IISE Transactions

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/uiie21

Order batching and picker scheduling in

warehouse order picking

İbrahim Muter & Temel Öncan

To cite this article: İbrahim Muter & Temel Öncan (2021): Order batching and picker scheduling in
warehouse order picking, IISE Transactions, DOI: 10.1080/24725854.2021.1925178

To link to this article: https://doi.org/10.1080/24725854.2021.1925178

View supplementary material

Published online: 14 Jun 2021.

Submit your article to this journal

Article views: 142

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=uiie21
IISE TRANSACTIONS
https://doi.org/10.1080/24725854.2021.1925178

Order batching and picker scheduling in warehouse order picking

_Ibrahim Mutera
€
and Temel Oncan b

a
Amazon Web Services, re:Invent Building, Seattle, WA, USA; bDepartment of Industrial Engineering, Galatasaray University, _Istanbul, T€
urkey

ABSTRACT ARTICLE HISTORY

This article focuses on the integration of order batching and picker scheduling decisions while tak- Received 31 May 2020
ing into account two objectives that have been considered in the literature, namely the minimiza- Accepted 22 April 2021
tion of both total travel time to collect all items and makespan of the pickers. This integrated
KEYWORDS
problem not only occurs naturally in wave picking systems in which the latest picking time of
Integer programming;
orders becomes the key performance metric, but also arises when there is a limit on the picker column generation;
operating time. We present models that result from combining these objectives and analyze their order batching
relationship through bounds. We propose a column generation-based exact algorithm for the inte-
grated problem. The novelty of the proposed approach lies in the ability of efficiently solving the
integrated order batching and picker scheduling problem to optimality by designing a column
generation subproblem based on the set of batches, which makes it a challenging optimization
problem due to its size. We alleviate this difficulty by reformulating this subproblem, which allows
efficient implicit enumeration of its variables. We have also devised a Variable Neighborhood
Search algorithm used as a subprocedure within the proposed exact solution algorithm. Finally,
we conduct experiments on randomly generated instances and show that the proposed algo-
rithms are capable of solving instances with up to 100 orders.

1. Introduction belonging to each batch in a single tour. For a categorization

of order picking methods we refer to de Koster et al. (2007).
Warehouses play a crucial role in determining the efficiency
We consider a rectangular parallel-aisle warehouse which is
of distribution networks by balancing the variation in cus-
illustrated in Figure 1. In this layout, there are 10 parallel
tomer demand, consolidating various products to common aisles numbered in increasing order starting from the left-
customers and rapidly responding to customer demand most aisle. Also, they are connected by cross aisles at the
(Tompkins et al. 2010). In addition, effective warehouse front and back of the parallel aisles, which are the so-called
management, which consists of the planning, optimization front aisle and back aisle, respectively. The input and output
and control of all warehousing activities, has a considerable operations are performed at the Input/Output (I/O) point
impact on warehouse performance (Faber et al. 2018). that is located at the left of the front aisle. L and w denote
Among the numerous warehousing activities, such as receiv- the vertical distance between the front and back aisles and
ing, put-away, storage and order picking, sorting and ship- the horizontal distance between the middle of two parallel
ping, the order picking activity is known to be the most aisles, respectively. We assume that a fixed number of identical
costly. It has been reported that the cost of the order pick- pickers with constant velocity are available at the I/O point, all
ing activities amounts to almost 65% of the total cost of requested items in the right quantities are available in the
warehousing operations (Frazelle 2001). Basically, order picking warehouse where their locations are known, and the items of a
is to retrieve the required items from their storage locations to customer order must be picked up in one picker tour.
fulfill independent customer orders (Petersen and Schmenner Broadly speaking, an order picker’s activities mainly
1999). On the other hand, an inefficient order picking per- involve searching, picking, traveling and setup for the routes
formance may cause late deliveries and incorrect shipments, (Tompkins et al. 2010). Among these activities, traveling is
which result in extra costs for the logistics systems and cus- the most time-consuming activity (de Koster et al. 2007).
tomer dissatisfaction (Henn and W€asher 2012). Order picking routing policies, which consist of finding the
In this study, we focus on a picker-to-parts system where pick sequence of the items in a batch, have been employed
human pickers walk or ride along the aisles and collect the to efficiently manage the traveling activity. In the literature,
requested items according to the sort-while-picking strategy. typical order picking routing policies are traversal, return,
We address a batch picking system that consists of batching midpoint, largest gap, combined and optimal. Although the
a set of orders such that a picker collects the items optimal order picking policy clearly outperforms other

