Supermercado 8

JID: EOR
ARTICLE IN PRESS [m5G;May 5, 2020;18:19]

European Journal of Operational Research xxx (xxxx) xxx
Contents lists available at ScienceDirect
European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor
Integrated order batching and vehicle routing operations in grocery

retail – A General Adaptive Large Neighborhood Search algorithm
Heinrich Kuhn a,∗, Daniel Schubert a, Andreas Holzapfel b
a
Department of Operations, Catholic University of Eichstaett-Ingolstadt, Germany
b
Department of Logistics Management, Hochschule Geisenheim University, Germany
a r t i c l e i n f o a b s t r a c t
Article history: In recent years, established and well-known grocery retailers have increasingly been investing in the
Received 30 April 2019 business of micro stores and petrol station shops. Supplying these stores with perishable and durable
Accepted 28 March 2020
goods leads to noticeable logistics challenges for the retailers. Since the total sales volumes of these shops
Available online xxx
are typically low and the respective sales areas are very limited, highly frequent deliveries of small sizes
Keywords: are required. These noticeably affect a number of operational planning problems. In the warehouse, the
Retail operations items requested have to be collected in small order sizes. In order to achieve efficient picking operations,
Order batching orders are therefore combined into larger picking orders, i.e., batches. Afterwards the orders have to be
Vehicle routing delivered to the stores at high frequency. In practice, all the planning problems mentioned are heavily
Adaptive Large Neighborhood Search interconnected due to the short planning horizon.
Despite the practical relevance, order batching, order picking and delivery operations have not so far
been investigated as an integrative planning problem. This paper therefore presents a novel modeling
and solution approach to solve practically relevant problem sizes. The combinatorial complexity of the
problem requires a heuristic solution approach. We propose an extension of the well-known Adaptive
Large Neighborhood Search (ALNS) metaheuristic that we call General ALNS (GALNS), and show that a
GALNS approach outperforms a similar ALNS algorithm in 96.35% of the problem instances generated.
Managerial insights from general problem data and a case study with a large German grocery retailer
support the applicability of the modeling and solution approach suggested in retail practice.
© 2020 Elsevier B.V. All rights reserved.
1. Introduction are no longer adequate due to the small order volumes per store.
Several orders are therefore combined to so-called picking orders.
Grocery retailers face diverse tasks along their supply chain in In the delivery domain, vehicle tours contain a larger number of
order to supply their stores. The products ordered by a store have stores that have to be served in an optimal sequence.
to be picked in a distribution center (DC) and prepared for delivery. As micro stores are also characterized by a very limited sales
Vehicles have to be scheduled and tours conducted to ultimately and backroom area, safety stock is almost nonexistent. Conse-
serve the stores. quently, micro stores rely on very frequent and punctual deliveries
While large supermarkets demand high-volume deliveries, mi- to be able to satisfy end consumer demand (Schubert, Scholz, &
cro stores, i.e., kiosks, petrol station shops, or canteens, require Wäscher, 2018). These factors together make the supply of micro-
much smaller deliveries. These small order volumes enhance the grocery stores a challenging field in practice, deserving of a scien-
need to ensure appropriate operations planning. The demand of tific investigation in its own right.
large supermarkets and the corresponding order volumes result in This paper treats the planning tasks of order picking and vehi-
picking operations that are processed order by order and delivery cle routing for the supply of micro stores in an integrative man-
operations that are conducted in a one-truck-to-one-store manner, ner, highlighting the interconnection between the planning tasks,
or by vehicle tours that contain only a small number of stores. Or- which are a result of the frequent deliveries of micro stores that
der picking and vehicle routing planning for micro stores, on the shorten the planning horizon. Previous work in the context of in-
other hand, is much more complex. Discrete picking operations tegrated order picking and vehicle routing has built on several as-
sumptions that do not cover the above-described case of micro
stores. In detail, a discrete picking strategy is applied consistently,
∗
Corresponding author. i.e., orders are picked one by one, which is an appropriate strat-
E-mail address: heinrich.kuhn@ku.de (H. Kuhn). egy if the picking process is technically restricted, e.g., for large
https://doi.org/10.1016/j.ejor.2020.03.075
0377-2217/© 2020 Elsevier B.V. All rights reserved.
Please cite this article as: H. Kuhn, D. Schubert and A. Holzapfel, Integrated order batching and vehicle routing operations in grocery
retail – A General Adaptive Large Neighborhood Search algorithm, European Journal of Operational Research, https://doi.org/10.1016/j.
ejor.2020.03.075
JID: EOR
2 H. Kuhn, D. Schubert and A. Holzapfel / European Journal of Operational Research xxx (xxxx) xxx
products (Schubert, Kuhn, & Holzapfel, 2019), or if the order vol- 2.1. Order picking area and order picking process
ume is large (De Koster, Le-Duc, & Roodbergen, 2007; Moons,
Braekers, Ramaekers, Caris, & Arda, 2019). In contrast, micro stores The order picking process takes place in the order picking area
demand only relatively few items that can be carried in one or of the retailer’s DC. We characterize this area before we detail the
only a few roll cage(s). These roll cages are then used for order order picking process.
picking as well as for delivery operations. However, picking de- Order Picking Area. The order picking area of a DC provides the
vices can transport multiple roll cages, so discrete picking opera- items that need to be collected to fulfill the store orders. Human
tions become inefficient. Order batching operations can be used to operators (order pickers) move through the picking area collect-
avoid these inefficiencies and lead to significant potential to reduce ing the items requested from their storage locations. Such so-called
the total processing time (pick time) in the warehouse (De Koster picker-to-part systems are still the most frequently applied type of
et al., 2007). order picking systems in grocery retailing (De Koster et al., 2007;
A favorable solution to the order batching problem may, how- van Gils, Ramaekers, Caris, & De Koster, 2018; Marchet, Melacini, &
ever, not be equally advantageous in the case of joint delivery. It is Perotti, 2015).
therefore vital to solve both problems – order picking and vehicle The storage locations are organized in a block layout, which is
routing – simultaneously. This is especially true if both processes the dominant layout design in practice (Roodbergen, 2012), and
overlap in time, which is very relevant in the application case con- also commonly assumed in order picking literature (e.g., De Koster
sidered. et al., 2007; van Gils, Braekers, Ramaekers, & Depaire, 2016; Scholz,
Integrated planning, however, is not an end in itself as it Schubert, & Wäscher, 2017). A block layout is defined by parallel
involves more complex planning and coordination. Independent picking aisles, which allow access to storage locations and orthog-
planning of order picking and delivery operations seems appro- onal cross aisles, which ensure entry to picking aisles. Storage lo-
priate if the majority of store orders can be picked and provided cations are arranged on both sides of a picking aisle and are acces-
sufficiently in advance of the delivery process starting. In practice, sible at grip height, while the aisles are broad enough for pickers
however, the time span between order availability and the desired to pass each other. Fig. 1 (a) illustrates such an order picking area.
delivery time is generally relatively short. In addition, intermedi- In the illustration, the area has two blocks; each block is restricted
ate storage areas at DC’s loading docks are limited, which requires by two neighboring cross aisles. Moreover, each block contains five
coordination between order picking and delivery operations. picking (sub-)aisles. Each item type is assigned to exactly one stor-
This paper contributes to the theory and practice of micro store age location.
deliveries by defining a practice-relevant planning problem and Fig. 1 (b), in addition, shows how the assigned orders of a batch
formulating a novel optimization model. A heuristic solution ap- affect the distance to be traveled and consequently the pick time.
proach is developed, due to the combinatorial complexity of the Black rectangles represent storage locations from which at least
given optimization problem. The solution approach applies the one item has to be picked (pick locations). While the pick loca-
Adaptive Large Neighborhood Search (ALNS) metaheuristic and ex- tions marked 1 and 3 are located close to each other, the locations
tends it to a General ALNS (GALNS) inspired by the concept of a of store order 2 require significant additional travel time.
General Variable Neighborhood Search (Hansen, Mladenović, Brim- In grocery retailing, DCs are usually designed for the supply
ber, & Pérez, 2010). We present extensive numerical experiments of large supermarkets. The large surface is therefore divided into
that are based on simulated data and practice data gained from a smaller zones, e.g., according to product types, to reduce travel
real-life case study of a large German grocery retailer. The results times. This means orders are split and an additional sorting ef-
provide meaningful and valuable insights into retail practice. fort is necessary after the picking process (De Koster et al., 2007).
The remainder of this paper is organized as follows. We first In the case of small order volumes, as for micro store deliveries,
detail the specific planning problem being considered (Section 2). only a few items are requested per zone and order. Order pick-
Relevant literature is discussed in Section 3. Section 4 presents ers therefore traverse the total picking area across zones collect-
a decision support model for the problem investigated. We de- ing the items of several store orders to achieve an efficient picking
velop the GALNS solution approach in Section 5. The subsequent tour. This batching strategy significantly reduces the total pick time
Section 6 provides approaches to benchmark the performance of (De Koster, van der Poort, & Wolters, 1999).
the solution approach suggested. Section 7 compares the result of Order Picking Process. Typically, order picking operations for
ALNS and GALNS, shows the advantage of an integrated planning micro stores are assigned to a fixed time slot at the DC in
approach, discusses the impact of order batching operations, and which no other picking operations (e.g., for supermarkets) take
presents the results of the real-life case study. Section 8 summa- place. Multiple employees (order pickers) work on these store or-
rizes the fundamental findings of the paper and presents an out- ders in the picking area. In practice, roll cages as depicted in
look on future research opportunities. Matusiak, de Koster, and Saarinen (2017) simplify the picking pro-
cess (Matusiak et al., 2017; Scholz et al., 2017). More precisely, roll
2. Problem description cages can carry a large number of items and also prevent addi-
tional sorting effort by unambiguously assigning store orders to
Micro stores are characterized by the limited assortment of- one or several cage(s). Order pickers use a picking device that can
fered, which limits the total sales volume even though individ- carry up to a certain number of roll cages (Matusiak et al., 2017),
ual items, e.g., snacks, sweets and drinks, feature high turnover. so store orders can be combined into so-called batches, i.e., picking
In addition, these shops only have limited sales areas and storage orders.
spaces. Micro stores therefore typically need small orders to be The order pickers usually obtain information on picking orders
supplied daily from the distribution center (DC). The supply pro- via devices such as mobile terminals, pick-by-light or voice-picking
cess starts by retrieving the items ordered from their storage lo- systems (Matusiak et al., 2017). This information gives a sequence
cations in the DC (order picking process), and is followed by de- (and path) according to which items of a batch need to be col-
livery to the micro store locations (vehicle routing process). Both lected. From this sequence, the processing time of a batch, i.e., its
processes are described in the following subsections. After that we pick time, results. The pick time of a batch comprises the setup
elaborate on the interdependencies between these processes that time for preparing a picking tour, the search time for identifying
require an integrated planning approach, considering both prob- an item at a storage location, the time for physically collecting
lems simultaneously. an item from its storage location, and finally the travel time for
ejor.2020.03.075
JID: EOR
H. Kuhn, D. Schubert and A. Holzapfel / European Journal of Operational Research xxx (xxxx) xxx 3
Fig. 1. Example layout of the picking area and S-Shape routing strategy.
moving through the order picking area (Tompkins, White, Bozer, ing operations can vary significantly. For example, a micro store lo-
& Tanchoco, 2010). With the exception of the travel time, all cated at a fuel station is easier to serve compared to a micro store
time values are a result of the composition of a batch. The travel integrated into a railway station.
time and thus the pick time of a batch results from the sequence Typically, subcontractors provide vehicles for the delivery oper-
according to which items are collected. The NP-hard picker routing ation. The number of vehicles is variable or fixed depending on
problem has to be resolved for this purpose. This is typically per- contract design. A frequently applied contract model in grocery
formed using heuristic routing strategies (De Koster et al., 2007). retail practice contains a fixed number of vehicles that are pro-
According to Roodbergen (2001), retail practice most commonly vided and costs that are calculated using a distance- and volume-
applies the S-Shape routing strategy. S-Shape routing follows the dependent transportation cost matrix, i.e., transportation costs are
simple idea that only sub-aisles are entered – and normally tra- charged based on the distance between the DC and the location
versed completely – if the sub-aisle contains at least one item to of the micro store and the number of roll cages ordered. The total
be picked (see Fig. 1 (b)). All aisles that contain at least one pick costs are therefore determined independently of the order picking
location are traversed completely. The fourth sub-aisle of the sec- and vehicle routing decisions. In connection with such a contract,
ond block represents an exception since it is the last aisle of block the retailer has the planning sovereignty and thus enables the re-
2, which contains items to be picked. tailer to strive for goals independently of the subcontractor’s pref-
After an order picker has collected all items of a batch, the cor- erences. This is an important issue as the requirements of the re-
responding roll cages are transported to an intermediate storage tailer and the subcontractor normally conflict with each other.
area that is located between the order picking area and the loading The retailer’s primary aim is to supply the stores punctually.
docks. In the case of small order volumes, the intermediate storage The limited number of employees in a micro store – normally one
area only has a bridging function between picking and delivery but – are usually responsible for serving customers as well as operat-
does not restrict the planning problems. ing the checkout. Replenishing operations are therefore performed
within a certain time slot in which shifts overlap or additional em-
2.2. Vehicle routing process ployees are available. A punctual supply of the stores is therefore
necessary to avoid overtime and idle time for employees.
After the orders have been prepared by the order pickers and Early deliveries cannot be processed due to employee capacity
made available at the DC’s loading docks, they can be delivered shortages. This defines the hard lower bound of the delivery time
to the stores. The resulting release dates of store orders, i.e., the windows for the associated vehicle routing problem.
point in time that the corresponding roll cages are provided at the In general, stores and the DC agree on certain delivery con-
