
empty container repositioning) are already partly integrated into subproblems of the
network design. Especially the work of Wang et al. (2013c) and Wang et al. (2013b)
are first steps towards the transit time consideration in liner shipping network de-
sign. Álvarez (2009) and Brouer et al. (2013) present heuristic solution methods to
solve large scale instances of the liner shipping network design. However, the state-
of-the-art analysis also indicate that no work considers the aspects transhipment,
empty container repositioning, transit times, slow steaming strategies and partners
in an integrated manner yet.
Based on the research gap presented in the preceding section, the goals of this thesis
are as follows:
1. Evaluate large scale real-world liner networks to simplify the computer sup-
ported manual planning process.
2. Formalize missing practical requirements of liner networks and generate opti-

mal solutions.
3. Develop optimization methods to automatically optimize medium-sized liner

networks in reasonable computational time.
4. Integrate the methods into a prototypical decision support system.
In the following paragraphs, each goal is presented in detail.
59
Evaluate large scale real-world liner networks

Evaluating network changes is a complex task that includes determining cargo allo-
cation, services’ speeds and empty container balancing on predetermined networks.
The challenge is the routing of thousands of cargo flows on arbitrarily capacitated
services. This routing determines the duration in the ports and thereby the overall
speed of the services’ vessels. Solving this problem to optimality is difficult, if not
impossible, if performed manually. The first goal of this thesis is to develop an
optimal solution approach to allocate containers to services’ legs to support the
manual planning. With the help of the model, carriers can evaluate global impacts
of regional changes to liner service and the performance of the overall network. The
model extends the state-of-the-art by integrating the following aspects:
1. Integrated service speed optimization
2. Detailed consideration of service capacities
3. Deadweight dependent vessel drafts
4. Partner networks
5. Empty container repositioning
The cargo allocation should be integrated with the speed optimization to respect
the interdependency between duration at the ports and the need to increase the
speed. So far, capacities are considered on a per slot basis. In real-world networks,
not only the (dry) slots limit the amount of transportable containers but also the
reefer container plugs and the maximum deadweight. The load dependent draft of
a vessel is another aspect that has not been integrated into the cargo allocation
problem before. This physical constraint is highly important to ensure the feasibility
and efficiency of a given network. Partner liner services should be integrated into
the cargo allocation on a slot charter basis because it represents the most common
form of cooperation. Finally, the consideration of empty container repositioning is
important to assure capacity on the vessels.
The cargo allocation problem is able to support the manual planning process of
liner network planners. This process consists of repetitively evaluating adjusted net-
works for optimization. The results of the cargo allocation provides information on
the leg’s utilization, cargo routing, services’ speed and vessel draft. These informa-
tion can be used to manually change the network structure. The second application
of the method is the quantitative evaluation of networks within optimization algo-
rithms. For both applications a short computational runtime is essential: On the
one hand, users expect to get the result for this subproblem quickly, on the other
hand, optimization algorithms rely on evaluating hundreds or thousands of networks.
To achieve this goal, the cargo allocation problem should be solved within a few
60
seconds, even for global scale liner networks. Thus, planners and heuristics can get
an almost immediate feedback on the network changes. The methods to solve the
cargo allocation problem are presented in Chapter 4.
Formalize real-world requirements for the LSNDP and generate optimal

solutions
Based on the research gap and the requirements in Chapter 2, the state-of-the-art
liner shipping network design problem is extended by the following major aspects:
1. Transit times
2. Embargo constraints
3. All cargo allocation aspects, namely:

Speed optimization
Service capacities
Deadweight dependent vessel drafts
Partner networks
Empty container repositioning
It has been shown in Chapter 2 that transit times are a major requirement for
practical networks. Thus, solution approaches developed in this thesis respect transit
times. Because this thesis aims to solve the strategic network design, transit times
are based on a port-to-port basis to ensure that the network structure is capable
to hold the maximum durations required by shippers.
Real-world liner services are often subject to cabotage constraints due to political
reasons. Cabotage means that specific ports must not be served together within
one service. Instead, the ports can be called via different services which implies
transhipment operations.
Considering transit times in the network optimization introduces a trade-off
between serving cargo (and respecting the maximum duration) and the cubic bunker
consumption. This trade-off should be integrated in the solution approaches by
enabling a per leg adjustment of the speed. The speed optimization should already
be considered in the cargo allocation. Further aspects from the subproblem of
cargo allocation are service capacities, deadweight scales, partner networks and
empty container repositioning. These must be integrated into the liner shipping
network design as well. Discussions with network planners from global carriers
showed that considering these aspects in the liner network planning is essential to
create practically applicable networks. The extensions are formalized as a mixed
integer program. The mathematical model to solve the liner shipping network design
problem is presented in Section 5.1.
61
Optimize the LSNDP for real-world instances in reasonable time

Brouer et al. (2013) show that the liner shipping network design problem is NP-
hard and thus difficult to solve. Plum et al. (2013b) present nearly optimal solutions
for small instances of the basic LSNDP without transit times, partners and slow
steaming. Thus, we expect the integrated LSNDP not to be solvable to optimality,
even for small instances in a reasonable amount of time and computational resources.
To be able to solve practical instances, a suitable solution method must be developed.
The runtime is not as constrained as in the cargo allocation problem due to its
strategic or tactical planning horizon. Discussions with liner network planners and
runtimes presented in literature indicate that several hours seem to be a reasonable
upper bound for finding good solutions. With the help of this method, medium-sized
instances with several thousand cargo flows, approximately 40 ports and the aspects
presented above can be optimized. The solution approach to solve the liner shipping
network design is presented in Section 5.2 and 5.3.
Integrate the developed Methods into a Decision Support System

The approaches to solve the cargo allocation and liner shipping network design prob-
lem heavily rely on the technical background of the users. To allow the applicability
of the methods for network planners, the methods are integrated into an interactive
decision support system (DSS) to support the manual and iterative planning process
of liner network planners. This requires the modeling of the planning process in
practice and a suitable mapping of the DSS according to this process. The usability
of the methods is improved by providing a graphical user interface for the cargo
allocation and network design methods.
The DSS is a proof of concept, answering the question of how to use Operations
Research methods for the liner network planning problems occurring in practice.
Minor questions are how to display liner service legs on world maps and define a
responsive user interface with little maintenance overhead. The decision support
system is presented in Chapter 6.
The remainder of this thesis is structured as follows according to the goals: Chapter
4 presents the concept and numerical results for the cargo allocation problem, Chap-
ter 5 exact and heuristic methods to automatically optimize liner shipping networks
in artificial and real-world networks. Chapter 6 presents the decision support system
to provide the developed mathematical methods to liner network planners.
62
4. Evaluating Networks - The Integrated Cargo
Allocation Problem
In this chapter, the integrated cargo allocation problem (CAP) for a given liner
shipping network is described in detail. The objective is to obtain an optimal
container allocation with maximized profit. The presented model extends the state-
of-the-art regarding cargo allocation problems by integrating the following aspects
into a single planning problem (see Chapter 2):
1. Speed optimization
2. Load dependent port durations
3. Load dependent vessel drafts
4. Multiple capacity types
5. Empty container repositioning
6. Support for all route types
First, a new approach to distinguish different calls of the same port and the
mathematical notation for all models used in this thesis is presented. Next, two
formulations of the same problem, namely an arc flow and a path flow formulation,
are presented. Although the models incorporate a linearization of the non-linear
bunker consumption function using integer variables, it is proved that the models
can be solved to optimality using the linear relaxation. Finally, numerical results
for the optimal and approximate solution approaches are shown.
4.1. Distinguishing Port Calls in Liner Services

Chapter 2 presents different route types used in real-world liner networks. The
butterfly and conveyor belt routes call identical ports two times per round trip. Thus,
cargo can be transshipped within the same service or between different services at
different port calls. The calls of the same port must be distinguishable for a liner
service to correctly calculate the transhipment costs. From the literature point
of view, the port calls could be distinguished by marking the butterfly port (see
Reinhardt and Pisinger (2010) and Brouer et al. (2013)), which allows butterfly, but
not conveyor belt routes. A second approach is to separate the service legs (see Plum
63
4. Evaluating Networks - The Integrated Cargo Allocation Problem
Layer 1 p1 p2 p3
p1 p2 p3 Layer 2 p1 p2 p3
(a) One layer representation. (b) Two layer representation.
Figure 4.1.: Layered service representation of a butterfly rotation to correctly handle

transhipments at port p2 .
p1 p2 p3
p1 p2 p3
Figure 4.2.: All edges required to create arbitrary valid route types for three ports.
et al. (2013b)) to allow all route types. Plum et al. (2013b) use binary variables to
distinguish the port call.
In the scope of this thesis another formulation is introduced that is based on dif-
ferent layers that enables all route types and correctly account for the transhipment
costs. The graph of service legs and ports for service s is denoted as Gs = (Es , Ns ),
with edges Es and nodes Ns . Ports are represented by nodes and legs by edges. To
distinguish the calls of a same port p within one service the network is extended
by layers Ls . The set of nodes for a specific service s is then a set of tuples (p, l)
with a port p ∈ Ns and a layer l ∈ Ls , Ps ⊆ {(p, l) : p ∈ Es , l ∈ Ls }. Ls is a set of
tuples with layered legs that define a valid round trip. A round trip requires the
edges to be connected, thus there must exist exactly one incoming and one outgoing
edge for all layered ports in Ps , ∀(p, l) ∈ Ps : |{(i, p, l0 , l) : (i, p, l0 , l) ∈ Ls }| = 1 and
|{(p, j, l, l0 ) : (p, j, l, l0 ) ∈ Ls }| = 1. Figure 4.1 shows a simple service with the port ro-
tation p1 → p2 → p3 → p2 → p1 , allowing the port call differentiation of p2 with a lay-
ered port rotation of (p1 , 1) → (p2 , 1) → (p3 , 1) → (p2 , 2) → (p1 , 1). The layered leg
structure for this service is Ls = ((p1 , p2 , 1, 1), (p2 , p3 , 1, 1), (p3 , p2 , 1, 2), (p2 , p1 , 2, 1)).
As can be seen in Figure 4.1, the service uses a second layer to distinguish the
two calls of port p2 . Figure 4.2 shows all possible edges for the three ports to allow
conveyor belt routes.
In Appendix B, an algorithm to create layered service legs based on a sequence of
non-layered ports is presented. The algorithm has a runtime of O(|Ls |) for a service
s and is thereby fast enough to transform large scale liner networks.
64
4.2. Common Notation
RG 1: Dry Slots Resource 1 Resource 2
RG 2: Reefer Plugs Resource 2
RG 3: Max. DWT Resource 3
Figure 4.3.: Resource groups for a vessel type.
4.2. Common Notation

This thesis uses a common notation for the mathematical models which is presented
in this section. Further model specific sets and parameters are described where
needed.
Network Structure
A liner shipping network is represented by a set of liner services S. Liner services

can be either operated by the carrier itself or by a partner. They are elements of
the set of operated S O or partner services S P , respectively, with S = S O ∪ S P
and S O \S P = ∅. Each service s is represented by the tuple (VT s , VC s , Ls ) where
VT s ∈ VT is the service’s vessel type, VC s ∈ N, VC s > 0 the number of vessels
deployed and Ls ⊂ Ps × Ps the set of connected layered edges (see previous section),
called legs in the shipping industry. Let L = P × P be the set of service independent
legs where P is the set of ports. Ps ⊆ P × Ls is the set of layered ports called by
service s. Let C = E ∪ N define the set of all transportable containers, either empty
containers of type E or cargo flows N . Cargo flows are distinguished by cargo flows
of the carrier N O and partner cargo flows N P . The set of all cargo flows consists
of N = N O ∪ N P , N O ∩ N P = ∅. Whereas the transport of cargo flows from the
operated carrier is optional, typically a fixed percentage of partner cargo flows must
be transported due to contracts. The parameter θ ∈ [0, 1] indicates the percentage
of partner cargo flows that must be transported.
The distance between port i and j for leg (i, j) ∈ L is denoted as li,j and given
in nautical miles (nm). For simplicity, let ls be the overall distance of service s,
calculated as ls = (i,j,l,l0 )∈Ls li,j .
P
Liner Service Capacities
As described in Chapter 2, a container vessel has different capacity types. We refer

to each capacity type as a resource group and introduce a general concept that is
shown in Figure 4.3 for container vessels.
65
Each of the resource groups contains one or more resources that utilize the specific
resource group. Dry containers can be placed in both dry slots or reefer plugs.
However, this reduces the plugs available for reefer containers that need a cooling.
The resource groups for the example vessel “CHARLOTTE MAERSK” are defined
as: Capacity of resource group RG1: DrySlots are 9,757 (for resource 1, i.e. dry slots)
plus 700 for resource group RG2: Reef erP lugs, together 10,457. Resource group
RG2 is utilized by the resource 2 only, and has a capacity of 700. Reefer containers
are utilizing both resource group RG1 and RG2, reducing the overall capacity
available. The third resource group RG3: M ax.DW T , the maximum deadweight,
is the upper bound for all transported container payload. A forty foot dry container
utilizes resource 1 by factor 2 and resource 3 by its specific weight. Beside the
three resource groups presented, further types can be introduced for real operations.
For example, length or commodity group capacities can be defined (due to security
reasons). For tactical planning, the three resource groups offer sufficient detail (based
on discussions with liner network planners). Note that the resource group concept
can also be used for further container types and for mixed cargo vessel capacities,
such as roll-on roll-off container vessels.
The resource group concept is formalized as follows to allow simple capacity
modifications: Let RG be the set of resource groups, R the set of resources such as
weight and Rrg ⊆ R the set of resources utilized by resource group r. In partner
services, each leg can have a specific capacity. To formalize the leg dependent
capacities, a set of segments SGs ⊆ Ls , s ∈ S P is defined. A segment is a list of
legs with specific resource capacities. Let the resource groups incident to segment
sg ∈ SGs of partner service s, be denoted as RGsg .
Cargo Flows
A cargo flow n ∈ N, n = (o, d, q Max , e, u, r) is defined as a container quantity with a
maximum of qnM ax ∈ R+ containers in the planning horizon between an origin and
destination port on ∈ P , dn ∈ P . A cargo flow has a specific equipment type en ∈ E
and has an associated resource utilization vector u. Let ur,c , r ∈ R, c ∈ C define the
utilization of resource r of container c, c ∈ C. The revenue gained for the transport
of one unit of cargo flow n is denoted as rn . All monetary parameters and variables
are in US$.
Costs and Revenues

The fixed port call cost when vessel type vt calls port p is defined as φPp,vt
C
and φD
vt is
the time charter or vessel depreciation cost for one vessel of type vt over the whole
planning horizon. φCHp is the container handling cost at port p at the cargo flow’s
origin or destination. φTp S is the transhipment cost at port p. φC e is an additional
container cost for leasing, depreciation and hinterland repositioning for one unit of
66
4.3. Arc-flow Formulation for the Cargo Allocation Problem
a cargo flow of type e. Let bc(vt, k) be the bunker consumption per day at sea in
metric tons (1000 kilograms) of a vessel of type vt that steams with k knots. Each
metric ton (mt) costs φT US$, so the bunker costs per day is given by bc(vt, k) · φT .
The slot cost on a partner service s ∈ S P ’s leg (i, j, l, l0 ) for container type e is given
by φSs,i,j,l,l0 ,e .
Timing Aspects
Let τ denote the planning horizon length in days. Let τp,vt E
denote the duration
in days to move (either load or unload) one container in port p with vessel type
vt. This value can depend on the specific vessel type, because larger vessels are
usually served by more cranes. Let fs define the frequency of service s in days. The
frequency determines how often a port is visited by an arbitrary vessel of service
s. Note that throughout this thesis a constant value of 7 days is used. τsRT is the
number of round trips a vessel can perform on service s in the planning horizon τ .
τpAdd is the additional time for vessels visiting port p in hours per port call. τpAdd
contains the pilotage and additional buffer in days that can be used to reserve
bunkering time or schedule delays.
Vessel Types
Each vessel type vt has a minimum and maximum speed in knots, kvt M in
and kvt
M ax
.
Furthermore, a vessel type has a capacity of Crg,vt for a specific resource group. Let
O
Crg,vt define the capacity per leg of resource group rg for vessel type vt of operated
services and Cs,rg,i,j,l,l
P
0 the capacity for leg segments for partner services s ∈ S . To
P
simplify the capacity usage (for operated and partner services) in the mathematical
models, the following parameter is defined:

O RT O
Crg,vt · VC s · τs , if s ∈ S

Cs,rg,i,j,l,l0 := Cs,rg,i,j,l,l
P
0 ·
τ
7
, if s ∈ S P (4.1)
0,

else

Each vessel type vt has a deadweight scale that defines the draft of a vessel subject
to its deadweight load. Let dwsSvt define the slope of the linearization and dwsIvt
the intercept of the y-axis. DpM ax defines the maximum depth in meters of port
p. It is assumed that the deployed vessel type V Ts for all services s respect the
compatibility of the vessel type’s lightship draft and the port depths.

The mathematical model for the integrated cargo allocation problem as a non-linear
mixed-integer program is presented. It is modeled as a flow formulation with flow
67
quantity decision variables for each services’ legs for each equipment type and cargo
flow. The model is based on the work of Álvarez (2009), but extended by empty
container repositioning, speed optimization, deadweight scales and port dependent
durations and partner services. The model is also presented in Guericke and Suhl
(2013).
4.3.1. Mathematical Model

The arc flow formulation uses the following additional parameter:
P≥0 Penalty costs for unserved partner cargo

in s,p,l ∈ Ls Stores the incoming leg for port p at layer l
out s,p,l ∈ Ls Stores the outgoing leg from port p at layer l
The arc flow formulation uses the following decision variables:
αn ∈ R+ Served quantity of cargo flow n, αn ≤ qnMax must hold

xs,c,i,j,l,l0 ∈ R+ Flow quantity of cargo flows or empty containers c ∈ C on
service s leg (i, j, l, l0 ) ∈ Ls
ls,c,p,l , us,c,p,l ∈ Loaded and unloaded quantity of cargo flows and empty con-
R+ tainers (c ∈ C) from/to liner service s at port (p, l) ∈ Ps
ks ∈ R+ Speed in knots of all vessels deployed on service s
τsB ∈ R+ Uniformly distributed buffer in days in the planning horizon
at all ports of service s
ρP ∈ R+ Slack variable for unserved contractually agreed partner cargo
ρK
s ∈ R+ Slack variable for vessel speeds above maximum speed
The non-linear cargo allocation problem presented in (4.2) – (4.18) is referred to

as CAParc.
CAParc
X
max = (rn − φC
en )αn (4.2)
n∈N
!
X X τ
− (φPp,VT
C
s
) − φD
VT s VC s (4.3)
O p∈Ps
f s
s∈S
X X
− p (us,n,p,l + ls,n,p,l )
φCH (4.4)
s∈S,(p,l)∈Ps n∈N :p=on ∨p=dn
X X
− φTp S (us,n,p,l + ls,n,p,l ) (4.5)
s∈S,(p,l)∈Ps n∈N :p6=on ∧p6=dn
68
X X
− φTp S · (us,e,p,l + ls,e,p,l ) (4.6)
s∈S O ,e∈E (p,l)∈Ps
X X X X
− φSs,p,ec ur,c ls,c,p,l (4.7)
s∈S P ,c∈C sg=(p,j,l,l0 )∈SG s rg∈RGsg r∈rg
X
− φT (τ VC s
s∈S O

X τ Add
Ts (us,c,p,l + ls,c,p,l ) + (τp + τs ) )bc(ks ) (4.8)
E B
− τp,V
fs
(p,l)∈Ps ,c∈C
!
X
+ P ρP + ρK
s (4.9)
s∈S O
The objective of model CAParc consists of seven different terms: revenue for each
transported cargo flow minus the container depreciation cost (term (4.2)). The fixed
port call and vessel time charter costs for the predetermined network’s services (term
(4.3)) are added for reference purposes only and can be removed from the model.
Next, container handling at the origin and destination ports for each cargo flow
(term (4.4)), transhipment costs for transhipping cargo flows (term (4.5)) and empty
containers (term (4.6)) from one service to another are imposed. Finally, partner
slot cost in term (4.7) are subtracted. The bunker cost for all services for the time
at sea during the planning horizon is calculated in term (4.8). If not enough partner
cargo can be transported in the given network, the last term (4.9) adds penalty
costs to the objective. Additionally, the service speed must be in the interval of
minimum and maximum speed, if possible.
The constraints are divided into flow balancing and transhipment amounts for
laden and empty containers, capacity constraints and speed constraints.
∀s ∈ S, c ∈ C,
xs,c,in s,p,l + ls,c,p,l = xs,c,out s,p,l + us,c,p,l
(p, l) ∈ Ps
(4.10)

X X αn ,
 if p = on
ls,n,p,l = us,n,p,l + −αn , if p = dn ∀p ∈ P, n ∈ N
0,

s∈S,l∈Ls s∈S,l∈Ls 
else
(4.11)
69
 P
− : αn , if p = on
P n∈N

en =e

X X
ls,e,p,l = us,e,p,l + n∈N : αn , if p = dn ∀p ∈ P, e ∈ E
 en =e
s∈S,l∈Ls s∈S,l∈Ls
0,

else

(4.12)
X
ur,c xs,c,i,j,l,l0 ≤ Cs,rg,i,j,l,l0 ∀s ∈ S, (i, j, l, l0 ) ∈ Ls , rg ∈ RG (4.13)
r∈rg,
c∈C
X τ
uweight,c dws SV Ts xs,c,i,p,l0 ,l + dws IV Ts ≤ DpM ax
fs
c∈C,(i,p,l0 ,l)∈Ls
∀s ∈ S O , (p, l) ∈ Ps (4.14)
X τ
uweight,c dws SV Ts xs,c,p,j,l,l0 + dws IV Ts ≤ DpM ax
fs
c∈C,(p,j,l,l0 )∈Ls
∀s ∈ S O , (p, l) ∈ Ps (4.15)
P
(i,j,l,l0 )∈Ls li,j
ks =
Ts τ /fs (us,c,p,l + ls,c,p,l ) + τp
E 1 Add + τ B
P
fs V Cs − (p,l)∈Ps ,c∈C τp,V s
∀s ∈ S O
(4.16)
Ts ≤ ks ≤ kV Ts + ρs
K
kVMax Max
∀s ∈ S O
(4.17)
X X
αn ≥ θ · qnM ax − ρP (4.18)
n∈N P n∈N P
Constraints (4.10) ensure the flow balance at each service port. Containers on a
service’s vessel are either continued to be transported or unloaded at a port. The
incoming leg to port (p, l) of service s is obtained by the left hand side. The right side
determines the outgoing flow of port (p, l). The unloaded containers either arrive at
their destination or must be picked up by another service’s vessel. Constraints (4.11)
and (4.12) ensure that unloaded containers are picked up again later. The explicit
modeling of unloaded and laden containers is required to associate transhipment
70
costs in the objective function. Furthermore, these constraints provide demand and
supply at the ports.
Constraints (4.13) limit the capacity for each service s and resource group rg.
Constraints (4.14) and (4.15) ensure that the draft of a vessel, determined by its
deadweight, does not exceed the depth of the port. Both, when a service enters and
leaves the port the deadweight scale constraints must be considered. Constraints
(4.16) set the required speed for each service operated by the carrier. The speed
is calculated by taking the distance of the service divided by the duration at sea
available to perform the round trip. Each service is operated by a fixed frequency
f , leading to a round trip time of f times the number of deployed vessels V Cs . The
time at sea can be calculated by subtracting the duration at all service’s ports that
is determined by fixed parameters (such as duration for pilotage and additional
buffer) and by the unloaded and loaded volume multiplied by the duration to move
one container.
Constraints (4.18) ensure that a fixed percentage of partner cargo is transported.
Otherwise the missed amount is penalized in the objective function by activating
variable ρP . Due to the speed calculation in constraints (4.16), the presented model is
a mixed integer non-linear program (MINLP) and could be solved using commercial
or open source non-linear solvers, such as BARON (2014) and Couenne (2014).
4.3.2. Bunker Cost Discretization
The MINLP CAParc is expected to be hard to solve (see for example (Byrd et al.,
2006, p. 35)). Thus, the non-linear constraints are linearized using the L01 method
described in Padberg (2000). The key idea is not to calculate the bunker cost based
on the speed of the services but rather based on the overall port duration per service
as sea. This duration can be calculated with a linear expression. The duration is
mapped to a piecewise port call duration linearization for which the resulting bunker
costs can be calculated in advance. Based on the overall duration, the service speeds
result from the remaining duration to travel the total service distance. The speed
times the duration is used to determine the overall bunker cost for the time at sea
in the planning horizon. The linearization approach is described in this section.
The denominator in constraints (4.16) is linearized and inserted into the objective
function CAParc to get a function that determines the bunker cost based on the
overall port duration per service. The duration is discretized into i ∈ D intervals.
The variables zs,0 , zs,1 · · · zs,D and ys,1
L L
, · · · ys,D per service s ∈ S O are introduced
according to a discretization approach (see Padberg (2000)). The approximated
function has D support points where it is piecewise linearized per interval. The
z-variables represent the overall port duration for all vessels and for all round trips
in the planning horizon for a specific service s. The optimization problem CAParc
is then extended by the following constraints (see Padberg (2000)):
71
zs,0 ≤ as,1 − as,0 ∀s ∈ S O (4.19)

zs,i ≥ (as,i − as,i−1 ) · ys,i ∀1 ≤ i ≤ D − 1, s ∈ S O
(4.20)
zs,i+1 ≤ (as,i+1 − as,i ) · ys,i ∀1 ≤ i ≤ D − 1, s ∈ S O (4.21)
zs,i ≥ 0 ∀0 ≤ i ≤ D, s ∈ S O
(4.22)
ys,i ∈ {0, 1} ∀1 ≤ i ≤ D, s ∈ S O
(4.23)
The constraints (4.19) – (4.23) successively activate ys,i until the required port
duration is reached. The duration, when using the discretization, is the sum of zs,i
variables and the x-offset as,0 . The total port duration of service s at the support
point i is as,i .
D
X
PCD s = as,0 + zs,i (4.24)
i=1
The bunker cost for service s is calculated as follows and must be added to the
objective function of the cargo allocation problem:
D
X bs,i − bs,i−1
s = bs,0 +
φB zs,i (4.25)
i=1
a s,i − a s,i−1
Using the L01 approach, the non-linear bunker cost are linearized. With increasing
port call duration, the speed for all vessels must be increased to hold the required
frequency fs . The port call duration is limited by the maximum speed of a vessel. The
bs,i parameters in equation 4.25 are precalculated function values of the nonlinear
bunker cost function for durations in the interval [ai−1 , ai ].
Using a linearization of the bunker cost overestimates the real costs. With an
increased number of support points D, the bunker cost decreases because the accuracy
of the real bunker consumption increases. Using artificial and real-world instances,
20 uniformly distributed support points seemed to be reasonable for a strategic
planning horizon.
To let the model determine the bunker cost correctly the port duration term
(4.24) must equal the port call duration denominator in constraints (4.16):
k
X X 1 1 τpAdd RT
P CDs = as,0 + zs,i = (us,c,p,l + ls,c,p,l ) · · + · τs · V Cs + τsB
i=1 p∈Ps ,
tpp 24 24
l∈Ls ,c∈C
72
4.4. Path-Flow Formulation for the Cargo Allocation Problem
With the help of the discretization, the original mixed integer non-linear program
CAParc is transformed into a mixed integer program (MIP) that can be solved
with commercial MIP solvers such as IBM CPLEX (2014) and Gurobi (2014). Note
that the transformed model is just an approximation of the original CAParc model
because the cubic bunker cost function is discretized. In the numerical results of
this chapter the approximation quality is analyzed.

The mathematical model CAParc, described in the previous section, is an arc flow
formulation. It contains decision variables that store the container flows for all legs,
services and commodities. The drawback of this formulation is the large number of
decision variables and constraints that can lead to long solution times. One method
that has been shown to work well in large scale CAP instances by Brouer et al.
(2011), is a delayed column generation approach. An adaption of this concept for
the integrated cargo allocation problem is presented in this section. With the help
of the column generation approach we aim to improve the runtimes to determine
an optimal container allocation. The reformulated model solves the same problem
as described above, including speed optimization and empty container repositioning
with respect to the capacity constraints.
With the help of Figure 4.4, the basic idea of the path flow formulation is shown.
Two services are given in this example, service one (black) and service two (gray)
(see Figure 4.4(a)).
p1 p2 p3 p1 p2 p3
p4 p5 p6 p4 p5 p6
(a) Liner network containing two (b) Container path between port
services, the back and the gray. p1 and p6 incident to two services.
Figure 4.4.: Container path for a cargo flow c1 incident to two services.
Cargo flow c1 should be routed from p1 → p6 , leading to the bold arcs in Figure
4.4(b). In the arc flow formulation, the variables xs1 ,c1 ,p1 ,p2 , xs1 ,c1 ,p2 ,p4 , xs2 ,c1 ,p4 ,p5 and
xs2 ,c1 ,p5 ,p6 would have the value 100 when transporting 100 containers of cargo flow
c1 . The path flow formulation for the routing in Figure 4.4(b) defines a container
path o = (s1 , p1 ), (s1 , p2 ), (s1 , p4 ), (s2 , p4 ), (s2 , p5 ), (s2 , p6 ) and one decision variable
λo . For this path, λo is set to 100 if 100 containers of commodity c1 are transported
73
on this path. Furthermore, a profit can be associated with the container path o,
i.e. the container and handling costs at port p1 and p6 and transhipment costs at
port p4 , because the commodity is transhiped from service s1 to s2 . Details on the
network structure to correctly obtain the costs associated with a container path
used in the delayed column generation are presented in Section 4.4.2.
4.4.1. Restricted Master Problem

Chapter 3.1.2 introduces the delayed column generation method. This method
iteratively solves a restricted master problem (RMP), denoted CAPrmp. The RMP
uses container paths to reduce the number of decision variables and constraints.
The path flow formulation requires the following sets in addition to the ones
presented in Section 4.2:
O Set of container paths in the restricted master problem

SLo Set of service legs used by container path o. SLo ⊆ {(s, i, j, l, l0 )|s ∈
S, (i, j, l, l0 ) ∈ Ls }
The container paths per cargo flow use the following parameters:
po Profit for transporting one container on container path o ∈ O

no Cargo flow of container path o ∈ O, no ∈ N
P
τo,s Overall port duration in days for service s when one container is
transported on path o
τsMin Minimum port call duration of service s
τsMin = 7 · V Cs − min(7 · V Cs , kMinls ·24 ) · τ
7
V Ts
τsMax Maximum port call duration of service s
τsMax = 7 · V Cs − min(7 · V Cs , kMaxls ·24 ) · τ
7
V Ts
in s,p,l ∈ Ls Stores the incoming leg for port p at layer l
out s,p,l ∈ Ls Stores the outgoing leg from port p at layer l
The path flow formulation uses the following decision variables:
λo ∈ R+
0 Amount of containers of cargo flow no routed on the container
path o. The variables are not constrained to be integer since we
allow the transportation of container fractions which should
be sufficient in the considered long planning horizons
xs,e,i,j,l,l0 ∈ R+
0 Flow quantity of empty containers e ∈ E on leg (i, j, l, l0 ) ∈ Ls
of service s
E
ls,e,p,l , uE
s,e,p,l ∈ Loaded and unloaded quantity of empty containers (e ∈ E)
R+ 0 from/to liner service s at port (p, l) ∈ Ps
τsB ∈ R+ 0 Additional buffer used for service s
74
ρP ∈ R+
0 Penalty cost for contractually agreed unserved partner cargo
ρK +
s ∈ R0 Penalty cost for steaming above maximum speed
The model below uses the bunker cost linearization approach presented in the
previous section.
CAPrmp
X
max po λo (4.26)
o∈O
 
X X τ
−  φPp,VT
C
s
− φD
VT s · VC s
 (4.27)
fs
s∈S O (p,l)∈Ps
X X
− (ls,p,l,e
E
+ uE
s,p,l,e )φp
TS
(4.28)
s∈S, (p,l)∈Ps ,e∈E
X X X X
− φSs,p,e ur,c ls,c,p,l
E
(4.29)
s∈S P ,e∈E sg=(p,j,l,l0 )∈SG s rg∈RGsg r∈rg
!
X X bs,ai − bs,ai−1
− bs,a0 + zs,i (4.30)
ai − ai−1
s∈S O i∈D:i≥1
!
X
+ P ρP + ρK
s (4.31)
s∈S O
Term (4.26) provides the profit of transporting one container on path o ∈ O. The
profit is calculated by taking the revenue and subtracting the container depreciation,
handling and transhipment costs. The container path profit is calculated in the
pricing subproblem (see Section 4.4.2). The term (4.27) subtracts the fixed port
and charter costs. The term (4.28) of the objective calculates the empty container
transhipment costs. Note that the empty container balancing is done in the restricted
master problem because it depends on the cargo flow volumes of the container
paths. Term (4.29) defines the slot cost for the empty containers on the partner
services. The term (4.30) calculates the bunker costs for the whole planning horizon
according to the previously introduced linearization. It is based on the duration for
transporting empty and laden containers. Term (4.31) penalizes unserved partner
cargo and vessel speeds above the type’s maximum speed.
X
λo ≤ qnM ax ∀n ∈ N (4.32)
o∈O:no =n
75
 
X X X ∀s ∈ S,
 ue,r · xE
s,i,j,l,l0 ,e + uno ,r λo  ≤ Cs,rg,i,j,l,l0 (i, j, l, l ) ∈ Ls ,
0
r∈rg e∈E o∈O:(s,i,j,l,l0 )∈SL o rg ∈ RGs

(4.33)
 
X X
s,i,j,l,l0 ,e +
ue,weight xE uno ,weight λo  dws SV Ts

e∈E o∈O:(s,i,j,l,l0 )∈SLo
τ ∀s ∈ S O ,
+ dws IV Ts ≤ DjM ax
fs (i, j, l, l0 ) ∈ Ls
(4.34)
 
X X
s,i,j,l,l0 ,e +
ue,weight xE uno ,weight λo  dws SV Ts

e∈E o∈O:(s,i,j,l,l0 )∈SLo
τ ∀s ∈ S O ,
+ dws IV Ts ≤ DiM ax
fs (i, j, l, l0 ) ∈ Ls
(4.35)
s,in s,p,l ,e + ls,p,l,e = xs,out s,p,l ,e + us,p,l,e

xE E E E
∀s ∈ S, (p, l) ∈ Ps , e ∈ E (4.36)
X X
E
ls,p,l,e = uE
s,p,l,e
s∈S,(p,l)∈Ps s∈S,(p,l)∈Ps
X X
+ λo − λo ∀p ∈ P, e ∈ E
o∈O:eno =e∧dno =p o∈O:eno =e∧ono =p
(4.37)
X τ X
τpAdd + E
Ts (us,p,l,e + ls,p,l,e ) + τs,p,l
E E B

τp,V
fs
(p,l)∈Ps (p,l)∈Ps ,e∈E
X X
+ P
τo,s λo = zs,i + ρK
s ∀s ∈ S O
o∈O:(s,i,j,l,l0 )∈SLo i∈D:i≥1
(4.38)
zs,i ≥ (ai + ai−1 )ys,i ∀s ∈ S O , i ∈ D : i ≥ 1
(4.39)
zs,i+1 ≤ (ai+1 − ai )ys,i ∀s ∈ S O , i ∈ D : i ≥ 1 ∧ i < |D| − 1
(4.40)
76
X
τsM in ≤ zs,i ≤ τsM ax ∀s ∈ S O
i∈D
(4.41)
X X
λo ≥ θ qnM ax − ρP (4.42)
n∈N P ,o∈O:no =n n∈N P
Constraints (4.32) limit the maximum amount of transported volume for each
cargo flow. Constraints (4.33) set the maximum transported quantity of empty and
laden containers for all services, layered legs and utilized resource groups. Constraints
(4.34) and (4.35) limit the cargo volumes that enter or leave a service’s port by the
vessel type’s deadweight scale. Constraints (4.36) balance the empty container flow
within the RMP. The empty container transhipment and empty container demand as
well as supply are calculated in (4.37). Constraints (4.38) calculate the overall port
duration per service by activating the zs,i variables. For these variables, constraints
(4.39) and (4.40) specify the variable activation range, according to the formulation
in Padberg (2000). Constraints (4.41) set the minimum and maximum port call
duration for the whole planning horizon according to vessel speeds. τsM in is the
minimum port duration according to the maximum speed and is calculated in the
parameter list. Respectively, τsM ax , is the maximum duration when steaming at
minimum speed. τsM ax < 0 means that not enough vessels are deployed on the
service to perform the round trip with maximum speed. In this case the model
makes use of the penalty variable ρK s to artificially increase the speed.
Constraint (4.42) ensures a minimum served amount of partner cargo, if possible.
Otherwise, the penalty variable ρP is activated.
4.4.2. Network Structure for determining Container Paths

Generating new container paths, i.e. extending the set of containers paths O, is
done in the pricing (sub)problem. The pricing problem is solved for each cargo flow
independently. The input for the cargo allocation problem is a fixed liner shipping
network, such as the example in Figure 4.5 with two services. The objective is to
find a path from an origin to a destination on which to route each cargo flow. On
that path, different costs can occur, in particular transhipment to other services,
transhipment within a service (due to butterfly or conveyor belt routes), container
handling and slot purchasing cost. To make use of efficient solution methods to
determine container paths in the pricing problem, the network structure makes use
of the layers presented in Section 4.1 and artificial connections between services and
ports.
Figure 4.6 shows the corresponding network to the original liner service network in
Figure 4.5. The pricing network structure is similar to the one presented by (Wang
et al., 2013c, p. 3) but is extended by multiple source and target ports. Thus, the
network structure can be created for all pricing problems and column generation
77
p1
p2 p3
Figure 4.5.: Flat network structure with two liner services as input for the pricing
network 4.6.
pT1 pS1
EO EI
p1,s1 ,1 p1,s2 ,1
ET
ES
ES
p2,s1 ,1 p3,s1 ,1 p3,s2 ,1

ET
ET
p2,s1 ,2 pS3 pT3
pS2 pT2
Figure 4.6.: Network structure for the pricing problem for the example in Figure
4.5 for one cargo flow n. The black edges are service legs, the remaining
edges are artificial edges for the pricing problem.
78
Dual Description Constraint LB UB

πn Duals for serving a cargo flow n (4.32) 0 ∞
δs,i,j,l,l0 ,rg Duals for the capacity constraint (4.33) 0 ∞
I
γs,i,j,l,l 0 Duals for the ingoing DWTS constraint (4.34) 0 ∞
O
γs,i,j,l,l 0 Duals for the outgoing DWTS constraint (4.35) 0 ∞
EC
φp,e Duals for the empty container balancing (4.37) −∞ ∞
πsP C Duals for the port call duration constraint (4.38) −∞ ∞
π P CF Duals for the unserved partner cargo flows (4.42) −∞ 0
Table 4.6.: Dual variables and ranges in the cargo allocation path flow formulation.
iterations at once. To account for different cargo flows and CG iterations, only the
edge costs must be updated.
For each port i with a cargo flow destination and for each cargo flow origin, an
artificial node pTi and pSi is introduced respectively. For each port that is called by
a service, a service specific node pi,s,l is introduced and connected to the source
port with edge e ∈ E I and to the destination port with edge e ∈ E O . The port
also depends on the layer on which the port is defined in the service to distinguish
several calls of the same port within a service round trip. The ports p2,s1 ,1 and p2,s1 ,2
are an example for the butterfly service in Figure 4.5.
Two or more services can call the same port. Thus, transhipments between services
can be performed at that port (see, for example, port p1 and p3 in network 4.6). Liner
service ports that are called by another service are connected by a transhipment
edge e ∈ E T . If a service calls a port more than once per round trip, a transhipment
edge is introduced between the different port layers (see for example p2 of service
s1 in Figure 4.6). Beside transhipment and cargo handling edges, an edge e ∈ E S
per service’s leg is introduced. Each edge of the pricing problem has a coefficient
that is described in the remainder of this section.
The restricted master problem CAPrmp defines the following dual variables that
are part of the edge costs. Duals are “shadow” costs resulting from increasing the
capacity of the corresponding constraint by one (marginal) unit. Table 4.6 shows
the dual variables from the restricted master problem that are relevant to determine
a new container path o ∈ O for a cargo flow n.
The duals can be used to calculate the reduced costs, i.e., the condition for
a container path to possibly lead to an objective improvement. Reduced costs
express the benefit for activating a column with one marginal unit, respecting the
interdependency of other variables. The equation for the reduced cost can also be
derived from transforming model CAPrmp to its dual formulation, determining the
constraints that are incident to a cargo flow n and summing up the negated left
hand sides of the constraints.
79
c¯o = po − πno
!
X X X
− uno ,r δs,i,j,l,l0 ,rg + uno ,weight dwsSV Ts (γs,i,j,l,l
I
0 + γs,i,j,l,l0 )
O
(s,i,j,l,l0 )∈SLo rg∈RG r∈rg