€
CONTACT Temel Oncan [email protected]
Supplemental data for this article is available online at https://doi.org/10.1080/24725854.2021.1925178

This research was accomplished prior to the author‘s employment at Amazon.
Copyright ß 2021 “IISE”
2 İ. MUTER AND T. ÖNCAN

Figure 1. Layout of a rectangular parallel-aisle warehouse.

policies in terms of total distance traveled, it has several dis- objective of minimizing the maximum duration of the
advantages in practice (de Koster et al. 1999). For instance, batches, since each batch can be assigned to one picker. This
this policy may be perceived as confusing by the pickers, objective is equivalent to the makespan objective prominently
and its implementation can be cumbersome, due to com- used in the scheduling literature. Note that in the order
patibility issues arising in the Warehouse Management batching context, makespan is the time when all batches are
Systems. In this study, we concentrate only on the traversal, collected by the pickers. In this study, we focus on a problem
return and midpoint policies which turn out to be useful in similar to the one addressed in the seminal work by
complex picking systems with several pickers, as they are Gademann et al. (2001). However, here we relax their unlim-
considered to be simpler than the other policies (Petersen ited pickers assumption and define our problem with a given
1997). Moreover, the algorithms proposed in this article can number of identical pickers that can be assigned to multiple
be implemented by considering other routing policies, e.g., batches. Consequently, the decision that entails this relaxation
largest gap, combined and optimum. is the assignment of batches to pickers. In addition to the
In the light of the definitions and assumptions introduced makespan objective as described above, time considerations
above, we now define the Order Batching Problem (OBP), also emerge as a limit on the operating times of the pickers.
which has been shown to be N P-hard for general instances This limit may stem from the due date of a set of orders that
(Gademann and van de Velde 2005). Given a set of customer are associated with a customer or from regulations that con-
orders, fixed locations of the items to be collected and a strain the operating hours of warehouse employees. Unlike
picker routing policy, the OBP is to construct order batches the wave picking case explained in the previous paragraph
such that the total distance traveled by the pickers collecting where the maximum operating times of the pickers is mini-
these orders is minimized. Assuming a constant travel speed mized, the limit on the operating times is generally imposed
of pickers, the OBP solution equivalently minimizes the total when the total time to retrieve all orders is minimized as an
duration to pick all orders in the batches. In the literature, objective. Recently, Matusiak et al. (2017) incorporated this
the OBP has been considered with several side constraints time limit on the operating time of the pickers, and the
(see e.g., Chabot et al. 2017) and/or the OBP has also been objective is to find an order batching and picker assignment
integrated with other problems such as the Order Sequencing solution that minimizes the total time required to pick all
Problem and Optimal Billing Sequencing (Henn and Schmid orders. In the literature, several studies have addressed
2013; Pinto et al. 2018; Pinto and Nagano 2020). restrictions on the operating times of the pickers, where each
The assignment of selected batches to the pickers is virtu- order is associated with a due date. For instance, in Chen
ally always made under time considerations. One such consid- et al. (2015), Henn (2015) and Scholz et al. (2017), the
eration arises in wave picking systems, in which a set of authors aim to minimize the tardiness of the orders, which
orders is simultaneously handled by a group of available pick- necessitates not only assignment of order batches to pickers,
ers. The wave picking has important implications in the deliv- but also sequencing them for each picker.
ery operations, especially when the orders to be collected The time considerations explained above bring about an
constitute a truck load (Gademann et al. 2001). In that case, assignment decision of order batches to pickers. The prob-
the completion time of the pickers that retrieve those orders lem is known as the integrated Order Batching, routing and
becomes extremely important, as the truck will not be dis- Picker Scheduling Problem (OBPSP). Formally speaking,
patched until all orders have been handed over by the pick- given a picker routing policy, the OBPSP consists of deter-
ers. Note that the maximum travel time of all pickers is the mining order batches, each of which is assigned to exactly
time when the truck is ready to leave the warehouse and one picker under operating time constraints such that the
deliver collected orders to their customers. Assuming an total travel time to pick all the orders is minimized.
unlimited number of pickers in a wave picking system, Matusiak et al. (2017) proposed an Adaptive Large
Gademann et al. (2001) formulated this problem with the Neighborhood Search algorithm to solve this problem with
 IISE TRANSACTIONS 3