loading docks, are taken as input for delivery planning. ditions, e.g., due dates or time windows (Schubert et al., 2018),
The roll cages for all store orders of one tour are consolidated which specify the delivery time desired (due dates) or the earli-
at their associated loading docks so that they can be sorted in the est and latest delivery time allowed (time windows), respectively.
sequence of the scheduled delivery tour. This saves unnecessary These store-associated due dates or delivery time windows are
sorting effort on arrival at the store. We therefore assume that the generally defined during a preceding mid-term planning stage and
loading of the vehicle happens when all orders to be loaded have are therefore assumed to be a given in our problem setting. Setting
been made available. The resulting truck loading time is assumed appropriate delivery time windows at a mid-term planning stage,
to be constant. e.g., consecutive time windows for stores located nearby, avoids
This time can be added to the travel time to reach the first store the scheduling of inefficient routes when following an aim at
of a delivery tour. At store locations, however, the time for unload- short-term planning horizon that minimizes tardiness times only.
ejor.2020.03.075
JID: EOR
Nevertheless, delivery tours have to be built on a daily basis since Zhang, Xuping, Chan, & Ruan, 2017). Henn (2015) deals with an
truck loads per store change and stores are added to or skipped order batching problem that includes multiple pickers. Moreover,
from the long-term delivery schedule from day-to-day. Total tar- each order has an order-specific due date by which it should be
diness, i.e., the sum of the (non-negative) differences between ac- released by the warehouse. In order to route pickers through the
tual delivery dates and upper bounds of the corresponding time picking area, the S-Shape and Largest-Gap strategies are applied.
windows, justifies an appropriate objective in our research setting This paper is extended by Scholz et al. (2017) by integrating picker
(Schubert et al., 2018). In addition, the aim fits the delivery con- routing decisions for each batch. Individual picker skills are consid-
tracts with carriers above described and confirms the general pro- ered by Matusiak et al. (2017) to minimize the sum of batch pro-
cedure in practice (van Gils et al., 2018; Ullrich, 2013). cessing times. Zhang et al. (2017) present an online variant of the
model presented by Henn (2015). A recent approach that integrates
2.3. Integrated decision problem order batching, picker routing, and picker scheduling is suggested
by van Gils, Caris, Ramaekers, and Braekers (2019).
In summary, two main problems have to be solved when sup-
plying micro stores from DCs. In the order picking area, picking
3.2. Vehicle routing problem
orders have to be conducted that contain one or several store or-
ders. These picking orders, i.e., batches, then have to be assigned
The vehicle routing problem in this research is mainly charac-
to order pickers and scheduled. In the delivery area, the planner
terized by the release dates at which store orders are available for
is faced with a vehicle routing problem. Store orders have to be
delivery operations and soft (semi-hard) time window constraints.
assigned to vehicles, and a route has to be planned for each ve-
Release Dates. Only a few papers deal with order-specific release
hicle. The interconnection of order picking and vehicle routing in
dates in the literature. Archetti, Feillet, and Speranza (2015a) study
the planning of the daily supply of micro stores is a result of the
the complexity of routing problems with release dates. A multi-
limited time slot for picking operations at the DC and the limited
period vehicle routing problem in which orders are characterized
time slot for micro store deliveries due to operational constraints
by release and due dates is investigated by Archetti, Jabali, and
at the stores.
Speranza (2015b). Cattaruzza, Absi, and Feillet (2016) deal with a
Order picking and vehicle routing operations are then intercon-
multi-trip vehicle routing problem with time windows and release
nected by the release dates of the store orders at the loading dock
dates.
of the DC and the start times of the vehicles in the problem setting
Soft Time Window Constraints. According to Balakrishnan (1993),
considered. A solution to the order picking problem determines the
the first research on a routing problem with soft time window con-
release dates, which are then taken as input for solving the ve-
straints, i.e., one or both time window bound(s) can be violated,
hicle routing problem. Vice versa, a solution to the vehicle rout-
leads back to Sexton and Choi (1986). Vehicle routing problems
ing problem provides release dates desired, i.e., due dates accord-
with soft time windows have been introduced by Koskosidis, Pow-
ing to which the orders should be provided by the warehouse to
ell, and Solomon (1992). A vehicle routing problem with a hard
ensure punctual deliveries via the start times of vehicles. The re-
lower bound and a soft upper bound was investigated by Taillard,
lease dates resulting from an independent solution of both prob-
Badeau, Gendreau, Guertin, and Potvin (1997), Gendreau, Guertin,
lems may contradict each other, which necessitates an integrated
Potvin, and Taillard (1999), and Fu, Eglese, and Li (2008). Addi-
solution of both problems.
tionally, some authors focus on VRPs in which the arrival time at
a store location is restricted by a maximum violation of the cor-
3. Literature review
responding soft time window, i.e., a hard time window embraces
the soft time window (Calvete, Galé, Oliveros, & Sánchez-Valverde,
Schmid, Doerner, and Laporte (2013) introduced the research
20 07; Chiang & Russell, 20 04; Figliozzi, 2010; Fu et al., 2008; Iqbal,
branch of integrated order picking and vehicle routing problems to
Kaykobad, & Rahman, 2015; Mouthuy, Massen, Deville, & Van Hen-
the literature, and it has increasingly been addressed since then.
tenryck, 2015).
Order picking and vehicle routing problems have, however, usually
been investigated independently in the related literature. In the
following we first analyze relevant literature regarding the specific 3.3. Integrated and related problems
problems of the entire problem setting. Afterwards we focus on in-
tegrated order picking and vehicle routing problems and show the The number of publications regarding integrated order picking
gap in the literature that is filled by the research at hand. and vehicle routing problems has increased recently. The paper of
Schmid et al. (2013) focuses on so-called rich routing problems and
3.1. Order picking problem proposes several research opportunities for extended vehicle rout-
ing problems, e.g., lot-sizing, container loading problems, and order
Order picking deals with the retrieval of customer-requested picking operations. They are the first studying an integrated order
items in a warehouse (De Koster et al., 2007). A comprehensive picking and vehicle routing problem and propose a model formu-
overview to order picking planning problems on a tactical level, lation for integrated order batching, assignment, sequencing and
e.g., storage location assignment and zone location, as well as on (picker) routing, and a vehicle routing problem with hard lower
an operational level, e.g., routing and workforce allocation, is given and soft upper time window bounds.
by De Koster et al. (2007) and van Gils et al. (2018). The order pick- In contrast to Schmid et al. (2013), problem settings including
ing problem considered in this paper combines order batching and discrete order picking operations have been studied since then. In
job assignment to multiple pickers. The literature review focuses other words, store orders are processed one after another. In the
on these planning problems in the following. retail-based application of Schubert et al. (2018), the number of
The Order Batching, (Batch) Assignment and Sequencing Prob- order pickers and (homogeneous) vehicles is strictly restricted. In
lem (OBASP) deals with consolidating orders to batches, assign- contrast, the number of order pickers and vehicles has to be de-
ing batches to a limited number of order pickers, and determining termined in Schubert et al. (2019). Moreover, heterogeneous vehi-
a schedule for the corresponding batches for each picker (Scholz cles satisfying different store requirements, i.e., vehicle-site depen-
et al., 2017). The OBASP has received increasing attention during dencies, are considered. Both papers contain model formulations,
recent years (Henn, 2015; Matusiak et al., 2017; Scholz et al., 2017; heuristic-based algorithms, and managerial insights regarding the
ejor.2020.03.075
JID: EOR
general problem settings and the value of an integrated planning presented assumes an order batching strategy for picking orders
perspective. including the picker routing problem in the warehouse and de-
A problem setting including a fixed number of regular or- livery time windows for orders with a hard lower and soft upper
der pickers, a limited number of pickers who can be hired tem- bound. Moreover, this model pursues the objective of minimizing
porarily, and a restricted number of heterogeneous or homoge- the distance covered both in the warehouse by order pickers and
neous vehicles in terms of capacities has been investigated by by vehicles during delivery as well as penalties for tardy orders.
Moons, Ramaekers, Caris, and Arda (2018), Moons et al. (2019), and This optimization model shows great similarities to our model. The
Ramaekers, Caris, Moons, and van Gils (2018). All these publica- present paper, however, focuses on a punctual delivery of stores
tions deal with an identical problem setting, but they propose a motivated by the order fulfillment process of grocery micro stores.
model formulation, a heuristic solution approach and subsequently The model developed here, therefore, aims to minimize the to-
apply the heuristic to gain managerial insights regarding time win- tal tardiness. The fundamental difference between the current pa-
dows. per and the paper of Schmid et al. (2013) is that the latter does
The above-mentioned publications propose neighborhood- not present any solution approaches. The current paper, however,
based algorithms and/or model formulations to deal with the un- develops a novel and sophisticated solution approach capable of
derlying problem settings. Schubert et al. (2018) and Schubert solving the optimization model developed for practically relevant
et al. (2019) each propose a model formulation and an iterated problem sizes. We additionally conduct extensive numerical exper-
local search approach or a variable neighborhood search algo- iments that are particularly based on data gained from a real-life
rithm, respectively. Moons et al. (2018) develop a model formu- case study. We derive meaningful and valuable recommendations
lation, Moons et al. (2019) propose a record-to-record approach. for grocery retailers operating a homogeneous or diverse set of mi-
Ramaekers et al. (2018) apply both the model formulation of cro stores using the results achieved.
Moons et al. (2018) and the heuristic of Moons et al. (2019).
The order picking problem shows similarities to the paral- 4. Mathematical model
lel machine scheduling problem when orders are picked dis-
cretely. In this case, integrated production and distribution plan- The upcoming section formulates the mathematical model for
ning problems implying a make-to-order production strategy be- the integrated Order Batching (OB) and Vehicle Routing Problem
come relevant. Chen (2010) and Moons, Ramaekers, Caris, and Arda with Semi-Hard Time Windows (VRPSTW) that we have illustrated
(2017) present detailed reviews about these problems. In addition, in detail in Section 2. We denote this problem OB-VRPSTW. First
Schubert et al. (2018) provide a recent overview of integrated ma- we describe the assumptions of the model and define the nota-
chine scheduling and vehicle routing problems containing single tion used; afterwards we formulate the mathematical optimization
or parallel machines and routing decisions. However, these ap- model.
proaches are of minor interest in our setting, since the present pa- General Assumptions and Notation. We assume that one DC (i =
per focuses on the possibility of batching store orders during the 0) is responsible for fulfilling the daily orders from a set of dedi-
picking process. cated micro stores (I = {1, . . . , i, . . . , |I|} ). Set I0 = I ∪ {0} then de-
Batch processes are rarely considered in integrated production notes the entire set of origin and destination points. Each store
and distribution planning problems within the context of a make- order contains a certain number of items that are stored on roll
to-order production strategy. At first glance, batching in produc- cages. The number of roll cages required to carry the items re-
tion and warehouse circumstances has some similarities. A closer quested for each store i is denoted by ci , i ∈ I.
look at the details of the batch processes, however, reveals no- Assumptions and Notation Related to Order Picking. In the
ticeable differences in both fields of application. Farahani, Grunow, order picking area of the DC, a set of order pickers M =
and Günther (2012), for example, consider an integrated produc- {1, . . . , m, . . . , |M|} is responsible for processing the store orders.
tion and distribution planning problem in the catering industry. The store orders can be combined into batches as long as the max-
The batch processing time equals the maximum processing time imum capacity (c) of the picking device is met. An order picker is
of the orders assigned to the same batch. In our case, however, the allowed to successively process several batches. Splitting of orders,
batch processing time depends on the associated routes for picking however, is not allowed.
all orders of one batch. Amorim, Belo-Filho, Toledo, Almeder, and We assume that the processing times of all feasible batches (set
Almada-Lobo (2013), on the other hand, consider the case where L) are calculated in advance. This is easily doable if a basic routing
identical products are batched. In our case, however, a batch con- heuristic is used, e.g., S-Shape routing (see Henn, 2015). Neverthe-
sists of a diverse set of identical and non-identical products. Gao, less, the number of feasible batches grows exponentially with the
Qi, and Lei (2015), furthermore, consider an integrated produc- number of orders, and thus the computation time (Ruben & Ja-
tion and distribution planning problem where the batch processing cobs, 1999). The set of all feasible batches (lots) is denoted with
time equals the sum of the processing times of all orders. Addi- L = {1, . . . , l, . . . , |L|}. Parameter bli ∈ {0, 1} indicates whether batch
tionally they assume that a batch has to be delivered immediately l, l ∈ L, contains the order of store i, i ∈ I. In addition, pl denotes the
after production, and only a single truck is available. All of these picking time of batch l, l ∈ L.
assumptions differ from our case. The order picking problem incorporates three decisions. It de-
Although there is an increasing set of contributions dealing cides which of the possible batches is chosen, the assignment of
with integrated order picking and vehicle routing problems, the the batches chosen to one of the pickers available, and in which
problem setting investigated in this research has not yet been dealt sequence the batches chosen and assigned are processed by the
with. In contrast with almost all of the above-mentioned papers, picker. The binary variable ylmn ∈ {0, 1} indicates all three decisions.
the current paper considers the possibility of batching orders when It equals 1 if batch l, l ∈ L, is assigned to sequence position n, n ∈ N,
picking delivery orders rather than assuming a discrete picking of picker m, m ∈ M, and 0 otherwise. Set N = {1, . . . , n, . . . , |N|} de-
strategy. notes the position assignable in the sequence of picking batches of
Closest related to the given research is the problem setting an order picker. Note that at most |I| orders (batches) can be as-
defined in Schmid et al. (2013). The paper of Schmid et al. signed to a picker; thus, the maximum cardinality of set N is |I|,
(2013) presents several problems that are relevant to practice but i.e., |N| ≤ |I|.
unexplored. They, however, present no specific approaches to solv- The decision variable ylmn defines several auxiliary variables.
ing the corresponding optimization models. One of their models The completion time of the batch that is assigned to position n
ejor.2020.03.075
JID: EOR
Table 1
Notation used formulating model OB-VRPSTW.
Sets
I set of stores, I = {1, . . . , i, . . . , |I|}