(
X π P CF , if n ∈ N P
+ φEC EC
on ,en − φdn ,en −
P PC
τo,s πs − (4.43)
s∈SoI
0, else
The reduced cost for a container path o serving cargo flow no is given in equation
(4.43). It assumes a routing on the legs o = ((s1 , i1 , j1 , l1 , l10 ), · · · (sn , in , jn , ln , ln0 )),
transhipments at ports PoT S with PoT S = {((sm , im , jm , lm , lm
0
), (sk , ik , jk , lk , lk0 )) ∈ o :
k = m + 1 ∧ sm 6= sk }. Container path o uses the services SoI ⊆ S.
X
P
τo,s = E
τi,V Ts + τono ,V Ts + τdno ,V Ts
E E
(4.44)
(s,i,j,l,l0 )∈PoT S
Equation (4.43) uses the profit po of container path o by taking the revenue for
transporting one unit of cargo flow n on the path, subtracting the container handling
cost at the origin and destination, the transhipment costs and slot purchasing costs.
The profit as follows:
X X
po (n) = rn − cCH CH
φTi S + φTi S − φS (s, n, p)

on − cdn −
((s1 ,i),(s2 ,i))∈PoT S p∈PoT S
(4.45)
For each edge in the pricing network (see Figure 4.6), the unit cost per cargo
flow n is defined as shown in equations (4.46) - (4.50) where the slot cost are given
separately in equation (4.51) for simplicity.
(
+φEC
p,en , if p = on ,
E I (p, (s, p, l)) :=cCH
p + τp,s
E
· πsP C + (4.46)
0, else
+ φSC (s, n, p)
(
p,en , if p = dn ,
−φEC
E O ((s, p, l), p) :=cCH
p + τp,s
E
· πsP C + (4.47)
0, else
X
E S (s, i, j, l, l0 ) := (un,r · δs,i,j,l,l0 ,rg )+
r∈rg
80
dws SVT s · un,weight · (γs,i,j,l,l

I
0 + γs,i,j,l,l0 )
O
(4.48)
E (s1 , s2 , i, l1 , l2 )
T
:=2φTi S + E
τp,s · (πsP1C + πsP2C ) +φ
SC
(s2 , n, o(l2 )) (4.49)
E (s, p, l, l )
IS 0
:=2φTp,sS (4.50)
(
0, if s 6= S P
φ (s, n, p) := P
S
SC
(4.51)
rg∈RG,r∈rg∩Rn φsgp ,r , else
Equation (4.46) calculates the edge costs for the cargo flow origin. These costs
occur for any unit of the transported cargo flow. It consists of the terms for the
container handling costs, the reduced costs for the port call (indirectly for the addi-
tional speed), the additional cost for transporting empty containers to the port of
origin and the slot costs if the service is operated by a partner. Respectively, (4.47)
calculates the reduced costs at the port of destination for one unit of cargo flow
n. Term (4.48) calculates the reduced costs for transporting one unit on leg i, j on
service s. The term consists of the indirect costs associated with the vessel dead-
weight scale constraints, which are mainly costs by decreasing the transportation
volume of other cargo flows. Term (4.49) calculates the transhipment cost between
different services by pricing the transhipment costs, the costs for empty container
repositioning and if necessary the costs for loading on a partner service. Finally,
equation (4.50) calculates the costs associated with the transhipment within a liner
service.
Algorithm 5 shows an overview of the delayed column generation method. The
RMP is initialized and the pricing network net for the liner shipping network NW
is instantiated. Afterwards, the initial container paths for all cargo flows are created
and added to the RMP. Since no dual information is available yet, the containers
are routed on their shortest (i.e. cheapest) path. Furthermore, information of non-
routeable cargo flows (for example when a port is not served by the network) is
stored. The RMP is solved using a commercial linear programming (LP) solver
and the solution, as well as duals, is stored. Afterwards, the algorithm tries to find
improving container paths for all routeable cargo flows and adds them to the model
if they have positive reduced cost.
The method CreateP ath in line 9 of Algorithm 5 first updates the edge cost which
depend on the current cargo flow according to equations (4.46) - (4.50). Afterwards
a shortest path algorithm is executed with the source node pSon and the target pTdn .
Note that the edge costs leaving the source node can be negative. This can be
explained by the fact that transporting one unit of a specific cargo flow can reduce
the number of empty containers required to balance the network. Thus, the duals
can be interpreted as opportunity costs for reducing the amount of empty containers.
81
They include the overall effects of the empty repositioning on the restricted master
problem’s objective. Because source nodes have only outgoing edges and the target
nodes only incoming, no negative cost cycles can occur. Thereby, shortest path
algorithms that cannot handle negative cost cycles can also be used, such as the
Bellman Ford or Dijkstra algorithm.
Algorithm 5: Overview of the delayed column generation for the integrated

cargo allocation problem.
Input: A liner shipping network NW
Output: Profit of the network NW
1 m ← RMP(NW );
2 net ← CreatePricingNetwork();
3 O ← InitialColumns(net, NW );
4 m.Add(O);
5 while O 6= ∅ do
6 solution ← LPSolve(m);
7 O ← ∅;
8 forall the n ∈ N do
9 o ← m.CreatePath(solution, net, NW );
10 if c¯o (n) > 0 then
11 O ← O ∪ {o};
12 end
13 end
14 end
15 return (objective, solution);
In general, the column generation solution approach does not lead to integer
solutions and a more advanced Branch & Price method must be used. In the next
section, we mathematically prove that relaxing the bunker cost linearization integer
variables still provides optimal integer solutions for the cargo allocation problem.
With this result, the linear relaxation of the restricted master problem provides the
optimal solution for the mixed integer program and efficient solution methods, such
as the dual simplex, can be used.
4.5. Relaxing the Integrality Constraints for the Bunker Cost

Padberg’s approximation model for separable non-linear functions used in the arc-
and path-flow formulations relies on binary zs,i variables (see (Padberg, 2000, p.
2)). This results in a mixed binary program instead of a pure linear program. In
this section we prove that a linear relaxation of the zs,i ∈ {0, 1} variables to be
0 ≤ zs,i ≤ 1 is possible in the scope of the specific assumptions for the cargo
82
allocation problem. For simplicity, only one service is used, leading to yi and zi
variables. The case for several services can be performed in a similar way because
the service discretizations are mutually independent.
The common function approximated in the arc- and path-flow formulation is the
bunker cost for a given overall port call duration in the whole planning horizon (see
equation 4.52). Details on the bunker cost are described in the previous sections.
!
l
φ(x) = φT · bc · (τ · V C − x), x ≥ 0 (4.52)
24(f · V C − VxC )
RT
The bunker consumption function bc is a function of speed to the power of a constant

factor a. Stopford suggests a = 2 for steam engines and a = 3 for diesel engines (see
(Stopford, 2009, p. 234)). It is shown in Chapter 2 that practical bunker profiles are
often approximated using cubic polynomial functions.
To be able to relax the integer constraint on the z-variables, the following as-
sumptions must hold:
Assumption 1 φ(x) is a bunker cost function such that an increase in the port
call duration cannot lower the function value. This means that increasing the
speed must always increase the total bunker cost, and the time saved at sea
by higher speed cannot compensate higher fuel consumption.
Assumption 2 The x-axis discretization of φ(x) is uniformly distributed with k
intervals with fixed width:
ai − ai−1 = ai+1 − ai = c > 0, ∀ 1 ≤ i ≤ k
Assumption 3 For the discretization of φ(x) using model 1 from Padberg (2000)
the following strict monotonicity must hold:
b 1 − b0 bk − bk−1
b0 < < ··· <
c c
where bi is the function value at ai , i.e. φ(ai ) = bi with bi > bi−1 .
Theorem In the optimal solution of a linear program (LP) relaxation that approx-
imates φ(x) as cost function, an integer solution with less or equal cost for all
relaxed zi variables can be found in constant time.
Proof : Let z ∗ , y ∗ ∈ Rk be the optimal solution for the approximated φ(x) function.
z ∗ is partitioned into:
Z I = {zi∗ : 1 ≤ i ≤ k, zi∗ = 1}
Z F = {zi∗ : 1 ≤ i ≤ k, 0 < zi∗ < 1}
Z Z = {zi∗ : 1 ≤ i ≤ k, zi∗ = 0}
83
The relevant constraints in model 1 in (Padberg, 2000, p. 2) for the z and y

variables are:
y1 ≤ c (4.53)
c · zi ≤ yi ≤ c · zi−1 ∀1 ≤ i ≤ k − 1 (4.54)
0 ≤ yk ≤ c · zk−1 (4.55)
By constraints (4.54) and (4.55) it follows that c · zi ≤ c · zi−1 and zi ≤ zi−1 .

Elements in Z F are connected as Z F = {zi∗ : n ≤ i ≤ p} for 1 ≤ n, p ≤ k. By the
partition it follows that zi∗ = 1 ∀i < n and zi∗ = 0 ∀i > p.
We prove the theorem by constructing an alternative solution for the LP with
integral z variables, show that it is feasible (Step one) and that the costs are not
larger than those of the optimal solution (Step two). Define z 0 , y 0 as follows:
(
1, i < n + a
zi0 = (4.56)
0, else

c,
 i<n+a
yi =
0
i=n yi − c(a − 1), i = n + a (4.57)
Pp+1 ∗
0,

else

with
& p+1 '
X y∗
a := i
(4.58)
i=n
c
In Figure 4.7 the variable values for y ∗ , y 0 are illustrated. Each interval’s maximum
value is c by assumption two. In the optimal solution, the z ∗ variables might be
activated fractionally, leading to y ∗ below the upper bound c. The new constructed
solution y 0 stacks y ∗ variables until the interval capacity c is reached. Thus, n+a ≤ p.
i 1 ··· n ··· p ···
i 1 ··· n ··· n+a ···
Figure 4.7.: Graphical illustration of the variable values in the optimal solution y ∗
(top) and the new constructed solution y 0 (bottom).
84
We show that for the y 0 variables the following holds:
yi0 ≥ 0 ∀1 ≤ i ≤ k (4.59)
To show that yn+a

0
≥ 0, it suffices to show that yn+a
0
≥ 0, because c ≥ 0 by assumption
two:
0
yn+a ≥0
p+1 p+1 p+1 ∗
X X X y
⇔ yi∗ − c(a − 1) ≥ 0 ⇔ yi∗ + c ≥ ca ⇔ i
+1≥a
i=n i=n i=n
c
p+1
& p+1 ' & p+1 ' p+1 ∗
X y∗ X y∗ X y∗ X y
⇔ i
+1≥ i
⇔ i
− i
≤1
i=n
c i=n
c i=n
c i=n
c
This is the definition of the ceiling function and true by definition.

Step one is to show that z 0 , y 0 is a valid solution for the LP. We divide the
constraint validation for z 0 , y 0 into three cases, namely 1 ≤ i < n + a, i > n + a and
i = n + a.
For all 1 ≤ i < n + a,
c · zi0 ≤ yi0 ≤ c · zi−1 ⇔ c ≤ c ≤ c
and for all i > n + a,
c · zi0 ≤ yi0 ≤ c · zi−1

0
⇔0≤0≤0
is always true. Let i = n + a. The lower bound

0
c · zn+a 0
≤ yn+a ⇔ c · 0 ≤ yn+a
0
holds because all yi0 ≥ 0 (see 4.59). For the upper bound on yn+a
0
,
0
yn+a 0
≤ c · zn+a−1 (4.60)
must hold because:

p+1
Def.z 0
X
0
yn+a ≤c· 0
zn+a−1 ⇔ 0
yn+a ≤c⇔ yi∗ − c(a − 1) ≤ c
i=n
p+1 p+1 ∗ p+1 ∗
& p+1 '
X X y X y X y∗
⇔ yi∗ ≤ ca ⇔ i
≤a⇔ i
≤ i
(4.61)
i=n i=n
c i=n
c i=n
c
85
which is always true for the ceiling function. Therefore, all cases for step one are
proved and z 0 , y 0 is a valid solution for the LP.
Step two is to show that φ(y 0 ) ≤ φ(y ∗ ). We show that the increased cost of the
changes to the y 0 variables for n ≤ i ≤ n + a is less or equal to the cost reduction
for n + a + 1 ≤ i ≤ p + 1 of the original solution y ∗ . The following inequality must
be valid:
n+a−1
X bi − bi−1 0 bn+a − bn+a−1 0
(yi − yi∗ ) + (yn+a − yn+a
∗
)
i=n
c c
p+1
X bi − bi−1 ∗
≤ (yi − yi0 ) (4.62)
i=n+a+1
c
The left hand side sums the increased cost due to the increased y 0 variables. The
right hand side sums the decreased y ∗ variable values. The additional cost must be
less or equal to the cost reduction. We prove this by contradiction:
n+a−1
X bi − bi−1 bn+a − bn+a−1 0
( yi0 −yi∗ ) + (yn+a − yn+a
∗
)
i=n
c |{z} c
=c
p+1
X bi − bi−1 ∗
> (yi − yi0 )
i=n+a+1
c |{z}
=0
n+a−1 p+1
X bi − bi−1 bn+a − bn+a−1 0 X bi − bi−1 ∗
⇔ (c − yi∗ ) + (yn+a − yn+a
∗
)> yi
i=n
c c i=n+a+1
c
n+a−1 p+1
X bn+a − bn+a−1 0 X bi − bi−1
⇔ (bi − bi−1 ) + yn+a > yi∗
i=n
c i=n
c
n+a−1 p+1
X bn+a − bn+a−1 0 X
⇔ 1
(bi − bi−1 ) + yn+a > (bi − bi−1 )
i=n
c i=n
bn+a − bn+a−1 0
⇔ 2 (bn+a−1 − bn−1 ) + yn+a > (bp+1 − bn−1 )
c
y0
⇔ 3 bn+a−1 + (bn+a − bn+a−1 ) n+a > bp+1
c
⇔ bn+a−1 > bp+1
yi∗
1
The upper bound for yi∗ is yi∗ ≤ c ⇔ c ≤1
2
Telescoping series, +bn−1
3
The lower bound for yn+a
0
is 0
86
4.6. Numerical Results for the Integrated Cargo Allocation Problem
This is a contradiction, because n + a ≤ p + 1 and n + a − 1 < p + 1 and by

assumption 3, bn+a−1 < bp+1 holds.
Thus, we have proved that the inequality (4.62) holds and therefore the con-
structed integer solution is not worse than the LP optimal solution. The constructed
integer solution cannot be better than the optimal LP solution (LP relaxation) be-
cause the mixed integer solution is at least as good as the LP solution.
To construct a valid integer solution from the fractional z values in the optimal
solution, the definitions for z 0 and y 0 can be used.
4.6. Numerical Results for the Integrated Cargo Allocation

Problem
In this section, numerical results for the models and solution methods developed
in this chapter are presented. The models are implemented in C# with Microsoft’s
.N ET framework 4.5. Gurobi 5.6.0 with the dual simplex method is used to solve
the arc flow cargo allocation formulation and the restricted master problem in the
path flow formulation. The tests are performed on a machine with Intel Core2 Quad
CPUs with 2.83 GHz, 8 GB RAM and Windows 7 64 Bit. The implementation
uses the .NET library Quickgraph 3.04 to solve the shortest path subproblems.
Quickgraph is inspired by the Boost Graph Library5 and provides a wide range of
different graph algorithms such as Dijkstra’s algorithm (Dijkstra (1959)) that is
used to solve the pricing problems of the cargo allocation problem. The library uses
an accelerated version of the basic Dijkstra algorithm using Fibonacci heaps (see
(Cormen et al., 2001, p. 476ff)).
This section is organized as follows: First, the problem instances and the reference
network generation are introduced. Afterwards, numerical results for the arc-flow and
path-flow formulation are presented and discussed. For the arc-flow formulation,
we present several key performance indicators that can be used to evaluate the
networks and their practical usefulness. Note that the path-flow formulation uses
the same data and would lead to identical results. Finally, the effects of solving the
path-flow formulation heuristically are shown.
4.6.1. Problem Instances

The problem instances used throughout this thesis are described in more detail.
The cargo allocation problem requires problem instance data, such as ports and
cargo flows, but also existing liner networks. These liner networks are created for
all instances of the LINER-LIB 2012 benchmark suite that provides the data (see
http://www.linerlib.org/ and Brouer et al. (2013)).
4
The source code and Nuget package is available at http://quickgraph.codeplex.com/.
5
See http://www.boost.org/libs/graph/
87
LINER-LIB Benchmark Instances

Brouer et al. (2013) present a set of seven benchmark instances for the liner shipping
network design problem. The benchmark suite contains a set of ports with geograph-
ical position (longitude and latitude), a set or legs with distances, vessel types for
each instance as well as available vessel types and demands (cargo flows) for each
instance. The instance’s ports are limited to ports where cargo origins or destinates.
This is a simplification because ports without cargo can still act as transhipment
hubs. However, in practice it is unlikely that a port is called without leading to any
profit. The cargo flows in the LINER-LIB have a deterministic weekly quantity, a
revenue for transporting one unit and a maximum transit time. As described in
Chapter 2, in the scope of this thesis the transit time is seen on a port-to-port basis
(see Notteboom (2006)). Of the set of transit times, only those are used that can
be served with the vessel types’ maximum speed on a direct trip.
In Table 4.7 an overview of the different instance sizes is given. Instances Baltic
and WAF (West Africa) are feeder networks, meaning all cargo either origins or
destinates at a single port (transhipment hub). Thus, no regional cargo flows exist
between the ports except the transhipment hub. This hub is based on the tranship-
ment port used by Maersk Line. The number of ports, demands and legs in these
feeder networks are relatively small. The second group of networks Mediterranean
and Pacific have a similar amount of ports and legs. However, they differ clearly by
the number of cargo flows and transit times because they cover different geographical
regions.
The last group of networks are WorldSmall, EuropeAsia and WorldLarge with
many ports, cargo flows, legs and vessel types. Thus, they are expected to be difficult
to solve. In the scope of this thesis, parameters are evaluated for different groups
of networks, with a focus on the small and medium sized networks.
Instance Ports Cargo Flows Transit Legs Vessel Equipment

Times Types Types
Baltic 12 22 22 132 2 1
WAF 20 37 33 380 2 1
Mediterranean 39 365 365 1,482 3 1
Pacific 45 722 719 1,980 4 1
WorldSmall 47 1,764 1,748 2,162 6 1
EuropeAsia 114 4,000 3,996 12,882 6 1
WorldLarge 201 9,622 9,599 40,200 6 1
Table 4.7.: LINER-LIB 2012 instance information (see Brouer et al. (2013)).
An important detail for the numerical results shown in the next sections is the
demand distribution. Figure 4.8 shows the percentage of cargo flows for a percentage
of different measures in the WAF instance. This is either the cumulative sum of
88
quantity times revenue (cum.Q ∗ R, i.e. the maximum total revenue per week), the
cumulative quantity (cum.Q) and the revenue per cargo flow (cum.R). Figure 4.8
illustrates that the revenue is distributed nearly uniformly, i.e. most of the cargo
flows have similar revenues. On the other hand, the quantity and possible total
revenue is not uniformly distributed. For example, 20 percent of all cargo flows leads
to approximately 60% of the possible total revenue per week. This information is
used to evaluate the influence of limited cargo flows on the networks profit.
100%
90%
80%
Cumulative percent of Metric
70%
60%
50% cum. Q*R

cum. Q
40% cum. R
30%
20%
10%
0%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Percent of Cargo Flows
Figure 4.8.: Distribution of the cumulative cargo flow quantity (cum.Q), revenue
(cum.R) and quantity * revenue (cum.Q ∗ R) for the LINER-LIB WAF
instance.
Note that all instances show a similar result for the quantity and maximum
weekly revenue distribution, and this is plotted in the figures in Appendix C.1.
Especially the Mediterranean (see Figure C.2) instance indicates a tendency to
have a non-uniform revenue distribution, probably due to several cargo flows that
actually originate in other regions and were simplified in the scope of the benchmark
suite. The Pacific, WorldSmall, EuropeAsia and WorldLarge instances (see Figures
C.3ff) indicate that the cargo flow quantity is distributed on fewer cargo flows. This
observation is used in the following chapters to create new networks.
For further details on the sets and parameters provided by the LINER-LIB, see
Brouer et al. (2013).
89
Reference Networks for the Cargo Allocation Problem

The cargo allocation problem determines the optimal profit for given networks.
Therefore, networks for the LINER-LIB instances presented above must be con-
structed. For each problem instance, random networks are generated with a con-
struction heuristic described in Section 5.2.3. To respect the stochastic fluctuations
of the runtime, 20 networks are created per instance.
180
166,0
160
ø Number of services, number of legs
140
122,7
120
100
80
ø(No. of services per network)
60,6
60 55,7 ø(No. of legs per network)
45,7
38,5 37,8
40
20,5
20 12,8 11,45 12,95 13,7
2,5 5,75
0
Mediterranean
Baltic
EuropeAsia
WorldSmall
WorldLarge
Pacific
WAF
Figure 4.9.: Average sum of services and legs for random networks per LINER-LIB
instance.
The method creates relatively small networks which might not be representative
of evaluate larger networks. Thus, the heuristic is extended to fill up the randomly
created networks with pendulum services at randomly served ports to extend tran-
shipment possibilities. The number of services for a random network s of LINER-LIB
instance i is picked uniformly from the interval [|Ports i |/7, |Ports i |/2]. Figure 4.9
shows the average sum of the services and legs for all random networks per LINER-
LIB instance. On average, the number of services are similar to those presented in
the solutions in (Brouer et al., 2013, p. 26 and 27).
4.6.2. Arc-Flow Formulation

First, the results for the arc flow formulation of the cargo allocation problem are
presented. The model is solved with the commercial LP solver Gurobi 5.6.0 with the
90
default parameters with the exception that the dual simplex LP solution method is
used. The dual simplex has been evaluated to be faster compared to interior point
methods. Recall that the objective is to allocate cargo flows on one or more services
to get one or more paths from the origin to destination. This results in durations
at each service’s port which determines the required speed to perform a round trip.
In Figure 4.10 the results for all seven instances of the LINER-LIB and 20 ran-
domly generated networks per instance are presented. On the horizontal axis the
instance name and the number of services per network is given to have a simplified
measure of the network size. On the vertical axis, the logarithmic (base 10) average
runtime in seconds is shown. One can observe a tendency that larger networks for a
specific instance lead to higher runtimes. However, the complexity of a network is
also determined by other factors such as the transhipment possibilities, the number
of legs and the tightness of the service capacities.
The results in Figure 4.10 show a runtime of less than 0.1 seconds for the small
Baltic and WAF instance. Gurobi needs between one to six seconds to solve the
medium sized Mediterranean and Pacific instances to optimality. The WorldSmall
instance is more difficult to solve, leading to runtimes of up to six minutes for the
largest instance with 22 services and 30 seconds on average. Note that there are
two times as many cargo flows in this instance compared to the Pacific instance.
The large EuropeAsia and WorldLarge instances have runtimes of up to 23 and 50
minutes for the largest networks and about two minutes on average.
To get an overview about the utilization of the services, Figure 4.11 shows the best
average utilization per service and the percent of served cargo flows in the optimal
solution. One can observe that most of the networks have highly utilized services,
probably those that follow the large trades (because the construction heuristic
focuses on these cargo flows). Furthermore, one can observe that only a fraction
of the cargo flow is actually served, either because the networks do not provide a
connection between the origin and destinations or because the available container
paths are not cost efficient enough to transport the cargo flow and gaining the
revenue. Optimizing networks can lead to an increase of the served cargo flows
and the utilization, especially on the long legs that connect different geographical
regions.
With the help of Figure 4.12, the port durations and the average service speed is
evaluated. On the vertical axis of Figure 4.12, the average port duration in hours and
the average speed in knots is given. A constant total pilotage duration of 3 hours and
no constant buffer per port call is assumed. Additionally, we have used a relatively
high container move rate of 50 containers/hours. Two important observations can be
made from Figure 4.12: First, the average port call duration varies strongly between
different networks (and within ports of a service). Especially large transhipment
hubs are called more than 30 hours. One should note that the cargo allocation
model formulation can compensate networks where services have too many vessels
deployed. In this case, the model adds a buffer to the services’ ports. This can
91
10000
1000
100
ø Runtime [s]
10
0,1
0,01
1
3
5
22
26
33
37
43
46
52
55
6
8
10
13
15
6
8
10
12
16
18
21
3
5
7
9
29
35
47
60
63
76
84
92
7
9
11
13
16
19
22
Baltic EuropeAsia Mediterranean Pacific WAF WorldLarge WorldSmall
Figure 4.10.: Runtime results for solving the cargo allocation to optimality using
Gurobi 5.6 with the arc-flow formulation.
100%
90%
80%
70%
60%
Utilization [%]
50%
ø BAU%
ø CF%
40%
30%
20%
10%
0%
1
4
22
31
37
44
52
5
8
12
15
7
10
14
18
2
5
8
29
46
60
68
84
6
9
12
16
21
Figure 4.11.: Best average utilization per network (BAU %) and average served cargo
flow percentage (CF %).
92
70 14
60 12
50 10
ø Duration per port call [h]
ø Speed [kts]
40 8
ø Port call duration

30 6 ø Speed
20 4
10 2
0 0
1
4
22
31
37
44
52
5
8
12
15
7
10
14
18
2
5
8
29
46
60
68
84
6
9
12
16
21
Figure 4.12.: Average speed in knots and duration per port call in hours.
increase the port call duration in some instances. Second, due to the large bunker
cost, the optimal solution tries to steam as slow as possible in the offered services.
All instances provide vessel types with a minimum speed of 10 knots, leading to slow
steaming strategies for the given networks. In the large instances, the average speed
can be higher because some vessel types (e.g. Panamax vessels) have a minimum
speed of 12 knots.
The results indicate the importance of including detailed load dependent port
durations and speed optimization in the cargo allocation.
4.6.3. Path-Flow Formulation

The results for the arc-flow formulation already indicate promising runtimes to solve
large scale instances of the cargo allocation problem. The model formulation can
thereby be used to support the manual network planning process by allowing the
evaluation on a network to optimality.
One goal of this thesis is to solve the more complex liner shipping network design
problem (see Section 3.4) that integrates the cargo allocation problem. Thereby, a
more sophisticated delayed column generation (CG) method has been proposed that
models the problem as a path-flow formulation. For each cargo flow, a container path
is generated. The results for the CG are presented in the succeeding sections and
use the same networks as for the arc-flow formulation results. The restricted master
93
16
15
14
13
12
11
10
ø Runtime [s]
9
8
Runtime CG (No Parallel.) [s]
7
Runtime CG (Parallel.) [s]
6
5
4
3
2
1
0
1
4
22
31
37
44
52
5
8
12
15
7
10
14
18
2
5
8
7
10
13
17
22
54
61
29
32
28
Baltic EuropeAsia Mediterranean Pacific WAF WorldSmall WorldLarge
Figure 4.13.: Average runtime in seconds of the parallel and non parallel implemen-
tation for networks of different sizes for the LINER-LIB instances.
problem is solved with Gurobi 5.6.0 and the pricing problem with the Quickgraph
implementation of the Dijkstra algorithm.
Figure 4.13 shows the runtimes for the CG method on different LINER-LIB
instances and different networks. On the horizontal axis the instance and the number
of services are shown. On the vertical axis the average runtime in seconds of all
networks of the specific instance and network size is given. All instances are solved to
optimality. Solving multiple shortest-path pricing subproblems has been parallelized
to further accelerate the solution process. The parallelization uses different physical
processors or cores on modern computers to solve the subproblem for different cargo
flows simultaneously. The computer used to solve the cargo allocation problem
has four cores that leads to four parallel shortest path pricing problems. Using
the standard threading classes of the .NET framework, the solution time could
be decreased, especially on the medium and large instances (see Figure 4.13). On
average, the parallel implementation is about 60 % faster compared to the non-
parallel version.
The cargo allocation problem for the small instances’ random networks could be
solved in a fraction of a second to optimality. The instance WorldSmall, EuropeAsia
and WorldLarge took several seconds to be solved. Figure 4.13 highlights that there
does not exist a direct correlation between the network size and the CG runtime.
This can be explained by the randomness of the networks: If the networks are highly
94
interconnected and a lot of transhipment possibilities exist, it is expected that more

paths between the cargo flow’s origin and destination will be created in the column
generation process. However, a slight tendency that larger networks lead to larger
runtimes can be observed.
To get a better insight into the column generation method, Table 4.8 shows
different metrics for container paths resulting from the solution. Recall that a
container path is the routing of a cargo flow from its origin to its destination using
legs from one or more services with transhipment operations performed at the service
transitions. If the shortest (cheapest) path is capacitated, the CG method would
create an equally or more expensive alternative container path. The restricted master
problem decides if paths should be used to route the cargo flow. For each instance, 20
different networks are solved to optimality and the minimum, average and maximum
number of activated container paths in the optimal solution per demand is stored.
Table 4.8 highlights the minimum, average and maximum number of used container
paths for all networks Ni of instance i. Furthermore, the average standard deviation
σnC is given in the last column.
Number of container paths

LINER-LIB Instance min average max standard deviation
Baltic 1 1.017 2 0.048
EuropeAsia 1 1.010 3 0.078
Mediterranean 1 1.002 2 0.012
Pacific 1 1.014 3 0.096
WAF 1 1.008 2 0.027
WorldLarge 1 1.007 3 0.057
WorldSmall 1 1.019 3 0.130
Table 4.8.: Minimum, average and maximum number of container paths as well
as the standard deviation in the optimal solution for different problem
instances and networks.
In Table 4.8, two important observations can be made: First, all instances indicate
an average number of container paths per cargo flow of nearly one. Supported by
the low standard deviation, this means that most of the cargo is served on one
single path from the origin to the destination. This information is used for the cargo
allocation heuristics presented in Section 4.6.5.
4.6.4. Comparison of Numerical Results

The comparison of the arc and path-flow formulation of the integrated cargo al-
location problem is given in Figure 4.14. The figure shows the average runtime
in seconds per instance (logarithmically scaled) for both formulations as well as
the ratio between the runtimes. Values larger than one indicate a benefit of the
95
column generation solution method, for example a ratio of 100 indicates that the
column generation’s runtime to solve the cargo allocation is 100 times lower (better)
compared to solve the arc flow formulation with Gurobi 5.6.0’s dual simplex method.
10000 800
700
1000
ø Runtime Ratio Arc-flow/Path-flow

600
100
500
ø Runtime [s]
10
400 ø Runtime Arc-flow

ø Runtime Path-flow
1
300 ø Runtime Ratio
0,1
200
0,01
100
0,001 0
1
4
22
31
37
44
52
5
8
12
15
7
10
14
18
2
5
8
29
46
60
68
84
6
9
12
16
21
Figure 4.14.: Comparison of the arc-flow formulation runtimes and the path-flow
formulation runtimes.
Figure 4.14 shows the superiority of the column generation method, especially
on the medium and large instances. The ratios vary between a factor of 20 and
240 for the medium instances. Similar results can be seen for the large EuropeAsia
and WorldLarge instances. Here the minimum speedup is about 70 for the smallest
networks and up to several hundred times faster runtimes compared to the arc-flow
formulation in the largest networks.
The results clearly show that column generation is a suitable method to solve large
scale instances of the integrated cargo allocation problem in less than 10 seconds.
This allows planners to evaluate the effects of local service changes on a global basis.
It must be noted that the results are based on random networks that are created by
construction heuristics and should be improved during an optimization. Therefore,
it is expected that the initial networks are unprofitable. One network property
that can influence the column generation approach is the deployed service capacity.
In case of too many vessels, cargo can be allocated on its cheapest path without
rejecting the path and creating alternative paths during the solution method. In
case of very tight capacities, it is expected that several paths should be used for
96
the cargo flow, leading to increased runtimes. This dependency has to be observed
during the automatic optimization runs in Chapter 5.
4.6.5. Choosing an appropriate Approximation Level

In the previous sections, the column generation method has been solved to optimality
within several seconds. To further speed up the methods to solve the cargo allocation
problem hundreds and thousands of times within metaheuristics, further ways to
decrease the runtime must be explored.
The drawback is that this usually comes along with decreased accuracy of the
resulting solution. Throughout the next sections, the following definition of an
optimality gap is used (whereas z ∗ is the optimal solution of the cargo allocation
problem and zi is any, not necessarily optimal, solution):
|z ∗ − zi |
gap(zi ) =
z∗
The gap can be larger than one (100%) if both objective values are negative. To
calculate the optimal solution z ∗ to compare the heuristics below, the column
generation is solved to optimality each time. Note that due to operating system
activities, relatively small deviations in the runtime ratios of two optimal solutions
can occur.
There are mainly three different approaches to solve the cargo allocation with
the CG heuristically:
1. Limiting the amount of cargo flows
2. Varying the bunker cost discretization accurateness
3. Limiting the number of CG iterations
The effects of these heuristic approaches on the gap and the runtime improvement
are analyzed in the following sections. For simplicity, the results are shown for the
instances WAF, Mediterranean, Pacific and EuropeAsia. Results for the remaining
instances can be found in Appendix C.2.
Limiting the Amount of Cargo Flows

Liner shipping networks are designed to follow trade patterns (see Notteboom and
Rodrigue (2008)), in particular the underlying cargo flows. According to discussions
with a global liner carrier the networks are furthermore aligned to high volume
trades. Thus it is expected that transporting a fraction of the available cargo flows
reduces the runtime and the profit only slightly.
In the scope of this thesis, four different strategies to select the most promising
cargo flows are analyzed:
97
450%
400%
350%
300%
ø Gap [%]
250%
Q*R %
cum. Q*R
200%
cum. Q
cum. R
150%
100%
50%
0%
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
Pacific WAF
Figure 4.15.: Gap in percent for varied cargo flow percentages and different cargo
flow selection strategies (for the Pacific and WAF instance).
1. Use p percent of the cargo flows, ordered descending by quantity * revenue

(Q ∗ R%).
2. Use the cargo flows that provide p percent of the maximum total revenue
(cum.Q ∗ R).
3. Use the cargo flows that provide p percent of the total quantity (cum.Q).
4. Use the cargo flows that provide p percent of the revenue (cum.R).
The first strategy selects the largest number of cargo flows of all strategies, because
it is independent of the cumulative quantity or revenue. The second strategy select
much less cargo flows because the quantity * revenue is not uniformly distributed
in any of the analyzed instances (see Appendix C.1). The same applies to the
third strategy that also uses the cumulative method, but with the quantity instead.
The last strategy is a mix of the first and the succeeding two because the revenue
is distributed nearly uniformly in half of the instances (see Appendix C.1). The
abbreviations of the strategies are used in Figures 4.15 - 4.17.
In Figures 4.15 and 4.16 the resulting average gap for the instances Pacific, WAF
and EuropeAsia and Mediterranean as well as different cargo flow percentage p and
selection strategies is shown. The results are averaged over 20 networks per instance.
For most of the strategies, the instances shown in Figure 4.15 are highly sensitive
98
70%
60%
50%
40%
ø Gap [%]
Q*R %
cum. Q*R
30% cum. Q
cum. R
20%
10%
0%
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
EuropeAsia Mediterranean
Figure 4.16.: Gap in percent for varied cargo flow percentages and different cargo
flow selection strategies (for the EuropeAsia and Mediterranean in-
stance).
to decreased cargo flow amount. Especially the WAF networks give very high gaps
of 100% and more when the cargo is limited to 70% or less. The Pacific networks
are slightly more robust, but also result in high gaps with 50% of the cargo flows.
In the WAF networks, a few cargo flows lead to a relatively small percentage of the
total revenue, compared to other instances (see Figure 4.8). Limiting the amount
as shown in Figure 4.15 leads to few cargo flows and thereby to low profit and large
gaps. One can observe that the Q ∗ R% strategy works superior to most of the other
strategies for most of the networks because the number of cargo flows is relatively
high, even with a high approximation.
The results for the EuropeAsia and Mediterranean instance, presented in Figure
4.16, show a slightly different picture. First, the maximum gap is less than half of the
previous instances, although only 20% of the cargo flows are available. Furthermore,
the cumulative strategies have a gap of less than 10% with a cargo flow amount of
70% or more. Although they work better in these instances, the Q ∗ R% strategy
performs again much better with gaps less than 8% with a cargo flow amount of
40% or more.
To sum up, the Q ∗ R% strategy works best with all instances up to an approxima-
tion of 60% for the small and medium sized instances and 30% for the large instance.
The cum.Q strategy works well for approximations up to 70% for the large scale
instance as well, the cum.Q ∗ R up to 70% for the medium sized and the cum.R
99
21
20
19
18
17
ø Runtime Ratio [Opt/ Approximation]
16
15
14
13
12
11 Q*R %
10 cum. Q*R
9 cum. Q
8
cum. R
7
6
5
4
3
2
1
0
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
EuropeAsia Mediterranean Pacific WAF
Figure 4.17.: Runtime improvement when solving Column Generation heuristically

by limiting the cargo flow amount according to different strategies.
strategy for the small instance up to an approximation of 60%. The disadvantage

of the gap goes along with a speedup of the runtime.
In Figure 4.17 the average ratio of the optimal to the approximation runtime for
all instances are shown. Limiting the cargo flows of the large EuropeAsia instance
to 30%, the Q ∗ R% strategy improves the runtime by a factor of four, whereas the
cum.Q strategy improved the runtime by factor 6.5 for approximations of 70%. The
Q ∗ R% strategy for approximations up to 60% for the medium and small instances
lead to a runtime improvement by a maximum factor of 2.5 for the Pacific instance.
Due to the very small solution times of less than a tenth of a second (see Figure
4.13), the WAF networks average speedup is very small. The cum.Q ∗ R strategy
for approximations not less than 70% for the medium sized instances reduced the
runtime by a factor of 2.5 (Mediterranean) to 6 (Pacific). The cum.Q ∗ R strategy
for the WAF instance did not lead to any clear speed up.
To conclude, routing not less than 70% of the cargo flows using the cum.Q selection
strategy can improve the runtime by a factor of 6 and more in the large instance.
An alternative with a much lower gap and a runtime improvement by factor 4 is to
use no less than 30% of the cargo flows with the Q ∗ R% strategy. Routing no less
than 70% of the cargo flows using the cum.Q ∗ R selection strategy for the small
and medium instances can improve the runtime by a factor of 2 and more.
100
18% 2,5
15,83%
16%
2
14%
ø Runtime Ratio [Opt/Approximation]