non-identical pickers. In another version of the OBPSP aris- total cost of the routes (Mingozzi et al. 2013). The main dis-
ing in the wave picking systems mentioned above, the make- tinction between the MTVRP and the OBPSP is the way the
span of the orders collected by the pickers is minimized. costs of a vehicle schedule and a picker schedule are calcu-
This problem will be referred to as the MOBPSP, and to the lated. In real life, the working time limits, as imposed in
best of our knowledge, in the literature there are very few both the MTVRP and the OBPSP, may be too short or large
studies addressing the MOBPSP. Two of them are the works compared with the actual picker operating times to collect
by Ardjmand et al. (2018) and Henn (2012). Ardjmand the given set of orders. In our approach to the OBPSP, we
et al. (2018) simultaneously tackle the order batching, the establish the ties between the makespan objective and the
assignment of batches to pickers and the picker routing in working time limit through the
-constraint method. In par-
order to minimize the makespan in a wave picking system. ticular, we impose the makespan to lie within a certain per-
Henn (2012) has addressed the online OBP with the object- centage gap from its lower bound, and show that this gap
ive of minimizing the maximum service times, namely also limits the deterioration in the objective value of OBP
makespan, of the batches collected by a single order picker. with the assignment of batches to the pickers. Therefore,
In this article, we address the integration of order batch- our approach provides insights into two objectives that per-
ing and picker scheduling decisions and take into account tain to the integrated problems such as the OBPSP and
the minimization of both total travel time to collect all the MTVRP.
orders by the pickers and makespan of the orders which Another problem that is similar in structure to the
arises in the wave picking systems. We will demonstrate MOBPSP is the Parallel Batch Processing Machine schedul-
through our proposed models that these two objectives are ing problem with makespan minimization (PBPM) (Muter
actually in conflict, i.e., the smaller the value of makespan 2020) where each machine is capable of processing a batch
gets, the larger the total travel time to retrieve all orders of jobs simultaneously under capacity constraints. The most
becomes. There are several methods designed for problems salient difference between the PBPM and the MOBPSP is
with multiple objectives, such as
-constraint, weighted sum that the former lends itself to a simple calculation of the
or lexicographic (Ehrgott 2005). In this article, we present total duration of a job batch, whereas the latter requires a
the models of both the OBPSP and the MOBPSP, analyze routing algorithm to calculate the total travel time of an
their bounds, and unravel the relation between them by the order batch. For this problem, Muter (2020) proposed an