I0 entire set of all origin and destination points, I0 = I ∪ {0} with i = 0 as DC
K set of vehicles, K = {1, . . . , k, . . . , |K |}
L set of possible batches including all feasible batches, L = {1, . . . , l, . . . , |L|}
M set of order pickers, M = {1, . . . , m, . . . , |M|}
N set of positions in the sequence of picking batches of an order picker, N = {1, . . . , n, . . . , |N|}
Parameters
αi (hard) lower time bound of store i’s time window, i ∈ I
βi (soft) upper time bound of store i’s time window, i ∈ I
blj parameter that indicates whether batch l, l ∈ L, includes the store order i, i ∈ I, (bli = 1 ) or not (bli = 0 )
c capacity of a picking device
ci capacity requirement of store order i, i ∈ I, in terms of picking device capacity, i.e., roll cages
Q sufficiently large number
pl processing time of batch l, l ∈ L
s0 loading time of a vehicle with its assigned store orders at the DC
si service time for store order i, i ∈ I
tij travel time between the locations i and j, i, j ∈ I0 , j = i
w maximum capacity of a vehicle, measured in volume
wi capacity requirement of store order i, i ∈ I, in terms of vehicle capacity
Decision Variables
ai arrival time at store location i, i ∈ I
dk departure time of vehicle k, k ∈ K, from the DC
ri release date of store order i, i ∈ I, for delivery from the DC
τi tardiness of store order i, i ∈ I
uki binary variable indicating whether location i, i ∈ I0 , is included in the tour of vehicle k, k ∈ K, (uki = 1 ) or not (uki = 0 )
xki j binary variable which indicates whether vehicle k, k ∈ K, visits location j, j ∈ I0 , immediately after location i, i ∈ N0 , j = i, xki j = 1 or not xki j = 0
ylmn binary variable which indicates wether batch l, l ∈ L, is processed at position n, n ∈ N, by picker m, m ∈ M, (ylmn = 1 ) or not (ylmn = 0 )
zmn completion time of the batch at position n, n ∈ N, of picker m, m ∈ M
of picker m is quantified by zmn , m ∈ M, n ∈ N. This quantity also Variable dk , k ∈ K, defines the departure time of vehicle k from the
defines the release date for delivery from the DC of all orders as- DC and variable ai , i ∈ I, quantifies the arrival time scheduled at a
signed to this batch. We denote this release date as ri , i ∈ I. store location i.
Assumptions and Notation Related to Vehicle Routing. We assume Notation Summary and Model Formulation. Table 1 summarizes
a set of homogeneous vehicles K = {1, . . . , k, . . . , |K |} available to the sets, indices, parameters, decision variables and auxiliary vari-
deliver the store orders from the DC to the respective store lo- ables defined. These symbols are afterwards used when formulat-
cations. Each vehicle features a limited capacity w and requires a ing the mathematical decision model OB-VRPSTW.
constant service (loading) time s0 for all store orders assigned to
Model OB-VRPSTW
it. Note that this time could also be included in the travel time
assumed to reach the stores directly from the DC. A store order i, minimize T = τi (1)
however, requires a dedicated capacity wi , i ∈ I, of the entire ve- i∈I
hicle capacity w and causes a certain service time si , i ∈ I, to un-
load the items at the corresponding store location. While the load- subject to
ing operations at the DC are usually very similar and constant for
the entire vehicle load, unloading operations at store locations can uki = 1 ∀i ∈ I (2)
vary due to different unloading processes and (un)loading docks. k∈K
For traveling between locations, i.e., DC and store locations, a cer-

tain travel time tij , i, j ∈ I0 , j = i, occurs. In addition, we assume that uk j = xki j ∀k ∈ K, j ∈ I (3)
all stores have to be served within a defined time window. We as- i∈I0 ,i= j
sume that the lower time-bound, α i , i ∈ I, is hard, and the upper

time-bound, β i , i ∈ I, is soft. Consequently, each route leads to a uk j = xkji ∀k ∈ K, j ∈ I (4)
feasible solution since all stores will be served – even with notice- i∈I0 ,i= j
able tardiness. We, however, assume sufficient vehicle capacity to
serve all stores. Please note that several delivery tours may lead to
uk0 ≥ uki ∀k ∈ K, i ∈ I (5)
the same objective contribution.
The vehicle routing problem involves two decisions: the as-

signment of a store order i, i ∈ I, to one of the vehicles (tours) k, xk0i ≤ 1 ∀k ∈ K (6)
k ∈ K, and on the sequence of store visits when executing the de- i∈I
livery tour. The binary variable uki defines the first decision. It in-
dicates whether the store order (location) i, i ∈ I0 , is assigned to ve-
ai ≥ dk + t0i − Q · (1 − uki ) ∀i ∈ I, k ∈ K (7)
hicle (tour) k, k ∈ K, (uki = 1 ) or not (uki = 0 ). Binary variable xki j ,
k, k ∈ K, i, j ∈ I0 , j = i, then defines the second decision. It indicates

whether vehicle k visits location i immediately before location j, a j ≥ ai + ti j + si − Q · 1 − xki j ∀i, j ∈ I, j = i, k ∈ K (8)
j = i, (xki j = 1 ) or not (xki j = 0 ).
Again, the decision variables include several auxiliary variables,
which are required to formulate the entire mathematical model. ai ≥ αi ∀i ∈ I (9)
ejor.2020.03.075
JID: EOR
part of several feasible batches. Consequently, Constraints (15) en-

sure that each order is finally processed in exactly one batch. The
τi ≥ ai − βi ∀i ∈ I (10)
completion time of a batch is calculated for the first and follow-
ing batches of a picker using Eqs. (16) and (17), respectively. The
release dates of store orders are set using Eqs. (18).
τi ≥ 0 ∀i ∈ I (11)
Model OB-VRPSTW consists of two classical decision problems
that are known to be NP-hard (Henn, 2015; Taillard et al., 1997).
Consequently, the integrated problem must be NP-hard as well.
w≥ wi · uki ∀k ∈ K (12) Schubert et al. (2018) show that a related problem is only solv-
i∈I
able to proven optimality for very small problem instances with a
commercial MIP solver. Thus, heuristic solution approaches to the
dk ≥ s0 + ri − M · (1 − uki ) ∀k ∈ K, i ∈ I (13) OB-VRPSTW are necessary to solve problems of a relevant size for
retail practice.
In the following we use the term sub-solution for the or-
ylmn ≤ 1 ∀m ∈ M, n ∈ N (14) der picking and vehicle routing subproblems, respectively. A sub-
l∈L solution includes all explicit decisions relating to the correspond-
ing subproblem. For the exclusive order batching problem we fix

bli · ylmn = 1 ∀i ∈ I (15) variables ylmn , which also determine the release dates of store or-
l∈L m∈M n∈N ders, i.e., variables ri . For the vehicle routing problem, however, we
assume fixed variables uki and xki j . Consequently, implicit decisions,
e.g., dk , ai , remain adaptable since those variables depend on the
pl · ylm1 ≤ zm1 ∀m ∈ M (16)
solution of the order batching problem. For example, vehicle de-
l∈L
parture times, dk , are adjusted if order release dates, ri , change due
to a modification of batch-picker assignments.
zm,n−1 + pl · ylmn ≤ zmn ∀m ∈ M, n ∈ N \ 1 (17)
l∈L
5. A general adaptive large neighborhood search algorithm

ri ≥ (bli · zmn − Q · (1 − ylmn )) ∀i ∈ I, m ∈ M, n ∈ N (18) We suggest an extension of the Adaptive Large Neighbor-
l∈L hood Search (ALNS) metaheuristic, i.e., the General Adaptive Large
Neighborhood Search (GALNS) metaheuristic as an integrative ap-
xki j ∈ {0, 1} ∀k ∈ K, i, j ∈ I0 (19) proach for solving the OB-VRPSTW. The ALNS metaheuristic was
proposed by Pisinger and Ropke (2007) and has been adapted
to an increasing number of optimization problems. For example,
uki ∈ {0, 1} ∀k ∈ K, i ∈ I0 (20) Pisinger and Ropke (2007) apply an ALNS approach to vehicle rout-
ing problems with multiple depots, vehicle-site dependencies and
multiple periods. Koch, Bortfeldt, and Wäscher (2018) use an ALNS
ylmn ∈ {0, 1} ∀l ∈ L, m ∈ M, n ∈ N (21) framework for an integrated vehicle routing problem with pickups
and deliveries and complex loading constraints. Moreover, Matu-
siak et al. (2017) propose an ALNS algorithm for an order batch-
zmn ≥ 0 ∀m ∈ M, n ∈ N (22) ing problem, integrating individual picker skills. Žulj, Kramer, and
Schneider (2018) similarly develop a hybrid ALNS including tabu
Model OB-VRPSTW aims to minimize the total tardiness of all search components for an order batching problem. The ALNS meta-
store orders. This is quantified by Eq. (1). The restrictions of model heuristic may thus provide a suitable basic strategy for the OB-
OB-VRPSTW are presented in two main parts. Constraints (2–12) VRPSTW as well. The framework, however, is extended to a GALNS,
define the vehicle routing problem. Afterwards, Constraints (14)– inspired by the General Variable Neighborhood Search (GVNS)
(18) define the order picking problem. Additional constraints, i.e., metaheuristic. Moreover, the metaheuristic developed is tailored
Constraints (13), link both problems. Finally Eqs. (19)–(22) deter- to the OB-VRPSTW. The subsequent Section 5.1 details the general
mine the domains of the decision and axillary variables used. Each framework of the GALNS metaheuristic. Afterwards, Sections 5.2–
store order has to be assigned to exactly one vehicle (2) and each 5.6 describe the individual components of the entire GALNS
vehicle can perform one tour at most (6). Moreover, Constraints algorithm.
(3) and (4) ensure that a store is visited exactly once and the
corresponding vehicle has to leave the store, respectively. The DC
(i=0) is included in all routes conducted, i.e., tours containing at 5.1. General framework of GALNS
least one store visit, by Constraints (5). Constraints (7) and (8) de-
termine the arrival time at a store location. These constraints also The GALNS approach is based on an ALNS algorithm. An ALNS
avoid the building of sub-tours. Note that waiting times are not algorithm is equipped with several destroy and repair heuristics,
modeled explicitly since the arrival time is bounded by the (hard) e.g., the five orders with the highest objective contribution are re-
lower bound of the delivery time window of the store served. The moved (Shaw, 1997) and included again according to greedy or
(non-negative) tardiness is calculated and defined in Eqs. (10) and regret logic. These heuristics are applied according to their indi-
(11). The restricted capacity of vehicles is taken into account by vidual weights, which are dynamically adjusted within the search
Constraints (12). As already mentioned above, Eqs. (13) connect process. The algorithm selects one destroy and one repair heuris-
the order picking and the vehicle routing problem. Eqs. (13) sat- tic in each iteration according to a roulette wheel selection con-
isfy that a vehicle cannot start its tour before all orders assigned sidering the current weight of each heuristic. The algorithm then
are released by the DC and loaded into the respective vehicle. Con- checks whether a new best solution is found after the incumbent
straints (14)–(18) then define the order picking problem. A feasible solution has been destroyed first and subsequently repaired. After-
batch can be assigned once at most (14), but a store order may be wards, the procedure may possibly update the incumbent solution
ejor.2020.03.075
JID: EOR
Fig. 2. General framework of the GALNS algorithm for solving OB-VRPSTW.
according to an acceptance/rejection criterion. Finally, the proce- heuristics, a VNDVRP is applied first and vice versa. Schubert et al.
dure updates the weights. The algorithm stops if a certain termina- (2018) propose a similar local search framework but vary the se-
tion criterion is met. In contrast to this procedure, (G)VNS applies quence of application.
random modifications according to systematically applied neigh- The proposed algorithm is capable of taking all decisions within
borhood structures (shaking) and a (sophisticated) local search af- one iteration. The integration of the problems within the algorithm
terwards. For example, it exchanges two randomly chosen orders of is performed by the combination of destroy and repair heuristic
different pickers (shaking) and then optimizes the individual pick- and the local search. The destroy and repair heuristics treat one
ing sequences (local search). subproblem at a time in an iteration and may lead to an improve-
The GALNS approach combines the general concept of the ALNS ment as a result. Additionally, they prepare the solution struc-
procedure with elements of a GVNS. More precisely, GALNS applies ture for the local search that optimizes the unaltered subsolution
a sophisticated local search procedure after the repair heuristic and first and then both subsolutions iteratively. This behavior induces
before the acceptance criterion, and includes random modifications a deep integration in decision making within one iteration since
of (sub-)solutions, e.g., random insertion or shaking. Fig. 2 shows a change in one subsolution that may possibly be slightly is in-
the framework of the GALNS algorithm to the OB-VRPSTW in pseu- tensively investigated by the local search in the other subproblem.
docode. This also considers the associated improvement potential of the in-
The problem structure of the OB-VRPSTW motivates the en- tegrated problem. The local search procedure therefore links the
hancement. For example, a modification of a solution to the order decision making between OPP and VRP.
picking problem (OPP) will possibly affect the release dates of We only apply local search procedures with a certain probabil-
store orders and thus impact the earliest possible start times of ity (in contrast to a GVNS) for several reasons. First, the additional
the delivery tours. The delivery tours and corresponding routes, local searches result in a significant amount of additional computa-
however, are strongly affected by the release dates of the asso- tion time per iteration. As a consequence, the total number of iter-
ciated store orders. Consequently, adjustments in the unaltered ations is greatly decreased compared to a similar ALNS. Moreover,
sub-solution could be necessary. The additional local search the advantages of the additional search opportunities may get lost
aims to derive advantages from these (possibly high-potential) due to the lower number of iterations for a certain amount of com-
temporary solutions, which could otherwise be missed. puting time. This is particularly relevant since ALNS already applies
We use two alternating Variable Neighborhood Decent algo- computationally intensive large neighborhood structures. Second, a
rithms (VND, see Hansen et al. (2010) and Section 5.5 for a detailed strong local search can block the acceptance mechanism of ALNS to
description) as a local search approach that optimize one sub- overcome local optima. Third, the local search may only be applied
solution each, i.e., VNDVRP for the VRP sub-solution and VNDOPP for high-potential solutions. However, high-potential solutions can
for the OPP sub-solution. Both VNDs are iteratively applied un- be hard to identify in the problem setting of the OB-VRPSTW.
til no improvement can be found. The local search considers the A random application allows for a controlled intensification and
previously applied destroy and repair heuristic and improves the diversification process. In pretests we found that an application
unchanged part of the solution first. More precisely, if the order probability of 30% leads to a good compromise between computing
picking sub-solution has been modified by the destroy and repair time, intensification and diversification for the OB-VRPSTW.
ejor.2020.03.075
JID: EOR
5.2. Initial solution Table 2