12%
10,92%
10,52% 1,5
10% 9,55%
ø Gap [%]
ø Gap
8%
ø Runtime Ratio
1
6%
4%
0,5
2,36% 2,46%
2,02%
2% 1,62%
0,73% 0,94%
0,46% 0,56%
0,10% 0% 0,15% 0% 0,05% 0% 0,10% 0%
0% 0
2
16
20
16
20
16
20
16
20
Figure 4.18.: Average effects of the different bunker cost linearization support points
on the objective function and the runtime.
Varying Bunker Cost Discretization’s Supporting Points
Another possibility to heursitically solve the cargo allocation problem is to decrease

the accuracy of the bunker cost function. Originally, this function is a polynomial
of roughly degree three, but we linearize it using discrete intervals (see Chapter 4).
In theory, infinity many discrete intervals must be used to get the actual result of
the function. This approach is not applicable when solving linear programs where
a finite set of decision variables must be used. Figure 4.18 shows the dependency
of the cargo allocation problem on the number of support points (discretization
intervals).
On the horizontal axis the LINER-LIB instance and the number of support points
(maximum of 20 points) is given. On the left vertical axis of Figure 4.18 the average
gap in percent of all networks relative to the formulation that uses 20 support points
is given with the bars. On the right vertical axis the runtime ratio between the
optimal path flow and the approximation is shown as points in the diagram. The
results indicate that all instances show a gap of up to 15.83% when approximating
the total bunker consumption using a linear function with two support points (a
single line with constant slope). This can be explained by the fact that a single
line always overestimates the real bunker costs. If the number of support points
increases by a factor of two, the gap is reduced in average by a factor of five. This
101
14% 5
12,2% 4,5
12%
4

10% 3,5
3
8%
ø Gap [%]
2,5
ø Gap
6%
5,2% 2 ø Runtime Ratio
4,8%
4% 1,5
1
2%
0,5
0,6% 0,5%
0,1% 0,0% 0% 0,00% 0% 0% 0% 0% 0,00% 0% 0% 0% 0% 0% 0%
0% 0
Opt.
Opt.
Opt.
Opt.
1
4
Figure 4.19.: Gap and runtime improvement when terminating the column genera-
tion solution approach prematurely.
leads to a very small gap when using close to 20 support points. In the scope of
this strategic planning problem 20 support points is a reasonable approximation.
When extensions of this problem are used for operational allocation problems, the
bunker cost can be linearized more accurately. Due to tighter formulations in the
restricted master problem, the results in Figure 4.18 indicate a slight runtime ratio
decrease when a linear discretization is used.
To conclude the cargo allocation approximation with bunker cost, using between
4 to 20 support points leads to reasonable small gaps for all instances and a runtime
improvement factor if 0.5 to 1. Compared to reducing the cargo flows, the runtime
improvement when decreasing the bunker cost accurateness has a much lower benefit.
Limiting the number of Column Generation Iterations

Another heuristic approach for the column generation method is to terminate the
solution process prematurely. For example, aborting the process after the first iter-
ation leads to shortest (cheapest) container paths in the restricted master problem.
This method should give the largest runtime improvement, but can lead to the
largest gaps especially if the networks are highly capacitated.
Figure 4.19 shows the maximum number of iterations performed in the column
generation solution approach per instance. The last column Opt. on the horizontal
102
4.7. Comparison and Interpretation of Results
axis indicates that the cargo allocation has been solved to optimality, if necessary
with more than 4 maximum iterations. On the left vertical axis the average gap
to the optimal solution (for 20 networks per instance), on the right vertical axis
the runtime ratio is shown as points in the diagram. The figure illustrates that
all instances except the Mediterranean, which is solved after the first iteration to
optimality, have an average gap of 4.8 to 12.2% after the first iteration. The gap
decreases by a factor of approximately 10 (for some instances even faster) in each
additional column generation iteration. The runtime ratio improvement increases
with less iterations. Compared to the optimal solution, the speed up factor is about
four to five in the medium and large instances (that are not solved to optimality
after the first iteration). The ratio is expected to decrease with more complex and
capacitated networks because more container paths would be created. All instances
except the EuropeAsia network are solved to optimality after the second iteration
(the third iteration must be used to verify that the solution is optimal).
Summing up, the results indicate a large runtime benefit of factor two to five
when limiting the amount of CG iterations to one and a relatively small gap of 0
to 12%.
4.7. Comparison and Interpretation of Results

With the help of the path flow formulation and a column generation approach, the
optimal cargo allocation on large scale liner networks can be determined within a few
seconds. This is a speed up of factor 50 to 300 compared to the arc flow formulation
solved with Gurobi 5.6.0 using the dual simplex method. The state-of-the-art is
extended by integrating empty container repositioning, speed optimization, load
dependent vessel drafts and capacity types into the cargo allocation.
The exact results can be approximated by limiting the amount of cargo flows, the
bunker cost discretization accuracy and the maximum number of iterations in the
column generation solution process. Limiting the cargo flow amount and the CG
iterations seem to be the most promising approximation. This leads to low gaps
and speedups of four to eight (while keeping relatively low gaps) and further 2 to
4.5 by the maximum number of CG iterations. Using eight support points for the
bunker cost linearization still provides runtime speed ups of factor 1.5 to 2.
With the help of the developed optimization methods, liner shipping networks
can be quickly evaluated and the manual planning process of liner network planners
is supported. The next chapter presents optimization methods to automatically
improve liner networks.
103
5. Improving Networks - The Liner Shipping Network
Design Problem
The second and third goals of this thesis are to develop optimization methods for
liner shipping networks. The assumption of fixed services as in the cargo allocation
problem is relaxed in this problem. The decision is to choose one or more liner
services, each consisting of a port rotation (port sequence), a vessel type and a
number of deployed vessels. The liner services must be chosen in such a way that the
profit, according to the cargo allocation, is maximized (see Chapter 4). The problem
formulation in this thesis extends the state-of-the-art by additionally incorporating
transit times and embargo constraints in the liner shipping network design problem.
For simplicity, weekly service frequencies are assumed throughout this chapter.
Furthermore, cabotage regions and other vessel specific constraints that would re-
quire specific vessels are not considered. These assumptions must be relaxed in
successive planning problems such as the fleet deployment.
A mixed integer formulation for the network design problem is presented. With
this model, optimal networks can be created for very small instances and the gap to
the optimal solution be calculated. Afterwards, two metaheuristics for the network
design problem are introduced. Therefore, first a decomposition concept that is
used for the construction and improvement heuristics is presented. After providing
numerical results for the metaheuristics, the impact of approximated fitness functions
(surrogates) on the metaheuristic convergence is analyzed. A sensitivity analysis on
uncertain bunker cost gives an outlook whether uncertainty should be included in
future solution approaches.
Finally, the developed metaheuristics are evaluated on a real-world liner shipping
network.
5.1. Mixed Integer Formulation

The liner shipping network design problem with transit times, partner services,
deadweight scales and speed optimization is presented as a mixed integer program
(MIP). Non-linear constraints are linearized using the L01 approach from Padberg
(2000) (used for the cargo allocation as well). The main difficulty when formulating
the problem as a MIP are the speed constraints and variable number of vessels. The
speed of the vessels deployed on a service is given by dividing the service distance
by the duration at sea. Both terms are decision variables which means the fraction
must be linearized. Therefore, several sets must be introduced that allow the linear
105
5. Improving Networks - The Liner Shipping Network Design Problem
formulation of the integrated speed optimization. After presenting the mathematical

model, numerical results for the small LINER-LIB instances are presented.
5.1.1. Mathematical Model

The liner shipping network design problem can either be based on a set of services or
vessels that are deployed on services (see e.g. Reinhardt and Pisinger (2010)). In the
scope of this thesis, the model is based on services because the number of services is
usually much lower compared to the vessels owned or operated by a liner carrier. The
model uses a maximum number of services in the given problem instance, referred
to as service candidates. An upper bound for the maximum number of vessels per
service, a discrete set of possible speeds and a number of port calls (see Chapter 4)
must be provided to linearize the non-linear bunker consumption constraints.
The model presented below incorporates the cargo allocation problem (CAP),
because it determines the service speed and the cargo allocation that defines the
bunker and container handling cost. Thus, several aspects are similar to the CAP.
In the following tables, the additional sets, parameters and variables required for
the mixed integer model are presented in addition to the sets and parameters from
Section 4.2.
Sets
K Set of discrete speeds in knots
S Set of all services, S = S O ∪ S P
SO Services (candidates) that can be operated by the carrier
SP Set of all partner services
L Set of network layers
LLs Legs used by partner service s ∈ S P
LL = L × L set of all layered legs that can be used to create new services
VC Set of deployable vessels per operated service
T Set of transit-time requirements, T ⊆ P × P
PL = L × P Set of layered ports
P̄ Set of port sets that cannot be called together on one service due
to embargo constraints
Parameters
f Fixed frequency in days of all operated services, f = 7
γ Fixed percentage of all partner cargo that must be transported
θt Maximum transit-time duration in days for transit-time requirement
t = (i, j) ∈ T
bcv,k Bunker consumption in metric tons per day for a vessel of type
v ∈ V T steaming at k knots
106
MF Number of containers that can be transported in the planning hori-

zon
MP C Maximum number of calls of the same port within one service round
trip
MP Maximum number of port calls per service
MK
k Maximum voyage duration in days between two ports
MSea
k Upper bound for the maximum duration at sea per service
MTrMax
i,j Upper bound for the maximum duration in days to steam leg (i, j)
Decision variables
αn ∈ R+ 0 Served quantity of cargo flow n ∈ N , αn ≤ qnM ax
xLs,n,i,j,l,l0 ∈ R+
0 Transported quantity of cargo flow n ∈ N on leg (i, j, l, l0 ) of
service s ∈ S
L
ls,n,p,l ∈ R+
0 Loaded quantity of cargo flow n ∈ N on service s ∈ S at port
(p, l)
uLs,n,p,l ∈ R+
0 Unloaded quantity of cargo flow n ∈ N from service s ∈ S at
port (p, l)
xE +
s,e,i,j,l,l0 ∈ R0 Transported quantity of empty containers of type e ∈ E on
leg (i, j, l, l0 ) of service s ∈ S
E
ls,e,p,l ∈ R+
0 Loaded quantity of empty containers of type e ∈ E on service
s ∈ S at port (p, l)
uE +
s,e,p,l ∈ R0 Unloaded quantity of empty containers of type e ∈ E from
service s ∈ S at port (p, l)
ysS ∈ {0, 1} Indicates whether service s ∈ S O is activated
0 ∈ {0, 1} Indicates whether service s ∈ S O uses leg (i, j, l, l0 )
L
ys,i,j,l,l
ys,i,l,v ∈ {0, 1}
P
Indicates whether service s ∈ S O calls port (i, l) ∈ P L with
vessel type v ∈ V T
VT
ys,v ∈ {0, 1} Indicates whether service s ∈ S O uses vessel type v ∈ V T
ws,v,vc ∈ {0, 1} Indicates whether service s ∈ S O uses vc ∈ V C many vessels
of type v ∈ V T
bs,v,vc,k ∈ R+
0 Bunker consumption of service s ∈ S O in the planning horizon
if vc ∈ V C many vessels of type v ∈ V T steam at k kts
S
τs,k ∈ R+
0 Duration at sea in days in the planning horizon when all vessels
of service s ∈ S O steam k ∈ K knots on all legs
K
ys,k ∈ {0, 1} Variable indicating whether all vessels deployed on service
s ∈ S O steam at k ∈ K knots
B
τs,p,l ∈ R+
0 Additional buffer time in days in the whole planning horizon
at port (p, l) of service s ∈ S O
I
δs,p,l,v ∈ R+
0 Incoming draft of vessel type v ∈ V T at service’s s ∈ S O port
call (p, l) ∈ P L
107
O
δs,p,l,v ∈ R+
0 Outgoing draft of vessel type v ∈ V T at service’s s ∈ S O port
call (p, l) ∈ P L
0 ∈ {0, 1} Variable indicating whether service s ∈ S leg (i, j, l, l0 ) ∈ LL
T
ys,t,i,j,l,l
is used to fulfill the transit time t ∈ T
T
τs,t,i,j,l,l +
0 ∈ R0 Duration in days to steam leg (i, j, l, l0 ) ∈ LL at service s ∈ S
in days of transit time requirement t ∈ T
ρP ∈ R+
0 Slack variable for not transported partner cargo flows that
must be transported
The objective is to maximize the carrier’s profit before interest and taxes (EBIT)
in the whole planning horizon:
X
max profit = rn − φC (5.1)

e n · αn
n∈N
 
X X X τ P X
− (φPp,vC ys,p,l,v
 )+ (φD
v vc ws,v,vc )
 (5.2)
f
s∈S O v∈V T p∈P L vc∈V C
X X X
− (φT bs,v,vc,k ) (5.3)
s∈S O v∈V T,vc∈V C k∈K
 ( ! 
p , if p = on ∨ p = dn
φCH
X X X
−  (ls,n,p,l
L
+ uLs,n,p,l )
s∈S O (p,l)∈P L n∈N
φTp S , else
(5.4)
 
X X X
−  φTp S (ls,e,p,l
E
+ uE
s,e,p,l )
 (5.5)
s∈S O (p,l)∈P L e∈E
X X X X
−
s∈S P sg=(i,j,l,l0 )∈SG s rg∈RGsg r∈rg
!
X X
L
φSs,i,en un,r ls,n,i,l + φSs,i,e ue,r ls,e,i,l
E
(5.6)
n∈N e∈E
− PρP (5.7)
Term (5.1) determines the revenue for transporting the cargo flows minus the
container depreciation. The revenue is gained for transporting both, own and partner
cargo. The second term (5.2) subtracts the port call cost for all used services and
the vessel depreciation in the planning horizon. The third term (5.3) calculates
the bunker cost for the deployed vessels in the whole planning horizon. In the
fourth term (5.4) the container handling cost at the port of origin and destination
as well as the transhipment costs at each port and service are calculated for the
108
laden containers (cargo flows) and for empty containers in term (5.5). Term (5.6)
calculates the slot cost when using partner services. The slot cost are modeled by
activating the service specific leg cost when loading a container onto the partner
service’s vessels. Finally, term (5.7) imposes penalty cost if partner cargo, on which
the partners have contractually agreed, is not transported.
The objective is subject to several constraints that are presented in the remainder
of this section. They are distinguished by flow balance and transhipment, draft
and capacity, liner service, round trip and port duration, transit time and business
constraints.
Flow Balance and Transhipment Constraints
The following constraints ensure the flow within the network and explicitly respect
transhipment operations to correctly associate the costs for the objective function.
X X ∀s ∈ S, n ∈ N,
xLs,n,i,p,l0 ,l + ls,n,p,l
L
= xLs,n,p,j,l,l0 + uLs,n,p,l
(p, l) ∈ P L
(i,p,l0 ,l)∈LL (p,j,l,l0 )∈LL
(5.8)
X X ∀s ∈ S, e ∈ E,
s,e,i,p,l0 ,l + ls,e,p,l =
xE s,e,p,j,l,l0 + us,e,p,l
E
xE E
(p, l) ∈ P L
(5.9)
Constraints (5.8) and (5.9) ensure the flow balance for services and ports. Incom-
ing containers must be either transported to the next called port of the service or
unloaded. Outgoing containers are either transported to the port by the service
or loaded at the port from another service (transhipment operation). The explicit
modeling of loading and unloading decisions is required to associate transhipment
costs with these operations.

X X αn ,
 if p = on
L
ls,n,p,l = us,n,p,l + −αn , if p = dn
L
∀n ∈ N, p ∈ P (5.10)
0,

l∈L,s∈S l∈L,s∈S 
else
Containers unloaded from a service at a port must be loaded by another service

at the same port. An exception is made when a laden container arrives at its port of
destination dn and is removed from the network (see constraints (5.10)). Containers
can also enter the network at their port of origin on .
109
X X X X
E
ls,e,p,l = uE
s,e,p,l − αn + αn ∀e ∈ E, p ∈ P
l∈L,s∈S l∈L,s∈S n∈N :on =p n∈N :dn =p
(5.11)
Similar constraints ensure that all transported laden containers are balanced with
empty containers (see constraints (5.11)). The sum of loaded empty containers
of type e at port p must equal the sum of unloaded empty containers (due to
transhipment) and the difference of unloaded and loaded laden containers that can
be used to serve other cargo flows with empty containers.
Vessel Draft and Capacity Constraints

The next group of constraints ensure the vessel type - port compatibility and vessel
capacity for all operated liner services.
P
dv · ys,p,l,v ≤ DpMax ∀s ∈ S O , (p, l) ∈ P L , v ∈ V T (5.12)
Constraints (5.12) ensure that the lightship1 draft of vessel type v used on service
s is less than the maximum depth at port p. This is a basic requirement to deploy
vessel type v to service s.
Beside this lightship draft load dependent drafts are considered as well in the
next constraints.
!
X X X
un,weight xLs,n,i,p,l0 ,l + ue,weight xE
s,e,i,p,l0 ,l dws Sv
(i,p,l0 ,l)∈LL n∈N e∈E
∀s ∈ S O ,
+ dws Iv = δs,p,l,v
I
(p, l) ∈ P L ,
v ∈VT
(5.13)
!
X X X
un,weight xLs,n,i,p,l0 ,l + ue,weight xE
s,e,i,p,l0 ,l dws Sv
(p,j,l,l0 )∈LL n∈N e∈E
∀s ∈ S O ,
+ dws Iv = O
δs,p,l,v (p, l) ∈ P L ,
v ∈VT
(5.14)
1
Lightship draft refers to a vessel’s draft without any payload.
110
Constraints (5.13) and (5.14) set the load dependent auxiliary draft variables for
all vessels used on service s. The draft is calculated by the linearization of the
deadweight scale (DWS), provided for all vessel types. The constraint uses the slope
dws Sv and the intersection dws Iv for the selected vessel type v of the service (see
Chapter 4) to calculate a draft of service s at port p, l for vessel type v. Note that
the draft is set for all cargo transported to and from the port in the planning horizon.
The draft auxiliary variable is set independent of the vessel type that is actually
selected by the model.
τ
I
δs,p,l,v ≤ DpMax + MDWS
v (1 − ys,p,l,v
P
) ∀s ∈ S O , (p, l) ∈ P L , v ∈ V T (5.15)
f
τ
O
δs,p,l,v ≤ DpMax + MDWS
v (1 − ys,p,l,v
P
) ∀s ∈ S O , (p, l) ∈ P L , v ∈ V T (5.16)
f
The load dependent vessel type - port compatibility is ensured by constraints

(5.15) for incoming and by constraints (5.16) for outgoing cargo to/ from port p. Note
that the deadweight scale constraints are imposed on operated services only, because
the actual load of the partner services are unknown. Both constraints deactivate the
draft upper bound when vessel type v and port p, l is not been selected for service s.
Otherwise, the maximum depth DpMax must hold at port p. We multiply the depth
with the number of calls in the planning horizon to account for the vessel draft in
the planning horizon.
X X
un,r xLs,n,i,j,l,l0 + ue,r xE
s,e,i,j,l,l0 ≤
n∈N,r∈rg e∈E,r∈rg
∀s ∈ S O ,
X τ VT
Cv,rg ys,v (i, j, l, l0 ) ∈ LL , (5.17)
f
v∈V T rg ∈ RG
Constraints (5.17) limits the capacity for each resource group rg and used service
leg (i, j, l, l0 ) for the vessels on service s operated by the carrier at hand in the
planning horizon. Laden and empty containers utilize different resources r, such
as the container weight, according to the coefficients un,r and ue,r . The capacity
constraint is imposed for the selected vessel type ys,v VT
in service s.
Liner Service Constraints
The next set of constraints ensure that the port rotation of each service performs a
round trip. These constraints are similar to the vehicle routing problem with the
exception that no predetermined depots exist.
111
X X
L
ys,i,p,l 0 ,l =
L
ys,p,j,l,l 0 ∀s ∈ S O , (p, l) ∈ P L (5.18)
X
L
ys,i,p,l 0 ,l ≤ 1 ∀s ∈ S O , (p, l) ∈ P L (5.19)
(i,p,l0 ,l)∈LL
X
L
ys,p,j,l,l 0 ≤ 1 ∀s ∈ S O , (p, l) ∈ P L (5.20)
(p,j,l,l0 )∈LL
X X ∀s ∈ S O ,
xLs,n,i,j,l,l0 + xE F L
s,e,i,j,l,l0 ≤ M ys,i,j,l,l0 (5.21)
(i, j, l, l0 ) ∈ LL
n∈N e∈E
X
0 ≤ 1 (5.22)
L
ys,i,j,l,l ∀s ∈ S O , l ∈ L
(i,j,l,l0 )∈LL :l0 6=l
X
L
ys,i,j,l,l 0 ≤ 1 ∀s ∈ S O , (i, j) ∈ L (5.23)
(i,j,l,l0 )∈LL
X
L
ys,p,j,l,l 0 ≤ M
PC
∀s ∈ S O , p ∈ P (5.24)
(i,p,l,l0 )∈LL
X
0 ≥ 2ys (5.25)
L S
ys,i,j,l,l ∀s ∈ S O
(i,j,l0 ,l)∈LL
Constraints (5.18) ensure the connectivity at each service’s port. If the port p
is called by operated service s, an incoming and outgoing leg must be activated.
Constraints (5.19) and (5.20) ensure that a port on a specific layer is not visited
more than once. Constraints (5.21) enable the routing of containers on the service’s
leg if it is activated. The upper bound MF can be set to the number of containers
that can be transported in the planning horizon, i.e. the sum of qnM ax variables. The
bounds can be improved if any information about the maximum service length is
known, such as the largest vessel type’s slot capacity times the weeks in the planning
horizon. Constraints (5.22) state that not more than one leg connecting two layers
per service can exist. This is required to avoid cycling between different network
layers. Constraints (5.23) ensure that a specific leg between port i and j is not used
more than once per service which is commonly found in practical networks. The
layer constraints (5.24) ensure that a port is not called more than MP C many times
per service round trip. Although theoretically not limited, MP C is set to two in
practical networks. Constraints (5.25) ensure that if a service is activated, at least
two legs for a pendulum service are used.
112
Liner Service Round Trip Constraints
The model now can create port rotations that allow complex route types. The next
step is to select exactly one vessel type per service and deploy vessels such that the
service can perform a weekly round trip.
X
P
ys,p,l,v ≤ MP ys,v
VT
∀s ∈ S O , v ∈ V T (5.26)
(p,l)∈P L
X X
P
ys,p,l,v = L
ys,i,p,l 0 ,l ∀s ∈ S O , (p, l) ∈ P L (5.27)
v∈V T (i,p,l0 ,l)∈LL
X
VT
ys,v ≤1 ∀s ∈ S O (5.28)
v∈V T
X
ws,v,vc = ys,v
VT
∀s ∈ S O , v ∈ V T (5.29)
vc∈V C
X
ws,v,vc = ysS ∀s ∈ S O , v ∈ V T (5.30)
vc∈V C
Constraints (5.26) ensure that the vessel type variable ys,v

VT
is activated if a port is
called by a service with vessel type v. M limits the maximum number of port calls
P
per service. Practical network services indicate that a value of 30 for the maximum
number of port calls per service is a reasonable upper bound. Long services using
large vessels are still limited by the panama canal’s locks, see Chapter 2. However,
the offering of long round-the-world services is performed in practice, although
several disadvantages are recognized (see (Stopford, 2009, p. 528)). Constraints
(5.27) state that if a port is visited by a vessel in service s, one outgoing leg must
exist. Constraints (5.28) limit the number of vessel types per service to one. To
allow more than one vessel type per service, the value can be increased. Due to
decreased time charter rates for medium sized panamax vessels, this could be an
interesting option in the future. After the vessel type is selected, a number of vessels
of this type must be determined. This is formalized in constraints (5.29), by forcing
ws,v be exactly one predetermined number vc of vessels per service. If service s is
used, at least one vessel must be deployed (see constraints (5.30)).
The next two groups of constraints deal with the timing to deploy enough services
to ensure weekly round trips. This is the major mode of operation in liner services
and highly relevant when designing practical networks.
X
K
ys,k = ysS ∀s ∈ S O (5.31)
k∈K
113
∀s ∈ S O ,
S
τs,k ≤ MSea ys,k
K
(5.32)
k∈K
X 1 τ ∀s ∈ S O ,
k (1 − ys,k ) (5.33)
S L
τs,k ≥ li,j · ys,i,j,l,l 0 − MK K
k24 f k∈K
(i,j,l,l0 )∈LL
 
X X 1  X X
vc f ws,v,vc = τS + P 
τs,p,l ∀s ∈ S O (5.34)
v∈V T vc∈V C
τ /f k∈K s,k L
(p,l)∈P
For each activated service, exactly one average speed must be selected (see constraints
(5.31)). Constraints (5.32) ensure no duration at sea is set if the speed is not activated.
MSea
k is an upper bound for the total maximum duration at sea. This value is set
to τ vc∈V
max
C vc and determines the duration when the vessel do not serve any cargo at
the ports.
The selected service speed leads to the duration at sea in days, calculated by
constraints (5.33). The duration is the distance of each activated leg divided by the
selected speed. To simplify the handling of the days at sea for different vessels in the
planning horizon, the value is multiplied by the number of weeks in the planning
horizon τ /f . This equals the number of vessels times the number of round trips by
the assumption of weekly port calls. The term at the right hand side of constraints
(5.33) disables the constraints if the speed is not selected. MK k can be set to the
maximum voyage duration in days between two ports, for example by taking the
maximum leg distance divided by the minimum speed of all vessel types.
Constraints (5.34) ensure that enough vessels are deployed to perform weekly port
calls. The round trip on the left hand side is calculated by taking the number of
deployed vessels vc times the constant frequency f (if vc vessels have been selected
for service s). The round trip time for each operated service must equal the duration
at sea τs,k
S
plus the duration in ports τs,p,l
P
. To determine the duration per round trip,
the variables are divided by the number of weeks in the planning horizon. In case
of non-predetermined frequencies, f would be a decision variable for each service
and would introduce a high degree of further complexity to the model.
Port Duration and Bunker Cost Constraints

Before ensuring the transit times are held, the port call duration and resulting
bunker cost are calculated.
!
X X
P
= τpE L
+ uLs,n,p,l + E
+ uE

τs,p,l ls,n,p,l ls,e,p,l s,e,p,l
n∈N e∈E
τ P ∀s ∈ S O ,
+ τpAdd ys,p,l + τs,p,l
B
(5.35)
f (p, l) ∈ P L
114
∀s ∈ S O ,
v ∈ V T,
k (1 − ys,k ) − Mk (1 − ws,v,vc ) (5.36)
S
bs,v,vc,k ≥ τs,k bcv,k − MSea S Sea
vc ∈ V C,
k∈K
Constraints (5.35) set the overall port call duration in days at each port based
on the duration in days to move one container at port p, τpE . The time to load and
unload laden and empty container is considered. The load and unload time is zero
if the port p, l is not called by service s because no containers can be transported.
The parameters τpAdd add the constant pilotage and strategic buffer in days per port
call and must be multiplied by the number of all port calls in the planning horizon.
This additional duration is only imposed if service s calls port p on layer l. We use
the model buffer variable τs,p,l
B
per service and port call to increase the flexibility of
the model for holding the round trip time. The model can decide to stay longer at
certain ports but increase the service speed for holding the transit times.
Constraints (5.36) determine the activity of the bunker consumption helper vari-
able bs,v,vc,k that stores the bunker consumption of service s in the planning horizon
when steaming with vc vessels of type v at k knots . The variable is used in the
objective to get the associated bunker cost. The constraint is deactivated by the
second and third term in case the speed, vessel type or vessel count is not selected
for the service.
Transit Time Constraints

The next constraints ensure the transit time between two ports is enforced if cargo
is transported. The idea is to determine a path between two ports and determine
the overall path duration by the sum of all used service legs.
X 1

li,j
T
τt,s,i,j,l,l 0 ≥ L
li,j ys,i,j,l,l 0 − (1 − ys,k )
K
k∈K
k24 k24
∀t ∈ T,
1 1 P
+ τs,i,l + τs,j,l
P
− MTrMax T
(5.37)

0 · yt,s,i,j,l,l 0 s ∈ S,
τ /f 2 i,j
(i, j, l, l0 ) ∈ LL
In constraints (5.37), the transit time specific’s duration τt,s,i,j,l,l

T
0 in days on leg
(i, j, l, l ) of service s is calculated. The duration consists of the duration at sea (first
0
term of the right hand side) plus the time at port i (τs,i,l P
) and port j (τs,j,l
P
0 ). The
port durations must be transformed to weekly durations (number of weeks in the

planning horizon) because they store the overall duration in the planning horizon.
At a port, the time when a specific container is unloaded or loaded is usually
unknown in the strategic planning horizon. It is assumed that it takes on average
115
half of the total port duration τs,i,l

P
. For partner services s ∈ S P , a constant duration
(such as 24 hours) and an average design speed can be assumed because the actual
speed is often not known for certain. MTrMax i,j is an upper bound for the maximum
duration to steam leg (i, j). For example, the duration when steaming with the
minimum speed and the maximum port duration can be used as an upper bound.
The following constraints ensure that if one unit of laden containers is transported
between ports o and d, the transit time between the ports must hold.
T
τt,s,i,j,l,l TrMax
0 ≤ Mi,j
T
· yt,s,i,j,l,l 0 ∀t ∈ T, s ∈ S, (i, j, l, l0 ) ∈ LL (5.38)
T
τt,s,i,j,l,l 0 ≤ MTrMax
i,j · L
ys,i,j,l,l 0 ∀t ∈ T, s ∈ S, (i, j, l, l ) ∈ L 0 L
(5.39)
Constraints (5.38) limit the duration of service s leg (i, j, l, l0 ) if it should not be
considered for the transit time. Similar, the duration on the leg can only be used if
it is served by service s (see constraints 5.39).
X X ∀t ∈ T, (p, l) ∈ P L ,
0 ,l = (5.40)
T T
yt,s,i,p,l yt,s,p,j,l,l
p 6= ot ∧ p 6= dt
0
(i,p,l0 ,l)∈LL ,s∈S (p,j,l,l0 ,s∈S)∈LL
X
T
yt,s,p,j,l,l 0 ≤ 1 ∀t ∈ T, (p, l) ∈ P L , p = ot (5.41)
(p,j,l,l0 )∈LL ,s∈S
X
T
yt,s,i,p,l 0 ,l ≤ 1 ∀t ∈ T, (p, l) ∈ P L , p = dt (5.42)
(i,p,l0 ,l)∈LL ,s∈S
The transit time duration is modeled by finding a path from ot to dt . Each

activated path variable must be part of a path (see constraints (5.40)) that does
not contain more than one incoming leg (see constraints 5.41) and outgoing leg (see
constraints (5.42)).
X X
T
τt,s,i,j,l,l 0 ≤ θt ∀t ∈ T (5.43)
s∈S (i,j,l,l0 )∈LL
X X
αn ≤ M F · T
yt,s,o t ,j,l,l
0 ∀t ∈ T (5.44)
n∈N :on =ot ∧dn =dt s∈S,(ot ,j,l,l)∈LL
Constraints (5.43) state that the duration of all activated legs on the path must not
be larger than the transit time duration θt in days. Finally, constraints (5.44) ensure
that the transit time between ports o and d holds when any cargo is transported
between o and d.
116
Business Constraints
The following constraints define further business rules that were identified through
discussions with liner carriers (see Chapter 2).
X X
αn ≥ θ · qnM ax − ρP (5.45)
n∈N P n∈N P
Constraints (5.45) ensure that a minimum required percentage of partner cargo

is served. If this is not possible, the penalty variable ρP is activated.
X
un,r · xLs,n,i,j,l,l0
n∈N O ,r∈Rn
X
+ ue,r · xE P
s,e,i,j,l,l0 ≤ Cs,i,j,l,l0 ,rg ∀s ∈ S P , (i, j, l, l0 ) ∈ Ls , rg ∈ RG
e∈E,r∈Re
(5.46)
Similar to the capacity constraint of operated services, constraints (5.46) ensure

the maximum partner capacity per leg for different resource groups. Network plan-
ners are able to set contractually defined slot capacities for partner services on a
per leg basis.
!
X X
P
ys,p,l,v ≤1 ∀s ∈ S O , p̄ ∈ P̄
p∈p̄ l∈L,v∈V T
(5.47)
Constraints (5.47) avoid ports on the same service that are subject to embargo.
Ports that are forbidden to be called together on a service are sets p̄ ∈ P̄ containing
two or more ports. The constraint ensures that out of each set only one port p is
called per service.
To include partner services in the network, the service, leg, vessel type and vessel
count variables are fixed in the model. The objective does not impose any costs for
the partner services in S P (except the slot cost) because the partner is in charge of
paying these. The draft constraints are not relevant for partners because the actual
tonnage is unknown.
5.1.2. Subtour Elimination Constraints

The liner shipping network design formulation in the previous Section 5.1 allows
subtours within a given service that are not allowed. Figure 5.1 illustrates an invalid
117
service with two subtours that are disconnected from each other. Note that this
configuration would be valid if the subtours were selected as individual services.
The removal of subtours must be assured for the final solution of the mixed integer
model.
Figure 5.1.: Invalid port rotation for one service due to existing subtours.
Several well-known formulations to avoid subtours exist in literature in the context

of the traveling salesman and vehicle routing problem (see (Suhl and Mellouli, 2013,
p. 244)). N denotes the set of all nodes (ports and layers), S the set of all subsets
of N . Three possibilities are:
X
xij ≥ 1, ∀ ∅ ⊂ S ⊆ N − 1 (5.48)
i∈S,j ∈S
/
X
xij ≤ |R| − 1, ∀ ∅ ⊂ R ⊆ {2, 3, · · · , |N |} (5.49)
i∈R,j∈R
zi − zj + n · xij ≤ n − 1, ∀ i, j := 2, · · · , n (i 6= j) (5.50)
Constraints (5.48) ensure that all nodes in each subset S is connected with every
other node that is not in the subset. This works by enforcing an activated edge
xij between these subsets. Variant (5.49) avoid cycles within each subset R of
{2, 3, · · · , |N |}. Each subset with |R| many nodes must not contain more than
|R| − 1 edges. Otherwise, the tour contains cycles. Constraints (5.50) introduce an
sequencing of the nodes to disable subtours that do not contain the starting node
which is assumed to be node 1. The formulation uses additional variables.
One can observe that constraints (5.48) and (5.50) are not applicable to the liner
shipping network design problem because not all ports must be served, however,
they can be served several times. The subtour elimination constraints proposed for
the liner shipping network design problem is based on constraints (5.49) and shown
in constraints (5.51). The optimal solution of the mixed integer program is denoted
as x∗ and S ∗ a set of activated service candidates, whereas each service s ∈ S ∗ uses
a set of legs L∗s and (layered) ports Ps∗ .
118
X
ys,i,p,l0 ,l ≤ |Ps∗ | − 1 ∀s ∈ S̄ ∗ (5.51)
(i,j,l,l0 )∈L∗s
The subtour elimination constraints (5.51) are added for all services that contain
subtours S̄ ∗ by checking each service s ∈ S ∗ . If the service contains subtours this
means that too many edges are used to serve the ports. We therefore force the
number of activated legs to be equal to the number of activated ports minus one. It
does not cut feasible solutions because the subtours can still be served by different
services, each not containing subtours.
The subtour elimination constraints can be added to the model before solving
it. However, enumerating all subsets of activated edges containing subtours (legs)
leads to an exponential number of constraints. To reduce this number, their creation
can be delayed. One way is to solve the mixed integer program to optimality, check
for subtours, exclude them from the model and resolve it. The drawback of this
approach is twofold: First, solving a model can be very time consuming and solving
it several times might not be applicable. Second, the gap to the optimal solution
provided by the MIP solver is not guaranteed to be the real gap because constraints
are missing.
To circumvent these problems, the subtour elimination constraints are inserted
within the solution process of the MIP solver. Most of the available solvers provide
callbacks to insert custom application logic at certain events such as when incum-
bents (new best integer solutions) are found. In the scope of this thesis, the callback
method of Gurobi (2014) is used to add the subtour elimination constraint when a
subtour in an incumbent solution is found. These lazy constraints are then propa-
gated in the solver’s threads that solve the B&B tree. The usage of lazy constraints
requires specific solver parameters. For example, dual information from the presolver
cannot be used anymore because the dual bounds are not valid when constraints
are missing in the initial model.
The mixed integer problem combined with the subtour elimination constraints can
be solved using commercial and non-commercial MIP solvers. In the next section,
runtime results for solving the model with Gurobi are presented.
5.1.3. Numerical Results for the Mixed Integer Program