-constraint method, in which the objective of total travel exact algorithm to minimize the makespan where the single
time to pick all orders is minimized while the makespan machine version of this problem is used to set upper and
objective is transformed into a constraint by imposing an lower bounds on makespan, and a search procedure is
upper bound. We propose the first exact algorithm in the devised to find the optimal solution. We have modified the
literature for the solution of the OBPSP, which is based on methodology proposed in Muter (2020) for the solution of
column generation. Furthermore, the application of the lex- MOBPSP, and additionally, we set the total travel time to
icographic method to the MOBPSP renders two objectives pick all orders as the secondary objective that is minimized
considered in the order of their priority in which the total at the minimum value of makespan.
travel time to collect all orders in the batches (secondary This article is organized as follows: the mathematical
objective) is minimized at the minimum makespan (primary models of both OBPSP and MOBPSP are constructed in the
objective). We propose a modification of the algorithm for next section, where the value of the picker time limit param-
the OBPSP to find the minimum makespan for the eter that influences the picker assignments is determined
MOBPSP, which is also the first exact algorithm for this through the relationship between the total travel time
problem. Therefore, the contribution of this article is math- required to collect all orders and the makespan objectives.
ematical models that are developed to combine two objec- This section also presents an overview of the methodologies
tives pertinent to the OBPSP, namely the total travel time developed to solve these two problems. Section 3 explains in
required to collect all orders by the pickers and makespan of detail the column generation algorithm to solve the OBPSP
the orders, and the first exact algorithms in the literature to for a given value of the picker time limit. In Section 4, we
solve these models to optimality. The novelty of the pro- introduce the outline of a tailor-made Variable
posed algorithm lies in the ability to efficiently solve the Neighborhood Search (VNS) algorithm for the OBPSP.
OBPSP formulation to optimality by designing and solving a Next, the performance of the proposed algorithms is eval-
column generation subproblem, which consists of a very uated in Section 5 in which the results are reported for three
large set of variables corresponding to batches. We integrate routing policies, namely the midpoint, return and traversal
this algorithm in a search mechanism to solve the MOBPSP policies. Finally, we summarize the findings in Section 6.
formulation. These approaches capitalize on the tight
bounds obtained for the formulations of both the OBPSP
2. Mathematical model and proposed methodology
and the MBPSP.
One of the problems reminiscent of OBPSP is the Multi- In this section, we first revisit the OBP and then develop the
Trip extension of the Vehicle Routing Problem (MTVRP) mathematical model for the OBPSP. This model will be aug-
where trips that are formed of a subset of customers are mented with an additional objective of makespan minimiza-
assigned to vehicles so as to deliver the customer orders tion, which results in a mathematical model of the
within the vehicle’s working time limit and minimize the MOBPSP. Analyzing the relationship between the objectives
4 İ. MUTER AND T. ÖNCAN

of these problems, we identify bounds on their models, of the batches assigned to each picker to T. Like OB, OBPS
which will be employed in developing solution approaches minimizes the total travel time required to collect all orders
to tackle them. in the batches, and its lower bound is characterized in the
Recall that the OBP is to pick a set orders I, where each next proposition.
i 2 I contains a number of items that use di units of the
Proposition 1. The optimal objective values of the Linear
capacity, from their storage shelves in batches such that total Programming (LP) relaxations of (1)-(3) and (4)-(7)
travel time to collect all orders is a minimum. This problem are equal.
can be formulated as the set partitioning problem by defin-
ing the set of batches B, each of which comprises a number Proof. See Mingozzi et al. (2013) for a similar proof showing
of orders whose total size is smaller than or equal to the the equality of LP relaxation bounds of VRP and MTVRP
batch capacity, which is Q in unit loads. Given an order that were modeled similar to OB and OBPS, respectively. w
picking policy, the total travel time to collect all orders in OBPS features a block diagonal structure in the constraint
batch b 2 B is denoted by cb. Whether an order i 2 I is set (6), which will be exploited in the next section.
picked in batch b is indicated P by binary parameter aib. Meanwhile, we focus on the OBPS in order to obtain bounds
Therefore, each batch satisfies i2I di aib  Q: The set parti- on not only the total travel time to collect all orders, but also
tioning formulation can then be written as follows: the makespan that will be incorporated to the OBPS to form
X the model of MOBPSP. As mentioned in Section 1, for a lim-
ðOBÞ zOB ¼ minimize cb xb (1)
b2B
ited number of pickers, the value of T determines the feasibil-
X ity of the problem. The bounds on these two objectives will
subject to aib xb ¼ 1, i 2 I, (2) reveal the relationship between the value of T and the make-
b2B span objective. We point out that T in (6) is a preset limit on
xb 2 f0, 1g, b 2 B, (3) the operating time of pickers, however, it also pertains to the
completion time to collect all orders by the pickers, which
where objective function (1) consists of the minimization of gives rise to the makespan objective. The multi-objective opti-
total travel time to pick all batches, constraints (2) ensure mization problem that incorporates the minimization of
each order resides in a batch, and (3) are the binary con- makespan to the OBPS, which we referred to as the MOBPS,
straints. This model requires an exhaustive enumeration of can be formulated as follows:
batches, which is prohibitive even for small instances. Muter
€ ðMOBPSÞ minimize Cmax (8)
and Oncan (2015) presented a column-and-cut generation XX
algorithm to find a lower bound z OB along with a set of minimize cb kkb (9)
dual variables ui, i 2 I associated with (2). In the same k2K b2B
paper, an Iterated Local Search Algorithm was devised to X
subject to cb kkb  Cmax , k 2 K, (10)
find an upper bound z OB : Using these bounds, the authors b2B
eliminated a subset of variables to create a reduced problem, X
which was then solved by an off-the-shelf Mixed-Integer cb kkb  T, k 2 K, (11)
Linear Programming (MILP) solver to find the optimal solu- b2B
XX
tion zOB. They conducted computational experiments on aib kkb ¼ 1, i 2 I, (12)
randomly generated instances and found the solution for k2K b2B
instances with up to 100 orders.
In the OBPSP, the batches are assigned to a set of pickers kkb 2 f0, 1g, b 2 B, k 2 K, (13)
K, each working T time units for picking a subset of orders in Cmax 2 Z, (14)
I. Let us define binary decision variable kkb for a batch b 2 B
which is an extension of (4)-(7). Furthermore, we impose
and a picker k 2 K, which is equal to one only if batch b is
integrality on Cmax in (14). Since batch durations are frac-
assigned to picker k. Thus, the total duration of the batches
tional, this imposition will facilitate the computations with-
assigned to each picker must not exceed the duration limit T.
out loss of generality. Note that the objectives (8) and (9)
The mathematical model can be written as follows:
XX are in conflict, as the smaller values of Cmax imposed in the
ðOBPSÞ zOBPS ¼ minimize cb kkb (4) right-hand-side of (10) never decrease the value of the total
k2K b2B travel time objective given in (9).
XX In tackling MOBPS, which is a multi-objective optimiza-
subject to aib kkb ¼ 1, i 2 I, (5)
tion problem, we apply the lexicographic method in which
k2K b2B
X we rank the objectives in the order given in the model,
cb kkb  T, k 2 K, (6) namely (8) as the primary objective and (9) as the secondary
b2B objective. Therefore, the minimum value of the makespan,