Neighborhood structures applied during the local search procedure.
To provide an initial solution, the following procedure is ap- VNDOPP

plied. First, the order picking problem (OPP) is solved with a sav- 1. order _shi f t _SP: shift an order to a different batch at the same picker
ings procedure (Clarke & Wright, 1964; Matusiak et al., 2017) such or build a new batch with the order on the same
that the total processing time is minimized. The batches deter- picker
mined are then successively assigned to pickers, taking into ac- 2. order _swap_SP exchange two orders assigned to the same picker
count the upper bound of the delivery time window. From all non- VNDVRP
assigned batches, the batch with the smallest upper delivery time
1. relocate: shift a store to a position on the same or another
window is assigned to the last position of the picker with the
vehicle
smallest current workload. To obtain a first solution to the vehi- 2. 2 − opt: remove two edges in a route and include two other
cle routing problem (VRP), the 3-regret heuristic is applied. This edges such that the route remains feasible
heuristic is also included as a repair heuristic in the GALNS (see 3. cross_exchange: exchange two stores assigned to different vehicles
Section 5.4).
5.3. Destroy heuristics 5.4. Repair heuristics
Vehicle Routing Domain. Five types of destroy operators are in- To include picking orders and store orders, respectively, that
cluded in the GALNS dealing with the composition of vehicle tours have been removed from a solution, three types of repair heuristics
and the corresponding routes (see also Pisinger & Ropke, 2007 and are integrated into the GALNS. The first type is a simple greedy-
Hemmelmayr, Cordeau, & Crainic, 2012). The first type is a straight- insertion procedure. For all picking (store) orders removed, the
forward remove heuristic that selects a number of stores randomly. best possible insertion position is calculated and the best move is
The second type of remove heuristic adapts the worst remove op- executed. This procedure is repeated until all picking (store) orders
erator by removing a number of stores with the greatest tardiness. are included in the solution. The second type is the well-known k-
Third, the stores with the longest waiting times are removed from regret operator (Hemmelmayr et al., 2012; Pisinger & Ropke, 2007;
an incumbent solution. If a vehicle has to wait at a store location Potvin & Rousseau, 1993). In contrast to a greedy-insertion op-
before the service can start, the corresponding store order is served erator, k-regret includes estimated information in the event that
in time, i.e., its tardiness is zero. However, waiting time may affect an order/store is not inserted in its currently best position. In
the arrival times at subsequent store locations and thus the corre- the GALNS, the 2-regret and 3-regret heuristics are included. Note
sponding tardiness. The fourth and fifth destroy operators remove that the greedy-insertion heuristic is a special type of the k-regret
stores with a certain relationship to each other (Shaw, 1997). First, heuristic with k = 1. The last type of repair heuristic, type three,
a seed order is chosen randomly. Afterwards, a number of orders inserts a picking (store) order randomly. Random insertions have
are selected that have the highest relationship value regarding the a similar character and task as shaking operations in GVNS. They
seed order. More precisely, the fourth remove heuristic selects or- usually do not improve the solution but give the local search the
ders that have the highest matching regarding arrival times. The opportunity to overcome local optima by starting from different
fifth remove heuristic selects orders with the shortest travel time solutions not covered in shaking operations. This idea adds an ad-
between the locations of the seed order and their respective store ditional search opportunity to the ALNS framework.
locations.
Order Picking Domain. We use five types of destroy heuristics in 5.5. Local search
the order picking domain. The first two types are similar to the
VRP and also applied by Žulj et al. (2018). First, a number of store Both VNDs applied in the local search, i.e., VNDVRP and VNDOPP ,
orders are randomly chosen and removed from the current solu- follow the standard VND framework. The general idea of the VND
tion. Second, a certain number of orders that are responsible for metaheuristic (see Hansen & Mladenović, 2001; Hansen et al.,
the greatest tardiness are removed. Moreover, two additional re- 2010) is a systematic and deterministic change of neighborhood
move heuristic types are introduced removing complete batches structures. In our application, a neighborhood is examined com-
from the incumbent solution. The third type chooses a number of pletely. If an improvement has been found within a neighborhood,
batches randomly, while the fourth removes a number of batches the first neighborhood for the new incumbent solution is applied
with the greatest tardiness. Type five removes randomly chosen or- again. If no improvement has been found, the next neighborhood
ders from the order picking solution assigned to a seed vehicle. is investigated. The algorithm terminates if all neighborhoods for
The heuristic first defines a random vehicle; it then selects orders the incumbent solution are examined without finding an improve-
of this vehicle and removes them from the picking solution. ment. Table 2 lists the neighborhood structures used when ap-
Additionally, two neighborhood-based shaking moves are com- plying the respective VND approaches (Hemmelmayr et al., 2012;
bined with a destroy heuristic. More precisely, the order-based ran- Henn, 2015).
dom remove operator is combined with a batch-shift move that
shifts a randomly chosen batch to a different picker. Finally, a 5.6. Weight updating mechanism, acceptance and termination criteria
batch-swap move that exchanges two batches assigned to differ-
ent pickers is performed after an application of the batch-based One fundamental property of an ALNS algorithm is the self-
random remove heuristic. As described in Section 5.1, GALNS com- adjustment mechanism enabling the algorithm to evaluate the per-
bines elements of ALNS and GVNS. At this stage, shaking opera- formance of the destroy and repair heuristics. According to the
tions (shift and exchange, GALNS) are combined with random re- weights assigned, a heuristic is chosen to be more or less likely
move operations (ALNS). The combination of both has the advan- within the search process. The weights are updated according to
tage that, especially at a late stage, the repair heuristics (and local their performance. For this purpose, the temporary solution is eval-
search) do not just repair (optimize) similar solutions frequently uated by an acceptance criterion that differentiates three states
but also slightly modified solution since shaking moves may not (Pisinger & Ropke, 2007). First, a new best solution is always ac-
be undone by repair heuristics. In total, this provides an additional cepted, and thus the current-best and the incumbent solution are
opportunity to escape local optima. updated, respectively. Second, if the temporary solution does not
ejor.2020.03.075
JID: EOR
Table 3 ble states for a temporary solution, Ttemp regarding the current-
States for a temporary solution within the GALNS.
best, Tbest , and the incumbent solution, Tinc .
State Consequence Weight The GALNS is terminated as soon as a certain time limit has
parameter been reached.
1) Ttemp < Tbest best and incumbent solution are updated ω1
2) Ttemp < Tinc · γ incumbent solution is updated 6. Benchmark approaches
2a) Ttemp < Tinc ω2
2b) Ttemp ≥ Tinc ω3
3) else temporary solution is rejected ω4
The solution quality of the GALNS approach developed is an-
alyzed by comparing the results achieved with those of suitable
benchmark approaches. We develop and suggest two alternative
heuristic approaches since we are not aware of specialized ap-
proaches from the literature that solve the OB-VRPSTW. The first
represent a new best solution, noise is applied to the objective approach is based on a problem-tailored standard ALNS algorithm
function value of the incumbent solution. The third and last state (Section 6.1). The second approach is a sequential approach that
rejects the temporary solution. solves both problems one after the other (Section 6.2). In addition,
In order to apply noise to the objective function value of the we use a standard MIP solver to solve the mathematical formula-
incumbent solution, it is multiplied by γ , which represents the tion of OB-VRPSTW for small instances at least.
realization of a random variable ∈ γ , γ , with γ ≤ 1 and γ ≥ 1.
Consequently, a superior (temporary) solution is not always ac- 6.1. ALNS benchmark approach
cepted (γ < 1) and an inferior solution not always rejected (γ > 1),
allowing for diversification and intensification within the search. The GALNS algorithm proposed in Section 5 is derived from
In contrast to Pisinger and Ropke (2007), the magnitude of the an ALNS algorithm and elements of the GVNS metaheuristic. We
noise directly depends on one or both objective values of the in- therefore suggest a similarly designed standard ALNS approach as
cumbent or best solution in our algorithm. Moreover, we combine one of the benchmark approaches. The ALNS benchmark algorithm
the noise with elements of threshold accepting (Dueck & Scheuer, includes the same destroy and repair heuristics, and also the iden-
1990) that have performed very well in similar problem settings tical acceptance and update mechanisms as the GALNS. However,
for neighborhood-based algorithms (Schubert et al., 2019; Schubert the local search is excluded. This seems reasonable for three rea-
et al., 2018). Compared to the simulated annealing acceptance cri- sons. First, ALNS has proved to provide excellent results when solv-
terion that decreases the probability of accepting weaker solutions ing both problems (see Section 5). Second, possible improvement
continuously, threshold accepting allows for intensification and di- of the enhanced ALNS, i.e., the GALNS, becomes visible when com-
versification depending on the near-term search history. More preparing it to its original version. Third, a GVNS algorithm that was
cisely, if a new best (or incumbent) solution has been found, the developed did not achieve the performance of ALNS.
adjacent search space is investigated intensively since the thresh-
old equals zero (or is close to it). After a certain number of iter- 6.2. Sequential benchmark approach
ations without improvements, the parameters γ and γ increase,
and thus weaker solutions may be accepted or slight improve- The sequential benchmark approach solves the OPP and the VRP
ments of the incumbent solution rejected. more or less independently of each other. This procedure is based
We propose a further segmentation of γ and γ into two sets on implementations in practice; however, it includes sophisticated
of terms each. For each parameter, one set of terms, i.e., γ inc and state-of-the-art algorithms to solve the OPP and VRP instead of
γ inc , depends on the search history regarding the current incum- manual decisions by dispatchers.
bent solution and the other set, i.e., γ best and γ best , is based on In practice, delivery operations are usually planned first while
the search history with respect to the best-known solution at run the picking orders are planned second, taking into account the de-
time. In detail, γ = 1 − γ −γ and γ = 1 + γ best + γ inc . Terms parture times of the associated delivery tours. However, solving the
best inc
γ inc and γ inc as well as γ best and γ best are updated after a cer- delivery problem, i.e., the VRP, requires at least a rough approxi-
tain number of iterations without finding an incumbent and best mation of the release dates of the delivery orders since the release
solution, respectively. If the incumbent or best solution has been dates are input parameters for the delivery plan. We approximate
improved, the corresponding parameters are reset to their initial these release dates via a preliminary solution to the order picking
values, e.g., zero. problem using a savings-based approach, which minimizes the to-
At the end of each iteration, both approaches, GALNS and tal processing times of all orders. Afterwards, batches are assigned
ALNS, adjust the weights of the destroy and repair heuristics to pickers using the largest processing time rule to minimize the
respectively applied. The weights of unused heuristics, how- makespan. The VRP and subsequently the OPP are solved after that.
ever, remain unchanged. We introduce four weight parameters The approximation proposed outperformed other precalculations
(ω1 > ω2 > ω3 > ω4 ) (Pisinger & Ropke, 2010). These weight pa- of release dates, i.e., neglecting picking operations when design-
rameters are chosen according to the states previously described. ing routes or assuming relatively late release dates when solving
To do this, the second state is further subdivided, i.e., Ttemp < Tinc the VRP, which are thus not further considered in the following.
(2a) and Ttemp ≥ Tinc (2b), where Ttemp refers to the temporary and The procedure of solving the OPP also considers the individual
Tinc to the incumbent solution. Correspondingly, weight parame- delivery sequences of all tours, which were generated as part of
ter ω1 is applied for state (1), weight parameter ω2 and ω3 for the previously solved VRP algorithm. To do this, the OPP algorithm
state (2a) and (2b), respectively. If the temporary solution is re- takes into account the effect of the release dates of delivery orders
jected, the weight parameter ω4 is chosen. Let be the weight on the departure time of the respective vehicles, and thus on the
parameter to apply according to the objective function value of arrival times and tardiness times of the associated orders at their
the temporary solution in a certain iteration and λ ∈ [0, 1] a de- store locations. Pretests showed that this procedure leads to sig-
cay parameter. Then, the weights of the destroy (repair) heuris- nificantly superior results compared to alternative sequential ap-
(i )
tic applied in iteration (i + 1 ) are adjusted as follows, with des j proaches.
as the weight of the chosen (jth) destroy heuristic in iteration (i): Solution for the Vehicle Routing Problem. We adapt the Unified
des(i+1 ) (i )
j
= λ · des
j
+ (1 − λ ) · . Table 3 summarizes the possi- Tabu Search (UTS) approach suggested by Fu et al. (2008) in or-
ejor.2020.03.075
JID: EOR
der to solve the VRP. The UTS approach represents a state-of-the- Table 4
Parameter settings.
art algorithm for VRPs with soft time window constraints (see
Section 3.2). Parameter Characteristic
We adjust the original UTS approach in respect of two issues. Small instances Large instances
First, we include the earliest possible starting times of the vehi-
Number of stores |I| ∈ {5, 7, 10, 15, 20, 25} {50, 100, 200}
cles, since the original UTS approach does not consider the release Number of pickers |M| ∈ {1, 2} {3, 6}
dates of orders. Second, we skip the moves that try to reduce the Batching opportunities fbatch ∈ {1, 2} {low, high}
number of vehicles for all neighborhood structures, since the num- Capacity vehicles w ∈ {5, 10} {30, 60}
ber of vehicles used is irrelevant in the OB-VRPSTW. Size of distribution area fsda ∈ {50} {50, 100, 200}
Tightness factor ft ∈ {tight} {tight, wide}
Solution for the Order Picking Problem. We adjust the solution
Closeness factor fc ∈ {narrow} {narrow, spread}
approach suggested by Henn (2015) in order to solve the OPP.
Henn (2015) developed a VNS algorithm that minimizes the total
tardiness of all picking orders. In our case, however, the tardiness Table 5
Improvements of GALNS approach related to diverse benchmark approaches for
of a picking order possibly affects the scheduled arrival time at the small instances.
store locations of the corresponding delivery tour and accordingly
the tardiness times of the associated store orders. The total tardi- Number of Percentage improvement of GALNS regarding Opt. GAP
ness therefore depends on the solution of the VRP as well. Conse- orders ALNS SEQ GALNSDP SEQDP CPLEX CPLEX
quently, we implement the VNS of Henn (2015) using the objective 5 0.00% 0.48% 6.26% 6.62% 0.00% 0.00%
function of the OB-VRPSTW. Note that a solution to the VRP is nec- 7 0.15% 14.04% 25.00% 32.67% 0.00% 0.00%
essary for this purpose. 10 0.00% 13.57% 22.87% 35.64% 0.00% 4.59%
15 1.60% 19.04% 38.31% 54.78% 15.33% 68.44%
20 8.55% 57.73% 65.23% 82.41% 51.49% 99.31%
7. Numerical experiments 25 17.26% 53.56% 65.03% 80.76% 54.63% 99.93%
Average 4.13% 28.92% 41.29% 60.45% 24.30% 45.38%
This section presents the numerical results achieved when
applying the various approaches developed for solving the OB-
VRPSTW. The analysis uses simulated data (Section 7.1) and data
Implementation Issues. The different algorithms analyzed are
from a real-life case study of a major German grocery retailer
coded in C++ and optimized with the g++ -O3 setting. The ex-
(Section 7.2). In both cases we derive recommendations, i.e., man-
periments were performed on a Linux cluster with Intel Xenon
agerial insights, for grocery retailers operating in the context as-
E5-2697 v3 CPUs. We used ILOG OPL CPLEX Optimization Studio
sumed.
12.9 (CPLEX) to implement and test the model formulation from
Section 4. These tests were performed on an Intel Core I7-6700HQ
7.1. Simulated data experiments
processor.
For all scenarios, we calculate the percentage improvements
The simulated data experiments are structured as follows. We
(reduction in total tardiness times) of the solution achieved by
first define the data set applied and describe additional details rel-
procedure GALNS and the results achieved by the respective
evant for the implementation of the algorithms applied. Second, T[bench] −TGALNS
we compare the results of the GALNS algorithm with those of the benchmark approaches ([bench]), i.e., T[bench] · 100 [%]. The
problem-tailored standard ALNS and the results of the sequential problem-tailored standard ALNS described in Section 6.1 will be
solution approach, SEQ. These analyses disclose the potential ad- denoted ALNS and the sequential solution approach presented in
vantages of an integrated planning approach. Finally, we evaluate Section 6.2 is termed SEQ. If SEQ or GALNS apply a discrete picking
different alternatives to design the planning and picking system. strategy, a corresponding subscript is added, i.e., GALNSDP , SEQDP .
7.1.1. Data structure and implementation details 7.1.2. Performance evaluation of GALNS for small instances
Problem Instances. The simulated problem instances follow the Small-size instances have been generated to evaluate the appli-
data sets applied in related studies (e.g., Schubert et al., 2019; cability of the model formulation proposed in Section 4. Moreover,
Schubert et al., 2018; Ullrich, 2013) and the data observed at our we obtain a first indicator for the performance of GALNS and the
case company. We only display the main set of parameters here. proposed benchmark algorithms. From the problem classes pre-
A technical appendix lists the domains of all additional parameters sented in Table 4 we fix |M| = 2, fbatch = 2, and w = 5 to test the
required. model formulation applicability while maintaining the typical char-
The main parameters defining the set of problem instances are acteristics of the OB-VRPSTW, i.e., batching opportunities and the
the following: the number of stores |I|, the number of order pick- sequencing and assignment decision in both order picking and ve-
ers |M|, the batching opportunities fbatch , the capacity of vehicles hicle routing. We use problem instances with 5, 7, 10, 15, 20, and
w, the size of the distribution area fsda , the width of delivery time 25 store orders. CPLEX ran for 30 minutes (12 hours) for instances
windows (tightness factor ft ), and the time gap between the start with 5, 7, and 10 (15, 20, and 25) store orders while the remaining
of the planning horizon and the delivery time windows (closeness algorithms ran 5 seconds. Table 5 shows the percentage improve-
factor fc ). We use these parameters for two different numerical ment of target values achieved by the GALNS approach compared
tests, i.e., small-size and large-size (practical-oriented) instances. to the results achieved by CPLEX and a variety of additional bench-
The ranges of these parameter values are summarized in Table 4. mark approaches (see Section 6).
Please note that fbatch is measured differently for small and large CPLEX results in optimal solutions for five and seven store or-
instances. For large instances, the capacity of the picking device is ders for all instances tested. For ten store orders, seven out of ten
varied in combination with the number of items per roll cage. In instances were solved to optimality. CPLEX then struggles to find
contrast, a roll cage carries up to one or two customer orders in any reasonable bound. This can be seen in the optimality gap for
small instances. Similarly, the vehicle capacity is measured in roll 15, 20 and 25 instances, which is close to or equal to 100%. Sim-