The previously described mixed integer program is applied to the LINER-LIB bench-
mark instances2 , already presented for the cargo allocation problem (see Chapter 4).
The numerical results in the remainder of this chapter are performed on identical
computers with the following configuration: Intel Core 2 Quad CPU 2.83 GHz, 8
2
The LINER-LIB instances are developed by Brouer et al. (2013).
119
GB RAM and Windows 7 64 Bit. The model is implemented in C# using .NET

framework 4.5 and compiled as a 64 Bit application. Gurobi 5.6.0 64 Bit is used as
MIP-solver. To get information on the solution progress, such as the current lower
and upper bound as well as the activities, callbacks are registered in Gurobi.
Table 5.4 shows the model sizes for the Baltic and WAF instance for different sets.
Recall that the Baltic instance contains 12 ports, 132 legs, 22 cargo flows and 22
transit times, whereas the WAF contains 20 ports, 37 cargo flows, 33 transit times.
Rows indicate the number of constraints in the model, Cols. the number of columns,
i.e. variables. Bin. indicate the number of binary variables. NZ presents the number
of non-zero elements in the problem matrix.
Maximum number of
Instance Services Vessels Layers Rows Cols. Bin. NZ
Baltic 1 1 1 6,261 8,711 2,829 3,1406
1 1 2 25,836 37,074 12,148 137,966
1 2 1 6,772 9,456 3,064 3,4548
2 1 1 13,352 19,196 6,088 71,387
2 2 2 51,074 74,450 24,401 283,701
WAF 1 1 1 13,884 25,712 6,252 73,207
1 1 2 82,039 144,661 38,921 479,836
1 2 1 28,091 43,042 13,167 348,674
2 1 1 34,492 57,793 15,647 194,046
2 2 2 252,427 392,596 122,807 1,517,454
Table 5.4.: Resulting model size using different input sets (including transit times).
The table indicates that especially the number of layers (L) and the maximum
number of services (S O ) influence the problem size and thus the duration to solve
the instance to optimality. The WAF MIP with not more than 2 services, vessels
and layers, already contains 122, 807 binary variables.
In Figure 5.2(a) to 5.2(d) example incumbents during the solution process of the
Baltic instance are shown graphically by extracting the variable activities. It can
be seen that the solver first identifies networks with unprofitable pendulum services.
Recall that all cargo flows in the Baltic instance originate or arrive at Bremerhaven,
Germany, because a feeder network should be created. As a result, incumbent two
(see Figure 5.2(b)) does not serve any cargo flows. Due to the shorter distances
the objective value is improved compared to Figure 5.2(a). The third incumbent
(see Figure 5.2(c)) is profitable and contains one pendulum and two circle services
serving Bremerhaven. The fourth incumbent further increases the profit and extends
the service length. The service calling Ålesund (Norway) and Kotka (Finland) is a
butterfly route and calls Bremerhaven two times during the round trip.
The mixed integer program relies mainly on three different sets: The maximum
number of services, the maximum number of vessels per service and the maximum
120
(a) First incumbent. Profit is -9.53 million US$. (b) Second incumbent. Profit is -6.16 million
US$.
(c) Third incumbent. Profit is 5.80 million US$. (d) Fourth incumbent. Profit is 6.81 million
US$.
Figure 5.2.: Example incumbents for the Baltic instance using not more than three
services, two layers and two vessels per service.
number of layers. Using more than one layer allows the model to create complex
route types. In Figure 5.3, the solution progress for a maximum of two vessels
per service and simple route types are shown. The maximum number of services
that the model can deploy is shown as different data sets indicated by Max S . The
maximum runtime of Gurobi is limited to 12 hours for the Baltic instance. Except
when limiting the model to create only one service, no instance could be solved to
optimality. In particular, a remaining gap of more than 10% can be observed. In
the largest model with up to four services, the first incumbent solution was found
after 2.5 minutes.
When enabling complex route types using two layers, the gap is closed even slower
(see Figure 5.4) and an optimal network could be determined using only one service.
Another problem is the increased duration until Gurobi finds the first incumbent
121
120%
100%
80%
Gap %
Max S=1
60% Max S=2
Max S=3
Max S=4
40%
20%
0%
0,01
0,02
0,03
0,04
0,05
0,08
0,09
0,11
0,12
0,13
0,14
0,15
0,16
0,18
0,31
0,33
0,34
0,39
0,44
0,46
0,48
0,54
0,57
0,58
0,59
0,64
0,66
0,71
0,73
0,86
0,93
0,96
0,98
1,11
1,33
1,36
1,44
1,49
1,71
2,54
5,64
6,21
6,24
6,72
0,1
0,6
0,7
1,5
5,4
12
Runtime [h]
Figure 5.3.: Gap in the Baltic instance per service upper bound allowing one layer
and a maximum of two vessels per service. The model is limited to
deploy a maximum of 1 to 4 services (MaxS ).
solution. For the largest model with four services, it takes about 30 minutes to find
the first integer solution.
Although the instances could not be solved to optimality, information can be
inferred from the non-optimal bounds. In Figure 5.5 the best incumbent (best lower
bound) and the best upper bound (LP relaxation) for different data sets are shown.
On the x-axis the number of vessels (1,..,4) and the maximum services (1,..,2) are
shown. In Figure 5.5(a), the network is limited to use one layer, in Figure 5.5(b)
two layers are allowed. The figures indicate, that solving the Baltic instance with
complex route type increases the complexity of the problem as the remaining gap
after 12 hours gets larger. The best solution found in the one layer case is 23.9 million
US$, in the two layer case 20.5 million US$ with a maximum of 4 and 2 vessels
per service respectively. This information provides a lower bound for heuristics. A
clear assessment of upper bounds is not possible. The results indicate that using a
maximum amount of four vessels per service and deploy not more than two services
cannot lead to a profit larger than 35 million US$. However, Figure 5.4 states a
remaining gap of more than 40%, so the actual upper bound can be around 21 million
US$ (in the two service case). Detailed numerical results without considering transit
times are given in Appendix D.1. Although the gap is smaller on average, most of
the instances could not be solved to optimality as well.
122
120%
100%
80%
Gap %
Max S=1
60% Max S=2
Max S=3
Max S=4
40%
20%
0%
0,01
0,02
0,05
0,11
0,23
0,28
0,54
0,75
1,43
1,57
1,66
1,92
2,06
2,24
2,35
2,49
2,62
2,65
2,84
3,33
3,48
4,59
4,73
5,99
6,71
7,74
7,9
9
9,08
9,19
9,39
10,82
12
Runtime [h]
Figure 5.4.: Gap in the Baltic instance per service upper bound allowing two layers
and a maximum of two vessels per service.
40 40
35 35
30 30
Profit [Mio. US$]
Profit [Mio. US$]
25 25
20 20
15 15
10 10
5 5
0 0
1 2 3 4 1 2 3 4 1 2 3 4 1 2 3 4
1 2 1 2
Max. vessel count, max services Max. vessel count, max services
best lower bound best upper bound best lower bound best upper bound
(a) One layer allowed. (b) Two layers (extended route types) allowed.
Figure 5.5.: Best bounds found for the Baltic LINER-LIB instance within 12 hours.
123
120%
100%
80%
Gap %
60%
Max S=1
Max S=2
40%
20%
0%
0 0,01 0,05 0,43 0,45 0,46 0,52 1,04 1,15 1,51 3,18 24
Runtime [h]
Figure 5.6.: Gap in the WAF instance per service upper bound allowing one layer
and a maximum of two vessels per service.
Figures 5.6 and 5.7 present the gap convergence and the resulting bounds for the
WAF LINER-LIB instance. To get results for the mixed integer model, we limit the
amount of layers to one and increase the maximum runtime to 24 hours. Figure
5.6 indicates that the duration to find a first integer solution increase to nearly 45
minutes when two services are allowed. The resulting gap for the data sets are still
more than 35%.
Figure 5.7 shows the lower and upper bound, found after a runtime of 24 hours.
The minimum number of vessels has been increased to two, because no profitable
network with services deploying not more than one vessel were found. This is a
result of the feeder network structure in the WAF instance where all cargo flows
have their origin or destination in Algeciras, Southern Spain. The distance to the
closest profitable port requires at least two vessels. The best network found has an
objective of 72.6 million US$. This is used as a lower bound for the heuristics as
well. The drawback is that the upper bound seems to be even more unreasonable
compared to the Baltic instance. The MIP solver reported a remaining gap of 65%
after 24 hours. Additionally, even three or four services with complex route types
might be beneficial for the WAF instance. However, within the runtime no integer
solution could be found.
The results from the mixed integer model show that the problem is hard in the
sense of runtime. This can be explained by four model properties:
124
200
180
160
140
Profit [Mio. US$]
120
100
80 best lower bound
best upper bound
60
40
20
0
2 3 4 5 2 3 4 5
1 2
Max. vessel count, max services
Figure 5.7.: Best bounds found for the WAF LINER-LIB instance within 24 hours.
1. The gap of the LP-relaxation solved at each Branch & Bound node is rela-
tively large to the integer solutions: Relaxing the integer variables does not
approximate the convex hull in a reasonable way due to the interdependency
between the constraints using binary variables. Thus, the constraints do not
provide a tight formulation regarding integer solutions.
2. A lot of inter-dependencies between the variables exist that lead to invalid

solutions. This results in enumerating many solutions individually which might
prevent efficient sub-tree pruning in the Branch & Bound process.
3. The model contains many binary variables to discretize the bunker cost, define
the liner services’ port rotations and ensure the transit times. Furthermore,
several Big-M constraints to model the logic associated with the linearization
are used. This can be disadvantageous for the runtime as well.
4. The model faces symmetry problems, because the port rotations and services
can be varied but still lead to identical solutions. Modeling the problem using
services (instead of specific vessels) and enabling MIP solver symmetry checks
reduce but does not eliminate the problem.
Our analysis for two of the LINER-LIB instances confirmed the results presented
in Plum et al. (2013b) about the complexity of this problem. The introduction of the
125
highly relevant transit times makes the problem even harder and results in large gaps
compared to previous publications. Although the instance sizes could be increased in
the last years (compare Reinhardt and Pisinger (2010)), the liner shipping network
design problem still provides a computational challenge for state-of-the-art mixed
integer solvers.
We conclude that the second goal of this thesis, to formalize real-world require-
ments of the liner shipping network design problem and solve small instances to
optimality, has been reached. However, real-world instances with more than 10 ports
require metaheuristics to obtain good integer solutions. The next sections present
two metaheuristics for the liner shipping network design problem and assess the
methods on the LINER-LIB instances and a real-world liner network.
5.2. Metaheuristics
The numerical results of the exact mixed integer program in the previous section
indicate that only very small instances can be solved to optimality. Álvarez (2009),
Mulder and Dekker (2013) and Brouer et al. (2013) show the successful application
of metaheuristics to the liner shipping network design problem. For solving larger
instances with respect to the practical requirements from Chapter 2, this section
introduces two metaheuristics: An hybrid evolutionary algorithm and a variable
neighborhood search.
In the scope of related vehicle routing problems many metaheuristics are evaluated
in the literature. Among others, genetic algorithms and variable neighborhood search
algorithms are shown to find good solutions (see for example Baker and Ayechew
(2003), Tasan and Gen (2012) and Nazif and Lee (2012) for genetic algorithms on the
VRP and Mladenović and Hansen (1997), Bräysy (2003) and Kytöjoki et al. (2007a)
for variable neighborhood search algorithms on the VRP). Due to the successful
use in vehicle routing and in liner shipping network design (see for example Mulder
and Dekker (2013) for a genetic algorithm for the LSNDP), we have selected two
promising metaheuristics to solve the network design problem: An evolutionary
algorithm and a variable neighborhood search. We extend the classic evolutionary
algorithm by local search algorithms from the variable neighborhood search (see
Mladenović and Hansen (1997)) to support the search space exploration. Thus, it is
referred to a hybrid evolutionary algorithm. The second algorithm that has become
popular and successful in vehicle routing is variable neighborhood search. These
algorithms are presented in the remainder of this section.
To cope with the different aspects in liner network design, first a decomposition
approach for both metaheuristics is presented. Afterwards, the fitness calculation
approach and numerical results are shown. A new approach to approximate the
objective function is presented. It shows promising results on the medium-sized
instances.
126
5.2. Metaheuristics
5.2.1. Decomposition Approach for the Metaheuristics

Chapter 2 presents the aspects of the liner network design problem and Chapter 3
the state-of-the-art solution approaches for the subproblems of the network design.
The subproblems considered in this thesis are:
1. Determine liner services (port rotation, vessel type and number of vessels)
2. Assure transit time and embargo constraints
3. Allocate cargo on the network to determine transhipment and bunker costs
4. Ensure a repositioning of empty containers
5. Respect the vessel draft subject to its deadweight
6. Consider capacity types
Table 5.5 shows the subproblems that must be covered by the cargo allocation
and the network design problem. Except the determination of liner services and
the transit time, each aspect from the network design is also relevant for the cargo
allocation and has been integrated into a column generation solution approach (see
Chapter 4). Subproblems 3 – 6 are therefore already solved with the previously
presented method.
Subproblem CAP LSNDP

Determine liner services (port rotation, vessel X
type and number of vessels)
Assure transit time and embargo constraints X
Allocate cargo on the network to determine tran- X X
shipment and bunker costs
Ensure repositioning of empty containers X X
Respect the vessel draft in respect to its dead- X X
weight
Consider capacity types X X
Table 5.5.: Subproblems of the cargo allocation presented in Chapter 4 and liner
shipping network design problem.
The previous Section 5.1 shows that the subproblems one and two are the main
challenges of the network design problem. Therefore, in the scope of this thesis, a
decomposition approach of the liner shipping network design problem is used (see
Figure 5.8).
127
(1) Solve the cargo allocation

Determine network us-
(2) perform tran-
ing the metaheuristic
sit time adjustment
Figure 5.8.: Decomposition of the overall liner shipping network design problem.
A metaheuristic is used to select and modify networks (i.e. the port rotation,
vessel type and number of vessels) using implementation specific operators.
Then, one or more networks (depending on the metaheuristic) are evaluated in
step (1) by solving the cargo allocation problem. Recall that the solution of this
optimization problem provides the duration at ports and at sea. Therefore, the
second part of the evaluation is to evaluate embargo and transit time constraints
(see step (2) in Figure 5.8). Service speeds can be increased to hold the transit
times but also lead to increased bunker prices. Under some circumstances, the
resulting schedule is not valid anymore. Therefore, a reparation of the schedule and,
if necessary, an invalid solution must be penalized.
The right hand side of the decomposition in Figure 5.8 is referred to as the fitness
of a solution for the further use in metaheuristics. The following sections show the
fitness calculation in more detail.
5.2.2. Determine the Fitness of a Solution

The fitness (i.e. the objective function value) of a solution (i.e. a liner shipping
network) is determined as follows:
1. Solve the cargo allocation problem
2. Try to increase the speed on incident legs to hold the transit times
3. Penalize transit times that cannot be held in the network
4. Penalize embargo constraints
The cargo allocation problem solution provides durations at the ports and at sea.
If a transit time between two port does not hold, the speed is increased on a per
leg basis (if possible). Afterwards, if necessary, the solution is penalized if the speed
is not sufficient to hold the requirements. Finally, the solution validates whether
embargo constraints are invalid.
To explain the implications of transit times on the cargo allocation, see Figure 5.9.
Two services are shown, one with the port rotation p1 → p2 → p3 → p1 and another
with p2 → p4 → p2 . After solving the cargo allocation problem (see Chapter 4),
the services’ vessels remain at a port for a determined duration dPi,l,s . Additionally,
the cargo allocation model might have introduced an artificial port buffer bPi,l,s to
128
5.2. Metaheuristics
dPi,l,s , bPi,l,s
p1
dS
i,j,l,l0 ,s
p3 p2 p4
Figure 5.9.: Durations resulting from the cargo allocation problem, for example
given at one port and one leg.
compensate for badly designed networks with too many deployed vessels that would
steam below its minimum speed. To hold the constant round trip time of each service,
the CAP sets the vessel’s leg independent speed ks and thereby the duration at sea
to dSi,j,l,l0 ,s .
The solution from the cargo allocation provides the duration at all ports and
all legs, for example given for one port and leg in Figure 5.9. If cargo is routed
from its origin i to its destination j, the transit time θi,j between the origin and
destination must hold. Given all durations by solving the cargo allocation problem,
the minimum duration to route a cargo is calculated by taking the shortest path
from i to j on the given network. In Figure 5.9, the shortest path for cargo between
p1 and p4 is the path p1 → p2 → p4 . The durations are fixed, because the transit
time is not incorporated into the cargo allocation problem. If the duration is larger
than the required transit time θi,j between port i and j different actions can be
performed:
1. Directly impose a penalty cost for the difference between the required and the
actual transit time
2. Set the cargo flow volume between port i and j to zero because the constraint
does not hold
3. Adjust the average speed of the overall service
4. Adjust the speed on the legs incident to the shortest path between port i and
port j
Each action has advantages and disadvantages. Directly imposing penalty cost would
speed up the fitness function in terms of computation time but make it more difficult
for the metaheuristics to find feasible solutions because the CAP is independent from
129
the transit times. This means, that evolved networks might not match all transit
times. Another disadvantage is the problem of actually finding valid solutions. Tests
indicate that the networks are systematically reduced, and afterwards expanded
again until a valid network is found. The second method requires the adjustment
of the durations at the ports, legs and buffers and is not considered in the scope of
this thesis.
The third method can simplify the solution finding process. Once the CAP identi-
fies a profit optimal network, the overall speed can be increased on the whole service
or single legs. In other words, the flexibility of the resulting cargo allocation can be
used to increase the overall networks speed. The main drawback of this approach
is the resulting invalid schedule timing: If the average (or leg specific) speed of a
service is increased and thereby the duration at sea decreased, its schedule becomes
invalid because the service’s round trip time stays constant. In case of adjusting
the whole service speed, an additional artificial buffer must be introduced at the
service’s ports. This can lead to transit time violations because the time at the
ports increase, making it again difficult to create valid solutions.
In this thesis, the fourth method is used to deal with the transit time constraints
which relaxes the leg independent speed per service assumption. This allows repa-
ration methods to use both, unaffected legs’ sea duration and port durations to fix
the services’ schedules. This method is visualized in Figure 5.10 for the example
presented above. The problem instance might have imposed a transit time constraint
between port p1 and p4 of θ1,4 = 5 days. In the example, the only route between
port p1 and p4 is the path p1 → p2 → p4 , with the incident services s1 and s2 . For
simplicity of the example, we assume that the duration of the path is six days using
service independent speeds ks1 = 10kts, ks2 = 12kts resulting from the CAP. The
duration can be decreased by increasing the speed on leg p1 , p2 of service s1 and leg
p2 , p4 of service s2 until the transit time is met or the maximum speed is reached.
For simplicity, we have set the speed to kp1 ,p2 ,1,1,s1 = kp2 ,p4 ,1,1,s2 = 15kts so that the
duration is 5 days now.
Unfortunately, the round trip durations of the overall service are not met anymore
when modifying the legs’ speeds. A repair procedure must either decrease the speed
on the remaining legs or increase the buffer time at port p3 or p4 . Due to the cubic
bunker cost, decreasing the speed down to the minimum is preferable, then, the new
bunker costs are applied to the solution. Finally, if the transit time could not be
reached due to the maximum speed constraints, penalty costs are imposed on the
fitness function.
Algorithm 6 formalizes the method described above. The functions incSpeed and
repairSchedule are described in more detail in the next sections. The algorithm is
initialized by setting the leg independent speed to each leg. All transit times are
evaluated where at least one container is routed from the transit time’s origin to
the destination. The reason is that in case a network does not offer a connection
between the origin and destination or no cargo flows are served, there is no need for
130
5.2. Metaheuristics
p1
kp1 ,p2 ,1,1,s1
kp2 ,p4 ,1,1,s2
p3 p2 p4
Figure 5.10.: Incident service legs for a transit time requirement between port p1
and p4 are marked gray. Service legs that have to be adjusted are
marked black.
Algorithm 6: Speed adjustment process to hold transit times.

Input: A liner shipping network N W and its cargo allocation problem
solution, specifically the port durations due to pilotage, unloaded and
loaded cargo flows and fixed strategic port buffer dPi,l,s , the artificial
buffer time bPi,l,s and the speeds per service ks
Output: Network evaluation that holds the transit time if possible
1 ki,j,l,l0 ,s = ks ∀s ∈ S , (i, j, l, l ) ∈ Ls ;
O 0
2 forall the (i, j) ∈ T : ∃n ∈ N : on = i ∧ dn = j ∧ αn > 0 do

3 pi,j = dijkstra(i, j, dPi,l,s , bPi,l,s , ki,j,l,l0 ,s );
4 if pi,j = nil then
5 continue;
6 end
7 di,j = duration(pi,j );
8 if di,j > θi,j then
9 ki,j,l,l0 ,s = incSpeed(di,j − θi,j , ki,j,l,l0 ,s , pi,j );
10 end
11 end
12 ki,j,l,l0 ,s , bi,l,s = repairSchedule(di,l,s , bi,l,s , ki,j,l,l0 ,s );
P P P
13 applyCosts(ki,j,l,l0 ,s )
131
the carrier to hold the transit times. Afterwards, the shortest path subject to the
sea, port and transhipment durations is calculated and stored in the variable pi,j .
For the whole path, di,j stores the duration in days. If the duration cannot be held,
the speed of all service legs on the path p is increased if possible (see next section).
Finally, the algorithm tries to repair the schedule and applies the new bunker cost
to the solution.
Increase Vessels’ Speed

Increasing the vessels’ speed to meet transit time requirements can impose large
bunker costs. Due to the cubic consumption function, it is advantageous to first
increase the speed on the legs with lower speed. In Figure 5.11, the example legs
incident to the shortest path for a transit time are given. In the first row, the speeds
equal the results from the cargo allocation problem and are identical over all service
legs. We try to respect the transit time by starting to increase the speed uniformly
on the legs with the lowest speeds (those of service s2 in the figure).
When the speed is increased further, other service leg speeds equal the ones
just increased (see Figure 5.11, row two). Then the legs of service s1 and s2 are
increased until the next higher service speed with 14 kn is reached. As indicated in
the last row, all but the first service’s vessel type’s maximum speed is 14 kn. Thus
the speed cannot be increased any further. The algorithm terminates if either all
incident service legs are at their maximum speed upper bound or the transit time
requirement holds for the new durations.
Repair the Proforma Schedule

Algorithm 6 increases the speed per leg, possibly making the proforma schedule
invalid. The following linear optimization model repairs all schedules of a network.
The main idea is to slow down the vessels on the legs that are not incident to any
transit times. The remaining gap to the service’s round trip time (RTT) is filled by
reducing artificial port buffers (if available) from the cargo allocation solution.
Sets
SO Services operated by the carrier.
Ps Ports visited by service s.
Ls Legs used in service s.
Parameters
dPi,l,s CAP durations in days due to cargo flows, strategic port buffer and
pilotage at port i of service s
bPi,l,s CAP buffer duration in days at port i of service s
132
5.2. Metaheuristics
dM in
i,j,l,l0 ,s Leg duration in days when steaming with adjusted speed according
li,j
to algorithm 6. dSO
i,j,l,l0 ,s = max{kM ax ,k 0 }·24 V Ts i,j,l,l ,s
dM ax
i,j,l,l0 ,s Maximum duration in days when steaming with minimum speed.
li,j
i,j,l,l0 ,s = max{kM in ,k
dM ax
0 }·24
V Ts i,j,l,l ,s
Variables
B
τi,l,s ∈ R+ 0 Artificial port buffer in days at port i’s lth visit in service s.
τi,j,l,l0 ,s ∈ Duration at sea on service s leg (i, j, l, l0 ) ∈ Ls in days.
S
R+0
τsP ∈ R+ 0 Penalty duration for service s if the round trip cannot be held
FIXSDLE
 
X X
max  S
τi,j,l,l P
0 ,s − τs (5.52)
s∈S O (i,j,l,l0 )∈Ls
s.t.
X X
RTT s = dPi,l,s + τi,l,s
B
+ S
0 ,s + τs
P
∀s ∈ S O

τi,j,l,l
(i,l,s)∈Ps (i,j,l,l0 ,s)∈Ls
(5.53)
i,j,l,l0 ,s ≤ τi,j,l,l0 ,s ≤ di,j,l,l0 ,s ∀s ∈ S , (i, j, l, l ) ∈ Ls (5.54)

0
dM in S M ax O
0≤ B
τi,l,s ≤ bPi,l,s ∀s ∈ S , (i, l) ∈ Ps
O
(5.55)
Objective (5.52) maximizes the durations at each leg, i.e. steam as slow as possible.
Constraints (5.53) ensure that after solving the model, all of the network’s services
are valid, if possible. Constraints (5.54) set the bounds for the duration at sea. The
duration’s lower bound is the duration after adjusting the speeds. If a lower duration
(higher speed) would be possible, the transit times can be violated again. Note that
the lower bound originates from the speed selected by the cargo allocation problem
that allows speeds larger than the maximum speed for badly designed networks.
The upper bound is the duration when steaming at minimum speed. The bounds
on the port buffer in constraints (5.55) are either zero or the artificial buffer that
has been introduced by the cargo allocation.
Model FIXSDLE can be solved within fractions of a second. Even in large net-
works, the upper bound on the rows are the number of services (150 - 200 for the
world’s largest carrier’s network). Assuming 15 legs per service, the model would
have up to 3000 columns (continuous variables).
133
ko,j,l,l0 ,s1 = 14 ki,j,l,l0 ,s2 = 10 ki,j,l,l0 ,s2 = 10 ki,d,l,l0 ,s3 = 12
Figure 5.11.: First row: Initial, leg independent speeds (in kn) per service after solv-
ing the cargo allocation problem for four legs incident to the shortest
path for transit time θo,d . Succeeding rows: Increasing leg dependent
speeds per service until transit times are met or maximum speed is
reached.
134
5.2. Metaheuristics
p1
p3 p2 p4
Figure 5.12.: Random network that uses pendulum services.
The presented approach is used to calculate the fitness for solutions (liner net-
works) within metaheuristics. Additionally, it can be used to evaluate the network
changes of liner network planners, for example by providing information on invalid
transit times.
5.2.3. Construction Heuristics

In this section, two construction heuristics to create initial networks are presented: A
pendulum service heuristic and a clustering heuristic. To the best of our knowledge,
these heuristics have not been presented in literature. The main challenge of these
construction heuristics is to provide solutions as good as possible in a reasonable
amount of time. For liner shipping networks, different aspects describe the perfor-
mance of initial solutions: First, the geographical coverage should be large enough
such that all trade regions are already served. Furthermore, the solutions should be
valid regarding the port drafts and transit times, and, finally, they should lead to a
reasonable profit.
Pendulum Service Heuristic

A simple method to create liner networks is to create services with pendulum routes
that contain only two ports. Because liner services are usually aligned with cargo
flows, services use legs where cargo flows exist. The cargo flows can be either taken
randomly or based on their possible revenue. After the alternating ports for a new
service are identified, a vessel type whose lightship draft holds the depth of the ports
is selected. The number of deployed vessels are selected based on the design speed
and the pilotage, buffer and constant port call duration of 24 hours per port. In
Figure 5.12, an example initial network with three services on four ports is shown.
The advantage of this construction heuristic is that networks can be created very
quickly and a large variety of ports are called, which leads to increased geographical
coverage.
However, the solution approach has several fundamental disadvantages:
135
p5 p3
p4 p1
p2
p6
p7
p8
Figure 5.13.: Clustering with three regions used for construction heuristic. The
reference ports in each region are marked bold.
• The networks impose large fixed costs for the vessels
• Beneficial transhipment operations to extend the cargo flow coverage are only
performed at random ports
• Invalid networks are possible due to high transit times when serving cargo
along different services
The construction heuristic presented in the next section covers the disadvantages
of high fixed costs, improves transhipment operations and reduces the risk of invalid
networks due to the transit times.
Clustering Heuristic
The clustering heuristic basically works in two steps: First, an initial network that
serves the largest ports within determined regions is created. Second, the created
services are extended locally with additional ports. The heuristic is outlined in
Algorithm 7.
Figure 5.13 shows the result of the first heuristic’s step of clustering the ports,
see line 1 in Algorithm 7. A parameter of the heuristic is to select the number of
regions (cc) in the problem instance. The k-means clustering algorithm (Hartigan
and Wong (1979)) is used in this thesis to cluster ports into regions using the L2
distance metric. For each cluster (or region), the port with the largest depth is
selected as a representative. This port is used to ensure that large vessels can enter
the region. Afterwards, a random list of region combinations is used to create initial
pendulum services (see Algorithm 7, line 9). A bound for the required vessels is
determined by dividing the weekly cargo flows by the average vessel capacity (its
136
5.2. Metaheuristics
Algorithm 7: Clustering heuristic to create an initial liner shipping network

design.
Input: Problem instance data, ports P , demands N , cluster count cc, service
count sc
Output: A liner shipping network design, ND
1 C ← KMeans(P, cc);
2 R ← ∅;
3 T SP ← ∅;
4 for i = 0; i < |C|; i + + do
5 for j = 0; j < |C|; j + + do
6 R ← R ∪ {(GetRepresentative(C[i]), GetRepresentative(C[j]))}
T SP ← T SP ∪ {GetRepresentative(C[i]), GetRepresentative(C[j])};
7 end
8 end
9 ND ← CreateInitialServices(R, sc);
10 f = Evaluate(ND);
11 for p ∈ sort(P \T SP ) do
12 ND 0 ← AddPort(ND, p);
13 f 0 = Evaluate(ND 0 );
14 if f 0 > f then
15 ND ← ND 0 ;
16 f = f;
17 end
18 end
19 return ND;
total nominal slots). Until this upper bound is reached, pendulum services between
two random regions are created. These services provide the basic network design.
In the second step of the heuristic in line 10, the network is evaluated by a cargo
allocation approximation described in Chapter 4. Afterwards, the algorithm tries
to insert all unserved ports into the closest service with method AddPort. If the
network’s fitness improves, the port is inserted and the next port is selected. Finally,
the best network found is returned. Considering the fitness usually leads to networks
that hold the transit time because invalid networks have a negative fitness.
The presented construction heuristics are used to provide a starting solution (or
solutions) for the metaheuristics in the next section.
5.2.4. Improvement Heuristics

In the scope of the liner shipping network design problem, several metaheuristics
have already been used successfully (see Chapter 3). This section proposes a hy-
137
brid evolutionary algorithm that combines the default approach with local search
methods. Then, a novel variable neighborhood search heuristic for the liner shipping
network design is presented.
Hybrid Evolutionary Algorithm

The general concept of a genetic algorithm, introduced by Holland (1992) is described
in Chapter 3. In the scope of this thesis, the algorithm is extended to cope with the
liner shipping network design problem. The extensions are:
• Problem specific crossover operators
• Mutation operators
• Local search method
The algorithm is based on a population of solutions (individuals). The population

evolves according to a fitness function. The algorithmic approach presented in
Section 3.1.3 is extended by custom crossover selection strategies and operators,
elitism, population variation operators and local search operators. These adjustments
to the basic version are required to deal with the complex problem of optimizing
liner networks. The hybrid evolutionary algorithm is described in Algorithm 8.
Each individual of a population is represented as an object in the memory. The
reason not to choose the classical chromosome-representation used in genetic al-
gorithms is the disadvantage of handling and validation. An individual is a liner
network design for a specific problem instance basically with a set of liner services.
Each liner service defines an ordered set of legs between ports, has a vessel type
attached and deploys a specific number of vessels. Each object has several properties,
such as a liner service’s route type or utilization. The properties can be used to sim-
plify the development of reporting and validation components. For each individual,
a method is called that evaluates the fitness function according to the approach
presented in Section 5.2.1 and determines how good a solution is compared to others.
The fitness is returned as a real value that can be negative.
In the scope of this thesis, two crossover selection strategies are evaluated (see line
7 of Algorithm 8): A binary tournament (BT) and roulette (wheel) selection (RS).
Binary tournament selects two parents for the crossover by selecting the best of
four randomly chosen parents. Roulette selection selects parents by selecting them
according to their unified fitness distribution. For more information on tournament
selection see Srinivas and Patnaik (1994) and Beasley and Chu (1996) for roulette
selection. Both are used to create an offspring based on two relatively good parents.
The crossover method decides how to combine the parents to the offspring, in
analogy to the child’s chromosomes based on its parents. The crossover method is
a core component of an evolutionary algorithm, because it must ensure diversity of
138
5.2. Metaheuristics
Algorithm 8: Hybrid Evolutionary Algorithm for the liner shipping network

design problem.
1 Initialize population P0 using the Pendulum and Clustering heuristic;
2 Initialize evaluations E = Evaluate(P0 );
3 Initialize iteration i = 0;
4 while runtime <= MaxRuntime do
5 Initialize offspring O ← ∅;
6 for i < (|Pi |/2) do
7 (A, B) = SelectP arents(Pi );
8 O ← O ∪ { Crossover(A, B) } ;
9 end
10 for o ∈ O do
11 for m ∈ M do
12 if mutation condition satisfied then
13 o ← M utate(o, m);
14 end
15 end
16 E ← E ∪ {EvaluateF itness(o)});
17 end
18 Pi+1 ← SelectP opulation(E, Pi , O);
19 i ← i + 1;
20 end
21 return BestN etwork
solutions and hold common properties. In the scope of this thesis, three operators
are implemented and tested:
1. Cycle crossover (CY) (for an application of the vehicle routing problem, see
for example Michalewicz and Fogel (2004)),
2. Informed cycle crossover (IC),
3. Clustering crossover (CL).

The cycle operator creates offspring by counting the number of services in each
parent and randomly selects services from each parent. Thus, the structure of the
parents is kept in the next population. The informed cycle operator can additionally
extract ports from single services and join them with the best position of a random
service in the offspring. The third operator identifies the served regions of the two
parents, loops through a random list of region combinations and selects a random set
of services between these regions. With a low probability, a service can be skipped
in this process, otherwise the network would be expanded too quickly.
139
After the offspring are created, random mutations are applied to the solutions
(see line 13 in Algorithm 8) with a specific probability. The operators used in this
thesis are:
1. Delete port
2. Delete service
3. Insert port sequence
4. Move port sequence between services
5. Insert pendulum service
Results indicate that very small changes in the solutions can lead to large changes
in the fitness. To reduce the negative impacts of modified networks, local search
strategies are applied to solutions. Strategies that only rely on distance measures,
such as a 2-opt operator (see Croes (1958)), have only worked well on small instances
because they do not consider the transit times. Therefore, a variable neighborhood
descent (VND) has been implemented to improve the networks after the mutation.
The VND is applied to a percentage of the population’s individuals and is described
in detail in the next section.
The succeeding population is a subset of the parents and the created offspring.
To limit the size of the population, one has to select a subset of all individuals that
continue to the next population (see line 18 of Algorithm 8). In the scope of this
thesis, the best ELM % of individuals are selected that have distinct fitness values.
This strategy is called elitism. To allow other individuals to survive this selection
process and evolve further to diversify the population, the remaining population is
tried to be filled with random individuals with distinct fitness. Taking the distinct
fitness into account allows to diverse the population. One should not select too many
similar individuals because this can lead to local minima. After a fixed number of
iterations, random individuals (created by the construction heuristics) are injected
into the population to diversify it and avoid premature local optima.
Variable Neighborhood Search

Variable neighborhood search (VNS) algorithms have been successfully used for solv-
ing vehicle routing problems in recent years (see for example Hansen and Mladenović
(2003)).
Summarizing, VNS starts with an initial network and randomly modifies (called
shaking in VNS terms) it to generate a new neighborhood solution (see Algorithm
9). Then, the best network within a specific range of possible changes is searched.
The process of modifying and performing a local search is repeated until a stopping
140
5.2. Metaheuristics
Algorithm 9: Variable Neighborhood Search for the liner shipping network

design problem using a metropolis criterion.
0 0
1 Select set of neighborhood structures Nk for k = 1, · · · , kmax ;
2 Create initial solution i using clustering heuristic;
3 repeat
4 k = 1;
5 repeat
6 x0 = Shake(Nk (i));
7 Apply local search
to x00 =VND(x

0
);
00 0
8 if rnd > exp −(f (x T)−f (x )) then
9 i = x00 ;
10 k = 1;
11 else
12 k = k + 1;
13 end
0
14 until k = kmax ;
15 until runtime > MaxRuntime;
16 return BestN etwork
criteria is matched. While no improving solutions are found, shaking operators

increasingly modify the solution to reduce the risk of local minima.
Several problem specific adaptions must be performed to apply the VNS to the
network design in liner shipping, see Algorithm 9. The key adjustments are the fitness
function and the neighborhood structures. VNS implementations rely on good local
search algorithms that help to identify the best solution within a neighborhood. In
the scope of the vehicle routing problem, the 2-opt local search (see Croes (1958)) has
been used in variable neighborhood search (see Chen et al. (2010) and Kytöjoki et al.
(2007b)). This thesis uses the Variable Neighborhood Descending (VND) algorithm
proposed by Mladenović and Hansen (1997) which is adapted to the liner network
design problem and used in the VNS (see line 7 in Algorithm 9). The algorithm
successively uses different operators, called neighborhood structures, in a specific
order. If the solution improves, VND restarts with the first neighborhood structure.
The algorithm terminates when a given runtime is reached and no neighboring
solution lead to any improvement. The general concept is presented in Chapter
3. For details on the implementation of the structures in the LSNDP and their
sequencing, see Hilger (2014). VND relies on the neighborhood structures presented
in Table 5.9.
The operator ChangeSinglePort, which moves a single port between two services,
is successfully used in vehicle routing problems as well (see Hemmelmayr et al.
141
Neighborhood Structure Description

ChangeSinglePort Changing a service’s port
ChangePortSequence Changing a sequence of ports
ChangeLinerService Joining two or separating one liner service
ChangeVesselType Changing a service’s vessel type
ChangeVesselCount Changing the number of deployed vessels
2-opt Performs a 2-opt local search based on the services’
distance and cargo flows
Table 5.9.: Neighborhood Structures used in Variable Neighborhood Descending.
(2009b)). Due to the vast solution space, several domain specific operators must be
implemented. These operators rely on the preferable high utilization (see Agarwal
and Ergun (2008)) and low bunker cost (see Wang and Meng (2012c)) in good liner
networks.
The neighborhood structure ChangePortSequence defines all neighborhoods of a
given network that can be created by moving a defined port sequence of service A
to another service B. Similar operators are used for the vehicle routing problem
as well (see Hemmelmayr et al. (2009b)). The operator selects the port sequence
based on the leg utilization and the geographical distance of the ports. It is assumed
that sequences with low utilization should be removed from a service and inserted
into another service with less than average utilization. Feasibility of the modified
services must be ensured by adjusting the vessel types due to draft limitations and
increasing/decreasing the number of vessels to ensure a weekly round trip time
based on the design speed.
The operator ChangeLinerService merges two services or splits one liner service
into two. The operator is required to enable the extension or contraction of networks.
A good selection of the splitting point can be achieved by selecting geographically
close ports. The operator ensures the connectivity of the network by calling a port
from both services that are created by the split-operation. In practical network most
of the services are connected with each other to allow transhipments. Exceptions
such as pendulum-services and networks with only one service must be considered
in the operator as well.
The problem specific neighborhood ChangeVesselType changes the vessel type of
one existing service in the network design. The operator is based on a preferring
high service utilization. Services with low average utilization probably have a vessel
type deployed where the capacity is too large for the transported cargo flows. Thus,
the operator selects a smaller vessel type. When using a larger vessel type, the
operator must ensure that all served ports’ depths are not smaller than the new
vessel’s lightship draft.
The neighborhood ChangeVesselCount defines the neighborhood of all networks
that can be created from a given network by increasing or decreasing the number
142
5.2. Metaheuristics
of deployed vessels for all services. Services that steam with a high speed (obtained
through the cargo allocation problem) may deploy another vessel to decrease the
overall service speed and thereby decrease the bunker costs of this service. Similar,
services that deploy too many vessels typically steam at their minimum speed and
may use artificial port buffer to obtain the constant round trip time (recall that
weekly services are assumed throughout the network design). For these services,
decreasing the number of vessel may reduce the time charter cost of the service.
The last operator performs a 2-opt local search on all liner services in the network
based on the distance and the transported cargo flows. The distance metric is impor-
tant because ports within a region should be served together. However, cargo flows
should be preferable routed on its shortest path from the origin to its destination.
Thus, the 2-opt swap operation considers both the distance and the legs with high
volume to perform the swap operations.
The described neighborhood structures can be called in arbitrary order. Hilger
(2014) performs an analysis of different sequences and concludes that the differences
are rather small. Nevertheless, the sequence presented in Table 5.9 performs best
on the learning instances and is thus used in the remainder of this thesis.
Another method to improve the convergence of variable neighborhood search
algorithms is the use of a Simulated Annealing (Kirkpatrick (1984)) acceptance
criteria (see line 9 in Algorithm 9). This extension has been successfully used in
vehicle routing problems (see for example Hemmelmayr et al. (2009b) and Chen
et al. (2010)). Numerical results indicate that the use of the acceptance criteria lead
to slightly improved convergence behavior in some instances of the LINER-LIB and
is used in the remainder of this thesis to diversify the search algorithms.
Numerical experiments performed by Hilger (2014) indicate that larger liner
network instances require a VNS shaking step that diversifies the neighborhood more
than in smaller instances. Therefore, the implemented shaking procedure executes
the ChangeSinglePort and ChangeLinerService operators with a probability of each
35%. Otherwise, no change on the network is done and the improvement is moved
to the descending step.
5.2.5. Numerical Results
The proposed metaheuristics are evaluated in this section. First, the results for
the hybrid evolutionary algorithm are presented. A detailed parameter tuning for
the evolutionary algorithm can be found in Appendix D.2. Then, the results for
the variable neighborhood search (VNS) are shown. Details on the VNS parameter
tuning are presented in Hilger (2014).
143
Evolutionary Algorithm
To simplify the evaluation of the evolutionary algorithm, the numerical results
for the basic parameters are moved to Appendix D.2. The parameters listed in
Table 5.10 are tested one after the other. Tools that automatically determine good
parameters rely on many instances, preferable hundreds (see for example Ansótegui
et al. (2009)), and are thus not applicable to the LSNDP using the LINER-LIB
instances due to overfitting. The default values listed in Table 5.10 are evaluated
on the WAF and Mediterranean instances (see Appendix D.2). Further information
on the parameters is given in Section 5.2.4.
Parameter Abbreviation (Default) Values

Instance I Baltic, WAF, Mediterranean, Pacific,
EuropeAsia, WorldSmall, WorldLarge
Runtime RT Instance dependent, 3, 5, 20, 30, 240, 240,
240 minutes
Crossover methodA CX Informed Cycle (IC), Clustering (CL)
and Cycle (CY) operator
Crossover selectA CXS Binary Tournament (BT), Roulette Se-
lection (RS)
Population sizeA PZ 10, 20, 50
Pendulum%A PDL 0%, 10%, 50%, 100%
Local search probability ∗ LSP 0%, 20%, 50%, 100%
ElitismA ELM 10%, 20%, 50%
Mutation rateF MR 1%, 2.5%, 5%
Network injectionA NI 10, 20, 50
Table 5.10.: Parameters for the hybrid evolutionary algorithm. Parameters marked
with A are evaluated in Appendix D.2. The local search probability is
analyzed in this section. The bold parameters have been selected by
the manual parameter tuning.
The following assumptions are made to simplify the algorithm evaluation for
the LINER-LIB: First, no additional port buffer is used. This means, that the
port duration is only determined by the pilotage and the duration resulting from
the cargo movement. Second, the tolerance for the transit times is set to one day
and no transhipment port duration is added. To deal with the randomness in the
evolutionary algorithm, each configuration’s result is averaged over five runs, always
using five identical initial networks for different configurations.
The EA uses a binary tournament selection strategy and the informed crossover
operator. A medium population size of 50 individuals per generation is used. The
initial population not only consists of networks created by the clustering heuristic,
but also 10% pendulum service networks to diversify the population. To obtain
convergence of the overall population, in every iteration the 20% best individuals
144
5.2. Metaheuristics
30 130
120
25 110
100
90
20
Fitness [Mio. US$]
Fitness [Mio. US$]