cages for large instances and in store orders for small instances. ilarly, the solution quality strongly decreases for 15 or more store
For parameter fSDA , the edge length in kilometers is given. We ran- orders. This is true as well for all remaining solution approaches
domly generate ten instances for each problem class. compared to GALNS. Especially the performance of ALNS indicates
ejor.2020.03.075
JID: EOR
Table 6 pickers has a more limited effect than if release dates are rather
Percentage improvement of target values achieved by CPLEX assuming the order
late.
batching (OB) strategy and the availability of two vehicles and two pickers com-
pared to different parameter settings, two picking strategies (DP and OB), and the A lower number of vehicles, i.e., one instead of two, increases
sequential solution approach (SEQ). total tardiness by 195.50%. The number of vehicles impacts the
problem structure. If two vehicles are available, order provision
Characteristic CPLEX Average SEQ Average
times have to be coordinated to ensure optimal start times. How-
DP OB DP OB ever, if there is only one delivery vehicle available, only the provi-
1 picker 85.59% 81.20% 83.39% 85.85% 81.35% 83.60% sion time of the last order, i.e., the makespan, is relevant. This re-
1 vehicle 2 pickers 77.94% 75.00% 76.47% 78.60% 75.70% 77.15% sult is also affected by order batching operations. More precisely, if
Average 81.77% 78.10% 79.93% 82.23% 78.52% 80.38% batching is allowed the objective value increases by 219.97%, while
1 picker 64.63% 47.70% 56.17% 69.91% 54.96% 62.43%
it increases by 178.23% if batching is prohibited. Allowing for or-
2 vehicles 2 pickers 22.87% 0.00% 11.43% 32.56% 13.57% 23.06% der batching operations makes it possible to provide several orders
(here: two) at the same time. This leads not only to a reduced total
Average 43.75% 23.85% 33.80% 51.23% 34.26% 42.75%
processing time but also offers the possibility of providing orders
before their scheduled delivery times.
Comparing objective values of SEQ and CPLEX shows that SEQ
that GALNS results in a superior coordination of both problems works very well for one-vehicle cases. In these cases, the mini-
if complexity increases. This is in particular a success of the ad- mum makespan in picking operations ensures an optimal vehicle
ditional local search. GALNS finds at least the same solutions as start time (independently of the actual route). In this case the or-
CPLEX for all instances tested. For 15 or more customer orders, der picking problem can be solved first, independently from rout-
GALNS and CPLEX only result in the same objective value in two ing determining reliable provision times for delivery planning. As
instances; for all remaining instances GALNS outperforms CPLEX by a consequence, SEQ performs much better if only one vehicle is
up to 604.88%. available since the approximation of order provision times mini-
We follow the experiment structure of Moons et al. (2018) and mizes the total processing time first and then the makespan. In
extend the ten-order instances to gain insights from the model detail, the deviation from the target value of CPLEX is only 2.16%
formulation. We evaluate the impact of the picking strategy (dis- if one vehicle is available, but 15.93% for two vehicles. This result
crete picking, DP, vs. order batching, OB), number of pickers (one does not depend on whether order batching is allowed or not. Note
vs. two), number of vehicles (one vs. two) and the solution ap- that this argumentation may not hold for independent planning of
proaches (CPLEX vs. SEQ). We use the ten-order instances assum- both problems. Providing orders such that they may arrive before
ing the order batching strategy and the availability of two vehi- their lower delivery time window bound leads to different rout-
cles and two pickers as a basis. Solving these instances by CPLEX ing solutions than if orders are provided after their upper delivery
leads to an average total tardiness value of 412.20 minutes, which time window bound.
denotes the minimum target value achievable within the current
analysis. We calculate the results achieved when assuming the dif- 7.1.3. Performance evaluation of GALNS for large instances
ferent parameter settings or applying one out of the two picking The present section analyzes the performance of the solution
strategies and solution approaches, respectively. We then quantify procedure GALNS. The combinatorial complexity is highly influ-
the respective percentage improvements when assuming the basic enced by the number of store orders, |I|. We therefore differenti-
scenario that is solved by CPLEX. Table 6 displays these percentage ate the allocated computational times for each algorithm accord-
improvements. ingly, i.e., 20, 40, and 60 minutes for instances with |I| ∈ {50, 100,
The improvements of CPLEX assuming the base scenario range 200} store orders, respectively. The sequential solution approach,
from 13.57%, i.e., compared to the SEQ solution approach, and in contrast, receives equal halves of these computational times for
85.85%, i.e., compared to the SEQ solution approach when assum- solving the OBP and the VRP, respectively. We compare the re-
ing one picker and one vehicle and applying the discrete picking sults of GALNS with the results achieved by ALNS and SEQ. Table 7
(DP) strategy. presents the results for a diverse set of problem parameters. The
The following general observations can be made from the re- parameters are varied according to the problem characteristics in-
sults achieved. Please note that the associated numerical values troduced in Section 7.1.1 and listed in Table 4.
mentioned below are not directly visible in Table 6. Prohibiting or- The GALNS improves the target values by 29.35% and 57.79%
der batching leads to an increase in tardiness of 32.37% due to the on average compared to the ALNS and SEQ approach, respectively.
greater order picking time. The number of vehicles available im- Thus, the GALNS outperforms the ALNS, and the ALNS, again, out-
pacts this value, i.e., one vehicle and discrete picking operations classes the SEQ approach. This general observation is valid for all
lead to a 23.13% higher total tardiness compared to order batching subsets of problem instances. These results particularly demon-
operations, while it amounts to 41.60% in the two-vehicle environ- strate the advantage of an integrated against a sequential planning
ment. approach.
The reduction of order pickers, i.e., from two to one, results The advantage of GALNS against ALNS is also based on im-
in an increasing objective value. The total tardiness is increased proved coordination between order picking and vehicle routing de-
by 73.82% due to the later order provision times. This result is cisions. The average total picking times, for example, are 0.15%
stronger in two-vehicle cases. If more vehicles are available, ur- smaller by ALNS than by GALNS solutions which, however, only
gent customers may be provided first by the pickers and can be means a difference of 1.05 minutes. It might be concluded that ve-
delivered on an early tour. If order batching is not allowed, the in- hicles can depart earlier and therefore shorter delivery tours can
crease in tardiness amounts to 118.08% (discrete picking) compared be found by ALNS than by GALNS. But the opposite is true. The av-
to 91.22% (order batching) since the total processing time is longer, erage total travel time per instance turns out to be 4.49% shorter
and picking operations therefore provide orders later. Moreover, in GALNS than in ALNS solutions. This supports the hypothesis that
the shorter total processing time when allowing order batching the additional local search procedure added to the ALNS, which re-
reduces the latitude for improvements, i.e., if order release dates sults in the GALNS approach, apparently improves this coordina-
are rather tight due to short processing times, additional reduction and leads to the reduction in total tardiness time mentioned
tion of the release dates by distributing the workload to different above.
ejor.2020.03.075
JID: EOR
Table 7 The results confirm the intuition that the value of integration
Percentage improvements of total tardiness time of GALNS compared to ALNS and
increases if the relative workload for picking and/or delivery in-
SEQ depending on diverse problem parameters.
creases during the planning horizon, i.e., the cases of a small num-
Parameter ALNS SEQ ber of order pickers, low batching possibilities, low vehicle capacity
50 10.39% 36.38% and large distribution areas. The results also show that increased
|I| 100 30.26% 58.91% planning complexity substantially drives the importance of an in-
200 47.39% 78.07% tegrated planning approach, i.e., a high number of orders, wide de-
3 31.73% 63.84% livery time windows and a long planning horizon. In contrast to
|M|
6 26.96% 51.74% the findings mentioned above, the latter findings are not those one
low 32.79% 65.43% would obviously expect, especially with respect to delivery time
fbatch
high 25.91% 50.15% windows and planning horizon.
30 36.96% 68.06%
Comparing the results of GALNS and ALNS in more detail leads
w to the following insights. Store orders reveal an average tardiness
60 21.74% 47.51%
of 54.41 and 64.35 minutes and a median of 28.49 and 39.27, re-
50 26.55% 55.20%
fsda 100 30.10% 58.95% spectively. This means the tardiness per store order decreases by
200 31.39% 59.21% 9.94 minutes on average. The average number of tardy orders is
also reduced from 53.66 to 46.99 by GALNS compared to ALNS.
tight 28.44% 56.14%
ft GALNS outperforms ALNS in 96.35% of all instances tested. Out of
wide 30.26% 59.44%
all 2880 instances, ALNS only generates superior solutions in 37
narrow 20.89% 47.97%
fc
spread 37.81% 67.60%
instances; in 68 instances both approaches yield equal results.
Correspondingly, we compare the results of GALNS and SEQ.
Average 29.35% 57.79%
The number of tardy orders is decreased by 22.22% when apply-
ing GALNS, although SEQ reduces the total picking time and the
total travel time compared to GALNS by 5.33% and 2.70%, respec-
tively. However, SEQ leads to longer vehicle waiting times: these
These general observations are supported when analyzing the
values are 13.19% higher. The fundamental drawback of SEQ is that
following parameter settings, in which coordination between order
the initial solution to the order picking problem only pursues the
picking and delivery planning, i.e., the value of integration, seems
objective of minimizing the total picking times. However the sub-
to be particularly relevant, and therefore GALNS leads to noticeably
sequent delivery problem, i.e., the structure of the store locations
better solutions than ALNS and SEQ:
and the corresponding time window distribution, is neglected. As a
result, store orders that could be favorably combined to a tour are
• A larger number of orders, |I| = 50, |I| = 100 and |I| = 200, ex- not taken into consideration. However, such an approach is widely
pands the combinatorial complexity of the problem considered used in both literature and practice. For example, the order batch-
and therefore the savings achievable also increase, i.e., 10.39%, ing problem is usually solved with the goal of minimizing the over-
30.26% and 47.39% compared with ALNS and 36.38%, 58.91% and all processing times. van Gils et al. (2018) state in their review re-
78.07 % compared with SEQ. garding combined order picking problems that 54 out of 73 papers
• A smaller number of order pickers, e.g., |M| = 3 compared use the order picking time as performance indicator. Our results
to |M| = 6, increases the release dates of delivery orders and indicate that the timely provision of delivery orders combined into
therefore postpones the earliest possible starting time of vehi- an appropriate tour – possibly at the expense of increasing pick
cles, and thus complicates the building of tours. times – is a more important objective.
• A reduced batching opportunity, e.g., fbatch = low vs. fbatch =
high, enlarges the picking times, which ultimately also post- 7.1.4. Development of solution quality over time
pones the earliest possible starting time of the vehicles. In this section we analyze the development of the solution
• A reduced vehicle capacity, e.g., w = 30 compared with w = 60, quality of GALNS and ALNS across the respective computational
requires more vehicles fulfilling the delivery tours, which in- times assigned. In doing so, we quadruple the computation times
creases the coordination effort. predefined in the analyses above. As a result, we set the time limit
• A larger size of the distribution area, e.g., fsda = 100 vs. fsda = to 80, 160, and 240 minutes for instances with |I| ∈ {50, 100, 200}
50, complicates the delivery planning, and therefore the coor- store orders, respectively. The enlarged computation times also an-
dination between order picking and delivery planning becomes swer the question, as to whether longer computational times can
more important. further improve the objective values attainable. GALNS and ALNS
• A larger delivery time window, e.g., ft = wide vs. ft = tight, in- further improve their objective values by 4.57% and 2.93%, respec-
creases the possibilities of building undelayed delivery tours, tively, compared to the results achieved in Section 7.1.3. In absolute
thus increasing the combinatorial complexity of the entire plan- values of tardiness, this corresponds to 1.30 and 2.62 minutes per
ning problem, which again fosters an integrated planning ap- order and instance, respectively.
proach. Fig. 3 shows the percentage deviation of the current best objec-
• An enlarged planning horizon, which leads to more spread de- tive value from the best-known objective value depending on the
livery time windows, e.g., fc = spread vs. fc = narrow, compli- actual percentage of the entire run time provided for GALNS (left-
cates the delivery planning due to the binding lower bounds of hand side) and ALNS (right-hand side). The best-known objective
the delivery time windows. This results in more room for rel- value represents the value of the best solution of both algorithm
ative improvement, i.e., 37.81% to 20.89% (GALNS to ALNS) and GALNS or – in very rare cases – of algorithm ALNS achieved at
67.60% to 47.97% (GALNS to SEQ). Closer time windows, how- the end of the entire computing time. It represents the benchmark
ever, offer significantly greater opportunities for absolute im- for both algorithms. Figure 3 does not include the solutions of
provements in respect of the absolute tardiness per order, i.e., the first percentage of the run time to obtain a meaningful graph.
4.46 to 18.77 min. (GALNS to ALNS) and 36.86 to 86.37 min. Both algorithms, GALNS and ALNS, start from the same initial solu-
(GALNS to SEQ). In this case it is much harder to meet the up- tions. The initial solutions, however, result in relatively unsatisfac-
per bounds of the delivery time windows. tory objective values. These objective values are 25.73, 43.79 and
ejor.2020.03.075
JID: EOR
Fig. 3. Percentage deviation of the current best objective value from the best-known objective value depending on the actual computation time.
71.65 times higher than the best target value achieved for |I| ∈ {50, An order batching strategy reduces the average total tardi-
10 0, 20 0} store orders, respectively. We exclude outlier instances ness compared to a discrete picking strategy, INTDP and SEQDP , by
here impacting the results disproportionately great, i.e., possess- 54.95% and 56.55%, respectively. The ability to batch orders there-
ing a very small final target value (less than 0.5 minutes per fore allows significantly increased performance of the entire order
order). fulfillment system in terms of punctual deliveries.
The results illustrated in Figure 3 indicate that both the GALNS The total picking times decrease in the case of an integrated
and ALNS algorithms significantly increase the solution quality and a sequential planning approach by 37.81% and 40.60%, respec-
within the first minutes of their run time. However, GALNS outper- tively, when applying order batching instead of discrete picking.
forms ALNS. GALNS provides a significantly lower target value after The reason behind this is that the entire walking distance to be
the first percent of the entire run time provided (see the start of covered by each picker significantly decreases. As a consequence,
the curves), which points to the success of the local search pro- orders are provided sooner, which allows the vehicles to depart
cedure that has been added. GALNS also improves the solutions earlier, which in turn decreases total tardiness.
achieved much faster than ALNS. GALNS therefore provides “ac- Nevertheless, the waiting times at the store locations are re-
ceptable” solutions within a much shorter run time than ALNS. duced when applying a discrete picking compared to an order
batching strategy, i.e., by 5.02 minutes per store (42.48%) in the
7.1.5. Analyzing different design alternatives case of an integrated planning approach, and by 0.85 minutes
The previous subsection especially analyzes the performance (6.04%) in the case of a sequential planning approach. Since the
capability of the solution algorithm developed, i.e., the GALNS pro- tours of the vehicles start later on average, they arrive later at store
cedure. In doing so, it also demonstrates the advantage of inte- locations and therefore are less often in conflict with the lower
grated planning over a sequential planning approach, but assumes bound of the delivery time window of the associated store orders.
a specific picking strategy in both cases, i.e., a multiple-order- Travel times are therefore reduced by 14.74% (INT) and 4.28% (SEQ)
picking (order batching) strategy. The present section continues on average, compared with the respective order batching strate-
this comparison by considering a further picking concept, i.e., a gies.
single-order-picking (discrete picking) strategy. Thus, we now focus Cross Comparison of Planning Approaches and Picking Strategies.
on different planning and picking alternatives simultaneously, i.e., The previous paragraph demonstrates the advantage of multiple-
sequential (SEQ) and integrated (INT) planning as well as discrete order-picking (order batching, OB) over single-order-picking (dis-
picking (DP) and order batching (OB). Combining these design al- crete picking, DP). However, each comparison assumes a certain
ternatives leads to four distinguishable scenarios that are denoted base scenario. It is thus of interest to know which improvements
as SEQDP , SEQOB , INTDP and INTOB . Note that the SEQ[...] concepts can be realized starting from different base scenarios and ap-
apply the procedure described in Section 6.2 and are so far de- plying both or possibly only one of these concepts, respectively.
noted as SEQ. The INT[...] concepts, however, apply the GALNS pro- Fig. 4 depicts the aggregated results of all possible comparisons
cedure as described in Section 5. The analysis considers the two between those concepts.
picking strategies. The sequential planning approach assuming a discrete picking
Order Batching versus Discrete Picking. This paragraph analyzes strategy (SEQDP ) constitutes the absolute base scenario. Starting
the impact of an order batching strategy compared to a discrete from here, applying an integrative planning approach (INTDP ) or
picking strategy. The capacity of the picking device in model OB- an order batching strategy (SEQOB ) reduces the total tardiness by
VRPSTW is therefore set to one store order. Now, an order picker an average of 52.17% and 56.55%, respectively. Applying both con-
is only allowed to process one order at a time while the other cepts simultaneously (INTOB ) leads to an even higher reduction of
assumptions remain unchanged. This drives discrete picking while 76.06%.
all other parameter values, e.g., store locations, time windows, or The results presented in Fig. 4 provide further insights. Apply-
items demanded per store order, are left unchanged. The proce- ing an order batching instead of a discrete picking strategy (down-
dures GALNS and SEQ are then applied to these modified instances. ward development in Fig. 4) leads to almost the same savings re-
The results are denoted as INTDP and SEQDP , respectively. The re- gardless of which planning approach is applied, i.e., 56.55% and
sults of INTDP (SEQDP ) are compared to the INTOB (SEQOB ) solu- 54.95%, respectively. However, applying an integrated rather than a
tions for the original problem instances, which allow order batch- sequential planning approach (development to the right in Fig. 4)
ing when picking the store orders in the warehouse. Please note offers slightly greater percentage savings when already starting
that the picking device can carry four or six roll cages in the orig- from an order batching rather than from a discrete picking strat-
inal instances. egy, i.e., 57.79% vs. 52.17%.
ejor.2020.03.075
JID: EOR
Table 8
Number of stores assigned to a DC depending on the delivery radius and the DC
network.
Radius 100 150 200
DC1 38 57 71
DC2 144 167 217
DC3 55 78 92
Table 9
Normalized results of different scenarios for real case instances, base case values
(SEQDP ) set to 100.
Performance measure SEQDP SEQOB INTDP INTOB
Total tardiness 100.00 55.50 44.84 31.68