80
70
15
60
50
10
40
30
5 20
10
0 0
1
5
9
13
17
21
25
29
33
37
41
45
49
53
57
61
65
69
73
77
81
85
89
93
97
101
105
109
113
117
121
125
129
133
137
141
145
149
153
157
161
165
169
173
177
1
8
15
22
29
36
43
50
57
64
71
78
85
92
99
106
113
120
127
134
141
148
155
162
169
176
183
190
197
204
211
218
225
232
239
246
253
260
267
274
281
288
295
Runtime [s] Runtime [s]
100 % 25 % 50 % no local search 100 % 25 % 50 % no local search
(a) Baltic instance. (b) WAF instance.
10 15
9 14
8 13
12
7
11
6 10
5 9
4 8
Fitness [Mio. US$]
Fitness [Mio. US$] 7

3
6
2 5
1 4
3
0
2
1
41
81
121
161
201
241
281
321
361
401
441
481
521
561
601
641
681
721
761
801
841
881
921
961
1001
1041
1081
1121
1161
1201
1241
1281
1321
1361
1401
1441
1481
1521
1561
1601
1641
1681
1721
1761
-1
1
-2 0
1
41
81
121
161
201
241
281
321
361
401
441
481
521
561
601
641
681
721
761
801
841
881
921
961
1001
1041
1081
1121
1161
1201
1241
1281
1321
1361
1401
1441
1481
1521
1561
1601
1641
1681
1721
1761
-3 -1
-2
-4
-3
-5 -4
-6 -5
100 % 25 % 50 % no local search 100 % 25 % 50 % no local search
(c) Mediterranean instance. (d) Pacific instance.
Figure 5.14.: Convergence of the evolutionary algorithm in different LINER-LIB

instances using the local search probabilities 0%, 25%, 50% and 100%.
(of the offspring and the parents) are selected for the next generation (referred to as
an elitism strategy), and every 10 iterations without any objective improvement 25
random networks are added to the population and selected according to the elitism
strategy.
The main challenge of the population based algorithm is avoiding local minima by
managing the randomness of the generation to get sufficient diversity. If too much
randomness is introduced, the networks’ fitness fluctuates too much. Thus, the
algorithms decide to shrink the networks to decrease the costs. Not until a sufficient
stable small network is found does the algorithm start to increase the network size.
Thus, a relatively small mutation rate of 1% is used in the EA. Furthermore, it is
important to start with relatively good, feasible, large networks to avoid the initial
shrinkage phase of the algorithm. Its runtime is limited to 30 seconds, although the
phase is often finished in less time for the small and medium sized instances. VND
is executed in parallel for many individuals which reduces the runtime (see Hilger
(2014)).
In Figure 5.14 the effects of different local search probabilities on the convergence
is presented for four LINER-LIB instances. In all tested instances, the algorithm’s
145
convergence is worst when not using any local search algorithm. In the smallest
instance, the improvement using a local search is relatively small (see Figure 5.14(a)).
The best fitness found within three minutes is about 10% higher when applying a
local search to 50% of the offspring. In contrast, the larger instances highly benefit
from the local search (see Figure 5.14(b) - 5.14(d)). The best fitness found for the
WAF instance is approximately 80% better and the Mediterranean’s about 170%.
On average, the algorithm cannot create a profitable network for the Pacific instance
without the local search in 30 minutes. Using the local search with a low probability
leads to profitable networks after five minutes and after nine minutes when using
a medium local search probability. This can be explained by the time needed to
execute VND for the individuals. The more often the VND is executed, the less
iterations are performed by the hybrid evolutionary algorithm. This explains why
performing a local search on only half of the population is better in all tested
instances compared to running VND on all offspring. The local search probability
is set to 50% for the remaining numerical results.
The LSNDP is by its nature very resilient to changes in its solution. For example
changing the port rotation by repositioning one port cannot only lead to reduced
profit, but also to invalid transit times and thus high penalty cost. Furthermore,
transhipment operations may become highly ineffective because the cargo has be
transported on many legs of the round trip. Removing a port can lead to the
breakdown of most of the cargo flows, because a major transhipment port is required
to distribute the cargo between different regions. Especially in the medium sized
instances, it becomes important to induce a certain degree of diversification, meaning
that not all of the offspring should be optimized by the local search.
Solving the mixed integer program in the previous section led to a lower bound of
approximately 23.93 million US$ after 12 hours. The best solution found with the
evolutionary algorithm after three minutes is 21.89 million US$ and has a gap of
approximately 8.5% to the lower bound of the mixed integer program. This can be
explained by the advantage of the mixed integer model of considering the interde-
pendency between serving cargo and respecting the transit time. The evolutionary
algorithm finds a network for the WAF instance with a profit of 110.35 million US$
after five minutes, compared to 72.59 million US$ after 24 hours using the mixed
integer model, which is an increase of more than 50%. The best network found for
the Mediterranean instance is 6.4 million US$, the Pacific instance’s is 10.99 million
US$. Within 24 hours, for none of the medium sized instances an integer solution
could be found using the mixed integer model.
The results indicate that the medium sized instances can be solved using the
hybrid evolutionary algorithm. It performs much better on the WAF instance com-
pared to the lower bound resulting from the mixed integer model. However, no
information on the solution quality can be stated for the Mediterranean and Pacific
instance. Therefore, another metaheuristic is introduced to challenge the results of
the hybrid evolutionary algorithm.
146
5.2. Metaheuristics
30 130
120
25 110
100
90
20
Fitness [Mio. US$]
Fitness [Mio. US$]

80
70
15
60
50
10
40
30
5 20
10
0 0
1
6
10
14
18
22
26
30
34
38
42
46
50
54
58
62
66
70
74
78
82
86
90
94
98
102
106
110
114
118
122
126
130
134
138
142
146
150
154
158
162
166
170
174
178
1
8
15
22
29
36
43
50
57
64
71
78
85
92
99
106
113
120
127
134
141
148
155
162
169
176
183
190
197
204
211
218
225
232
239
246
253
260
267
274
281
288
295
(a) Baltic instance. (b) WAF instance.
10 1
9 0
4
44
84
124
164
204
244
284
323
363
403
443
483
523
563
603
643
683
723
763
803
843
883
923
963
1003
1043
1083
1123
1163
1203
1243
1283
1323
1363
1403
1443
1483
1523
1563
1603
1643
1683
1723
1763
8 -1
7 -2
6 -3
5 -4
4 -5
Fitness [Mio. US$]
Fitness [Mio. US$]

3 -6
2 -7
1 -8
0 -9
2
42
82
122
162
202
242
282
321
361
401
441
481
521
561
601
641
681
721
761
801
841
881
921
961
1001
1041
1081
1121
1161
1201
1241
1281
1321
1361
1401
1441
1481
1521
1561
1601
1641
1681
1721
1761
-1 -10
-2 -11
-3 -12
-4 -13
-5 -14
-6 -15
(c) Mediterranean instance. (d) Pacific instance.
Figure 5.15.: Convergence of variable neighborhood search in selected LINER-LIB

instances.

The numerical results for the variable neighborhood search are presented in the
remainder of this section. Similar to the evolutionary algorithm, the VNS has several
parameters that can be configured. Details on the manual parameter tuning can
be found in Hilger (2014). The configuration for the numerical results presented
below uses a simulated annealing acceptance criteria and several parallel competing
variable neighborhood descending steps. The shaking step is performed based on
the ChangeLinerService and ChangeSinglePort neighborhood structures.
In Figure 5.15, the convergence of the VNS in selected LINER-LIB instances is
shown. When compared to the evolutionary algorithm (EA), shown in Figure 5.14,
one can observe that the evolutionary algorithm converges slower in the small Baltic
and WAF instance compared to the VNS. Furthermore, the best fitness found in
the Baltic instance using VNS is about 0.5% worse than the one found with the
EA. The best network found for the WAF instance has a profit of 117.01 million
US$ which is about 6% better than the best network found with the EA. In the
Mediterranean instance, the evolutionary algorithm converges faster and reaches a
fitness of close to 6 million US$. However, the best fitness found within 30 minutes
147
is about 13% better using the VNS. In the Pacific instance, the convergence is again
slower in the VNS compared to the EA, and results in marginal profitable networks
after 30 minutes, whereas the EA has a fitness of about 11 million US$.
We conclude that both algorithms can solve the small instances within several
minutes. Comparing the VNS with the EA, the VNS finds similar solutions regarding
the profit in the same amount of time for the small instances. However, the differences
are within a 6% range and could be explained by the stochastic components of the
algorithms. In the Mediterranean instance, the convergence of the EA is faster
compared to the VNS. However, the VNS finds an overall much better solution. In
the Pacific instance, the EA outperforms the VNS in both the convergence and the
best solution found. This can be explained by many individuals that diversify the
whole population. The results of the hybrid evolutionary algorithm for the Pacific
instance show that it becomes more important to execute local search operators more
seldomly because they require a large runtime portion. The experiments indicate
that an important factor is the number of iterations that the metaheuristics can
perform. For the larger instances it can thus be relevant to further speed up the
evaluation to increase the number of iterations.
5.3. Surrogate Extensions to Metaheuristics

The metaheuristics presented in the previous sections rely on a fitness function that
is time-consuming to compute. Chapter 4.6 introduces several promising heuristics
for the real fitness, such as limiting the number of cargo flows. These heuristic
fitness functions can be used in the scope of metaheuristics. Heuristic (or approx-
imate) fitness values are called surrogates (or proxies) and are used frequently in
the engineering context (see for example Forrester et al. (2008), Shahrokhi and
Jahangirian (2010) and Jin (2011)). In these applications, evaluating a complicated
engineering model using simulation approaches is often too time-consuming to be
included within metaheuristics. Thus, approximation functions with an adaptive
level of accuracy are used. In the context of discrete optimization, the application of
surrogate models is relatively rare. Successful applications can be found for example
in Gendreau et al. (1996) where a surrogate has been used to quickly evaluate
operators within a tabu-search which is an example of an informed operator (see
Jin (2011)).
5.3.1. Metaheuristic Metacontrol

One approach to use surrogates is to replace the time-consuming fitness function
with an approximation. The results in Section 5.3.2 indicate that this usually leads
the heuristic to an optimal solution of the surrogate, called a false optimum. This
148
observation is also found by Jin et al. (2000). They, together with Jin (2011), high-
light the importance of managing the surrogate models by combining the surrogates
with the real fitness function. Jin (2005) divides the strategies of applying surro-
gates in fixed and adaptive evolution control. Fixed evolution control can be divided
into individual and generation based approaches. Individual-based approaches select
individuals randomly or fitness-based and afterwards reevaluate them with the real
fitness. Generation-based approaches reevaluate the whole population according to
a strategy, such as a fixed number of generations.
The adaptive evolution control adapts the surrogate accuracy (or parameters) of
the fixed evolution control (such as the number of individuals to reevaluate) during
the solution process.
The strategies presented in Jin (2005) apply to evolutionary algorithms but can
also be used within other metaheuristics, see Gendreau et al. (1996). Although
the concept of using surrogates within metaheuristics is not new, relatively little
attention has been paid to it in the context of discrete optimization so far. This can
be explained by the following reasons:
1. The theoretical impact on the solution approach has not been analyzed in
detail.
2. The impact of the surrogate quality (see Jin et al. (2003)) on the solution
process has not been studied in detail.
3. Many problems, such as the distance based vehicle routing problem, provide
a compact and easy way to calculate the fitness (objective) function.
Surrogates in discrete optimization problems are useful for problems with time-
consuming fitness functions, such as multi-commodity flow problems with design
decisions. Especially when no delta-evaluation of the solution changes is possible
within a reasonable amount of time, the use of surrogates can help to explore the
search space.
In Figure 5.16, a concept to manage different surrogate models in metaheuristics
is given as a UML class diagram. Metaheuristics attach to a MetaControl class
that handles the evaluation of one or more individuals, depending on whether the
algorithm is population (genetic or evolutionary algorithms, ant-colony optimization
etc.) or individual based (for example VNS, simulated annealing, tabu search etc.).
The MetaControl must be configured to use a specific strategy to evaluate the
solutions. In Figure 5.16, strategies for the fixed evolution control are given that
can be either used in individual or population based metaheuristics and operate in
one or two phases. The following strategies are implemented and evaluated in the
numerical results:
Constant (CS) Evaluates (the/all) solution(s) according to a fixed surrogate.
149
Figure 5.16.: Concept for a heuristic meta control for population and individual
based heuristics.
Fixed Generation (FGS) Evaluates the whole population using the real fitness
after a fixed number of generations.
Fixed Population Percentage (FPS) Evaluates a fixed percentage of random in-

dividuals using the real fitness, the remaining individuals using the surrogate.
Best Reevaluation (BRS) This two phase strategy first evaluates the whole pop-
ulation using a constant surrogate. Afterwards, the best b% of the individuals
in the population are reevaluated using the real fitness.
The use of surrogates can have the following advantages compared to always
using the optimal or real fitness:
1. Speed-up of the fitness calculation and thereby a faster search space explo-
ration.
2. Diversification of the search because the surrogate can lead to false but diver-
sified decisions in metaheuristics.
Although there are advantages of using surrogates, it is reported that their appli-
cation can lead heuristics to false optima (see Jin (2011)). Surrogates continuously
150
evaluate a solution worse (or better) than the solution actually is. The impact of
surrogates on the metaheuristic and surrogate quality measures are analyzed in Jin
(2011). In the next section, we apply the (static) surrogate concept to the evolution-
ary algorithm and the variable neighborhood search. We show, that especially the
larger instances in the VNS benefit from the use of surrogates.
5.3.2. Numerical Results

The surrogate concept presented above is applied to the proposed metaheuristics.
The results for the hybrid evolutionary algorithm are presented first. The fitness
function presented in this section is an optimization algorithm: the cargo allocation
problem.
As presented in Section 4, three different approximation parameters exist which
are used as surrogates here: reduce the amount of transportable cargo flows (CF),
reduce the accuracy of the bunker cost function (SP) and decrease he maximum
number of column generation (CG) iterations (It). The transportable cargo flows
can be selected using different strategies: In the scope of this analysis we focus on
the cumQ ∗ R and Q ∗ R. cumQ ∗ R orders the cargo flows by their possible total
revenue (quantity times revenue) and selects the cargo until a required percentage
X (based on the overall possible revenue) is reached (see Chapter 4). In the Q ∗ R
strategy, the cargo flows are ordered descending by their quantity times revenue.
The Q ∗ R strategy takes X percent of all cargo flow definition, leading to a higher
amount of cargo flows for the larger instance (see Appendix C.1). The surrogate
strategies presented in the previous section are evaluated using different surrogate
accuracies, denoted as (CF, SP, It). CF is the percentage of cargo flows, SP the
percentage of support points (20 support points are referred to as 100%) and It the
maximum number of column generation iterations.
Evolutionary Algorithm
The evolutionary algorithm provides several opportunities for static surrogates that
are backed by the literature (see Jin (2011)).
The most simple usage of surrogates is to replace the real fitness (the cargo allo-
cation problem) with a surrogate, using the constant surrogate (CS). This method
has the largest benefit to speedup heuristics because only the surrogate is used.
Figure 5.17 presents the results for this strategy for the WAF LINER-LIB instance
for different cargo flow volumes, bunker cost accuracies and column generation iter-
ations. The results are averaged on five runs to reduce the stochastic effects of the
metaheuristics.
In general, one can observe that more accurate fitness functions also find better
solutions. The number of column generation iterations combined with the cargo
flow volume highly influence the quality of the solutions. The configuration with
151
130
120
110
100
Fitness [Mio. US$] 90
80
70 CF 100%_SP 100%_It 1
60 CF 100%_SP 40%_It 1
50 CF 100%_SP 40%_It 2
40 CF 70%_SP 40%_It 1
30
20
10
0
1
15
29
43
57
71
85
99
113
127
141
155
169
183
197
211
225
239
253
267
281
295
Runtime [s]
Figure 5.17.: Fitness convergence in the evolutionary algorithm on the WAF LINER-
LIB instance using a constant surrogate with different accuracies.
100% cargo flows and 40% bunker cost discretization accuracy (CF 100%, SP 40%, *)
performs better when two column generation iterations instead of one are performed.
This can be explained by the high utilization of the improved networks that lead
to beneficial alternative paths for the cargo flow due to the capacity constraints.
On the other hand, modifying the accuracy of the bunker cost discretization does
not influence the convergence clearly. Configuration CF 100%, SP 100%, It1 behaves
almost identical to CF 100%, SP 40%, It1. This is backed by the observations in
Chapter 4.6, where the bunker cost discretization is relatively good with few support
points. Limiting the cargo flows’ volume (for example to 70% most profitable cargo)
clearly converges to a false global optimum (this also has been observed by Jin
(2005)). The reason is that the opportunity to route the remaining 30% is missing
and the heuristic focuses on transporting the most profitable cargo only.
The fitness values in Figure 5.17 refer to the surrogate values and are expected
to be lower compared to evaluating the cargo allocation problem to optimality, i.e.
using all cargo flows and an exact bunker cost function. In Figure 5.18 the average
real fitness of the best solutions found after five minutes using the surrogate CS
strategy is presented. The reevaluated networks have fitness values similar to those
presented in Figure 5.17, whereas the 100% cargo volume and one column generation
iteration are very similar. This indicates that the networks are designed to serve
the cargo on the cheapest path. Allowing alternative paths does not increase the
profit by using more than one CG iteration. Compared to the surrogate fitness, the
fitness of configuration CF 70%, SP 40%, It1 increased most because more revenue
can be gained through the available cargo flows.
To conclude the constant surrogate evaluation, the results are not better compared
to the real fitness in the previous section. The evaluation of larger instances is
152
130
120
110
100
90
Fitness [Mio. US$]
80 CF 100% SP 100% It 1
70 CF 100% SP 40% It 1
60 CF 100% SP 40% It 2
50 CF 70% SP 40% It 1
40
30
20
10
0
Figure 5.18.: Average real fitness for the best solutions found for different constant
surrogates. Note that configuration CF 70%, SP 40%, It1 leads to one
transit time invalid solution when reevaluated with the real fitness
function.
omitted because it is expected that algorithms that only rely on a constant surrogate
show a similar undesirable behavior.
To avoid the problem to converge to a false global optima according to the
surrogate, different strategies that make use of the real fitness can be used (see
previous section). These are: Fixed Individual Percentage Strategy (FPS), Fixed
Generation Strategy (FGS) and Best Reevaluation Strategy (BRS). The numerical
results for these strategies are presented in the remainder of this section.
The convergence of the evolutionary algorithm in the WAF instance is shown in
Figure 5.19. In each subfigure, the convergence when evaluating 10%, 25%, 50%,
75% and 100% random population’s individuals using the real fitness is plotted for
the fixed population percentage strategy (FPS). Because the evaluation runtime of
the remaining individuals’ should be decreased, the relatively inaccurate surrogate
with one column generation iteration, 70% and 30% of cargo flows and 40% and 30%
of bunker cost discretization support points are used. In the top row of Figure 5.19,
the cum.Q∗R, on the bottom the Q∗R cargo flow selection strategy are shown. The
results indicate, the cumQ ∗ R strategy works slightly better than Q ∗ R. The main
drawback is that all results indicate that evaluating the whole population using the
real fitness is superior compared to the surrogate model in the WAF instance.
Very similar results are available for the Mediterranean instance (see Figure 5.20).
Nevertheless, a slightly faster convergence is observed at the beginning of the solution
approach. This can be explained by the quicker fitness calculation.
Overall, we can conclude that randomly picking individuals from the population
that are evaluated exactly is not superior to the classic metaheuristic approach and,
usually, even leads to solutions with worse objective values.
153
130 130
120 120
110 110
100 100
90 90
Fitness [Mio. US$]

Fitness [Mio. US$]
80 80
70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
101
111
121
131
141
151
161
171
181
191
201
211
221
231
241
251
261
271
281
291
101
111
121
131
141
151
161
171
181
191
201
211
221
231
241
251
261
271
281
291
1
11
21
31
41
51
61
71
81
91
1
11
21
31
41
51
61
71
81
91
FPS 10 % FPS 25 % FPS 50 % FPS 75 % FPS 100% FPS 10 % FPS 25 % FPS 50 % FPS 75 % FPS 100%
(a) 70% CF (cum. Q*R), 40% Support Points (b) 30% CF (cum. Q*R), 30% Support Points
and 1 CG-Iteration. and 1 CG-Iteration.
130 130
120 120
110 110
100 100
90 90
Fitness [Mio. US$]
80 Fitness [Mio. US$] 80

70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
1
11
21
31
41
51
61
71
81
91
101
111
121
131
141
151
161
171
181
191
201
211
221
231
241
251
261
271
281
291
1
11
21
31
41
51
61
71
81
91
101
111
121
131
141
151
161
171
181
191
201
211
221
231
241
251
261
271
281
291
FPS 10% FPS 25% FPS 50% FPS 75% FPS 10% FPS 25% FPS 50% FPS 75%
(c) 70% CF (Q*R), 40% Support Points and 1 (d) 30% CF (Q*R), 30% Support Points and 1
CG-Iteration. CG-Iteration.
Figure 5.19.: Fitness convergence using different surrogate configurations for the
FPS strategy in the evolutionary algorithm for the WAF instance.
10 10
9 9
8 8
7 7
6 6
5 5
Fitness [Mio. US$]
Fitness [Mio. US$]
4 4
3 3
2 2
1 1
0 0
1
111
166
221
276
331
386
441
496
551
606
661
716
771
826
881
936
991
111
166
221
276
331
386
441
496
551
606
661
716
771
826
881
936
991
1046
1101
1156
1211
1266
1321
1376
1431
1486
1541
1596
1651
1706
1761
56
1046
1101
1156
1211
1266
1321
1376
1431
1486
1541
1596
1651
1706
1761
56
-1 -1
-2 -2
-3 -3
-4 -4
-5 -5
-6 -6
FPS 25% FPS 50% FPS 25% FPS 50%
(a) 70% CF (cumQ*R), 40% Support Points and (b) 70% CF (Q*R), 40% Support Points and 1
1 CG-Iteration. CG-Iteration.
Figure 5.20.: Fitness convergence using different surrogate configurations for the
FPS strategy in the evolutionary algorithm for the Mediterranean
instance.
154
130 130
120 120
110 110
100 100
90 90
Fitness [Mio. US$]
Fitness [Mio. US$]

80 80
70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
101
111
121
131
141
151
161
171
181
191
201
211
221
231
241
251
261
271
281
291
101
111
121
131
141
151
161
171
181
191
201
211
221
231
241
251
261
271
281
291
1
11
21
31
41
51
61
71
81
91
1
11
21
31
41
51
61
71
81
91
FGS-2 FGS-5 FGS-10 FGS-20 FGS-2 FGS-5 FGS-10 FGS-20
(a) cumQ ∗ R cargo flow selection. (b) Q ∗ R cargo flow selection.
Figure 5.21.: Evolutionary Algorithm fitness convergence using the Fixed Genera-
tion Strategy (FGS) with a 70% CF, 40% Support Points and 1 CG
iteration surrogate with the cumQ ∗ R and Q ∗ R strategy for the WAF
instance.
The second static surrogate is the fixed generation strategy (FGS) for which the
numerical results are shown in Figure 5.21 for the WAF, in Figure 5.22 for the
Mediterranean instance. For every instance and cargo flow selection strategy, every
2, 5, 10 and 20 iterations the whole population is evaluated using the real fitness
function. The remaining individuals are evaluated with a surrogate that allocates
70% of the cargo flows, uses 40% of the support points and terminates after the first
delayed column generation iteration. The cum.Q ∗ R cargo flow selection in Figure
5.21(a) converges better than the Q ∗ R strategy in Figure 5.21(b). Compared with
the evolutionary algorithm using the real fitness, the best fitness found is about 4 -
5% better using the FGS surrogate strategy and the cum.Q ∗ R cargo flow selection.
The Q ∗ R selection is about 3% better compared to the real fitness evaluation. The
results for the Mediterranean instance indicate that in some configurations the runs
with less exact evaluation, in others the ones with more exact evaluations perform
better. A reason can be the diverse search space exploration of the evolutionary
algorithm. This has been considered when choosing the crossover operators. In the
case of evaluating the whole population approximately, the operators might not
work as expected.
Finally, the results for the best reevaluation strategy (BRS) in the hybrid evolu-
tionary algorithm are presented. The whole population is evaluated using a surrogate
that routes 70% of the cargo flows according to the cum.Q ∗ R strategy, 40% of
the support points and the column generation is terminated after the first iteration.
In a second phase, the top 5%, 10%, · · · , 75% of the population is reevaluated
using the real fitness. The results are shown in Figure 5.23 for the WAF instance.
Compared to the real fitness evaluation, the best fitness is 1-2% higher using the
BRS surrogate on 75% of a population’s individuals. Decreasing the percentage
155
10 10
8 8
6 6
Fitness [Mio. US$]

Fitness [Mio. US$]
4 4
2 2
0 0
1
111
166
221
276
331
386
441
496
551
606
661
716
771
826
881
936
991
1046
1101
1156
1211
1266
1321
1376
1431
1486
1541
1596
1651
1706
1761
56
111
166
221
276
331
386
441
496
551
606
661
716
771
826
881
936
991
1046
1101
1156
1211
1266
1321
1376
1431
1486
1541
1596
1651
1706
1761
56
-2 -2
-4 -4
-6 -6
FGS-2 FGS-5 FGS-10 FGS-20 FGS-2 FGS-5 FGS-10 FGS-20
Figure 5.22.: Evolutionary Algorithm fitness convergence using the Fixed Genera-
tion Strategy (FGS) with a 70% CF, 40% Support Points and 1 CG
iteration surrogate with the cumQ ∗ R and Q ∗ R strategy for the
Mediterranean instance.
130
120
110
100
90
FItness [Mio. US$]
80
70
60
50
40
30
20
10
0
101
111
121
131
141
151
161
171
181
191
201
211
221
231
241
251
261
271
281
291
1
11
21
31
41
51
61
71
81
91
Runtime [s]
BRS 5% BRS 10% BRS 25% BRS 50% BRS 75%
Figure 5.23.: Evolutionary Algorithm fitness convergence using the Best Individuals
Reevaluation Strategy (BRS) with 70% CFs, 40% Support Points and
1 CG iteration surrogate with the cumQ ∗ R strategy for the WAF
instance.
leads to solutions that are much worse. The numerical results are similar for the
Mediterranean instance. In Figure 5.24, the best found fitness equals the fitness of
the exact function.
The results of the surrogate fitness functions in the hybrid evolutionary algorithm
are not significant. Although the best reevaluation strategy performs slightly better
than the real fitness, the changes are within a one-digit percentage and are rather
explained with the stochastic nature of the algorithms. One reason for the bad
156
10
9
8
7
6
5
Fitness [Mio. US$]
4
3
2
1
0
1
56
111
166
221
276
331
386
441
496
551
606
661
716
771
826
881
936
991
1046
1101
1156
1211
1266
1321
1376
1431
1486
1541
1596
1651
1706
1761
-1
-2
-3
-4
-5
-6
Runtime [s]
BRS 5% BRS 10% BRS 25% BRS 50% BRS 75%
Figure 5.24.: Evolutionary Algorithm fitness convergence using the Best Individu-
als Reevaluation Strategy (BRS) with 70% CFs, 40% Support Points
and 1 CG iteration surrogate with the cumQ ∗ R strategy for the
Mediterranean instance.
performance of the surrogate fitness function can be the missing normalization

between the surrogate and the exact fitness values. The best individuals are selected
according to the elitism strategy that is based on the fitness. When some individuals
are evaluated using a surrogate, they depend on the random part of the elitism
function to survive the selection process and evolve different generations. Thus, the
metaheuristic should rely on fitness values, evaluated with the same fitness method.
In the next section, the constant surrogate strategy is applied to the variable
neighborhood search.

The variable neighborhood search (VNS) heavily relies on the local search algorithm
variable neighborhood descent (VND) that is used in this thesis. Due to the high
fitness sensitivity to network changes, the crossover operator might be the main
drawback in the hybrid evolutionary algorithm. Thus, the VNS has been proposed
that already indicated promising results for the medium sized instances. The draw-
back in larger instances is the relatively slow convergence. However, within the VNS
one can easily adjust the VND surrogate accuracy to decrease the runtime. The
results indicate that decreasing the accuracy can lead to false decisions within the
whole algorithm.
Regarding the acceptance decision, the quality of the surrogate is of great im-
portance. The VNS metaheuristic differs in an important point to the evolutionary
157
algorithm: Within the VND, a modified solution is accepted if its fitness is better
than the previous solution. This means that the single decision that has to be
supported by the surrogate is whether the solution is better (or worse) than before
according to the real fitness.
Table 5.11 presents an overview of the surrogate quality (for a detailed evaluation,
based on each single operator, see Appendix D.1 and D.2). The table shows the
number of neighborhood structure executions for one VNS run for the Mediterranean
instance. Beside the number of calls, the table indicates the number of correct
decisions (according to the real fitness function), the number of wrong decisions and
the number of no changes for different surrogate configurations. In all configurations,
the number of iterations of the delayed column generation is fixed to one and only
40% of the support points are used. The different surrogate configurations limit the
amount of transportable cargo flows to increase the speed of the evaluation. For
example, the second configuration in Table 5.11, CF 75% − Q ∗ R, indicates that
75% of the cargo flows, ordered descending by the most profitable cargo, is used in
the surrogate. One can observe that on average 85% of the operations performed in
the VNS are correct in the sense that evaluating them with the real fitness function
would lead to the same decision. The right decision can be either to accept or reject
a neighboring solution.
Configuration/ correct wrong

operator operations decisions [%] decisions [%]
P
CF 100%, Q*R 26907 88,26% 11,74%

CF 75%, Q*R 30177 84,90% 15,10%
CF 50%, Q*R 34461 84,60% 15,40%
CF 100%, cum. Q*R 25969 86,65% 13,35%
CF 75%, cum. Q*R 38989 83,99% 16,01%
CF 50%, cum. Q*R 52314 83,56% 16,44%
Table 5.11.: Accuracy of surrogate evaluation in the variable neighborhood descent

solving the Mediterranean instance. The CG is aborted after the first
iteration and 40% (8) support points for the bunker cost discretization
is used. Data is recorded in five runs of the variable neighborhood
descent.
The configuration CF 100%, Q*R and CF 100% - cum. Q*R indicates that the
number of iterations and support points still have a strong effect on the correctness
of the decisions. With the decreased number of cargo flows, the executed operations
increase, especially for the cum.Q ∗ R cargo flow selection, which relies on taking
the cargo flows that lead to X% of the possible overall revenue. Due to the different
cargo flow distributions (see Appendix C.1), the speed up is larger compared to the
Q ∗ R strategy and more than twice as many VNS iterations could be performed
when using 50% of the CFs instead of 100% in the cum.Q ∗ R strategy. However,
158
130 130
120 120
110 110
100 100
90 90
Fitness [Mio. US$]
Fitness [Mio. US$]

80 80
70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
1
101
111
121
131
141
151
161
171
181
191
201
211
221
231
241
251
261
271
281
291
101
111
121
131
141
151
161
171
181
191
201
211
221
231
241
251
261
271
281
291
11
21
31
41
51
61
71
81
91
1
11
21
31
41
51
61
71
81
91
CF 75 %, SP 40 %, CumQR CF 50 %, SP 40 %, CumQR CF 100 %, SP 40 %, CumQR CF 75 %, SP 40 %, QR CF 50 %, SP 40 %, QR CF 100 %, SP 40 %, QR
Figure 5.25.: Fitness convergence using different static surrogates in the VND local
search in the WAF LINER-LIB instance. CG is terminated after the
first iteration.
130 130
120 120
110 110
100 100
90 90
Fitness [Mio. US$]
Fitness [Mio. US$]
80 80
70 70
60 60
50 50
40 40
30 30
20 20
10 10
0 0
1
11
21
31
41
51
61
71
81
91
101
111
121
131
141
151
161
171
181
191
201
211
221
231
241
251
261
271
281
291
101
111
121
131
141
151
161
171
181
191
201
211
221
231
241
251
261
271
281
291
1
11
21
31
41
51
61
71
81
91

second iteration.
the percentage of wrong decisions only increases slightly from 13.35% to 16.44%.
The results indicate that the surrogate approximation in the variable neighborhood
descending step is promising because the vast majority of decisions are still correct.
Figure 5.25 shows the convergence of the VNS with the configurations presented
in Table 5.11 for the WAF instance. The results indicate that the convergence is
highly influenced by the static surrogates: Limiting the surrogate to 50% of the
cargo flows leads to objective values that are much worse than the best ones found
using the real fitness function (compare Figure 5.15(b)). The best fitness found
using the surrogate in the VND is about 4% worse compared to the real fitness for
all cargo flow selection strategies.
The results change slightly when the accuracy of the surrogate is increased by
limiting the number of column generation iterations to two. In Figure 5.26 the best
159
13 13
12 12
11 11
10 10
9 9
8 8
7 7
Fitness [Mio. US$]

Fitness [Mio. US$]
6 6
5 5
4 4
3 3
2 2
1 1
0 0
111
166
221
276
331
386
441
496
551
606
661
716
771
826
881
936
991
1
111
166
221
276
331
386
441
496
551
606
661
716
771
826
881
936
991
1
56
1046
1101
1156
1211
1266
1321
1376
1431
1486
1541
1596
1651
1706
1761
56
1046
1101
1156
1211
1266
1321
1376
1431
1486
1541
1596
1651
1706
1761
-1 -1
-2 -2
-3 -3
-4 -4
-5 -5
-6 -6
search in the Mediterranean LINER-LIB instance. CG is terminated
after the first iteration.
objective found with two CG iterations and 40% of support points is more than
120 million US$ (about 2.5%, see Figure 5.15(b)). However, the differences in the
surrogates are relatively small for the WAF instance.
As expected, the advantage of faster evaluations in the VND local search increases
with larger problem instances. In Figure 5.27, the VNS convergence is shown for the
same surrogate configurations. Recall that the best fitness found for the Mediter-
ranean instance using the real fitness was about 7 million US$ (see Figure 5.15(c)).
The results shown in Figure 5.27 indicate a much faster convergence for all
surrogates for the maximum runtime of 30 minutes. The best fitness found is at
least 30% better compared to the VNS using the real fitness VND evaluation. The
best fitness found using the surrogate with the worst accuracy of 50% of cargo flows,
40% of support points and one delayed column generation iteration is about 70%
better. In contrast to the much smaller WAF instance, increasing the number of
column generation iterations to two slows the convergence down because the fitness
calculation becomes too time consuming (see Figure 5.28).
These results indicate that the runtime improvements using surrogate models can
be beneficial, especially in medium sized instances (see Figures 5.28(a) and 5.28(b)).
To back these findings, Figure 5.29 shows the same surrogate configurations for
the Pacific instance. The best solution found with the VNS without a surrogate was
on average 100,000 US$ after 30 minutes. Using the surrogate results in a fitness
of more than 8 million US$ using the cumQ ∗ R strategy and more than 14 million
US$ using the Q ∗ R strategy. Furthermore, the convergence of the VNS clearly
improves because better solutions were found much faster. This holds for all tested
LINER-LIB instances.
Compared to the hybrid evolutionary algorithm that found Pacific networks with
a profit of 11 million US$, VNS improves the best solution found as well. It is
160
5.4. Interpretation of Results
13 13
12 12
11 11
10 10
9 9
8 8
7 7
Fitness [Mio. US$]
Fitness [Mio. US$]

6 6
5 5
4 4
3 3
2 2
1 1
0 0
111
166
221
276
331
386
441
496
551
606
661
716
771
826
881
936
991
1
56
111
166
221
276
331
386
441
496
551
606
661
716
771
826
881
936
991
1046
1101
1156
1211
1266
1321
1376
1431
1486
1541
1596
1651
1706
1761
56
1046
1101
1156
1211
1266
1321
1376
1431
1486
1541
1596
1651
1706
1761
-1 -1
-2 -2
-3 -3
-4 -4
-5 -5
-6 -6
after the second iteration.
about 45% higher which clearly shows the superiority of the VNS using a constant
surrogate strategy.
17 17
16 16
15 15
14 14
13 13
12 12
11 11
10 10
Fitness [Mio. US$]
Fitness [Mio. US$]
9 9
8 8
7 7
6 6
5 5
4 4
3 3
2 2
1 1
0 0
-1 -1
1
56
111
166
221
276
331
386
441
496
551
606
661
716
771
826
881
936
991
1046
1101
1156
1211
1266
1321
1376
1431
1486
1541
1596
1651
1706
1761
111
166
221
276
331
386
441
496
551
606
661
716
771
826
881
936
991
1046
1101
1156
1211
1266
1321
1376
1431
1486
1541
1596
1651
1706
1761
56
-2 -2
-3 -3
-4 -4
-5 -5
search in the Pacific LINER-LIB instance. CG is terminated after the
first iteration.
5.4. Interpretation of Results

Numerical results for the exact mixed integer model and metaheuristics for the liner
shipping network design problem are presented in this chapter.
They indicate that the exact model can only be solved for very small instances of
the liner shipping network design problem to optimality. For the smallest LINER-
LIB instances large gaps are still observed after 24 hours. The main reason is the
relative bad bounds that result from the formulation presented in this thesis which
161
could be improved by introducing cutting plane techniques. This technique to solve

the problem to optimality is also recognized by Plum et al. (2013b).
To be able to solve the problem for medium sized instances, two metaheuristics
are presented: An hybrid evolutionary algorithm and a variable neighborhood search.
Both methods are able to provide reasonable results up to the Pacific instance of the
LINER-LIB benchmark suite. Using the default configuration, VNS works better
for the small instances. Its drawback is the slow convergence in the medium sized
instances where the hybrid evolutionary algorithm outperforms VNS.
To improve the convergence of both heuristics by speeding up the fitness calcula-
tion, the surrogate concept is applied to the liner shipping network design problem.
The hybrid evolutionary algorithm provides a broad range for surrogate applica-
tion strategies. However, none of the methods clearly outperforms the exact fitness
function. The reason is that the population evolves based on the individuals fitness
which is hard to normalize with difficult fitness functions.
The surrogate results for the VNS clearly outperforms the evolutionary algorithm
and the VNS without surrogate in the medium sized Mediterranean and Pacific
LINER-LIB instances. The reason lies in the acceptance criteria of the variable
neighborhood local search. A solution that is evaluated with the surrogate is com-
pared with a neighbor solution evaluated with the surrogate as well. The comparison
shows that using the surrogate is sufficient to identify whether to accept or reject
a candidate solution. Additionally, surrogates introduce diversification which is
important for the heuristics as well.
In the remainder of this chapter, a sensitivity analysis is performed for the so far
fixed bunker prices. Then, the methods are evaluated in a real-world case study.