Fig. 4. Cross comparison of different concepts by percentage improvements of av- Number of tardy orders 100.00 83.35 89.08 73.07
erage total tardiness time.
An additional question of interest is which move – downwards of stores assigned to a DC depending on delivery radius and the
or to the right – should be preferred when starting from the ab- DC network chosen, respectively.
solute base scenario, i.e., from the sequential planning approach, In addition, the retailer provides historical data of the daily de-
assuming a discrete picking strategy (SEQDP ). Comparing the aver- mands of the respective micro stores for a period of four weeks.
age savings, 56.55% vs. 52.17%, or the average tardiness per order, The data contain the number of articles and cages per store order
105.89 min. vs. 107.22 min., leads one to expect a slight advan- as well as discrete picking times. The generation of batch picking
tage if an order batching strategy is implemented first, and an in- times requires an adequate model of the warehouse and the pick-
tegrated planning approach second. The median tardiness per or- ing process. This model includes the setup time per picking tour,
der, however, is lower when applying an integrated planning ap- the distances between storage locations, the movement speed of
proach (INTDP ) instead of an order batching strategy (SEQOB ), i.e., order pickers, the time per pick, and the storage assignment policy
77.05 min. vs. 80.61 min. As a consequence INTDP generates su- used by the retailer, who essentially applies a class-based ABC pol-
perior solutions in 59.06% of all problem instances, i.e., 1701 out icy. From this model we quantify all relevant picking times, includ-
of 2880 cases. Two cases lead to identical objective values. How- ing the discrete picking times. We verify these times by the times
ever, the order batching strategy (SEQOB ) reveals a lower variance provided by the retailer. Fig. 6 shows the distribution of discrete
of the objective values achieved than the integrated planning ap- picking times generated, which in general are in the magnitude of
proach (INTDP ) in our problem setting. Thus, the decision as to minutes. However, picking times of almost half an hour can also
which concept – integrated planning or order batching – should be observed.
be implemented first highly depends on the specific circumstances Altogether we generated 90 distinguishable problem instances
(parameter settings), and on the particular preferences of the man- representing a realistic and broad spectrum of observable planning
agement concerned. situations at our case company.
Results. Similar to above, we analyze the real-life case data us-
7.2. Real-life case ing the two different planning approaches, i.e., sequential plan-
ning (SEQ) and integrated planning (INT), and the two different
The last part of the numerical experiments is dedicated to a picking strategies, i.e., discrete picking (DP) and order batching
real-life case study of a large German grocery retailer operating (OB), which results in four different scenarios denoted as SEQDP ,
micro stores, classical supermarkets as well as hypermarkets. The SEQOB , INTDP and INTOB , respectively. Table 9 presents the results
study, however, focuses on the fulfillment of micro-store orders. achieved, whereby the results of the sequential planning approach
The retailer currently applies a discrete picking strategy even when assuming a discrete picking strategy, i.e., the absolute base scenario
picking micro-store orders. These orders typically contain one to a (SEQDP ), are set to 100, and the other scenarios are normalized ac-
maximum of three roll cage(s) only. The retailer investigated the cordingly. Note that the base scenario represents the actual situa-
benefits of an order batching strategy for those store orders. In ad- tion at our case company.
dition, the retailer currently schedules the picking and delivery of In general, the given results confirm the findings achieved in
orders independently of each other. Analyzing the advantage of an the previous subsection analyzing simulated data. Applying an or-
integrated planning approach is therefore also of great value for der batching strategy instead of a discrete picking strategy leads to
the retailer. In the following we first describe the data provided by an improvement of the total tardiness time by 44.50%, even though
the retailer in an anonymous manner. Second, the results of the we still apply the sequential planning approach. Moving from a se-
solution approaches are presented and managerial insights are de- quential to an integrated planning approach but keeping the dis-
rived. crete picking strategy will improve total tardiness time by 55.16%.
Data Set. The data set comprises historical data of three regional Applying both concepts simultaneously, i.e., INTOB , leads to a re-
DCs of the retailer including demand, delivery and order picking duction in total tardiness of 68.32%.
information. Fig. 5 illustrates the distribution network of the three The differences of the INTOB to other approaches are slightly
DCs investigated in the case study. The store locations, i.e., the mi- smaller in the experiments with practical data than in the cases
cro stores, are depicted by dots and the DCs by triangles. using simulated data. This is caused by a different ratio of pick-
We define three delivery radius around the DC, i.e., 100, 150, ing time to total workload. The proportion of total picking time for
and 200 kilometers, in order to generate different problem classes simulated data (Section 7.1) amounts to 23.68% of the total work-
from the actual network configuration of the retailer. Each prob- load, i.e., the sum of total picking time, total travel time, and total
lem class only considers those stores that are located within the waiting time, for INTOB solutions. This ratio is roughly half con-
respective delivery radius. The travel distances between stores are sidering the instances here. The reduced share of picking times on
calculated using the Euclidean metric. Table 8 displays the number the total workload will generally decrease the release dates of or-
ejor.2020.03.075
JID: EOR
Fig. 5. Distribution networks considered in the real-life case study.
sult. Consequently, management has to carefully balance which so-

lution seems more appropriate if both aims cannot be fulfilled si-
multaneously.
The above results generally emphasize the need for a sophisti-
cated planning, picking and delivery system in the context of the
supply of grocery micro stores. The problem setting investigated is
characterized by a close time-related dependency of order picking
and order delivery, i.e., vehicles need to be available to start their
delivery tours as soon as order picking begins and punctual deliv-
ery within a narrow time frame is requested. An efficient picking
Fig. 6. Distribution of discrete picking times observable in the case study. system that allows for reduced picking times is therefore just as
important as sophisticated coordination between order picking and
delivery planning.
ders for delivery operations with the result that tour building is
less restricted by picking operations than in cases using simulated 8. Summary and future areas of research
data. Consequently, the interdependencies between order picking
and delivery are slightly less important here. Furthermore, shorter Summary. This paper focuses on the supply of micro grocery
discrete picking times also impact the potential for reducing total stores from their associated DC offering convenience goods. Mi-
picking times by order batching. While order batching reduces to- cro stores are characterized by small order volumes and, in ad-
tal tardiness by 54.95% compared to discrete picking in the case dition, by limited storage and selling space. These stores there-
of an integrated planning approach when analyzing simulated data fore require daily delivery in small volumes, i.e., usually only one
(see Fig. 4), it only reduces total tardiness by 44.84 −31.68
44.84 = 29.34% or two and at a maximum three roll cage(s) per day. These re-
in the real case study. quirements considerably differ from the supply of classical super-
An additional aspect of relevance that should be discussed here markets, which mostly receive full truck-load deliveries. Moreover,
is the possible tradeoff between total tardiness time and the num- the deliveries need to arrive punctually within store-specific de-
ber of tardy jobs. For example, the number of tardy orders is livery time windows, since the limited number of store employ-
smaller when applying SEQOB compared to INTDP (83.35 vs. 89.08) ees – often only one – are not in a position to handle incoming
although INTDP generates less total tardiness on average (55.50 vs. products at any time. Accordingly, processes in the warehouse also
44.84), cf. Table 9. A target conflict thus exists between minimizing change compared to the processes performed when fulfilling or-
the number of tardy orders and minimizing total tardiness, which ders of ordinary supermarkets. Small orders are typically combined
can be viewed from two perspectives. On the one hand, a solution into larger picking batches in the warehouse and afterwards newly
where tardiness is more spread across stores as a whole appears combined to appropriate delivery tours. Both “grouping” processes,
fairer and customers may be faced with empty shelves less often i.e., grouping orders for joint picking tours and grouping orders for
if the in-store safety stock level is sufficiently high to cover de- joint delivery tours, however, may contradict each other. The plan-
mand during relatively contained delivery tardiness. On the other ning of both processes therefore needs to be coordinated, particu-
hand, a higher number of tardy orders increases waiting time for larly because of the short time frame available to conduct the pick-
the orders at the receiving stores and thus the number of em- ing and delivery of store orders. As a result, the supply of micro
ployees needing to serve overtime and/or find themselves faced stores constitutes a considerable challenge in grocery retailing.
with idle time, which may result in higher penalties or compen- We develop a novel modeling and solution approach solving
sation payments. In addition, more out-of-stock situations may re- both planning problems simultaneously. Model OB-VRPSTW fol-
ejor.2020.03.075
JID: EOR
lows the objective of minimizing the total tardiness time of store vehicles and the retailer retaining planning sovereignty, i.e.,
deliveries, which is the primary aim in this context and firmly ver- the retailer pays a fixed amount per order according to a
ified by our case company. We extend the well-established Adap- predefined classification of distances. The retailer can thus
tive Large Neighborhood Search (ALNS) metaheuristic, which we only focus on punctual delivery, which is highly important
denote as General ALNS (GALNS), and show the dominance of this when considering the store-specific requirements, e.g., lim-
enhanced procedure compared to a classical ALNS in our prob- ited access times, provision of shelf stackers, and/or time-
lem setting. In fact, the GALNS approach shows superior results for availability of goods. The literature shows that adhering to
nearly 97% of the problem instances generated and reduces total time windows and minimizing the distance traveled may
tardiness by almost 30% on average. be conflicting objectives [Punakivi et al., 2001]. In scenarios
The numerical analyses using simulated data clarify the dom- where punctuality is less critical, the focus is on target mea-
inance of a simultaneous planning approach compared to a se- sures such as distance, which is not covered by our model.
quential planning approach. Sequential approaches are still com- Further research could thus define a multi-criteria objective
mon practice in grocery retailing. This paper also demonstrates function that combines distance and tardiness goals.
the advantage of a multiple-order-picking strategy (order batching)
over a single-order-picking strategy (discrete picking). We also ap- Technical appendix – Generation of problem instances and
ply the modeling and solution approach suggested in a real-world algorithm parameter settings
case study originating from a leading German grocery retailer. We
again achieve remarkable improvements compared to a sequential This technical appendix details the generation of problem in-
approach and/or a discrete picking strategy. All in all, the numeri- stances (store order and delivery configuration as well as ware-
cal analyses demonstrate the relevance and the applicability of the house and order picking configuration) and the parameter settings
suggested approach in retail practice. of the two solution algorithms GALNS and ALNS.
Future areas of research. The suggested modeling and solution
approach could be extended in several directions, leading to new Customer and Delivery Configurations
challenges for further research: The number of stores is set to 50, 100, or 200 (Desaulniers,
Madsen, & Ropke, 2014; Scholz et al., 2017; Schubert et al., 2019;
(a) The available approaches solving model OB-VRPSTW are still Schubert et al., 2018). Store locations are uniformly distributed
limited, although the present paper develops three alterna- within a square of 50 × 50, 100 × 100 and 200 × 200 kilometers
tives, i.e., GALNS, ALNS and SEQ. Thus, a matter of particu- (Holzapfel, Hübner, Kuhn, & Sternbeck, 2016; Schubert et al., 2019;
lar interest could be to develop additional benchmark proce- Schubert et al., 2018). The DC is always located at the intersection
dures or to develop at least methods providing sharp lower of the square diagonals. Travel times between two locations are de-
bounds to further benchmark the current favorite, GALNS. termined according to Euclidean distances and rounded to the next
(b) In addition to above, the enhanced ALNS, i.e., the GALNS, integer (Ullrich, 2013). Loading times at the warehouse are not ex-
algorithm shows very promising results in our case. The plicitly modeled and assumed to be included in the travel times t0i ,
GALNS approach could therefore possibly be used to solve i ∈ I. The service times for unloading a truck at a store’s location are
related combinatorial problems, e.g., selected vehicle routing uniformly distributed within the interval [s, s + 1, . . . , s] with s = 15
problems. minutes and s = 40 minutes.
(c) Model OB-VRPSTW implies certain assumptions that could The homogeneous vehicles can carry up to 30 or 60 roll cages,
be extended to meet retail practice in comparable settings. which represent small and large trucks in practice. We set the
For example, Model OB-VRPSTW assumes an unlimited in- number of vehicles |K| depending on the total number of store or-
termediate storage area and an unlimited number of load- ders, |I|, the maximum number of roll cages per store order, wUB ,
ing docks for the vehicles at the DC. In other settings, how- and the capacity of the vehicles, w, in order to ensure a feasible
ever, the intermediate storage area may be limited and/or solution:
the number of available loading docks could be smaller than
the number of vehicles ready to be loaded (Schubert et al.,
|K | = (|I| · wUB )/w (23)
2019). Additional contributions could take up and analyze The determination of the time windows is orientated to Ullrich
these extensions. (2013) and Schubert et al. (2018). Ullrich (2013) introduces a close-
(d) The modeling and solution approach developed could be ness and a tightness factor, fc and ft , respectively, to set the posi-
transferred to other application areas that show great sim- tion and the length of the delivery time windows of store order i.
ilarities to our problem setting. For example, the fulfillment Eq. (24) quantifies the lower bound, α i , while Eq. (25) defines the
of e-commerce orders displays equivalent conditions, e.g., upper bound of a time window, β i .
small order volumes, large article ranges, and short delivery
times (Zhao & Yang, 2017). However, e-commerce uses pack-
αi = pickdiscrete
i + t0i + η (24)
ages instead of roll cages when fulfilling customer orders. with
In addition, other objective functions beyond minimizing to- ⎧ ⎡ ⎛ |I| ⎞⎤⎫
tal tardiness could be relevant. These differences would re- ⎪
⎨ i=1
pickindividual
i
|I |
t
i=1 0i ⎪
⎬
⎢ ⎜ |I | |I | ⎟⎥
quire appropriate adjustment of the modeling and solution η ∼ U 0, 1, . . . , ⎢ f c · ⎝ + ⎠⎥
⎪
⎩ ⎢ |M | |K | ⎥⎪
approach suggested. ⎢ ⎥⎭
(e) The automotive industry operates so-called supermarkets to
supply the assembly line with specific subsets of parts (part βi = αi + θ + s (25)
kits) to fulfill the part requirements of customized vehicles with
(Lehmann & Kuhn, 2019). These part kits have to be picked
θ ∼ U {1 , . . . , f t · s }
in the respective supermarkets and delivered to different
line segments on time. This also has great similarities to our We determine the lower bound of the delivery time window, α i ,
problem setting. depending on the discrete pick time of a store order i, pickdiscrete
i
,
(f) The underlying objective function is based on a popular pay- and the travel time to the location of store i, t0i . Moreover, general
ment structure between subcontractors providing delivery problem data are taken into account using the average (discrete)
ejor.2020.03.075
JID: EOR
pick time and the average travel time from the warehouse to a 0.005, respectively, after 50 iterations without updating the incum-
store location as well as the number of order pickers and vehicles. bent solution. If γ best (γ best ) is updated, γ inc (γ inc ) is reset to zero.
Closeness factor fc controls the time gap between the start of the The weight-adjustment parameters ω1 , ω2 , ω3 , and ω4 are set to
planning horizon and α i . A small closeness factor results in time 16, 8, 4 and 1, respectively, and the decay parameter λ to 0.8. Even-
windows closer to the beginning of the planning horizon that are tually, the local search of GALNS is performed with a chance of
consequently harder to meet. Accordingly, β i is determined. The 30%.
width of a time window, i.e., its temporal extent, is determined
depending on minimum service time and the tightness factor ft
Acknowledgments
(see Eq. (25)). The average width of the time window increases as
the tightness factor goes up, and vice versa. The parameters ft and
The authors gratefully acknowledge the computation and data
fc are set to 0.5 (tight) and 1 (wide), and 0.75 (narrow) and 1.25
resources provided by the Leibniz Supercomputing Centre (www.
(spread), respectively.
lrz.de). In addition, we would like to thank the anonymous review-
ers and one of the Guest Editors of the special issue for their valu-
Warehouse and Order Picking Configurations
able recommendations, which have significantly improved our pa-
We apply the same warehouse configuration as Scholz et al.
per.
(2017), considering a layout with three blocks. Additionally, we
make use of the class-based storage assignment policy applied in
Henn (2015) and Scholz et al. (2017). Three or six order pickers References
fulfill orders in the picking area.
Amorim, P., Belo-Filho, M., Toledo, F., Almeder, C., & Almada-Lobo, B. (2013). Lot
From the order picking point of view, a store order is charac- sizing versus batching in the production and distribution planning of perishable
terized by a given set of items to be picked. The number of roll goods. International Journal of Production Economics, 146(1), 208–218.
cages that have to be picked in the picking device is therefore also Archetti, C., Feillet, D., & Speranza, M. G. (2015a). Complexity of routing problems
with release dates. European Journal of Operational Research, 247(3), 797–803.
known in advanced. A roll cage can carry up to 15 items. We differ- Archetti, C., Jabali, O., & Speranza, M. G. (2015b). Multi-period vehicle routing with
entiate between two scenarios regarding the opportunity to com- due dates. Computers & Operations Research, 61, 122–134.
bine orders into batches, which is denoted by parameter fbatch . In Balakrishnan, N. (1993). Simple heuristics for the vehicle routeing problem with soft
time windows. The Journal of the Operational Research Society, 44, 279–287.