Fuel consumption is the biggest cost factor (compare Section 2.10). We analyze the
effects of the bunker price fluctuations on the networks and discuss whether future
research should be performed by including the uncertainties in the network design.
To analyze the effects of different bunker prices on the network design the LINER-
LIB 2012 Baltic instance is used as an example. The reference bunker price of 600
US$ per metric ton has been varied by +/- 25%, leading to an upper price of 750 US$
and a lower one of 450 US$ per ton. All presented results are based on three runs
of the variable neighborhood search using the surrogate with a maximum runtime
of three minutes. Figure 5.30 shows the average best profit of the three runs for
different bunker price per ton.
The bunker prices plotted on the x-axis represents a 5% change in the price per
ton (based on the reference of 600 US$). The profit, shown on the y-axis, decreases
with increased bunker price. The profit’s difference within a 5% change is largest
between 450US$ and 480 US$, meaning 2% profit loss. The overall difference between
162
25,5
25,0
24,5
Profit [Mio. US$]
24,0
23,5
23,0
22,5
22,0
21,5
450 480 510 540 570 600 630 660 690 720 750
Bunker price per ton [US$]
Figure 5.30.: Average best profit of three VNS runs for different bunker prices per
ton.
7,0 90
80
ø revenue
6,0 70
Revenue/ THC [Mio. US$]

ø CC 60
Cost [Mio. US$]
5,0
50
ø THC
40
4,0
ø PCC 30
3,0 ø BC 20
10
2,0 0
450 480 510 540 570 600 630 660 690 720 750
Figure 5.31.: Cost components and revenue in the best networks for different bunker
prices per ton.
750US$ and 450US$ are decreased profits of about 9.1%. This of course assumes
that the network can be always freely adjusted to the new parameter at no costs.
The average profit at a bunker price of 630 US$ is slightly lower than expected. This
can be explained by the stochastic component of the algorithm.
Figure 5.30 indicates that the networks adapt to the different bunker cost per ton,
however, it does not indicate whether a structural change has taken place. Changes
in the network design (such as modified vessel types or changed port rotation)
means the method should consider the bunker cost variation directly. The first step
to identify the reasons for profit changes is to look at the cost components.
163
5,5 14
5,0 13
12
4,5
Depl. vessels and ø port calls/wk
Called ports and used services

11
4,0 10
3,5 9
8
3,0 ø depl. vessels
7
2,5 ø port calls/wk
6
ø ports
2,0 5
ø services
1,5 4
3
1,0
2
0,5 1
0,0 0
450 480 510 540 570 600 630 660 690 720 750
Figure 5.32.: Structural network changes with different bunker prices per ton.
In Figure 5.31, the average revenue and average costs of the best networks are
shown. One can observe, that the average revenue, port call cost (PCC) and charter
cost (CC) decrease with increased bunker cost. The terminal handling cost (THC)
stay relatively constant. The decreased profit is caused by 6reduced port calls, which
avoid the transportation of unprofitable cargo. The nearly constant terminal han-
dling cost can be explained by an increased transhipment tendency in the network.
The average total bunker cost (BC) increase with increased bunker cost per ton.
The charter costs indicate that less vessels are deployed on average in the network.
This is reasonable because a network with less vessels leads to an acceleration of
the vessels to match the round trip time. Removing ports allows higher slack time
at sea and a lower average speed.
The reduced number of called ports and deployed vessels can also be seen in
Figure 5.32. The figure indicates a nearly steady number of services used throughout
different bunker costs. In the Baltic instance, on average 2.5 services are deployed,
meaning that the best networks found by the VNS contained two to three services.
Similarly, on average six to seven ports are called per round trip, whereas slightly
less ports are called per round trip under large bunker costs.
As indicated in Figures 5.31 and 5.32, the optimization reacts to the higher
bunker cost by decreasing the called ports and deployed vessels. The average speed
on the services is not clearly reduced (see Figure 5.33). The speed results from the
port rotation and the number of deployed vessels. Because both are decreased with
higher bunker prices, the model can avoid high speeds. Only in case of relatively
low bunker prices are slightly higher speeds observed. This means that under all
circumstances, the speed optimization is highly relevant for the shipping industry
due to the cubic consumption function. The served cargo flows are also decreased
in the high bunker price scenarios. The reason is that either the port of origin or
164
76% 75,2% 12,0
11,8
74%
11,6
72% 71,7%
Served Cargo Flows [%]
70,5% 11,4
69,9% 69,7%
70%
11,2
Speed [kts]
68% 11,0
ø served CFs
66,4% 66,1% 66,2%
66% 65,7% 66,0% 10,8 ø speed
10,6
64%
62,4% 10,4
62%
10,2
60% 10,0
450 480 510 540 570 600 630 660 690 720 750
Figure 5.33.: Served cargo flows and average speed with different bunker prices per
ton.
destination are not served anymore or the cargo becomes unprofitable. Vessels must
increase their speed to compensate the duration to load and unload the cargo, which
becomes unattractive for high bunker cost. The exception for the 630 US$ solutions
can be seen again. VNS leads to networks that decrease the number of vessels and
port calls more than similar solutions (see Figure 5.31). This results in less cargo
served (see Figure 5.33).
Table 5.12 shows the effects on the networks’ utilizations. The second column
indicates the best peak utilization per service and per leg (BPU), the third the
average worst utilization (WPU) and the last the maximum average utilization per
service (BAU). The WPU remains at values between 78 - 87%, indicating that all
legs have a good lower utilization bound. BPU shows that leg utilization up to
100% can be reached on nearly 40% of the networks. Due to the decreased served
cargo flows and the reduced speed, unprofitable cargo is skipped in the optimization
under relatively high bunker cost per ton. This is indicated by a slightly decreased
average best utilization. However, since the network can perform adjustments to
their structure, the BAU remains relatively high.
Concluding the results, structural network design changes occur for the port
rotation and number of vessels. In the low bunker price scenario, the optimization
model deploys more vessels and extends the geographical coverage. Additionally,
more cargo is transported at a higher speed. With increased bunker prices per ton,
less vessels and ports are served in the networks and the average speed is decreased.
Typically, liner networks are designed and implemented for several years. In this
time period, the bunker price is subject to uncertainty and the liner carriers have
to adjust to the increased cost.
On the one hand, the sensitivity analysis indicates that a relatively high profit can
165
Bunker Cost US$ BPU% WPU% BAU%

450 100.00% 81.79% 100.00%
480 100.00% 86.71% 99.68%
510 100.00% 86.54% 99.68%
540 98.81% 79.13% 95.83%
570 97.63% 85.88% 97.31%
600 100.00% 87.37% 100.00%
630 100.00% 85.42% 99.36%
660 95.46% 79.17% 95.14%
690 95.46% 79.71% 95.46%
720 96.83% 79.67% 96.83%
750 95.46% 79.71% 95.46%
Table 5.12.: Average BPU, WPU and BAU for networks optimized to different
bunker prices per ton.
8%
7%
6%
5%
4%
Profit changes
3%
2% Δ adjustment
1% Δ no-adjustment
benefit of adjustment
0%
-1% 450 480 510 540 570 600 630 660 690 720 750
-2%
-3%
-4%
-5%
Bunker price scenario [US$]
Figure 5.34.: Benefit of adjusting network for different scenarios that is optimized
for a bunker price of 600 US$ per metric ton.
still be reached under high bunker cost scenarios by implementing slow steaming
strategies and structurally adjusting the network. An increase in the bunker price
by more than 60% leads to decreased profits of less than 10%. On the other hand,
however, this adjustment is only possible if a carrier is able to implement the changes.
Long term vessel charter contracts or long term cargo flow contracts would fix both
the capacity and the required ports, which limits the freedom of action.
In Figure 5.34, the networks optimized with a bunker price of 600 US$ are used
166
5.6. Numerical Results from a Global Liner Carrier
as a reference. On the x-axis different scenarios that vary the bunker cost with a
step size of 5% is given. The y-axis indicates the profit change in percent with a
basis of the profit at 600 US$. A liner carrier can either change the network (∆
adjustment) or remain in the network and use the slack time (both to the maximum
and minimum speed) to react on the new bunker price (∆ no-adjustment).
When the bunker price decreases, the network is more profitable in both reactions
due to the decreased total cost. The difference in the benefit slightly increases
when the new bunker price increasingly differs from the planned price. Figure 5.34
indicates that the benefits gained from redesigning the network instead of keeping
the network can be up to 3% for this relatively small feeder network. When the
bunker price changes by more than 20%, the network should be redesigned or at
least adjusted to the new parameters.
In times of increased bunker prices, liner carriers also benefit from redesigning
their network and becomes more important with increasing prices. However, the
benefit is smaller compared to decreased bunker prices. Figure 5.32 indicates that
in these cases fewer ports per round trip are called. We argue that a critical network
size is reached, where serving even less ports becomes less useful. Of course, the
development of increased bunker cost indicated in Figure 5.31 will only work out
as long as transporting the cargo flows is still profitable.
The sensitivity analysis in this section shows that it can be efficient to adjust
liner networks to new bunker prices. In particular, in scenarios of large bunker price
changes the profit can be increased when adjusting to the new price realization.
However, before implementing these changes, further cost must be considered. Tier-
ney and Jensen (2012) indicate that repositioning the vessels to service the adjusted
liner services can lead to large costs. Therefore, the vessel repositioning cost must
be considered before adjusting the network to the new bunker price.
The analysis also shows that the structure of liner shipping networks changes
under different bunker prices. Especially during high volatile bunker price periods,
such as in 2009 (see Chapter 2), the scenarios should be included to create solutions
as stable as possible for different bunker prices. This could be an interesting aspect
in future research.

The methods described in the previous chapters are applied to data of a global liner
carrier. To challenge the algorithms, a single region within the global network has
been selected. Table 5.13 shows the size of the problem instance. Approximately 40
ports (where in about 31 ports have cargo flow origins or destinations), 1400 legs
and 1600 cargo flows of five different equipment types are selected: 20 and 40 foot
dry and reefer as well as high cube containers. For the cargo flows, 140 transit time
requirements between ports are given. The carrier operates a global network and
167
Set Size
Ports 40
Cargo Flows 1600
Vessel Types 4
Resource Groups 3
Equipment Types 5
Legs 1400
Embargo Constraints 12
Transit Time 140
Table 5.13.: Approximate size of the problem instance used for the realistic case
study.
cargo flows are transported into and out from the region to different destinations.
To reduce the network size for optimization the port of loadings and destinations
are set to the current transhipment ports outright the region South America for
example. However, the model is still allowed to tranship containers between different
services at all ports within the region.
The instance defines three available vessel types, ranging from nominal 800 TEU
to 2800 TEU vessels. With an average weight of 14 tons per container, the capacity
is nearly half. Each of the vessel types has a bunker profile attached and uses three
capacity types, namely dry slots, reefer plugs and maximum weight. Each cargo
flow and empty container utilizes the weight resource and either the slot resource.
In case of reefer containers the plug resource is utilized as well.
The case study uses different embargo constraints that define ports that cannot
be served within one liner service due to political reasons. The instance size is
related to the LINER-LIB Pacific and Europe-Asia instance and thereby imposes a
large computational challenge on the algorithms. For improving the given network
structure and determining practical valid networks, some adjustments on the solution
methods must be done (see Figure 5.35): The overall process starts with loading the
instance’s data from a database. Then, the cargo flows are aggregated to improve
the solution time. The aggregation transforms all container types into 20 foot Dry
and Reefer containers between the origin and destination and averages the revenue
per container and the resources’ utilization. Thus, we reduced the amount of cargo
flows by removing the 40 foot and high cube containers. This aggregation reduces
the amount of cargo flows by approximately 50%. The metaheuristics are used
to improve the network. We use the carrier’s current network as a starting point.
The carrier pointed out that the maximum runtime can be set to several hours or
even days due to the strategic decisions and the large potential of the optimization.
However, we have limited the algorithm to terminate after eight hours. Then, the
network must be reevaluated using the disaggregated cargo flows to obtain the
optimal profit according to the cargo allocation (see Figure 5.35). Depending on the
168
Start
Load instance
Aggregate cargo flows
Load Create initial solutions using current and

Database network random networks
Run metaheuristic
Disaggregate cargo flows
Port call
Prune port calls yes
duration<6
no
Reduce vessels yes Buffer used
yes
no
Improved
no yes
Restore Store
Reevaluate
End
Figure 5.35.: Adjusted optimization process to handle practical constraints for the
best liner shipping network found.
data, the gap between the actual cargo flows and the aggregation can be quite large.
However, the tests indicate that the error concerning the evaluation of the current
network is not more than 10%. After the aggregation, the network is modified to
meet practical requirements:
• Remove ports that are not called more than six hours per visit.
• Adjust deployed vessels in case of services with large buffers.

First, the resulting network might still contain ports where the port call duration
is less than six hours which is unreasonable for real-world networks. To remove
these ports from the solution, we remove all ports that are called less than a specific
amount of hours (if the pruning leads to an increased profit). Second, the process
evaluates whether the cargo allocation determines a high additional buffer for all
169
Cons. Cargo Type Utilization (TEU) Utilization (Reefer)

Cargo Flows 75.09% 73.77%
Empty Containers 9.51%
Overall 84.60%
Table 5.14.: Current average utilization (with the optimal cargo allocation). The
TEU utilization is split to laden and empty containers. Additionally.
the average utilized reefer plugs are shown.
ports. This might be the case if the metaheuristic has too many vessels deployed
on the service. Then, the number of vessels is tried to be decreased. Finally, the
whole solution is reevaluated, the proforma schedules that contain the port durations
(including pilotage, buffer and moving time), sea slack time (duration difference from
the current to the maximum speed), detailed cargo flows as well as transhipment
operations are printed out and the method terminates.
The results for this case study are presented in the remainder of this section. One
specific trade has been selected for optimization. The specifics of that trade are: A
high number of competitors, a high level of maturity of the provided liner services
and a connection of two subregions. On the trade route, the operator uses different
services, whereas some operate as feeders from the existing transhipment hubs but
others serve cargo flows on the trade. Table 5.14 indicates the average utilization
of the existing network using the cargo allocation method presented in Chapter 4
(note that this is a profit maximal cargo allocation that is not done in practice yet).
It can be seen that the network’s average utilization is about 84% (including dry,
reefer and empty containers). The table shows that trade imbalances exist on the
services, leading to an average utilization of about 10% due to empty containers.
The utilization of reefer container plugs is less than 74%.
Figure 5.36 provides an overview of the cost structure of the carrier in the sub-
network. The bunker and container handling cost make up about 30% each. The
third largest cost component are the fixed vessel costs at 19%, followed by the port
call costs of 13%. The smallest component is the container depreciation at 5%.
The existing network is optimized using the variable neighborhood search pre-
sented in Section 5.2.4. To evaluate the intermediate results from the heuristic and
be able to compare different solutions, the current system is compared with three
alternatives: Intermediate is a network that results after a runtime of 1.5 hours using
a random start solution. Network A and Network B are the two best solutions found
within one run by the metaheuristic using the current network as initial solution and
referred to as the optimized networks. The intermediate network uses one service
more than the current, the optimized use as many services as before, but modified
the services clearly.
Table 5.15 provides the utilization for the existing, intermediate and optimized
170
13%
Bunker costs
34%
Container Handling
costs
19%
Container costs
Vessel fix costs
5% Port call costs
29%
Figure 5.36.: Distribution of the average cost structure in the current liner network
at hand.
networks. The average utilization of all services is increased by more than 10% on all
optimized networks. The metaheuristic first tries to optimize the network according
to its utilization, leading to a close to full utilization for the intermediate network.
Afterwards, the network’s utilization is decreased compared to the intermediate
solution.
Type Current Intermediate Network A Network B
Laden containers 75.09% 93.52% 82.81% 85.03%
Empty containers 9.51% 5.85% 12.15% 13.02%
Overall 84.60% 99.37% 94.96% 98.04%
Table 5.15.: Average service utilization for all legs by laden (dry and reefer cargo
flows) and empty containers
The utilization indicates that the modifications performed by the metaheuristic

are suited to solve the problem. The differences of the solutions can be explained
using Table 5.16: The provided utilization only considers cargo that can be trans-
ported in the network, however the volume of served cargo is relatively low in the
intermediate solution. During the optimization process, the network is successively
extended so that more ports and thereby more cargo flows can be served. The
resulting utilization also depends on the capacity deployed in the network. Table
5.16 indicates that the intermediate solution serves more ports than the best net-
works found, however it still deploys relatively few capacity compared with the other
networks. This leads to a high utilization, but only 62.48% of served cargo flows.
Another observation is made regarding the average speed on all service legs per
network. The optimization clearly tries to modify the networks in such a way that
171
the vessels can steam with a relatively low average speed on the services (shown in
the first row of Table 5.16). The speed is decreased by more than 25% in the best
solution found. The decreased speed can either be the result of more deployed vessels,
less port calls or rearranged port sequences. The intermediate and network A still
have less capacity (and fewer vessels) deployed, leading to a relatively high speed.
Network B finds a suitable trade-off between network extension and compensation by
more vessels. Table 5.16 indicates that increasing the number of port calls does not
directly result in an increased speed. Instead, the port sequence is highly important.
Property Current Intermediate Network A Network B

System
Served cargo 96.18% 62.48% 95.90% 96.61%
flows [TEU]
∅ speed [kts] 16.15 15.55 14.36 13.14
Max. speed difference 1.20 10 3.70 4.00
per service [kts]
Deployed capacity [%] 100.00% 46.15% 98.29% 103.42%
Port calls [%] 100.00% 123.53% 114.71% 111.76%
Total slack time [d] 23.92 43.51 24.16 41.59
∅ slack time [d] 5.98 8.70 6.04 10.40
Table 5.16.: Properties of the current, an intermediate and two alternative networks.
The results also indicate that the current system is designed to steam on a
relatively constant speed on all legs, indicated by the row Max. speed difference that
shows the maximum difference between the maximum and minimum leg speed of
all services. Small values indicate that the overall network is designed for relatively
constant speeds but still keeps the transit time. High values indicate that the vessels
must accelerate to keep the required transit times, which can lead to high bunker
cost. Especially the intermediate network has a very high speed difference value,
indicating that the solution is not yet suited to reduce the overall average speed.
Network A and B have a higher speed difference compared to the current network.
This indicates that it is useful to let the network have different leg speeds to deal
with the transit time. However, they must not be too large due to the cubic bunker
consumption.
The last indicators presented in Table 5.16 are the total and average slack time
at sea. These represent the total (or average) difference between the duration when
steaming at the schedule’s speed compared to the maximum speed. Thereby, the
time can be used in operations as a buffer to hold the schedule in case of delays. The
slack time depends on the length of the leg and the planned speed on the leg. The
case study’s current network has a total slack time of approximately 24 days, with six
days on average per service. This value is highly increased in the intermediate and
172
network B and indicates that the networks serve either longer distances (intermediate
network) or implement a super slow steaming strategy (network B). The slack time
of network A is relatively similar to the current network, indicating a relatively
similar network structure regarding the distances. The solution methods presented
in this thesis are designed to potentially increase the slack time because the bunker
costs play the most important role in optimizing existing networks.
Intermediate (%) Network A (%) Network B (%)

Profit -47.91% 5.62% 7.66%
Revenue -32.30% 0.22% 1.70%
Bunker cost 44.55% -30.65% -33.06%
Time charter cost -39.06% -1.74% 3.47%
Port call cost -19.58% 7.53% 0.73%
Handling cost -24.51% -1.54% 1.93%
Container cost -30.43% 0.20% 1.24%
∅ cost change 13.81% -5.24% -5.14%
Total cost change 9.39% -7.70% -7.05%
Table 5.17.: Changes in the profit, revenue and cost structure for the intermediate
solution and the two alternative networks.
In Table 5.17 the changes of the profit, revenue and different costs for the three
alternative networks are given in percent of the current system. Values less than zero
indicate that the value is decreased compared to the current system. Values larger
than zero indicate an improvement (on the revenue part) or a worsening (in case of
the cost side). One can observe that the intermediate result’s profit is nearly 50%
worse compared to the current system. This can be explained by the low revenue
and the increased bunker cost by more than 40%. The intermediate network deploys
more than 50% less capacity in the network (and also smaller vessels), which leads
to decreased port call and time charter costs. The handling and container cost are
decreased due to the decreased served cargo flows. Overall, the intermediate network
increases the cost by 9.39% compared to the existing network (due to the bunker
cost).
The optimized networks found by the metaheuristic indicate an improvement of
the current network between 5.6% to 7.7%. The networks transport slightly different
cargo flows that are more profitable. However, the total number of cargo flows is
only changed marginal (see Table 5.16). This can be explained by the fact that the
current system already routes most of the cargo. The optimization tries to increase
the cargo volume and decrease the costs. However, Table 5.17 indicates that the
optimization mainly focuses on the bunker cost, which is decreased by more than
30%. The handling costs are only changed slightly because the revenue of the current
system should be also reached. Overall, network A results in 7.7% decreased cost,
173
whereas B improves the cost by 7.05%.

The case study shows that the developed methods can be used to optimize real-
world liner shipping networks. In particular, it indicates that already relatively small
subnetworks can provide optimization potential.
174
6. Integration into a Decision Support System
The scope of this chapter is to integrate the developed mathematical methods
into a decision support system (DSS). Little (1970) defines a DSS as a “model-
based set of procedures for processing data and judgments to assist a manager in
decision making”. Decision support systems can help planners to investigate more
alternatives and get a higher level of confidence in their decision (see Sharda et al.
(1988)). The purpose of the DSS presented in this chapter is a proof of concept that
shows the technical possibility of supporting planners in adjusting and improving
liner networks. The DSS is referred to as the Liner Network Web Optimizer (LinWo).
Planners can manually create networks for different problem instances (such as the
LINER-LIB) and evaluate their network designs using the proposed cargo allocation
problem (see Chapter 4). The software checks model constraints that help to validate
the manually planned networks (see Section 5.2.1). For example, the transit times
and port drafts are checked automatically and the output is given via the web
interface. The planners thereby get a direct and fast feedback about the global
impacts of regional liner service changes. Furthermore, planners can automatically
optimize liner networks using the metaheuristics presented in Chapter 5.
The chapter starts with applying a general DSS structure proposed in litera-
ture to the liner shipping network planning. A planning process is presented and
implemented in the DSS. Finally, the software architecture and user interface are
presented.
6.1. Decision Support System Components

Decision support systems in general consist of the following components (see (Turban
and Aronson, 2007, p. 100)):
1. Data management
2. Model management
3. Knowledge-based management
4. User interface
These subsystems are shown in Figure 6.1 and presented for the Liner Network
Web Optimizer in the next paragraphs. The data management subsystem reads and
stores the data of the DSS. For LinWo, the data is read directly from the format
175
Other Internet,
computer-based intranets,
systems extranets
Data: external
and internal
Data Model External
management management models
Knowledge-based
subsystem
User interface
Organizational Manager
Knowledge Base (user)
Figure 6.1.: Components of a decision support system, from (Turban and Aronson,
2007, p. 100). The implemented components for the DSS at hand are
marked bold.
provided by the LINER-LIB1 . However, the data can also be read from a database,
for example to provide an external interface to the data.
The data management subsystem is used by the model management and the user
interface. In the scope of the DSS developed in this thesis, the knowledge-based
management subsystem is not implemented due to the missing business specific
requirements. In general, the knowledge-based management component can either
work independently or support the subsystems such as the organizational knowledge
base with knowledge interfered from the quantitative models.
The model management component contains quantitative models to provide ana-
lytical capabilities for the overall system. For LinWo, all methods developed in this
thesis are part of the model management component. The cargo allocation problem
presented in Chapter 4 uses ports, demands and vessel types on a given network to
evaluate it. The data is accessed by the data management subsystem. The cargo
allocation problem is extended to calculate the fitness of a network (see Chapter
1
To be able to frequently update the data, it is directly read from the comma separated files
provided by the LINER-LIB, see http:\\linerlib.org and Brouer et al. (2013).
176
6.2. Process Overview
5.2.2). This extension can be used to provide information on the transit times and
embargo constraints for the designed liner network. Finally, the optimization ap-
proaches presented in Section 5.2 are also integrated into LinWo. The methods are
accessible by the user interface and the data management subsystem that handles
the storage of intermediate results.
The users (planners and managers) access the decision support with the user
interface. For the liner shipping network planning, users are at least the network
planners. Depending on the level of organizational integration of a DSS product, the
sales or vessel charter department can be involved as well. Users interact with the
software through a web-based user interface. The communication is done via a web
server that distributes the websites to the client. Implementation details are given
in Section 6.3.
In the next section, the business processes arising within liner shipping network
planning are described in detail. These processes are supported by the graphical
user interface presented in Section 6.4.
6.2. Process Overview

The business process model2 in Figure 6.2 provides an overview of the process to
plan, evaluate and adjust liner networks using the decision support system.
The first step is to select a problem instance for which networks are designed.
The problem instances, such as the LINER-LIB benchmark instances, contains all
information required to evaluate and optimize a network (see Chapters 4 and 5).
For example, port depths, costs, vessel types, cargo flows etc. are contained in each
instance.
After selecting an instance, the network planner creates a new network or selects
an existing from the database. Networks can be edited by changing the port rotation,
number of vessels or the vessel type of an existing liner service. New liner services
within a liner shipping network can be created as well. At each step, the network
can be evaluated using the cargo allocation problem, which provides the user with
information on the profit and costs of the configuration (see Section 6.4).
Based on random initial networks or a set of existing networks, automatic im-
provement methods can be executed. The user is informed on intermediate networks
resulting from the mathematical optimization methods. The improvement process
can be terminated at any point in time. Afterwards, he is able to manually adjust
parts of the network and reevaluate them using the cargo allocation problem.
With the help of this iterative process, further business constraints can be re-
spected. In the strategic network planning it is assumed that not all aspects can be
covered by formal optimization methods. For example, a liner carrier can require to
2
Business Process Model and Notation (BPMN) is a standard of the Object Management Group
to model internal business procedures, see http:\\www.bpmn.org.
177
Figure 6.2.: User processes for the liner shipping network planning.
always serve a specific port within a service that has either not been explored by the
metaheuristics or evaluated as not profitable to service. These manual adjustments
can be easily done in the process presented in Figure 6.2.
6.3. Client-Server Communication

The software is implemented as a web-application, whereas data can be manipulated
in the client’s browser. This architecture allows the simple deployment of software
changes in case of a multi-user environment. The software components are selected
to support a wide range of web browsers and operating systems. Thereby, we suc-
cessfully could run the application also on mobile devices. The cargo allocation
and network optimizations are performed on the server side to reduce the clients’
hardware requirements.
The used architecture is visualized in Figure 6.3 and uses the Model-View-Control
(MVC) paradigm that distinguishes between the (business) model that stores the
data, the view that presents this data and the connecting controller layer that is
accessed by the view. The architecture differentiates relatively lean data that is sent
to the web site, database objects used to persist networks and objects used by the
models. Both, the web server and the client work and exchange data based on the
same model definition. The server translates between the different models.
The view component in Figure 6.3 uses a single-page application that dynamically
loads the data from the server. Unlike dynamic websites, the server provides static
178
6.3. Client-Server Communication
Model View
HTML
Controller
Client Angular Controller
Controller
Data Management
Model Storage
Server
Figure 6.3.: Client-server communication of the decision support system.
HTML websites that are loaded by the client. The website is not dynamically
rendered on the server to decrease the computation requirements on the server
and to improve the duration to load the site, because clients can cache the static
HTML pages. Inside the static HTML site, several partial views exist that are
dynamically activated and filled with data from the server using the AngularJS
framework from Google (see AngualarJS (2014)). Other JavaScript model view
libraries such as KnockoutJS could be used instead of AngularJS. AngularJS 7
ensures the synchronization between JavaScript objects and the visualization in the
view. Thus, complex DOM3 manipulation is done automatically by Angular which
reduces the required work to adjust the user interface (UI).
The client controller in Figure 6.3 implements functions used by the website’s
JavaScript and control functions in the user interface. The server side’s controller is
responsible to provide CRUD (create, read, update and delete) operations for each
object in the model (such as a liner service, a network etc.).
The server provides interfaces based on the Representational State Transfer
3
Document object model (DOM) is a specification to access HTML objects.
179
(REST) paradigm (see Fielding (2000)), implemented with the ASP4 WebAPI frame-
work. With the help of this API, HTTP requests can be done on unique URLs. For
example, a call on /api/network/2 returns the existing network with the unique ID
2. The result format can be XML or JSON for example. The controller implements
methods for the HTML methods GET, POST, PUT and DELETE. The methods
are called asynchronously and a callback function on the client side is executed as
soon as the method execution is finished. The client controller then processes the
data and the view is updated. The advantage of this technology and communication
is the operation system independence. For most of the platforms and programming
language, libraries to retrieve and parse XML or JSON exist. Thereby, the server’s
data can also be retrieved by a thin clients (such as smart phones or tablet comput-
ers), by a Java Swing or .NET WCF client application. This provides a high degree
of extensibility. In the scope of the proof of concept, messages have to be send from
the server to the client, for example to inform about new optimization results or the
solution progress. To simplify the asynchronous communication between the web
browser and the server, SignalR5 is used. It automatically uses the best technology
available on the client side to send messages from the server. When recent browsers
are used, the new technology Websockets is used.
The graphical user interface is implemented using the Twitter Bootstrap 36 frame-
work. The framework provides a wide range of controls required for modern web
applications. Besides, several extensions for smart phones and table computers exist
that enhance the usability of the web application. Furthermore, the framework is
adjusted to support all modern browsers. To simplify the CSS7 handling, LESS 4 is
used as preprocessor and Leaflet 58 with OpenStreetMap data to visualize the world
map and the liner services. On the server side, a ASP.NET MVC 4 application is
used that communicates with a MS SQL 2008 database using the Entity Framework
5. The ASP sites can be hosted on an IIS9 server that supports server side events.
In Figure 6.4, an example liner service calling three ports is shown. The drawing
of routes between two ports is not supported in OpenStreetMaps or in any other
open source solution by the time of this thesis. Therefore, a system to calculate
the routes between two ports had to be developed. The algorithm used is presented
in the next section. Basically, the algorithm overlays the world map with a node
grid of variable accuracy, removes the nodes that intersect with land masses and
identifies the nodes closest to the ports. Afterwards, a shortest path between all
4
Active server pages (ASP) is a technology developed by Microsoft to dynamically generate
websites.
5
SignalR can be downloaded at http://signalr.net/
6
The Bootstrap framework is available at http://getbootstrap.com
7
Cascading style sheets (CSS) is a standardized language to define the layout of documents, such
as HTML websites.
8
Leaflet is available at http://leafletjs.com
9
Internet Information Services (IIS) is a web server and application platform from Microsoft.
180
6.4. Graphical User Interface
port pairs is calculated using the algorithm of Dijkstra (1959). Each shortest path
can be used to calculate the distance in nautical miles but also the way points on
the world grid between two ports. The shortest paths between all port pairs are
calculated once, stored in a database and automatically read when a liner service
is manipulated on the website.

An important feature of the decision support system is the graphical visualization
of the networks. Figure 6.4 provides an overview of the user interface for the liner
shipping network decision support system LinWo.
In Figure 6.4, the ports of the LINER-LIB Baltic instance are shown. The user
can get information for these ports, such as depth, cost and UN location code, by
clicking on the markers. Furthermore, the view point of the map can be arbitrarily
changed with drag and drop. An example network has been created for the Baltic
region that consists of two liner services: The east Baltic service (EBS) and the west
Baltic service (EBS). The legs of the liner services are created using the method
described in appendix E. The way points for the legs drawn in Figure 6.4 indicate
the approximate routes of the vessels and are used to distinguish the legs. Note that
the distances between the ports are calculated using a much higher resolution of
the map.
An existing liner shipping network can be either modified or evaluated. Figure
6.5 shows the solution of the cargo allocation problem presented in Chapter 4.
Different costs are shown to the user, providing an immediate feedback of the network
quality on a monetary basis. The results of the cargo allocation can be transferred
to other IT systems using the Microsoft Excel format. The document contains
detailed information on the container paths through the network, transhipment
volumes and vessel speeds per leg. This results in a proforma vessel schedule whose
implementation can be discussed with other involved departments, such as the vessel
charter department.
Figure 6.6 shows the graphical support to modify existing liner services. For
example, the port of Gdynia in Poland should be added to the north bound direction
of service EBS, before the call of Kaliningrad in Russia. Therefore, the user clicks
on the edit button of the EBS service and selects Gdynia as new port. The port
can be found using the UN location codes or port names. After selecting the new
port, it is added to the last position of the selected service. Note that the legs are
automatically refreshed on the website to provide an immediate feedback of the
changes in the port rotation. The user can arbitrarily rearrange the ports within
the port rotation using drag and drop. Again, the legs are automatically refreshed.
Finally, network planners can commit their changes by saving them using the data
management subsystem (see Figure 6.6).
181
Figure 6.4.: Graphical user interface for the decision support system.
Figure 6.5.: Evaluating an existing liner shipping network.
182
(a) Port added.
(b) Modified liner shipping network (sequence changed).
Figure 6.6.: Modification of the port rotation of an existing liner service.
183
Editing a liner service can also involve changing the vessel type. Figure 6.7(a)
shows the user interface to select another type using a combo box that provides the
vessel types defined in the current problem instance. Again the user commits the
modified service by saving the changes. Now, one can reevaluate the liner network
using the cargo allocation. Note that extensive validation methods are implemented
within this thesis to ensure the validity of the resulting liner service. As a result,
the cargo allocation problem not only returns the solution and its profit and costs,
but also results from the instance validation. In Figure 6.7(b) an validation error
is shown: The selected vessel type’s lightship draft (draft without any load, see
Chapter 2) is too high to enter the port of Kaliningrad. This intermediate feedback
on network changes is supposed to help planning liner shipping networks.
In the scope of this chapter, a proof of concept for a liner shipping decision support
system has been developed. For a real-world application, several extensions should be
considered: Providing timing and utilization aspects directly in the network drawn
on the map or visualize the intermediate solutions from optimization algorithms.
Furthermore, more planning problems from the literature, such as the empty con-
tainer repositioning or bunker optimization, can be included in the decision support
system.
The design of the architecture respects the scalability and extendability of the
decision support system. The back-end is currently a Microsoft IIS Webserver and
not able to perform concurrent network optimizations. It is expected that already
two parallel optimizations will slow down the server. To avoid this problem, the
architecture can be expanded to delegate the optimization requests to different
optimization processes, preferably to different physical machines. This could be
implemented using an agent-based back-end architecture, for example with a message
queuing system for the communication. In .NET, the agents could use the library
MassTransit10 together with the messaging system RabbitMQ11 . This allows high
scalability in multi-user, multi-optimization scenarios. Optimization-agents could
be started on demand and the messaging system could automatically perform a
load balancing of the optimization requests.
Regarding the process definition, further analysis on the different roles of the
decision support system can be done. Planning liner shipping networks is a highly
interwoven problem, connected with different liner carrier departments: Charter,
sales and line managers provide data for this problem. This aspect can be integrated
into a knowledge base proposed by Turban and Aronson (2007). It is expected that
implementing such as decision support system for liner carriers can highly improve
the planning and optimization process in practice.
10
MassTransit is available for .NET using NuGet or at http://masstransit-project.com
11
The message queuing system RabbitMQ can be downloaded at http://www.rabbitmq.com
184
(a) Changing vessel type.
(b) Reevaluating network.
Figure 6.7.: Changing a liner service’s vessel type and reevaluating the whole net-
work.
185
7. Conclusion
This chapter concludes the thesis by summarizing the work, performing a critical
assessment of the goals and providing an outlook on future research opportunities.
7.1. Summary
Chapter 2 introduced real-world requirements to the liner shipping network design
problem. These requirements were specified based on the current state-of-the-art in
literature but also with experts from the liner shipping industry. Relevant planning
aspects include complex route types, demand properties, timing aspects, cooperation
in liner shipping and empty container repositioning. The chapter also defines the
scope of the network design problem used in this thesis.
Chapter 3 reviews existing research. It showed that the network design problem is
based on fundamental problems that exist in literature since decades. Following,
work in the field of maritime research was analyzed. In particular, literature on
the liner shipping network design, the cargo allocation (container routing), speed
optimization and empty container repositioning problem were reviewed.
An increasing research interest in the liner shipping network design problem could
be observed. The publications to tackle this problem are twofold: Some authors
solve very small instances of the network design problem to optimality, others focus
on optimizing large scale networks using metaheuristics without the possibility of
evaluating the solution quality due to missing optimal solutions. Altogether, the
methods still lack of practical applicability because important components, such
as cargo transportation durations (transit times), speed optimization and partner
network integration, are not considered yet.
Optimization methods to automatically improve liner shipping networks must
deal with the cargo allocation (sub)problem. The container paths resulting from
the cargo allocation problem determine some of the important costs, more specif-
ically, the transhipment and fuel consumption costs. Chapter 3 showed that the
state-of-the-art in cargo allocation already includes some of the real-world require-
ments (such as empty container repositioning) but also lacks of the integration of
major aspects such as integrated speed optimization, draft constraints and partner
integration. The chapter concludes the state-of-the-art by outlining the research
gap and deriving the goals of this thesis.
Chapter 4 presented two mixed integer models with a linearized cubic fuel con-
187
7. Conclusion
sumption function to solve the cargo allocation problem. The models integrated
speed optimization, draft constraints, capacities of operated and partner services
and empty container repositioning. To the best of our knowledge, this has not been
done before in literature.
This thesis proved that the mathematical model can be solved to optimality using
a linear programming relaxation of the mixed integer model. For further speed up,
a novel column generation solution approach was proposed for the cargo allocation
problem. The methods were evaluated on 140 networks of the LINER-LIB bench-
mark suite1 .
Chapter 5 introduced solution approaches for the more complex liner shipping
network design problem. The approaches consider transit times between different
ports to create competitive liner networks. To the best of our knowledge, multiple
transit times across different services have not been considered in the liner network
design before.
First, a mixed integer formulation to solve the liner shipping network design
problem to optimality was presented. Second, two novel metaheuristics were pro-
posed and evaluated on four of the LINER-LIB instances: A hybrid evolutionary
algorithm (EA) and a variable neighborhood search (VNS). To further speed up
the algorithms’ convergence behavior, a fitness approximation (surrogate) approach
known from the engineering context has been applied to the metaheuristics. To
the best of our knowledge, surrogates have not been used in a linear optimization
problem in the field of logistics or maritime research before. After the metaheuris-
tics were evaluated, a sensitivity analysis to assess the impact of different bunker
prices on the liner shipping networks design was performed. Closing Chapter 5, the
solution approaches were evaluated in a real-world case study of a global liner carrier.
Finally, Chapter 6 integrated the solution approaches into a decision support sys-
tem for the liner shipping network planning. Using the web system, networks can
be adjusted using drag and drop functions. Then, using the cargo allocation prob-
lem, the adjustments can be evaluated. At any point of time, the network can
be automatically optimized using the developed metaheuristics for the liner ship-
ping network design problem. Afterwards, manual changes can be made on these
optimized solutions to respect qualitative constraints.
7.2. Critical Assessment

At the beginning of this thesis in Section 1.1 we stated four research objectives
that were defined in more detail in Section 3.5. Summing up, the goals were (1) to
1
The LINER-LIB benchmark suite is introduced by Brouer et al. (2013) and contains data for
ports, demands and vessel types based on Maersk Line
188
7.2. Critical Assessment
evaluate large scale, real-world liner networks, (2) to formalize practical requirements
of liner networks and generate optimal solutions, (3) to develop metaheuristics to
automatically optimize medium-sized liner networks in reasonable computational
time, and (4) to integrate the methods into a prototypical decision support system.
From our perspective these objectives have been achieved.
In Chapter 4 we presented a column generation based solution approach to tackle
the first goal. The numerical results showed that the method can be used to solve
large scale cargo allocation problems subject to real-world constraints in a few
seconds to optimality. With the help of the proposed solution methods, planners are
able to quickly assess a network’s profit by determining the optimal cargo allocation
and vessel speed. This highly relevant problem can be used by liner carriers to assess
alternative networks very quickly. The container path generation subproblem, as
part of the column generation method, is very flexible regarding future extensions,
such as cabotage limitations.
With respect to the second and third goal, Chapter 5 presented optimal and
heuristic approaches to solve the liner shipping network design problem. Practical
requirements to optimize liner networks were successfully formalized as a mixed
integer program and only very small instances were solved to optimality. Therefore,
metaheuristics were developed to optimize medium-sized liner networks in a rea-
sonable amount of computing time. In Chapter 5 is was shown that both, medium
sized artificial instances from the LINER-LIB and real-world networks, could be
optimized within several hours. The numerical results indicated that small networks
in even relatively mature markets can still be optimized within several hours. The
results of the solution methods were discussed and evaluated with network planners
from a global liner carrier.
The developed planning methods were successfully integrated into a prototypical
decision support system in Chapter 6 to realize goal four. We successfully developed
an algorithm to enable the visualization of liner shipping networks on open source
maps. This enables the use of the mathematical models to the real-world application
and enhances the usability of the proposed methods.
Finally, we would like to conclude the goal achievement with a few critics. First,
the mixed integer program was not able to solve the smallest instances of the LINER-
LIB with a sufficient number of services and layers to optimality. The reasons have
been discussed in Section 5.1.3. This made it difficult to assess the solution quality of
our metaheuristics. However, with the help of the lower bounds and the application
to the real-world network we claim that liner shipping networks can be optimized
with the proposed methods.
Second, in reality there exist an interdependency between network changes and
the available cargo flow. Calling different ports within a region can attract new
cargo flows that are not provided as input for the model. Thus, once a new liner
network has been designed the cargo flows should be reevaluated with experts from
the liner carrier.
189
7. Conclusion
7.3. Future Research Opportunities

The solution methods in this thesis are developed for the context of liner shipping.
Some of the aspects, for example the cargo allocation and speed optimization might
be applicable to other industries. However, fuel consumption and speed optimization
play a secondary role in many planning problems, such as public transportation. Not
until fuel prices increase a substantial amount, decreasing the speed on roadways is
of interest.
Extensions to the cargo allocation problem, such as time periods, berth windows
and cargo flow dependent transit times would be an interesting future research. In
operations, the port arrival and destination time is fixed and time window constraints
exist. Furthermore, tide dependent drafts often impose a challenge on when to enter
or leave a port to keep the schedule. These constraints require the modeling of
liner service schedules with specific departure days. This enables the development
of methods to create robust schedules.
First steps towards the consideration of cargo flow dependent transit times and
integrated speed optimization are done by Guericke and Tierney (2014). They
enable leg dependent speed optimization which might support the model to hold
the transit times. Integrating their model into the fitness calculation of liner network
is an interesting research opportunity and should be accompanied by developing
heuristic solution methods.
The liner shipping network design problem could be solved with a reasonable
objective value for medium sized (sub)networks of a global liner network in this
thesis. Future research should improve the mixed integer model formulation or use
advanced solution methods to solve the liner shipping network design problem with
transit times to optimality. Additionally, large scale, real-world liner networks still
provide a computational challenge for the future.
Another research opportunity is to further evaluate the fitness approximation
approach. To obtain a more profound knowledge whether the approach is beneficial
in general, it should be evaluated in other planning problems in the field of discrete
optimization. Additional research can be done in the theoretical analysis of fitness
approximations.
190
Bibliography
R. Agarwal and Ö. Ergun. Ship scheduling and network design for cargo routing in
liner shipping. Transportation Science, 42(2):175–196, 2008.
R. Agarwal and Ö. Ergun. Network design and allocation mechanisms for carrier
alliances in liner shipping. Operations research, 58(6):1726–1742, 2010.
Alphaliner Newsletter, Volume 1, 2011. Operated Capacity January 2010 and 2011,
4th August 2014. URL http://www.alphaliner.com.
Alphaliner Newsletter, Volume 22. Carriers’ Operating Margins vs. Revenue: 1st
Quarter 2014, 13th August 2014. URL http://www.alphaliner.com.
Alphaliner Newsletter, Volume 26, 2013. East-West Carrier Alliances: How do they
stack up?, 4th August 2014. URL http://www.alphaliner.com/.
J.F. Álvarez. Joint routing and deployment of a fleet of container vessels. Maritime
Economics & Logistics, 11(2):186–208, 2009.
M.W. Andersen. Service Network Design and Management in Liner Container

Shipping Applications. PhD thesis, Department of Transport, Technical University
of Denmark, 2010.
AngualarJS. Google inc., 4th August 2014. URL https://angularjs.org.
C. Ansótegui, M. Sellmann, and K. Tierney. A gender-based genetic algorithm

for the automatic configuration of algorithms. In Principles and Practice of
Constraint Programming-CP 2009, pages 142–157. Springer, 2009.
L.H. Appelgren. A column generation algorithm for a ship scheduling problem.