the first scenario, i.e., low batching opportunities, fbatch = low, a
Calvete, H. I., Galé, C., Oliveros, M.-J., & Sánchez-Valverde, B. (2007). A goal program-
store order contains between six and 45 items (wUB = 3) and thus ming approach to vehicle routing problems with soft time windows. European
has a capacity requirement of 2.125 roll cages on average, while Journal of Operational Research, 177(3), 1720–1733.
Cattaruzza, D., Absi, N., & Feillet, D. (2016). The multi-trip vehicle routing problem
the picking device can carry up to four roll cages. In the second
with time windows and release dates. Transportation Science, 50, 676–693.
scenario, i.e., high batching opportunities, fbatch = high, the average Chen, Z. L. (2010). Integrated production and outbound distribution scheduling: Re-
number of items per order ranges from ten to 20 items per order view and extension. Operations Research, 58, 130–148.
(wUB = 2) and the capacity of the picking device is increased to Chiang, W.-C., & Russell, R. A. (2004). A metaheuristic for the vehicle-routeing prob-
lem with soft time windows. The Journal of the Operational Research Society, 55,
six cages. As result, 1.45 cages are required to fulfill a store order 1298–1310.
on average. Taking the average number of cages per store order, Clarke, G., & Wright, J. W. (1964). Scheduling of vehicles from a central depot to a
an order picker can carry up to 1.88 orders per batch in scenario number of delivery points. Operations Research, 12, 568–581.
De Koster, M. B. M., van der Poort, E. S., & Wolters, M. (1999). Design and control of
one and 4.14 in scenario two. Note that we focus on the context warehouse order picking: A literature review. International Journal of Production
of micro stores in grocery retailing, but capacity metrics can easily Research, 37, 1479–1504.
be defined differently, e.g., number of items, weight and/or volume De Koster, R., Le-Duc, T., & Roodbergen, K. J. (2007). Design and control of ware-
house order picking: A literature review. European Journal of Operational Re-
requirements, respectively. search, 182(2), 481–501.
Desaulniers, G., Madsen, O. B. G., & Ropke, S. (2014). The vehicle routing problem
Parameters for GALNS and ALNS with time windows. In P. Toth, & D. Vigo (Eds.), Vehicle routing: Problems, meth-
ods and applications (2nd edition) (pp. 119–160). Philadelphia: Society for Indus-
The parameters chosen have been determined according to trial and Applied Mathematics and the Mathematical Optimization Society.
pretests with 50 randomly generated instances not included in the Dueck, G., & Scheuer, T. (1990). Threshold accepting: A general purpose optimization
final test setup. We generally focused on parameter settings used algorithm appearing superior to simulated annealing. Journal of Computational
Physics, 1, 161–175.
in literature, e.g., Hemmelmayr et al. (2012); Koch et al. (2018);
Farahani, P., Grunow, M., & Günther, H. O. (2012). Integrated production and distri-
Žulj et al. (2018); Pisinger and Ropke (2007). However, in the in- bution planning for perishable food products. Flexible Services and Manufacturing
tegrated problem setting and in combination with the local search Journal, 24(1), 28–51.
procedure, larger solution perturbations (e.g., 10% to 40% (Pisinger Figliozzi, M. A. (2010). An iterative route construction and improvement algorithm
for the vehicle routing problem with soft time windows. Transportation Research
& Ropke, 2007), or 17.5% to 35% of customer orders (Žulj et al., Part C, 18, 668–679.
2018) lead to long computing times. Please note that the destroy Fu, Z., Eglese, R., & Li, L. Y. O. (2008). A unified tabu search algorithm for vehicle
and repair heuristics additionally work similar to shaking operators routing problems with soft time windows. Journal of the Operational Research
Society, 59, 663–673.
in GALNS due to the subsequent local search. The perturbations Gao, S., Qi, L., & Lei, L. (2015). Integrated batch production and distribution schedul-
have therefore been reduced until a good tradeoff was found being with limited vehicle capacity. International Journal of Production Economics,
tween intensification, diversification and computing time per iter- 0(16), 13–25.
Gendreau, M., Guertin, F., Potvin, J.-Y., & Taillard, E. (1999). Parallel tabu search for
ation. Conversely, the local search may prevent diversification. We real-time vehicle routing and dispatching. Transportation Science, 33, 381–390.
tested application probabilities between 10% and 100% as a result. van Gils, T., Braekers, K., Ramaekers, K., Depaire, B., & Caris, A. (2016). Improving
We randomly set the number of orders, batches and stores that order picking efficiency by analyzing combinations of storage, batching, zon-
ing, and routing policies. In A. Ruthmair, & M. Voß (Eds.), Lecture notes in com-
have to be removed in the destroy heuristics of GALNS and ALNS
putational logistics. no. 9855 in lecture notes in computer science (pp. 427–442).
according to the following rules. Between one and up to 10% of all Springer International Publishing.
orders are removed from the corresponding subsolution. The num- van Gils, T., Caris, A., Ramaekers, K., & Braekers, K. (2019). Formulating and solv-
ing the integrated batching, routing, and picker scheduling problem in a re-
ber of batches removed varies between 1 and 10 % of all batches.
al-life spare parts warehouse. European Journal of Operational Research, 277(3),
Initially, γ best , γ best , γ inc , and γ inc are set to zero. After 250 it- 814–830.
erations without finding a new best solution, γ best (γ best ) is in- van Gils, T., de Ramaekers, K., Caris, A., & De Koster, R. B. (2018). Designing effi-
creased by 0.01 (0.1). However, if certain limits are exceeded, i.e., cient order picking systems by combining planning problems: State-of-the-art
classification and review. European Journal of Operational Research, 267(1), 1–15.
0.1 (γ best ) and 0.4 (γ best ), respectively, the values are reset to 0.01 Hansen, P., & Mladenović, N. (2001). Variable neighborhood search:. Principles and
and 0.1, again. Accordingly, γ inc and γ inc are increased by 0.01 and applications. European Journal of Operational Research, 130(3), 449–467.
ejor.2020.03.075
JID: EOR
Hansen, P., Mladenović, N., Brimber, J., & Pérez, J. A. M. (2010). Variable neighbor- Potvin, J.-Y., & Rousseau, J. M. (1993). A parallel route building algorithm for the
hood search. In M. Gendreau, & J. Y. Potvin (Eds.), Handbook of metaheuristics, vehicle routing and scheduling problem with time windows. European Journal
2nd edition (pp. 61–86.). Springer. International Series in Operations Research & of Operational Research, 66(3), 331–340.
Management Science 146. New York et al. Ramaekers, K., Caris, A., Moons, S., & van Gils, T. (2018). Using an integrated order
Hemmelmayr, V. C., Cordeau, J.-F., & Crainic, T. G. (2012). An adaptive large neigh- picking-vehicle routing problem to study the impact of delivery time windows
borhood search heuristic for two-echelon vehicle routing problems arising in in e-commerce. European Transport Research Review, 56(10), 1–11.
city logistics. Computers & Operations Research, 39, 3215–3228. Roodbergen, K. J. (2001). Layout and routing methods for warehouses. Ph.D. thesis,
Henn, S. (2015). Order batching and sequencing for the minimization of the total RSM Erasmus University, the Netherlands Ph.D. thesis.
tardiness in picker-to-part warehouses. Flexible Services and Manufacturing, 27, Roodbergen, K. J. (2012). Storage assignment for order picking in multiple-block
86–114. warehouses. Manzini, R. (Ed.), Warehousing in the Global Supply Chain. Springer
Holzapfel, A., Hübner, A., Kuhn, H., & Sternbeck, M. G. (2016). Delivery pattern and London, pp. 139–155.
transportation planning in grocery retailing. European Journal of Operational Re- Ruben, R. A., & Jacobs, F. R. (1999). Batch construction heuristics and storage as-
search, 252(1), 54–68. signment strategies for walk/ride and pick systems. Management Science, 45,
Iqbal, S., Kaykobad, M., & Rahman, M. S. (2015). Solving the multi-objective vehi- 575–596.
cle routing problem with soft time windows with the help of bees. Swarm and Schmid, V., Doerner, K. F., & Laporte, G. (2013). Rich routing problems arising in
Evolutionary Computation, 24, 50–64. supply chain management. European Journal of Operational Researchxs, 224(3),
Koch, H., Bortfeldt, A., & Wäscher, G. (2018). A hybrid algorithm for the vehicle rout- 435–448.
ing problem with backhauls, time windows and three-dimensional loading con- Scholz, A., Schubert, D., & Wäscher, G. (2017). Order picking with multiple pick-
straints. OR Spectrum, 40(4), 1029–1075. ers and due dates - simultaneous solution of order batching. Batch assignment
Koskosidis, Y. A., Powell, W. B., & Solomon, M. M. (1992). An optimization-based and sequencing, and picker routing problems. European Journal of Operational
heuristic for vehicle routing and scheduling with soft time window constraints. Research, 263(2), 461–478.
Transportation Science, 26, 69–85. Schubert, D., Kuhn, H., & Holzapfel, A. (2019). Same-day deliveries in Omni-channel
Lehmann, M., & Kuhn, H. (2019). Modeling and analyzing sequence stability in flex- retail - integrated order picking and vehicle routing with vehicle-site dependen-
ible automotive production systems. Flexible Services and Manufacturing Journal cies. submitted to Naval Research Logistics.
First Online: 28 January 2019, 1–29. Schubert, D., Scholz, A., & Wäscher, G. (2018). Integrated order picking and vehicle
Marchet, G., Melacini, M., & Perotti, S. (2015). Investigating order picking system routing with due dates. OR Spectrum, 40(4), 1109–1139.
adoption: a case-study-based approach. International Journal of Logistics Research Sexton, T., & Choi, Y. (1986). Pickup and delivery of partial loads with soft time win-
and Applications, 18, 82–98. dows. American Journal of Mathematical and Management Sciences, 6, 369–398.
Matusiak, M., de Koster, R., & Saarinen, J. (2017). Utilizing individual picker skills Shaw, P. (1997). A new local search algorithm providing high quality solutions to
to improve order batching in a warehouse. European Journal of Operational Re- vehicle routing problems. Working paper, APES Group, Department of Computer
search, 263(3), 888–899. Science, University of Strathclyde, Glasgow, Scotland.
Moons, S., Braekers, K., Ramaekers, K., Caris, A., & Arda, Y. (2019). The value of in- Taillard, E. D., Badeau, P., Gendreau, M., Guertin, F., & Potvin, J. Y. (1997). A tabu
tegrating order picking and vehicle routing decisions in a B2C e-commerce en- search heuristic for the vehicle routing problem with soft time windows. Trans-
vironment. International Journal of Production Research, 57(20), 6405–6423. portations Science, 31, 170–186.
Moons, S., Ramaekers, K., Caris, A., & Arda, Y. (2017). Integrating production schedul- Tompkins, J. A., White, J. A., Bozer, Y. A., & Tanchoco, J. M. A. (2010). Facilities plan-
ing and vehicle routing decisions at the operational decision level: A review and ning (4th ed.). New Jersey: John Wiley & Sons.
discussion. Computers & Industrial Engineering, 104, 224–245. (February), Ullrich, C. A. (2013). Integrated machine scheduling and vehicle routing with time
Moons, S., Ramaekers, K., Caris, A., & Arda, Y. (2018). Integration of order picking windows. European Journal of Operational Research, 227(1), 152–165.
and vehicle routing in a B2C context. Flexible Services and Manufacturing Journal, Zhang, J., Xuping, W., Chan, F. T. S., & Ruan, J. (2017). On-line order batching and se-
30, 813–843. quencing problem with multiple pickers: A hybrid rule-based algorithm. Applied
Mouthuy, S., Massen, F., Deville, Y., & Van Hentenryck, P. (2015). A multi-stage Mathematical Modelling, 45, 271–284.
very large-scale neighborhood search for the vehicle routing problem with soft Zhao, Z., & Yang, P. (2017). Improving order-picking performance by optimizing or-
time-windows. Transportation Science, 29, 223–238. der batching in multiple-cross-aisle warehouse systems: A case study from e–
Pisinger, D., & Ropke, S. (2007). A general heuristic for vehicle routing problems. commerce in china. In Proceedings of the 4th international conference on indus-
Computers & Operations Research, 34, 2403–2435. trial engineering and applications (ICIEA): 1 (pp. 158–162).
Pisinger, D., & Ropke, S. (2010). Large neighborhood search. Gendreau, M., Potvin, Žulj, I., Kramer, S., & Schneider, M. (2018). A hybrid of adaptive large neighborhood
J.-Y. (Eds.), Handbook of Metaheuristics, 2nd edition. International Series in Op- search and tabu search for the order-batching problem. European Journal of Op-
erations Research & Management Science 146. New York et al.: Springer, pp. erational Research, 264(2), 653–664.
399–420.
ejor.2020.03.075

Supermercado 8

Uploaded by

Copyright:

Available Formats

Supermercado 8

Uploaded by

Document Information

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

Supermercado 8

Uploaded by

Copyright:

Available Formats

JID: EOR

ARTICLE IN PRESS [m5G;May 5, 2020;18:19]

Contents lists available at ScienceDirect

European Journal of Operational Research

Integrated order batching and vehicle routing operations in grocery

I set of stores, I = {1, . . . , i, . . . , |I|}

For traveling between locations, i.e., DC and store locations, a cer-

part of several feasible batches. Consequently, Constraints (15) en-

Fig. 2. General framework of the GALNS algorithm for solving OB-VRPSTW.

5.2. Initial solution Table 2

To provide an initial solution, the following procedure is ap- VNDOPP

5.3. Destroy heuristics 5.4. Repair heuristics

Radius 100 150 200

Performance measure SEQDP SEQOB INTDP INTOB

Total tardiness 100.00 55.50 44.84 31.68

Fig. 5. Distribution networks considered in the real-life case study.

sult. Consequently, management has to carefully balance which so-

You might also like