Transportation Science, 3(1):53, 1969.
L.H. Appelgren. Integer programming methods for a vessel scheduling problem.

Transportation Science, 5(1):64, 1971.
C. Archetti, M.G. Speranza, and A. Hertz. A tabu search algorithm for the split
delivery vehicle routing problem. Transportation Science, 40(1):64–73, 2006.
C. Audet, V. Béchard, and S. Le Digabel. Nonsmooth optimization through mesh

adaptive direct search and variable neighborhood search. Journal of Global Opti-
mization, 41(2):299–318, 2008.
191
Bibliography
M. Avriel, M. Penn, N. Shpirer, and S. Witteboon. Stowage planning for container

ships to reduce the number of shifts. Annals of Operations Research, 76:55–71,
1998.
A.L. Ayre. The Theory and Design of British Shipbuilding. Thomas Reed and Co.,
LTD., 1885.
A.J. Baird. Containerization and the decline of the upstream urban port in europe.
Maritime policy and Management, 23(2):145–156, 1996.
B.M. Baker and M.A. Ayechew. A genetic algorithm for the vehicle routing problem.
Computers & Operations Research, 30(5):787–800, 2003.
R.H. Ballou. Business Logistics Management, volume 4. Prentice-Hall, 1999.
BARON. Baron nonconvex minlp solver, 4th August 2014. URL http://
archimedes.cheme.cmu.edu/?q=baron.
C.B. Barrass. Ship Stability for Masters and Mates, volume 6. Butterworth-
Heinemann, 2006.
J. Basedow, U. Magnus, and R. Wolfrum. The Hamburg Lectures on Maritime

Affairs 2009 and 2010. Springer, 2012.
S. Basu. Tabu search implementation on traveling salesman problem and its varia-
tions: A literature survey. American Journal of Operations Research, 2(2):163–173,
2012.
J.E. Beasley and P.C. Chu. A genetic algorithm for the set covering problem.
European Journal of Operational Research, 94(2):392–404, 1996.
T. Bektas. The multiple traveling salesman problem: an overview of formulations

and solution procedures. Omega, 34(3):209–219, 2006.
M. Bellmore and G.L. Nemhauser. The traveling salesman problem: a survey.

Operations Research, 16(3):538–558, 1968.
H.B. Bendall and A.F. Stent. On measuring cargo handling productivity. Maritime
Policy and Management, 14(4):337–343, 1987.
P.R. Bevington and D.K. Robinson. Data reduction and error analysis for the
physical sciences, volume 2. McGraw-Hill, 1969.
M. Boile, S. Theofanis, A. Baveja, and N. Mittal. Regional repositioning of empty

containers: Case for inland depots. Transportation Research Record: Journal of
the Transportation Research Board, 2066(1):31–40, 2008.
192
Bibliography
A.J. Booker, J.E. Dennis Jr, P.D. Frank, D.B. Serafini, V. Torczon, and M.W. Tros-
set. A rigorous framework for optimization of expensive functions by surrogates.
Structural optimization, 17(1):1–13, 1999.
O. Bräysy. A reactive variable neighborhood search for the vehicle-routing problem

with time windows. INFORMS Journal on Computing, 15(4):347–368, 2003.
M.R. Brooks. Sea Change in Liner Shipping. Pergamon, 2000.
B.D. Brouer and G. Desaulniers. A matheuristic for the liner shipping network
design problem. In Electronic Proceedings of LOGMS 2012, 2012.
B.D. Brouer, D. Pisinger, and S. Spoorendonk. Liner shipping cargo allocation with
repositioning of empty containers. INFOR: Information Systems and Operational
Research, 49(2):109–124, 2011.
B.D. Brouer, J. Dirksen, D. Pisinger, C.E.M. Plum, and B. Vaaben. The vessel
schedule recovery problem (vsrp)-a mip model for handling disruptions in liner
shipping. European Journal of Operational Research, 2012.
B.D. Brouer, Alvarez J., Plum C.E.M., Pisinger D., and Sigurd M. A base inte-
ger programming model and benchmark suite for liner-shipping network design.
Transportation Science, 48(2):281 – 312, 2013.
Bunker Index. Bunker price index, 4th August 2013. URL http://www.
bunkerindex.com.
Bunker Index. Bunker index: Price index, 3rd August 2014. URL http://www.
bunkerindex.com/prices/bixfree.php?priceindex_id=1.
A. Busch and W. Dangelmaier. Integriertes Supply Chain Management. Gabler,

2002.
I.L. Buxton and G.H. Stephenson. Evaluating design for upgradeability: A simula-
tion based approach for ships and marine products. In Practical Design of Ships
and Other Floating Structures. Proceedings of the Eighth International Symposium
on Practical Design of Ships and Other Floating Strucutres, volume 1, 2001.
R.H. Byrd, J. Nocedal, and R.A. Waltz. Knitro: An integrated package for nonlinear
optimization. In Large-scale nonlinear optimization, pages 35–59. Springer, 2006.
Canal de Panamá. Panama canal expansion program progress report, 1st August
2014. URL http://www.pancanal.com/eng/expansion/index.html.
C. Chen and Q. Zeng. Designing container shipping network under changing demand
and freight rates. Transport, 25(1):46–57, 2010.
193
Bibliography
P. Chen, H. Huang, and X.Y. Dong. Iterated variable neighborhood descent algo-
rithm for the capacitated vehicle routing problem. Expert Systems with Applica-
tions, 37(2):1620–1627, 2010.
S.C. Cho and A.N. Perakis. Optimal liner fleet routeing strategies. Maritime Policy
and Management, 23(3):249–259, 1996.
C. Chou, S. Kimbrough, J. Sullivan-Fedock, C.J. Woodard, and F.H. Murphy. Using

interactive evolutionary computation (iec) with validated surrogate fitness func-
tions for redistricting. In Proceedings of the fourteenth international conference on
Genetic and evolutionary computation conference, pages 1071–1078. ACM, 2012.
M. Christiansen, K. Fagerholt, and D. Ronen. Ship routing and scheduling: Status

and perspectives. Transportation Science, 38(1):1, 2004.
M. Christiansen, K. Fagerholt, B. Nygreen, and D. Ronen. Ship routing and schedul-

ing in the new millennium. European Journal of Operational Research, 228(3):
467–483, 2012.
V. Chvátal. Linear programming. Macmillan, 1983.
J.F. Cordeau, G. Laporte, J.Y. Potvin, and M.W.P. Savelsbergh. Transportation

on demand. Transportation, Handbooks in operations research and management
science, 14:429–466, 2004.
T.H. Cormen, C.E. Leiserson, R.L. Rivest, C. Stein, et al. Introduction to algorithms,
volume 2. MIT press Cambridge, 2001.
Couenne. Couenne nonconvex minlp solver, 4th August 2014. URL https://
projects.coin-or.org/Couenne.
T.G. Crainic, M. Gendreau, and P. Dejax. Dynamic and stochastic models for the
allocation of empty containers. Operations Research, 41(1):102–126, 1993.
G.A. Croes. A method for solving traveling-salesman problems. Operations Research,

6(6):791–812, 1958.
G.B. Dantzig and J.H. Ramser. The truck dispatching problem. Management
science, 6(1):80–91, 1959.
M. Davarynejad, J. Vrancken, J. van den Berg, and C.A. Coello. A fitness gran-
ulation approach for large-scale structural design optimization. In Variants of
Evolutionary Algorithms for Real-World Applications, pages 245–280. Springer,
2012.
194
Bibliography
A. Delgado, R.M. Jensen, K. Janstrup, T.H. Rose, and K.H. Andersen. A constraint
programming model for fast optimal stowage of container vessel bays. European
Journal of Operational Research, 220(1):251–261, 2012.
G. Desaulniers, J. Desrosiers, and M.M. Solomon. Column generation, volume 5.

Springer, 2005.
M. Di Francesco, T.G. Crainic, and P. Zuddas. The effect of multi-scenario policies

on empty container repositioning. Transportation Research Part E: Logistics and
Transportation Review, 45(5):758–770, 2009.
E.W. Dijkstra. A note on two problems in connexion with graphs. Numerische

mathematik, 1(1):269–271, 1959.
M. Dorigo and L.M. Gambardella. Ant colony system: A cooperative learning

approach to the traveling salesman problem. Evolutionary Computation, IEEE
Transactions on, 1(1):53–66, 1997.
R. Epstein, A. Neely, A. Weintraub, F. Valenzuela, S. Hurtado, G. Gonzalez, A. Beiza,

M. Naveas, F. Infante, F. Alarcon, et al. A strategic empty container logistics
optimization in a major shipping company. Interfaces, 42(1):5–16, 2012.
A.L. Erera, J.C. Morales, and M. Savelsbergh. Robust optimization for empty
repositioning problems. Operations Research, 57(2):468–483, 2009.
EUROGATE Container Terminal Hamburg GmbH. Prices and conditions from 2013,
4th August 2014. URL http://www1.eurogate.de/en/Terminals/Hamburg.
K. Fagerholt. A computer-based decision support system for vessel fleet scheduling–

experience and future research. Decision Support Systems, 37(1):35–47, 2004a.
K. Fagerholt. Designing optimal routes in a liner shipping problem. Maritime Policy

& Management, 31(4):259–268, 2004b.
K. Fagerholt, T.A.V. Johnsen, and H. Lindstad. Fleet deployment in liner shipping:

a case study. Maritime Policy & Management, 36(5):397–409, 2009a.
K. Fagerholt, G. Laporte, and I. Norstad. Reducing fuel emissions by optimizing

speed on shipping routes. Journal of the Operational Research Society, 61(3):
523–529, 2009b.
L. Fan, M. Luo, and W.W. Wilson. Pricing joint products in liner shipping. In
International Forum on Shipping, Ports and Airports (IFSPA) 2012: Transport
Logistics for Sustainable Growth at a New Level, 2012.
195
Bibliography
C.E. Fayle. A short history of the world’s shipping industry. Taylor & Francis US,
2005.
Federal Maritime Commission. The impact of the ocean shipping reform act of 1998.
Technical report, Federal Maritime Commission, 2001.
C.M. Feng and C.H. Chang. Empty container reposition planning for intra-asia
liner shipping. Maritime Policy & Management, 35(5):469–489, 2008.
R.T. Fielding. Architectural styles and the design of network-based software archi-
tectures. PhD thesis, University of California, 2000.
S.D.P. Flapper, J.A. van Nunen, and L.N. Van Wassenhove. Managing closed-loop
supply chains. Springer, 2005.
L. Fleischer and M. Skutella. The quickest multicommodity flow problem. In

Integer Programming and Combinatorial Optimization, volume 2337, pages 36–53.
Springer, 2006.
R.W. Floyd. Algorithm 97: shortest path. Communications of the ACM, 5(6):345,
1962.
A. Forrester, A. Sobester, and A. Keane. Engineering design via surrogate modelling:

a practical guide. John Wiley & Sons, 2008.
Frontline Solvers. Xpress solver engine, 4th August 2014. URL http://www.solver.
com/.
R.A. Gatica and P.A. Miranda. Special issue on latin-american research: A time
based discretization approach for ship routing and scheduling with variable speed.
Networks and Spatial Economics, 11(3):465–485, 2011.
S. Gelareh and D. Pisinger. Fleet deployment, network design and hub location of
liner shipping companies. Transportation Research Part E: Logistics and Trans-
portation Review, 47(6):947–964, 2011.
S. Gelareh, S. Nickel, and D. Pisinger. Liner shipping hub network design in a

competitive environment. Transportation Research Part E: Logistics and Trans-
portation Review, 46(6):991–1004, 2010.
S. Gelareh, R. Neamatian Monemi, P. Mahey, N. Maculan, and D. Pisinger. Single

string planning problem arising in liner shipping industries: A heuristic approach.
Computers & Operations Research, 40(10):2357–2373, 2013.
M. Gendreau and J.Y. Potvin. Handbook of metaheuristics, volume 2. Springer,

2010.
196
Bibliography
M. Gendreau, G. Laporte, and R. Séguin. A tabu search heuristic for the vehicle
routing problem with stochastic demands and customers. Operations Research,
44(3):469–477, 1996.
P.C. Gilmore and R.E. Gomory. A linear programming approach to the cutting-stock
problem. Operations research, 9(6):849–859, 1961.
F. Glover and G.A. Kochenberger. Handbook in Metaheuristics. Kluwer Academic

Publishers, 2003.
M. Goetschalckx and C. Jacobs-Blecha. The vehicle routing problem with backhauls.

C.K. Goh and Y. Tenne. Computational intelligence in expensive optimization

problems, volume 2. Springer, 2010.
J.F. Gonçalves, J.J. de Magalhães Mendes, and M. G.C. Resende. A hybrid genetic
algorithm for the job shop scheduling problem. European Journal of Operational
Research, 167(1):77–95, 2005.
S. Guericke and L. Suhl. Liner network design under consideration of transit times
and partner networks. In International Conference on Logistics and Maritime
Systems 2013, 2013.
S. Guericke and K. Tierney. Liner shipping cargo allocation with service levels and
speed optimization. Technical report, University of Paderborn, DS&OR working
paper no. 1403, Submitted to EJOR, 2014.
Gurobi. Gurobi optimization, 4th August 2014. URL http://www.gurobi.com/.
P. Hansen and N. Mladenović. Variable neighborhood search. Springer, 2003.
J.A. Hartigan and M.A. Wong. Algorithm as 136: A k-means clustering algorithm.
Applied statistics, 28(1):100–108, 1979.
V.C. Hemmelmayr, K.F. Doerner, and R. F. Hartl. A variable neighborhood search

heuristic for periodic routing problems. European Journal of Operational Research,
195(3):791–802, 2009a.
V.C. Hemmelmayr, K.F. Doerner, and R.F. Hartl. A variable neighborhood search
heuristic for periodic routing problems. European Journal of Operational Research,
195(3):791–802, 2009b.
H. Hernández-Pérez and J.J. Salazar-González. The one-commodity pickup-and-

delivery travelling salesman problem. In Combinatorial Optimization—Eureka,
pages 89–104. Springer, 2003.
197
Bibliography
H. Hernández-Pérez and J.J. Salazar-González. Heuristics for the one-commodity

pickup-and-delivery traveling salesman problem. Transportation Science, 38(2):
245–255, 2004.
H. Hernández-Pérez and J.J. Salazar-González. The multi-commodity one-to-one

pickup-and-delivery traveling salesman problem. European Journal of Operational
Research, 196(3):987–995, 2009.
I. Hilger. Konzeption und Implementierung von Variable-Neighborhood-

Suchstrategien für das Netzwerkdesignproblem in der Linienschifffahrt. Master
Thesis, University of Paderborn, 2014.
J.H. Holland. Adaptation in natural and artificial systems: An introductory analysis

with applications to biology, control, and artificial intelligence. A Bradford Book,
1975.
J.H. Holland. Adaptation in natural and artificial systems. MIT Press, Cambridge,
MA, 1992.
C.I. Hsu and Y.P. Hsieh. Direct versus terminal routing on a maritime hub-and-
spoke container network. Journal of Marine Science and Technology, 13(3):209–
217, 2005.
L.M. Hvattum, I. Norstad, K. Fagerholt, and G. Laporte. Analysis of an exact

algorithm for the vessel speed optimization problem. Networks, 82(2):132–135,
2013.
IBM CPLEX. IBM ILOG CPLEX Optimizer, 4th August 2014. URL http://
www-01.ibm.com/software/commerce/optimization/cplex-optimizer.
G.B. Ihde. Transport, Verkehr, Logistik, volume 3. Franz Vahlen, 2001.
A. Imai, K. Sasaki, E. Nishimura, and S. Papadimitriou. Multi-objective simulta-

neous stowage and load planning for a container ship with container rehandle in
yard stacks. European Journal of Operational Research, 171(2):373–389, 2006.
A. Imai, K. Shintani, and S. Papadimitriou. Multi-port vs. hub-and-spoke port calls

by containerships. Transportation Research Part E: Logistics and Transportation
Review, 45(5):740–757, 2009.
International Organization for Standardization. ISO 55.180.10 - general purpose

containers, 4th August 2014. URL http://www.iso.org.
J.J. Jaw, A.R. Odoni, H.N. Psaraftis, and N.HM. Wilson. A heuristic algorithm
for the multi-vehicle advance request dial-a-ride problem with time windows.
Transportation Research Part B: Methodological, 20(3):243–257, 1986.
198
Bibliography
Y. Jin. A comprehensive survey of fitness approximation in evolutionary computa-

tion. Soft computing, 9(1):3–12, 2005.
Y. Jin. Surrogate-assisted evolutionary computation: Recent advances and future

challenges. Swarm and Evolutionary Computation, 1(2):61–70, 2011.
Y. Jin, M. Olhofer, and B. Sendhoff. On evolutionary optimization with approximate

fitness functions. In GECCO, pages 786–793, 2000.
Y. Jin, M. Hüsken, and B. Sendhoff. Quality measures for approximate models in

evolutionary computation. In GECCO, pages 170–173, 2003.
L. Jourdan, C. Dhaenens, and E.G. Talbi. Using datamining techniques to help

metaheuristics: A short survey. In Hybrid Metaheuristics, pages 57–69. Springer,
2006.
B. Kallehauge, J. Larsen, O.B.G. Madsen, and M.M. Solomon. Vehicle routing

problem with time windows. Column Generation, pages 67–98, 2005.
A.J. Keane. Experiences with optimizers in structural design. In Proceedings of the

conference on adaptive computing in engineering design and control, volume 94,
pages 14–27, 1994.
H.J. Kim. A lagrangian heuristic for determining the speed and bunkering port of
a ship. Journal of the Operational Research Society, 65:747–754, 2013.
H.J. Kim and J.G. Kim. A heuristic algorithm for determining the speed and
bunkering port of a ship considering greenhouse gas emissions. In Proceedings
of the Asia Pacific Industrial Engineering and Management Systems Conference,
2012.
J. Kim and B.R. Moon. A genetic algorithm for linear ordering problem using an
approximate fitness evaluation. In Proceedings of the 2014 conference companion
on Genetic and evolutionary computation companion, pages 1461–1462. ACM,
2014.
S. Kirkpatrick. Optimization by simulated annealing: Quantitative studies. Journal

of statistical physics, 34(5-6):975–986, 1984.
K.H. Kjeldsen. Classification of ship routing and scheduling problems in liner

shipping. INFOR: Information Systems and Operational Research, 49(2):139–152,
2011.
K.H. Kjeldsen. Routing and Scheduling in Liner Shipping. PhD thesis, Department
of Economics and Business, Aarhus University Denmark, 2012.
199
Bibliography
F. Klug. Logistics implications of global production networks in car manufacturing.

In Institute for Manufacturing (Hrsg.): Cambridge International Manufacturing
Symposium, Symposium Proceedings, Cambridge, pages 13–24. 2010.
J.W. Konings. The Future of Intermodal Freight Transport: Operations, Design and
Policy. Edward Elgar Publishing, 2008.
R. Kruse, C. Borgelt, F. Klawonn, C. Moewes, M. Steinbrecher, and P. Held. Fun-

damental evolutionary algorithms. In Computational Intelligence, pages 227–274.
Springer, 2013.
J. Kytöjoki, T. Nuortio, O. Bräysy, and M. Gendreau. An efficient variable neigh-

borhood search heuristic for very large scale vehicle routing problems. Computers
& Operations Research, 34(9):2743–2757, 2007a.
J. Kytöjoki, T. Nuortio, O. Bräysy, and M. Gendreau. An efficient variable neigh-

borhood search heuristic for very large scale vehicle routing problems. Computers
& Operations Research, 34(9):2743–2757, 2007b.
K.K. Lai, K. Lam, and W.K. Chan. Shipping container logistics and allocation.
Journal of the Operational Research Society, 46(6):687–697, 1995.
E.L. Lawler, J.K. Lenstra, A.H.G.R. Kan, and D.B. Shmoys. The traveling salesman
problem: a guided tour of combinatorial optimization, volume 3. Wiley, 1985.
M. Levinson. The box: How the shipping container made the world smaller and the
world economy bigger. Princeton University Press, 2006.
T.C. Lirn, H.A. Thanopoulou, M.J. Beynon, and A.K.C. Beresford. An application
of ahp on transhipment port selection: a global perspective. Maritime Economics
& Logistics, 6(1):70–91, 2004.
Lloyd’s List. The chinese conundrum, 30th July 2014. URL http://www.
lloydslist.com/ll/sector/containers/article445059.ece.
J. Little. Models and managers: The concept of a decision calculus. Management

Science, 16(8):466–485, 1970.
C. Liu, Z. Jiang, F. Chen, X. Liu, L. Liu, and Z. Xu. Empty container repositioning-
a review. In Intelligent Control and Automation (WCICA), 2010 8th World
Congress on, pages 3028–3033. IEEE, 2010.
H.A. Lu, C.W. Chu, P.Y. Che, et al. Slot allocation planning for an alliance service
with ship fleet sharing. Asia Pacific Management Reviews, 15:325–339, 2010.
200
Bibliography
M.E. Lübbecke and J. Desrosiers. Selected topics in column generation. Operations

Research, 53(6):1007–1023, 2005.
Y.H.V. Lun, K.H. Lai, and T.C.E. Cheng. Shipping and Logistics Management.
Springer, 2010.
Maersk Line. The world’s largest ship, 4th August 2014. URL http://www.
worldslargestship.com/facts/the-worlds-largest-ship-2/.
M. Malchow and A. Kanafani. A disaggregate analysis of factors influencing port

selection. Maritime Policy & Management, 28(3):265–277, 2001.
D.C. Mattfeld. The Management of Transshipment Terminals:Decision Support for

Terminal Operations in Finished Vehicle Supply Chains. Springer, 2006.
Q. Meng and S. Wang. Liner shipping service network design with empty container
repositioning. Transportation Research Part E: Logistics and Transportation
Review, 47(5):695–708, 2011a.
Q. Meng and T. Wang. A scenario-based dynamic programming model for multi-

period liner ship fleet planning. Transportation Research Part E: Logistics and
Transportation Review, 47(4):401 – 413, 2011b.
Q. Meng, S. Wang, H. Andersson, and K. Thun. Containership routing and schedul-

ing in liner shipping: overview and future research directions. Transportation
Science, 2013.
J. Meyer, R. Stahlbock, and S. Voß. Slow steaming in container shipping. In

System Science (HICSS), 2012 45th Hawaii International Conference on, pages
1306–1314. IEEE, 2012.
U. Mhaky and C.Y. Lee. Service level assignment and container routing for liner
shipping service networks. In Proceedings of the International Multi-Conference
of Engineers and Computer Scientists, volume 2, 2014.
Z. Michalewicz and D.B. Fogel. How to Solve It: Modern Heuristics, volume 2.
Springer, 2004.
R. Midoro and A. Pitto. A critical evaluation of strategic alliances in liner shipping.

Maritime Policy & Management, 27(1):31–40, 2000.
R. Midoro, E. Musso, and F. Parola. Maritime liner shipping and the stevedor-
ing industry: market structure and competition strategies. Maritime Policy &
Management, 32(2):89–106, 2005.
201
Bibliography
C.E. Miller, A.W. Tucker, and R.A. Zemlin. Integer programming formulation of
traveling salesman problems. Journal of the ACM (JACM), 7(4):326–329, 1960.
S. Mitrovic-Minic and G. Laporte. The pickup and delivery problem with time
windows and transshipment. Infor-Information Systems and Operational Research,
44(3):217–228, 2006.
N. Mittal. Regional empty marine container management. ProQuest, 2008.
N. Mladenović and P. Hansen. Variable neighborhood search. Computers & Opera-

tions Research, 24(11):1097–1100, 1997.
P. Moscato et al. On evolution, search, optimization, genetic algorithms and martial

arts: Towards memetic algorithms. Caltech concurrent computation program, C3P
Report, 826:1989, 1989.
J. Mulder and R. Dekker. Methods for strategic liner shipping network design.
H. Nazif and L.S. Lee. Optimised crossover genetic algorithm for capacitated vehicle
routing problem. Applied Mathematical Modelling, 36(5):2110–2117, 2012.
G.L. Nemhauser and L.A. Wolsey. Integer and combinatorial optimization. Wiley-
Interscience series in discrete mathematics, 1999.
E.K. Norlund and I. Gribkovskaia. Reducing emissions through speed optimization

in supply vessel operations. Transportation Research Part D: Transport and
Environment, 23:105–113, 2013.
I. Norstad, K. Fagerholt, and G. Laporte. Tramp ship routing and scheduling with
speed optimization. Transportation Research Part C: Emerging Technologies,
2010.
I. Norstad, K. Fagerholt, L.M. Hvattum, H.S. Arnulf, and A. Bjørkli. Maritime

fleet deployment with voyage separation requirements. Flexible Services and
Manufacturing Journal, pages 1–20, 2013.
T. Notteboom and P. Cariou. Fuel surcharge practices of container shipping lines:

Is it about cost recovery or revenue-making. In Proceedings of the International
Association of Maritime Economists Conference, Copenhagen, pages 24–26, 2009.
T. Notteboom and J.P. Rodrigue. Re-assessing port-hinterland relationships in

the context of global commodity chains. Ports, Cities and Global Supply Chains,
pages 51–66, 2007.
202
Bibliography
T. Notteboom and J.P. Rodrigue. Containerisation, box logistics and global supply
chains: The integration of ports and liner shipping networks. Maritime Economics
& Logistics, 10:152–174, 2008.
T.E. Notteboom. The time factor in liner shipping services. Maritime Economics
& Logistics, 8(1):19–39, 2006.
T.E. Notteboom and B. Vernimmen. The effect of high fuel costs on liner service
configuration in container shipping. Journal of Transport Geography, 17(5):325 –
337, 2009.
A. Orden. The transhipment problem. Management science, 2(3):276–285, 1956.
I.H. Osman and G. Laporte. Metaheuristics: A bibliography. Annals of Operations
Research, 63(5):511–623, 1996.
D. Pacino. An lns approach for container stowage multi-port master planning. In
Computational Logistics, pages 35–44. Springer, 2013.
M. Padberg. Approximating seperable nonlinear functions via mixed zero-one
programs. Operations Research Letters, 27(1):1–5, 2000.
P.M. Panayides and D.W. Song. Maritime logistics as an emerging discipline.
P.M. Panayides and R. Wiedmer. Strategic alliances in container liner shipping.
Research in Transportation Economics, 2011.
V. Papapanagiotou, D. Weyland, R. Montemanni, and L.M. Gambardella. A
sampling-based approximation of the objective function of the orienteering prob-
lem with stochastic travel and service times. Lecture Notes in Management
Science, 5:143–152, 2013.
V. Papapanagiotou, R. Montemanni, and L.M. Gambardella. Objective function
evaluation methods for the orienteering problem with stochastic travel and service
times. Journal of Applied Operational Research, 6(1):16–29, 2014.
U. Pape. Ein heuristisches verfahren zur ermittlung von fahrplänen in der container-
schiffahrt. Operations-Research-Spektrum, 1(3):169–181, 1980.
S.N. Parragh, K.F. Doerner, and R.F. Hartl. Variable neighborhood search for the
dial-a-ride problem. Computers & Operations Research, 37(6):1129–1138, 2010.
C.E.M. Plum, D. Pisinger, J.J. Salazar-González, and M.M. Sigurd. The multi-
commodity one-to-one pickup-and-delivery traveling salesman problem with path
duration limits. In Proceedings of the International MultiConference of Engineers
and Computer Scientists, volume 2, 2012.
203
Bibliography
C.E.M. Plum, D. Pisinger, J.J. Salazar-González, and M.M. Sigurd. Single liner
shipping service design. Computers & Operations Research, 2013a.
C.E.M. Plum, D. Pisinger, and M.M. Sigurd. A service flow model for the liner
shipping network design problem. European Journal of Operational Research,
2013b.
O. Polat. Designing Liner Shipping Feeder Service Networks in the New Era of
Mega Containerships. PhD thesis, Technical University Berlin, 2013.
B.J. Powell and A.N. Perkins. Fleet deployment optimization for liner shipping:
an integer programming model. Maritime Policy & Management, 24(2):183–192,
1997.
H.N. Psaraftis and C.A. Kontovas. Speed models for energy-efficient maritime trans-
portation: A taxonomy and survey. Transportation Research Part C: Emerging
Technologies, 26:331–351, 2013.
X. Qi and D.P. Song. Minimizing fuel emissions by optimizing vessel schedules

in liner shipping with uncertain port times. Transportation Research Part E:
Logistics and Transportation Review, 48(4):863–880, 2012.
K. Rana and RG Vickson. A model and solution algorithm for optimal routing of
a time-chartered containership. Transportation Science, 22(2):83, 1988.
K. Rana and RG Vickson. Routing container ships using lagrangean relaxation and
decomposition. Transportation Science, 25(3):201, 1991.
A. Ratle. Accelerating the convergence of evolutionary algorithms by fitness land-

scape approximation. In Parallel Problem Solving from Nature—PPSN V, pages
87–96. Springer, 1998.
C.R. Reeves. Modern heuristic techniques for combinatorial problems. John Wiley
& Sons, Inc., 1993.
L.B. Reinhardt and D. Pisinger. A branch and cut algorithm for the container
shipping network design problem. Flexible Services and Manufacturing Journal,
24(3):1–26, 2010.
L.B. Reinhardt, B. Kallehauge, A. Nørrelund, and A. Olsen. Network Design Models

for Container Shipping. Centre for Traffic and Transport, Technical University
of Denmark, 2007.
J. Renaud, G. Laporte, and F.F. Boctor. A tabu search heuristic for the multi-depot
vehicle routing problem. Computers & Operations Research, 23(3):229–235, 1996.
204
Bibliography
Drewry Maritime Research. Reefer Shipping Market Annual Review and Forecast
2013/14. Drewry, 2013.
J.P. Rodrigue. Factors impacting North American freight distribution in view of

the panama canal expansion. Van Horne Institute, 2010.
J.P. Rodrigue. The Geography of Transport Systems, volume 3. Routledge, 2013.
D. Ronen. The effect of oil price on the optimal speed of ships. Journal of the
Operational Research Society, 33(11):1035–1040, 1982.
D. Ronen. Short-term scheduling of vessels for shipping bulk or semi-bulk commodi-

ties originating in a single area. Operations Research, pages 164–173, 1986.
D. Ronen. Ship scheduling: The last decade. European Journal of Operational

Research, 71(3):325–333, 1993.
N Saeidi, H Jafari, A Ameli, and N Zaersoleymani. Container repositioning man-

agement in liner shipping industry. Management Science Letters, 3(6):1795–1804,
2013.
M. Şahin, G. Çavuşlar, T. Öncan, G. Şahin, and D. Tüzün Aksu. An efficient

heuristic for the multi-vehicle one-to-one pickup and delivery problem with split
loads. Transportation Research Part C: Emerging Technologies, 27:169–188, 2012.
Y. Sano and H. Kita. Optimization of noisy fitness functions by means of genetic

algorithms using history of search. In Parallel Problem Solving from Nature PPSN
VI, pages 571–580. Springer, 2000.
A. Schieck. Internationale Logistik: Objekte, Prozesse und Infrastrukturen gren-

züberschreitender Güterströme. Oldenbourg Verlag, 2008.
G. Schmidt and W.E. Wilhelm. Strategic, tactical and operational decisions in

multi-national logistics networks: a review and discussion of modelling issues.
International Journal of Production Research, 38(7):1501–1523, 2000.
A. Schönknecht. Maritime Containerlogistik: Leistungsvergleich von Containerschif-

fen in intermodalen Transportketten. Springer, 2009.
A. Shahrokhi and A. Jahangirian. A surrogate assisted evolutionary optimization

method with application to the transonic airfoil design. Engineering Optimization,
42(6):497–515, 2010.
J.S. Shang and C.K. Cuff. Multicriteria pickup and delivery problem with transfer
opportunity. Computers & industrial engineering, 30(4):631–645, 1996.
205
Bibliography
R. Sharda, S.H. Barr, and J.C. MCDonnell. Decision support system effectiveness:
a review and an empirical test. Management science, 34(2):139–159, 1988.
W.S. Shen and C.M. Khoong. A DSS for empty container distribution planning.
Decision Support Systems, 15(1):75–82, 1995.
X. Sheng, L.H. Lee, and E.P. Chew. Dynamic determination of vessel speed and
selection of bunkering ports for liner shipping under stochastic environment. OR
Spectrum, 36(2):455–480, 2013.
K. Shintani, A. Imai, E. Nishimura, and S. Papadimitriou. The container shipping

network design problem with empty container repositioning. Transportation
Research Part E: Logistics and Transportation Review, 43(1):39–59, 2007.
Shipping Watch. Analyst firms examine cartel ban, 1st August 2014. URL http:
//shippingwatch.com/carriers/article5709489.ece.
H.K. Singh, T. Ray, and W. Smith. Surrogate assisted simulated annealing (sasa) for
constrained multi-objective optimization. In Evolutionary Computation (CEC),
2010 IEEE Congress on, pages 1–8. IEEE, 2010.
B. Slack. Containerization, inter-port competition, and port selection. Maritime

Policy and Management, 12(4):293–303, 1985.
D. Song, J. Zhang, J. Carter, T. Field, J. Marshall, J. Polak, K. Schumacher, P. Sinha-

Ray, and J. Woods. On cost-efficiency of the global container shipping network.
D.P. Song and J. Carter. Empty container repositioning in liner shipping. Maritime
Policy & Management, 36(4):291–307, 2009.
D.P. Song and J.X. Dong. Effectiveness of an empty container repositioning policy
with flexible destination ports. Transport Policy, 18(1):92–101, 2011.
D.P. Song and J.X. Dong. Long-haul liner service route design with ship deployment
and empty container repositioning. Transportation Research Part B: Methodolog-
ical, 55:188–211, 2013.
M. Srinivas and L.M. Patnaik. Genetic algorithms: A survey. Computer, 27(6):

17–26, 1994.
M. Stopford. Maritime economics, volume 3. Taylor & Francis, 2009.
Suez Canal Authority. Canal characteristics, 1st July 2014. URL http://www.
suezcanal.gov.eg/sc.aspx?show=12.
206
Bibliography
L. Suhl and T. Mellouli. Optimierungssysteme: Modelle, Verfahren, Software, An-

wendungen, volume 3. Springer, 2013.
C. Sys. Is the container liner shipping industry an oligopoly? Transport policy, 16

(5):259–270, 2009.
H. Takagi and H. Iba. Interactive evolutionary computation. New Generation

Computing, 23(2):113–114, 2005.
A.S. Tasan and M. Gen. A genetic algorithm based approach to vehicle rout-
ing problem with simultaneous pick-up and deliveries. Computers & Industrial
Engineering, 62(3):755–761, 2012.
K. Tierney and R.M. Jensen. The liner shipping fleet repositioning problem with
cargo flows. In Computational Logistics, volume 4555, pages 1–16. Springer, 2012.
K. Tierney, B. Áskelsdóttir, R. M. Jensen, and D. Pisinger. Solving the liner shipping

fleet repositioning problem with cargo flows. Transportation Science, 2014.
P. Toth and D. Vigo. An exact algorithm for the vehicle routing problem with
backhauls. Transportation science, 31(4):372–385, 1997.
P. Toth and D. Vigo. Models, relaxations and exact approaches for the capacitated
vehicle routing problem. Discrete Applied Mathematics, 123(1):487–512, 2002.
N.K. Tran and H.D. Haasis. Literature survey of network optimization in container
liner shipping. Flexible Services and Manufacturing Journal, pages 1–41, 2013.
E. Turban and J.E. Aronson. Decision Support Systems and Intelligent Systems,
volume 7. Prentice Hall, 2007.
UN. United nations commodity trade 2012, 4th August 2014. URL http:
//comtrade.un.org.
UN Conference on Trade and Development. Review of maritime transport 2013.

Technical report, United Nations, 2013.
P. Vansteenwegen, W. Souffriau, and D.V. Oudheusden. The orienteering problem:

A survey. European Journal of Operational Research, 209(1):1–10, 2011.
C. Vilhelmsen, R. Lusby, and J. Larsen. Tramp ship routing and scheduling with
integrated bunker optimization. EURO Journal on Transportation and Logistics,
pages 1–33, 2013.
S. Wang. Essential elements in tactical planning models for container liner shipping.
Transportation Research Part B: Methodological, 54:84–99, 2013.
207
Bibliography
S. Wang. A novel hybrid-link-based container routing model. Transportation Re-

search Part E: Logistics and Transportation Review, 61:165–175, 2014.
S. Wang and Q. Meng. Schedule design and container routing in liner shipping.
Transportation Research Record: Journal of the Transportation Research Board,
2222(-1):25–33, 2011.
S. Wang and Q. Meng. Liner ship route schedule design with sea contingency time
and port time uncertainty. Transportation Research Part B: Methodological, 46
(5):615–633, 2012a.
S. Wang and Q. Meng. Liner ship fleet deployment with container transshipment
operations. Transportation Research Part E: Logistics and Transportation Review,
48(2):470–484, 2012b.
S. Wang and Q. Meng. Sailing speed optimization for container ships in a liner
shipping network. Transportation Research Part E: Logistics and Transportation
Review, 48(3):701–714, 2012c.
S. Wang and Q. Meng. Robust schedule design for liner shipping services. Trans-
portation Research Part E: Logistics and Transportation Review, 48(6):1093–1106,
2012d.
S. Wang and Q. Meng. Liner shipping network design with deadlines. Computers
& Operations Research, 41:140–149, 2014.
S. Wang, Q. Meng, and Z. Liu. Bunker consumption optimization methods in ship-
ping: a critical review and extensions. Transportation Research Part E: Logistics
and Transportation Review, 53:49–62, 2013a.
S. Wang, Q. Meng, and Z. Liu. Containership scheduling with transit-time-sensitive
container shipment demand. Transportation Research Part B: Methodological, 54:
68–83, 2013b.
S. Wang, Q. Meng, and Z. Sun. Container routing in liner shipping. Transportation
Research Part E: Logistics and Transportation Review, 49(1):1–7, 2013c.
D. Weyland, R. Montemanni, and L.M. Gambardella. A metaheuristic framework
for stochastic combinatorial optimization problems based on gpgpu with a case
study on the probabilistic traveling salesman problem with deadlines. Journal of
Parallel and Distributed Computing, 73(1):74–85, 2013a.
D. Weyland, R. Montemanni, and L.M. Gambardella. An improved heuristic for
the probabilistic traveling salesman problem with deadlines based on gpgpu. In
Computer Aided Systems Theory-EUROCAST 2013, pages 332–339. Springer,
2013b.
208
Bibliography
V. Windeck. Environmental routing. In A Liner Shipping Network Design, pages

39–78. Springer, 2013.
L.A. Wolsey. Integer programming. IIE Transactions, 32(273-285):2–58, 2000.
P. Wong and C. Bamford. A study of market structure in liner shipping under the
influence of government policies. In Proceedings of 19th International Business
Research Conference, 2012.
World Port Source. World port source, 3rd August 2014. URL http://www.
worldportsource.com.
World Shipping Council. Container ship design, 1st August 2014. URL
http://www.worldshipping.org/about-the-industry/liner-ships/
container-ship-design.
www.sea-distances.com. Sea distances - voyage calculator (nautical miles), 4th

August 2014. URL http://sea-distances.com/.
www.sea-rates.com. Transit time/ distance calculator, 3rd August 2014. URL

http://www.searates.com/reference/portdistance/.
Y. Xu, G. Lin, and H. Sun. Accounting for the china–us trade imbalance: An
ownership-based approach. Review of International Economics, 18(3):540–551,
2010.
S. Yan, C.Y. Chen, and S.C. Lin. Ship scheduling and container shipment planning
for liners in short-term operations. Journal of marine science and technology, 14
(4):417–435, 2009.
Z. Yao, S.H. Ng, and L.H. Lee. A study on bunker fuel management for the shipping
liner services. Computers & Operations Research, 39(5):1160–1172, 2012.
H. Yi. Topics in Contract Pricing and Spot Markets. PhD thesis, Georgia Institute
of Technology, 2008.
S. Zurheide and K. Fischer. A revenue management slot allocation model for liner
shipping networks. Maritime Economics & Logistics, 14(3):334–361, 2012.
209
List of Figures
2.1. Selected planning problems arising in the (liner) shipping industry

in the context of service planning and operating. . . . . . . . . . . . 10
2.2. Example liner network with three services. . . . . . . . . . . . . . . 12
2.3. Main global maritime shipping routes, Rodrigue (2013). . . . . . . . 14
2.4. Liner services with different route types. . . . . . . . . . . . . . . . 15
2.5. Round trip of example service with approximate leg distances. . . . 18
2.6. Duration of example service with two vessels serving the ports on a
weekly basis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.7. Example transit time between Rotterdam and Shanghai using service
A (black) and B (gray). . . . . . . . . . . . . . . . . . . . . . . . . 21
2.8. East-West Carrier Alliances 2013 . . . . . . . . . . . . . . . . . . . 23
2.9. Example deadweight scale for a Panamax vessel. . . . . . . . . . . . 26
2.10. Trade imbalances on the main trades . . . . . . . . . . . . . . . . . 27
2.11. Empty container repositioning using depots . . . . . . . . . . . . . 28
2.12. Voyage costs per vessel type . . . . . . . . . . . . . . . . . . . . . . 30
2.13. Bunker consumption in tons per day for different vessel types, (Not-
teboom and Cariou (2009)). . . . . . . . . . . . . . . . . . . . . . . 31
2.14. Example bunker profile for a 4300 TEU vessel. The y-axis is scaled
logarithmically to the basis of 10. . . . . . . . . . . . . . . . . . . . 32
2.15. Bunker Index, aggregated data from Bunker Index (2014). . . . . . 33
2.16. Monthly BIX changes, data from Bunker Index (2014). . . . . . . . 34
4.1. Layered service representation of a butterfly rotation to correctly

handle transhipments at port p2 . . . . . . . . . . . . . . . . . . . . 64
4.2. All edges required to create arbitrary valid route types for three ports. 64
4.3. Resource groups for a vessel type. . . . . . . . . . . . . . . . . . . . 65
4.4. Container path for a cargo flow c1 incident to two services. . . . . . 73
4.5. Flat network structure with two liner services . . . . . . . . . . . . 78
4.6. Network structure for the pricing problem . . . . . . . . . . . . . . 78
4.7. Graphical illustration of the variable values in the optimal solution
y ∗ (top) and the new constructed solution y 0 (bottom). . . . . . . . 84
4.8. Distribution of the cumulative cargo flow quantity . . . . . . . . . . 89
4.9. Average sum of services and legs for random networks per LINER-LIB
instance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
4.10. Runtime results for solving the cargo allocation to optimality using
Gurobi 5.6 with the arc-flow formulation. . . . . . . . . . . . . . . . 92
211
List of Figures
4.11. Best average utilization per network (BAU %) and average served
cargo flow percentage (CF %). . . . . . . . . . . . . . . . . . . . . . 92
4.12. Average speed in knots and duration per port call in hours. . . . . . 93
4.13. Average runtime in seconds of the parallel and non parallel imple-
mentation for networks of different sizes for the LINER-LIB instances. 94
4.14. Comparison of the arc-flow and path-flow formulation . . . . . . . . 96
4.15. Gap in percent for varied cargo flow percentages and different cargo
flow selection strategies (for the Pacific and WAF instance). . . . . 98
4.16. Gap in percent for varied cargo flow percentages and different cargo
flow selection strategies (for the EuropeAsia and Mediterranean in-
stance). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.17. Runtime improvement when solving Column Generation heuristically
by limiting the cargo flow amount according to different strategies. . 100
4.18. Average effects of the different bunker cost linearization support
points on the objective function and the runtime. . . . . . . . . . . 101
4.19. Gap and runtime improvement when terminating the column gener-
ation solution approach prematurely. . . . . . . . . . . . . . . . . . 102
5.1. Invalid port rotation for one service due to existing subtours. . . . . 118
5.2. Example incumbents for the Baltic instance using not more than
three services, two layers and two vessels per service. . . . . . . . . 121
5.3. Gap in the Baltic instance . . . . . . . . . . . . . . . . . . . . . . . 122
5.4. Gap in the Baltic instance per service upper bound allowing two
layers and a maximum of two vessels per service. . . . . . . . . . . . 123
5.5. Best bounds found for the Baltic LINER-LIB instance within 12 hours.123
5.6. Gap in the WAF instance per service upper bound allowing one layer
and a maximum of two vessels per service. . . . . . . . . . . . . . . 124
5.7. Best bounds found for the WAF LINER-LIB instance within 24 hours.125
5.8. Decomposition of the overall liner shipping network design problem. 128
5.9. Durations resulting from the cargo allocation problem, for example
given at one port and one leg. . . . . . . . . . . . . . . . . . . . . . 129
5.10. Incident service legs for a transit time requirement between port p1
and p4 are marked gray. Service legs that have to be adjusted are
marked black. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.11. Service speed adjustment . . . . . . . . . . . . . . . . . . . . . . . . 134
5.12. Random network that uses pendulum services. . . . . . . . . . . . . 135
5.13. Clustering with three regions used for construction heuristic. The
reference ports in each region are marked bold. . . . . . . . . . . . . 136
5.14. Convergence of the evolutionary algorithm . . . . . . . . . . . . . . 145
5.15. Convergence of variable neighborhood search in selected LINER-LIB
instances. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
212
List of Figures
5.16. Concept for a heuristic meta control for population and individual
based heuristics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5.17. Fitness convergence in the evolutionary algorithm on the WAF LINER-
LIB instance using a constant surrogate with different accuracies. . 152
5.18. Average real fitness for EA surrogate . . . . . . . . . . . . . . . . . 153
5.19. Fitness convergence using different surrogate configurations for the
FPS strategy in the evolutionary algorithm for the WAF instance. . 154
5.20. Fitness convergence using different surrogate configurations for the
FPS strategy in the evolutionary algorithm for the Mediterranean
instance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
5.21. EA with FGS surrogate strategy in the WAF instance . . . . . . . . 155
5.22. EA with FGS surrogate strategy in the Mediterranean instance . . 156
5.23. EA with BRS surrogate strategy in the WAF instance . . . . . . . . 156
5.24. EA with BRS surrogate strategy in the Mediterranean instance . . 157
5.25. Fitness convergence using different static surrogates in the VND local
first iteration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
second iteration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
after the first iteration. . . . . . . . . . . . . . . . . . . . . . . . . . 160
after the second iteration. . . . . . . . . . . . . . . . . . . . . . . . 161
search in the Pacific LINER-LIB instance. CG is terminated after
the first iteration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
5.30. Average best profit of three VNS runs for different bunker prices per
ton. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
5.31. Cost components and revenue in the best networks for different
bunker prices per ton. . . . . . . . . . . . . . . . . . . . . . . . . . 163
5.32. Structural network changes with different bunker prices per ton. . . 164
5.33. Served cargo flows and average speed with different bunker prices per
ton. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
5.34. Benefit of adjusting network for different scenarios that is optimized
for a bunker price of 600 US$ per metric ton. . . . . . . . . . . . . . 166
5.35. Adjusted optimization process to handle practical constraints for the
best liner shipping network found. . . . . . . . . . . . . . . . . . . . 169
5.36. Distribution of the average cost structure in the current liner network
at hand. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
213
List of Figures
6.1. Components of a decision support system . . . . . . . . . . . . . . . 176

6.2. User processes for the liner shipping network planning. . . . . . . . 178
6.3. Client-server communication of the decision support system. . . . . 179
6.4. Graphical user interface for the decision support system. . . . . . . 182
6.5. Evaluating an existing liner shipping network. . . . . . . . . . . . . 182
6.6. Modification of the port rotation of an existing liner service. . . . . 183
6.7. Changing a liner service’s vessel type and reevaluating the whole
network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
C.1. Distribution of the cargo flow quantity (cum. Q), revenue (cum. R)
and quantity * revenue (cum. Q*R) for the LINER-LIB Baltic instance.225
and quantity * revenue (cum. Q*R) for the LINER-LIB Mediter-
ranean instance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
and quantity * revenue (cum. Q*R) for the LINER-LIB Pacific instance.226
and quantity * revenue (cum. Q*R) for the LINER-LIB WorldSmall
instance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
and quantity * revenue (cum. Q*R) for the LINER-LIB Europe-Asia
instance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
and quantity * revenue (cum. Q*R) for the LINER-LIB WorldLarge
instance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
C.7. Gap in percent for varied cargo flow amount and different cargo flow
selection strategies (for the Baltic, WorldSmall and WorldLarge). . 227
C.8. Effects of the support points on the objective function and the run-
time (for the Baltic, WorldSmall and WorldLarge instances). . . . . 227
C.9. Gap and runtime improvement when terminating the Column Gen-
eration solution approach prematurely (for the Baltic, WorldSmall
and WorldLarge instance). . . . . . . . . . . . . . . . . . . . . . . . 228
D.1. Baltic LINER-LIB 2012 instance with maximum of one vessel per
service, no complex route type and no transit times. . . . . . . . . . 229
D.2. Baltic LINER-LIB 2012 instance with maximum of two vessel per
service, no complex route type and no transit times. . . . . . . . . . 229
D.3. Baltic LINER-LIB 2012 instance with maximum of two vessel per
service, complex route types allowed and no transit times. . . . . . 230
D.4. Baltic LINER-LIB 2012 instance with maximum of one vessel per
service, no complex route type but transit times. . . . . . . . . . . . 230
214
List of Figures
D.5. Convergence of the evolutionary algorithm for LINER-LIB instances

using the clustering CX(CL), the cycle CX(CY) and informed cycle
CX(IC) crossover methods (averaged over five runs). . . . . . . . . 230
D.6. Convergence of Evolutionary Algorithm in different LINER-LIB in-
stances using the roulette (RS) and binary tournament (BT) selection
for the informed cycle crossover (IC) method (averaged over five runs).231
D.7. Convergence of Evolutionary Algorithm in different LINER-LIB in-
stances using different population sizes (10, 20, 50 individuals per
generation) for the informed cycle crossover method with a binary
tournament strategy (averaged over five runs). . . . . . . . . . . . . 231
D.8. Convergence of Evolutionary Algorithm in the WAF instance using
different initial pendulum percentage (10%, 50%, 100%) for the in-
formed cycle crossover method with a binary tournament strategy
and a population size of 50 (averaged over five runs). . . . . . . . . 232
D.9. Convergence of Evolutionary Algorithm in the WAF instance us-
ing different elitism ranges (10%, 20%, 50%) for the informed cycle
crossover method with a binary tournament strategy, a population
size of 50 and 10% pendulum services in the initial population (av-
eraged over five runs). . . . . . . . . . . . . . . . . . . . . . . . . . 232
D.10.Convergence of Evolutionary Algorithm in different LINER-LIB in-
stances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
E.1. Visual process to calculate way points between ports for the user
interface. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
215
List of Tables
1.1. Largest liner operators . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2. Operated capacity of the top 15 carriers . . . . . . . . . . . . . . . 4
2.1. Typical container vessel types . . . . . . . . . . . . . . . . . . . . . 13

2.2. Selection of frequently used container types. . . . . . . . . . . . . . 17
2.3. Example slot sharing agreement between three carriers. . . . . . . . 24
3.1. Classification of pickup and delivery problems. . . . . . . . . . . . . 45

3.2. Literature on the liner shipping network design problem . . . . . . . 49
3.3. Literature on the CAP within the LSNDP and the ECRP . . . . . . 53
3.4. Literature on the speed optimization . . . . . . . . . . . . . . . . . 56
4.6. Dual variables and ranges in the cargo allocation path flow formulation. 79
4.7. LINER-LIB 2012 instance information . . . . . . . . . . . . . . . . 88
4.8. Container paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.4. Resulting model size using different input sets (including transit times).120
5.5. Subproblems of the cargo allocation and liner shipping network design
problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.9. Neighborhood Structures used in Variable Neighborhood Descending. 142
5.10. Parameters for the hybrid evolutionary algorithm . . . . . . . . . . 144
5.11. Accuracy of surrogate evaluation . . . . . . . . . . . . . . . . . . . 158
5.12. Average BPU, WPU and BAU for networks optimized to different
bunker prices per ton. . . . . . . . . . . . . . . . . . . . . . . . . . 166
5.13. Approximate size of the problem instance used for the realistic case
study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
5.14. Current average utilization . . . . . . . . . . . . . . . . . . . . . . . 170
5.15. Average service utilization for all legs . . . . . . . . . . . . . . . . . 171
5.16. Properties of the current, an intermediate and two alternative networks.172
5.17. Changes in the profit, revenue and cost structure for the intermediate
solution and the two alternative networks. . . . . . . . . . . . . . . 173
D.1. Detailed accuracy of surrogate evaluation using the Q*R strategy . 234
D.2. Detailed accuracy of surrogate evaluation using the cum. Q*R strategy235
217
List of Algorithms
1. Delayed Column Generation Procedure . . . . . . . . . . . . . . . . . 38

2. Genetic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3. Variable neighborhood descent . . . . . . . . . . . . . . . . . . . . . 40
4. Basic variable neighborhood search . . . . . . . . . . . . . . . . . . . 41
5. Overview of the delayed column generation for the integrated cargo

allocation problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6. Speed adjustment process to hold transit times. . . . . . . . . . . . . 131

7. Clustering heuristic to create an initial liner shipping network design. 137
8. Hybrid Evolutionary Algorithm for the liner shipping network design
problem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
9. Variable Neighborhood Search for the liner shipping network design
problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
10. CreateLayeredLegs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
219
Appendix A.
Glossary
Bunker Name for vessel fuel
Cargo flow A container demand between two specific ports

with a specific equipment type, revenue and re-
source utilization (such as weight)
Deadweight tonnage (tdw) Measure how many tons a vessel can carry safely,
including load, fuel, crew, ballast water
Feeder service A service used for local or coastal transport and

usually connected with main services
Forty foot equivalent unit Standardized container type of 12.19m length,

(FEU) 2.59m height and 2.44m width
Knot (kn) Unit of speed, equals one nautical mile per hour
Leg Connection between two ports
Lightship draft Lightship draft refers to a vessel’s draft without

any payload
Liner service Fleet of vessels operating between fixed ports on

a regular basis. Consists of a port rotation and a
vessel deployment
Long-haul operations Transocean (such as transatlantic or transpacific)

transportation
Longshoreman Port employee to load and unload vessels
221
Appendix A. Glossary
Nautical mile (nm) Nautical distance unit, approximately 1.852 kilo-

meter
Port call Vessels call at different ports within a liner service
Port duration Duration of a vessel’s port call
Port rotation Sequence of ports that perform a round trip. A

port rotation can call a port more than once per
round trip
Proforma schedule Weekly port schedule with fixed arrival and depar-
ture days offered to the customer by the liner car-
rier
Shipper The person or company who is usually the supplier

or owner of commodities shipped
Short-haul operations Local or coastal transportation
Subtour Round trip that includes two disconnected sub

round trips
Trade A commercial link between two or more markets
Transit time The duration to transport cargo from one port to

another using the underlying network
Twenty foot equivalent unit Standardized container type of 6.06m length, 2.59m
(TEU) height and 2.44m width
Vessel deployment Part of a liner service that specifies the vessel type
and the number of vessels
222
Appendix B.
Transformation Algorithm for the Layered Network
Structure
With the help of Algorithm 32, arbitrary liner services (defined by their leg list
L) can be transformed in O(|L|) since the algorithm iterates through all legs and
performs constant time operations inside the f or loop. Note that o(l), d(l) gives the
origin and destination port of the leg l.
Algorithm 32 works as follows: first, the result list LA and current origin and
destination network layer lo , ld and a temporary set LT , used to check whether a
port is visited twice on one layer, is initialized. Afterwards all legs L of the service
are iterated. Three different cases must be distinguished to create layered legs from
leg i: first, the last visited leg’s layer on the round trip must be connected to the
first leg’s origin layer, here zero. If the destination port of leg l is already visited on
the current layer (see line 14), the leg connects the next layer with the current and
a layer crossing leg lC is created and inserted into the return list. Afterwards, the
origin is set to the target to continue with legs on the same layer. In the last case
(see line 26), a new leg is created on the layer indices lo , ld and added to the result
list. Finally the result list is returned.
223
Appendix B. Transformation Algorithm for the Layered Network Structure
Algorithm 10: CreateLayeredLegs

Input: An ordered list of legs,
L = {(i, i + 1) : i, i + 1 ∈ P }, (i, j), (k, l) ∈ L : j = k
Output: The list of layered legs LL = {(i, j, l, l0 ) : (i, j) ∈ L}
1 L = ∅;
A
2 lo = ld = 0;
3 L = ∅;
T
4 for i = 0; i < |L|; i + + do

5 l = Li ;
6 if i = 0 then
7 LT ← LT ∪ o(l);
8 end
9 if i = |Ls | − 1 then
10 lB = (o(l), d(l), ld , 0);
11 LA ← LA ∪ {lB };
12 continue;
13 end
14 if LT \{d(l)} = 6 ∅ then
15 ld ← ld + 1;
16 lC = (o(l), d(l), lo , ld );
17 LA ← LA ∪ lC ;
18 lo = ld ;
19 LT = {o(l)};
20 continue;
21 end
22 if LT \{o(l)} = 6 ∅ then
T T
23 L ← L ∪ {o(l)};
24 end
25 LT ← LT ∪ {d(l)};
26 l = (o(l), d(l), lo , ld );
27 if lo = ld then
28 LA ← LA ∪ {l};
29 end
30 lo = ld ;
31 end
32 return L ;
A
224
Appendix C.
Extended Numerical Results for the Integrated Cargo
Allocation Problem
C.1. Problem Instances
100% 100%
90% 90%
80% 80%
70% 70%
60% 60%
50% cum. Q*R 50% cum. Q*R

cum. Q cum. Q
40% cum. R 40% cum. R
30% 30%
20% 20%
10% 10%
0% 0%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Percent of Cargo Flows Percent of Cargo Flows
Figure C.1.: Distribution of the cargo flow Figure C.2.: Distribution of the cargo flow
quantity (cum. Q), revenue quantity (cum. Q), revenue
(cum. R) and quantity * revenue (cum. R) and quantity * revenue
(cum. Q*R) for the LINER-LIB (cum. Q*R) for the LINER-LIB
Baltic instance. Mediterranean instance.
225
Appendix C. Extended Numerical Results for the Integrated Cargo Allocation Problem
100% 100%
90% 90%
80% 80%

70% 70%
60% 60%

cum. Q cum. Q
40% cum. R 40% cum. R
30% 30%
20% 20%
10% 10%
0% 0%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Pacific instance. WorldSmall instance.
100% 100%
90% 90%
80% 80%
70% 70%
60% 60%

cum. Q cum. Q
40% cum. R 40% cum. R
30% 30%
20% 20%
10% 10%
0% 0%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Europe-Asia instance. WorldLarge instance.
226
C.2. Cargo Allocation Approximations
C.2. Cargo Allocation Approximations
1200%
1000%
800%
ø Gap [%]
Q*R %
600%
cum. Q*R
cum. Q
cum. R
400%
200%
0%
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1
Baltic WorldSmall WorldLarge
Figure C.7.: Gap in percent for varied cargo flow amount and different cargo flow
selection strategies (for the Baltic, WorldSmall and WorldLarge).
80% 2,5
69,18%
70%
2
60%
54,10%
50%
1,5
ø Gap [%]
40%
ø Gap
ø Runtime Ratio
1
30%
20%
0,5
11,47%
9,76%
10% 6,41%
1,55% 0,22% 1,69% 0,29% 2,92%
0,00% 0,00% 0,95% 0,13% 0,00%
0% 0
2
16
20
16
20
16
20
Figure C.8.: Effects of the support points on the objective function and the runtime
(for the Baltic, WorldSmall and WorldLarge instances).
227
Appendix C. Extended Numerical Results for the Integrated Cargo Allocation Problem
80% 7
70% 67,21% 6

60%
5
50%
42,91% 4
ø Gap [%]
40%
ø Gap
3
ø Runtime Ratio
30%
2
20% 18,77%
10% 7,73% 1
0,00% 0,00% 0,00% 0,44% 0,01% 0,00% 1,15% 0,10% 0,00% 0,00% 0,00%
0% 0
Opt.
Opt.
Opt.
1
4
Figure C.9.: Gap and runtime improvement when terminating the Column Gener-
ation solution approach prematurely (for the Baltic, WorldSmall and
WorldLarge instance).
228
Appendix D.
Extended Numerical Results for the Liner Shipping
Network Design Problem
D.1. Liner Shipping Network MIP Formulation
120% 120%
100% 100%
80% 80%
Max S=1 Max S=1
Gap %
Gap %
Max S=2 Max S=2

60% 60%
Max S=3 Max S=3
Max S=4 Max S=4
40% Max S=5 40% Max S=5

Max S=6 Max S=6
20% 20%
0%
0%
0,0001
0,0008
0,0013
0,002
0,0028
0,0042
0,0052
0,0066
0,012
0,0154
0,0206
0,0242
0,0322
0,0417
0,0447
0,0471
0,0598
0,0781
0,0911
0,1098
0,1903
0,2587
0,2932
0,4195
1,3643
2,1227
2,8842
3,1755
4,7893
6,9867
8,4404
10,9465
0
0,0006
0,001
0,0015
0,0026
0,0037
0,0048
0,0061
0,0076
0,0127
0,0143
0,0146
0,0341
0,0355
0,0418
0,0484
0,0533
0,0557
0,0818
0,0965
0,1086
0,112
0,1225
0,1479
0,2062
0,2806
0,3691
0,396
0,5284
0,9654
5,5785
Runtime [h] Runtime [h]
Figure D.1.: Baltic LINER-LIB 2012 in- Figure D.2.: Baltic LINER-LIB 2012 in-
stance with maximum of one ves- stance with maximum of two
sel per service, no complex route vessel per service, no complex
type and no transit times. route type and no transit times.
229
Appendix D. Extended Numerical Results for the Liner Shipping Network Design Problem
120%
120%
100%
100%
80%
80%
Gap %
Max S=1
60% Max S=2
Gap %
Max S=3 Max S=1
60% Max S=2
Max S=4
40% Max S=5 Max S=3
Max S=4
40%
20%
20%
0%
0,001
0,0037
0,0066
0,0108
0,028
0,0293
0,0357
0,0687
0,0833
0,1126
0,2456
0,2648
0,293
0,3414
0,373
0,4444
0,7813
0,9256
1,0945
1,1841
1,351
1,5912
1,762
1,9929
2,0757
2,2612
2,281
3,3558
3,6782
5,4358
0%
0
0,01
0,02
0,03
0,05
0,06
0,07
0,08
0,09
0,1
0,11
0,13
0,17
0,19
0,24
0,26
0,33
0,34
0,62
0,79
0,96
2,14
5,86
6,31
6,67
8,13
8,63
11,19
12
Runtime [h]
Runtime [h]
Figure D.3.: Baltic LINER-LIB 2012 in-

Figure D.4.: Baltic LINER-LIB 2012 in-
stance with maximum of two
stance with maximum of one ves-
vessel per service, complex route
sel per service, no complex route
types allowed and no transit
type but transit times.
times.
D.2. Parameter Tuning for the Evolutionary Algorithm
130 10
9
120
8
110 7
100 6
5
90
4
Fitness [Mio. US$]
80 3
Fitness [Mio. US$]
70 2
CX(CL) 1 CX(CL)
60
CX(CY) 0 CX(CY)
1
31
61
91
121
151
181
211
241
271
301
331
361
391
421
451
481
511
541
571
601
631
661
691
721
751
781
811
841
871
901
931
961
991
1021
1051
1081
1111
1141
1171
50 CX(IC) -1 CX(IC)
40 -2
-3
30
-4
20 -5
10 -6
-7
0
-8
1
9
17
25
33
41
49
57
65
73
81
89
97
105
113
121
129
137
145
153
161
169
177
185
193
201
209
217
225
233
241
249
257
265
273
281
289
297
-9
(a) WAF instance. (b) Mediterranean instance.
Figure D.5.: Convergence of the evolutionary algorithm for LINER-LIB instances

using the clustering CX(CL), the cycle CX(CY) and informed cycle
CX(IC) crossover methods (averaged over five runs).
230
130 10
9
120
8
110 7
100 6
5
90
4
Fitness [Mio. US$]
80 3
Fitness [Mio. US$]

70 2
1
60 Roulette Roulette
0
BinTourn BinTourn
1
31
61
91
121
151
181
211
241
271
301
331
361
391
421
451
481
511
541
571
601
631
661
691
721
751
781
811
841
871
901
931
961
991
1021
1051
1081
1111
1141
1171
50 -1
40 -2
-3
30
-4
20 -5
10 -6
-7
0
-8
1
9
17
25
33
41
49
57
65
73
81
89
97
105
113
121
129
137
145
153
161
169
177
185
193
201
209
217
225
233
241
249
257
265
273
281
289
297
Runtime [s] -9
Runtime [s]
Figure D.6.: Convergence of Evolutionary Algorithm in different LINER-LIB in-

stances using the roulette (RS) and binary tournament (BT) selection
for the informed cycle crossover (IC) method (averaged over five runs).
130 10
9
120
8
110 7
100 6
5
90
4
Fitness [Mio. US$]
80 3
Fitness [Mio. US$]
70 2
10 1 10
60
50 0 50
1
30
59
88
117
146
175
204
233
262
291
320
349
378
407
436
465
494
523
552
581
610
639
668
697
726
755
784
813
842
871
900
929
958
987
1016
1045
1074
1103
1132
1161
1190
50 20 -1 20
40 -2
-3
30
-4
20 -5
10 -6
-7
0
-8
1
9
17
25
33
41
49
57
65
73
81
89
97
105
113
121
129
137
145
153
161
169
177
185
193
201
209
217
225
233
241
249
257
265
273
281
289
297
-9

stances using different population sizes (10, 20, 50 individuals per
generation) for the informed cycle crossover method with a binary
tournament strategy (averaged over five runs).
231
130 10
9
120
8
110 7
100 6
5
90 4
Fitness [Mio. US$]
80 3
Fitness [Mio. US$]

2
70
1
0,1 0,1
60 0
0,5 0,5
1
30
59
88
117
146
175
204
233
262
291
320
349
378
407
436
465
494
523
552
581
610
639
668
697
726
755
784
813
842
871
900
929
958
987
1016
1045
1074
1103
1132
1161
1190
-1
50 1 1
-2
40 -3
30 -4
-5
20
-6
10 -7
-8
0
-9
1
9
17
25
33
41
49
57
65
73
81
89
97
105
113
121
129
137
145
153
161
169
177
185
193
201
209
217
225
233
241
249
257
265
273
281
289
297
-10
Figure D.8.: Convergence of Evolutionary Algorithm in the WAF instance using

different initial pendulum percentage (10%, 50%, 100%) for the in-
formed cycle crossover method with a binary tournament strategy and
a population size of 50 (averaged over five runs).
130 10
120 9
8
110
7
100
6
90
5
Fitness [Mio. US$]
80
4
Fitness [Mio. US$]
70 3
0,1 0,1
60 2
0,2 0,2
50 0,5 1 0,5
40 0
1
30
59
88
117
146
175
204
233
262
291
320
349
378
407
436
465
494
523
552
581
610
639
668
697
726
755
784
813
842
871
900
929
958
987
1016
1045
1074
1103
1132
1161
1190
-1
30
-2
20
-3
10
-4
0 -5
1
9
17
25
33
41
49
57
65
73
81
89
97
105
113
121
129
137
145
153
161
169
177
185
193
201
209
217
225
233
241
249
257
265
273
281
289
297
-6
Figure D.9.: Convergence of Evolutionary Algorithm in the WAF instance using

different elitism ranges (10%, 20%, 50%) for the informed cycle crossover
method with a binary tournament strategy, a population size of 50 and
10% pendulum services in the initial population (averaged over five
runs).
232
130 10
120 9
8
110
7
100
6
90
5
Fitness [Mio. US$]
80
4
Fitness [Mio. US$]

70 10 3 10
60 20 2 20
50 50
50 1
No Inject No Inject
40 0
1
31
61
91
121
151
181
211
241
271
301
331
361
391
421
451
481
511
541
571
601
631
661
691
721
751
781
811
841
871
901
931
961
991
1021
1051
1081
1111
1141
1171
-1
30
-2
20
-3
10
-4
0 -5
1
9
17
25
33
41
49
57
65
73
81
89
97
105
113
121
129
137
145
153
161
169
177
185
193
201
209
217
225
233
241
249
257
265
273
281
289
297
-6

stances injecting new networks after different iterations for the in-
formed cycle crossover method with a binary tournament strategy, a
population size of 50, 10% pendulum services in the initial population
and an elitism of 10% (averaged over five runs).
233
D.3. Accuracy of VNS Surrogate Evaluation
Configuration/ operator \sum operations correct decisions [%] wrong decisions [%]
CF 100%, Q*R
Insert Port 4761 82,15% 17,85%
Change Vessel Type 3122 95,32% 4,68%
Change Vessel Count 3012 91,17% 8,83%
Change Single Port 5766 84,91% 15,09%
Change Port Sequence 3996 87,41% 12,59%
Change Liner Service 3507 90,08% 9,92%
2-opt 2743 93,58% 6,42%
sum/average 26907 88,26% 11,74%
CF 75%, Q*R
Insert Port 5305 77,49% 22,51%
2-opt 3086 88,72% 11,28%
sum/average 30177 84,90% 15,10%
CF 50%, Q*R
Insert Port 6074 76,84% 23,16%
2-opt 3279 90,58% 9,42%
sum/average 34461 84,60% 15,40%
Table D.1.: Accuracy of surrogate evaluation in the variable neighborhood descent

solving the Mediterranean LINER-LIB instance. The column generation
method is aborted after the first iteration and 40% (8) support points
for the bunker cost discretization is used. Data is recorded in five runs
of the variable neighborhood descent. The cargo flows are selected using
the Q ∗ R strategy.
234
D.3. Accuracy of VNS Surrogate Evaluation
Configuration/ operator \sum operations correct decisions [%] wrong decisions [%]
CF 100%, cum. Q*R
Insert Port 4549 82,48% 17,52%
2-opt 2579 91,82% 8,18%
sum/average 25969 86,65% 13,35%
CF 75%, cum. Q*R
Insert Port 6760 75,00% 25,00%
2-opt 3990 89,00% 11,00%
sum/average 38989 83,99% 16,01%
CF 50%, cum. Q*R
Insert Port 8757 76,09% 23,91%
2-opt 5510 86,26% 13,74%
sum/average 52314 83,56% 16,44%
Table D.2.: Accuracy of surrogate evaluation in the variable neighborhood descent

solving the Mediterranean LINER-LIB instance. The column generation
method is aborted after the first iteration and 40% (8) support points
for the bunker cost discretization is used. Data is recorded in five runs
of the variable neighborhood descent. The cargo flows are selected using
the cum.Q ∗ R strategy.
235
Appendix E.
Calculating Waypoints and Sea Distances
To visualize liner networks with services on a website, using libraries such as the
open source software OpenLayers1 , the path between two ports must be calculated.
To the best of the author’s knowledge, no free software exists that can draw the
route between two ports by using seaways. This section describes a method that
can be implemented easily and helps to visualize liner services.
The basic idea of the approach works as follows (see Figure E.1 for a graphical
overview of the process): The network is initialized and the polygon data for the
land masses loaded2 (see E.1(b)). Afterwards, a graph is constructed and mapped
to the polygon data, see E.1(c). The nodes that would lie on land are removed from
the graph. In the next step, the nodes are connected with each neighbor horizontally,
vertically and diagonal, see E.1(d). Afterwards, users can add further connection
between nodes. This can be useful to add small canals to the graph or compensate
inaccurate land polygons. The next step is to add cuts that remove connections, for
example on headland (see E.1(f)). Now, the ports are added to the graph by inserting
additional nodes and connecting them with edges to the closest two nodes. Now,
either a shortest path algorithm between all port nodes (for example the Dijkstra
algorithm, see Dijkstra (1959)), or a more specialized all-pairs shortest path (such
as the Floyd-Warshall, see Floyd (1962)) can be used to calculate the distance and
the way points. After this step, the unused nodes can be removed from the graph.
The remaining edges are reduced to decrease the number of way points that have
to be stored (see Figure E.1(i) and E.1(j)). Finally, the paths and distances can be
smoothed and stored for later use (see E.1(k) and E.1(l)).
1
Available under http://openlayers.org/
2
Polygon data is available in different level of detail under http://openstreetmapdata.com/
data/land-polygons
237
Appendix E. Calculating Waypoints and Sea Distances
(a) Initial state. (b) Loaded polygon (c) Created network (d) Create connected
data for land masses. nodes for sea areas. graph.
(e) Add connections (f) Add cuts to the (g) Connect ports to (h) All-pairs shortest
between graph com- graph. closest nodes. paths for ports.
ponents.
(i) Remove unused (j) Further path re- (k) Optionally path (l) Store final way
nodes and edges and duction. smoothing. points between ports.
prune graph on hor-
izontal and vertical
edges.
Figure E.1.: Visual process to calculate way points between ports for the user inter-
face.
238

Liner Shipping Network Design: Dissertation

Uploaded by

Copyright:

Available Formats

Liner Shipping Network Design: Dissertation

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

Liner Shipping Network Design: Dissertation

Uploaded by

Copyright:

Available Formats

Dissertation

Liner Shipping Network Design

M.Sc. Stefan Guericke

Schriftliche Arbeit zur Erlangung des akademischen Grades

Paderborn, Dezember 2014

2. Liner Shipping Network Planning 9

3. State-of-the-Art and Research Opportunities 35

4. Evaluating Networks - The Integrated Cargo Allocation Problem 63

5. Improving Networks - The Liner Shipping Network Design Problem 105

List of Figures 211

List of Tables 217

List of Algorithms 219

B. Transformation Algorithm for the Layered Network Structure 223

C. Extended Numerical Results for the Integrated Cargo Allocation Prob-

D. Extended Numerical Results for the Liner Shipping Network Design

E. Calculating Waypoints and Sea Distances 237

1.1. Research Goals

1. Evaluate large scale real-world liner networks

3. Optimize the LSNDP for real-world instances in reasonable time

4. Integrate the developed methods into a decision support system

2.1. Planning Process Overview

Fleet size and Alliance Liner network

Fleet Buffer Fleet

Speed Vessel schedule Bunkering Stowage Revenue Environmental

2.2. Basic Liner Shipping Network Design Problem

”A liner service is a fleet of ships, with a common ownership or manage-

Figure 2.2.: Example liner network with three services.

Vessel type Vessel type TEU capacity

Main Maritime Shipping Routes

Cape of Good Hope

Figure 2.3.: Main global maritime shipping routes, Rodrigue (2013).

2.3. Route Types and Network Structure

(a) Pendulum. (b) Cyclic. (c) Butterfly. (d) Conveyor belt.

Figure 2.4.: Liner services with different route types.

2.4. Transportation of Containerized Cargo

2.4.1. Demand Structure

2.4.2. Container Sizes and Types

Table 2.2.: Selection of frequently used container types.

Especially perishable commodities transported in ventilated and reefer containers

These durations are described in the next section.

2.5. Timing Aspects

2.5.1. Port Calls and their Impact on the Vessels’ Speed

Round trip time 14 days Time

Alexandria and Rotterdam with two vessels and a weekly frequency.

4. Bunkering (i.e., fueling the vessel)

5. Port buffer to compensate uncertainties (such as port congestion etc.)

2.5.2. Transit Times in the Maritime Context

2. Vessel speed and the duration at sea

3. Port call durations

2.6. Cooperative Agreements

2.6.1. Liner Conferences, Consortium Agreements and Global Alliances

transportation conditions with shippers individually (see (Basedow et al., 2012, p.

2.6.2. Slot Sharing Agreements

Table 2.3.: Example slot sharing agreement between three carriers.

2.7. Liner Service Capacities

2.7.1. Container Vessel Capacities

2.7.2. Port Depth and Vessels’ Deadweight Scale

Figure 2.9.: Example deadweight scale for a Panamax vessel.

2.8. Empty Container Repositioning