which is denoted by Cmax , is achieved first, and then, we
kkb 2 f0, 1g, b 2 B, k 2 K, (7)
solve the OBPS in order to minimize (9) by setting T ¼
where the constraint set (5) ensures each order is picked in

Cmax : Note that the left-hand sides of both (10) and (11)
a batch, and the constraint set (6) limits the total duration represent the operating time of the pickers, and for any
 IISE TRANSACTIONS 5

feasible value of T  Cmax , (11) is redundant and can be 2.2. Bounds of OBPS
discarded from MOBPS without changing its optimal
As in MOBPS, a subset of OBPS formed of (4), (5) and (7)
solution. Next, we set bounds on the makespan objective
is equivalent to OB, and similarly, Bþ denotes the optimal
of MOBPS, namely (8), and the total travel time objective
batches and C max denotes the optimal objective of (15)-(18).
of OBPS.
For a given T in OBPS, C  max  T attests to the optimality
of batches Bþ and ^k b , b 2 Bþ and k 2 K for this problem
k

2.1. Bounds on makespan with the objective value zOBPS ¼ zOB : In order to link T to
A subset of MOBPS induced by (9), (12) and (13) corre- the makespan objective, we apply the
-constraint method to
sponds to the OB formulation. Let ^x b , b 2 B be the solution MOBPS where we restrict the makespan objective (8) in a
of OB, and Bþ denote the set of selected batches in this constraint while minimizing (9). We derive this constraint
solution, i.e., ^x b ¼ 1, b 2 Bþ and ^x b ¼ 0, b 2 BnBþ : Given on (8) using the lower bound on Cmax as
X
this solution, the rest of MOBPS becomes cb kkb  Cmax  aC max , k 2 K (19)
C max ¼ minimize Cmax (15) b2Bþ
X where a  1 is the control parameter that drives the quality
subject to kkb ¼ 1, b 2 Bþ , (16) OB
k2K of the makespan in the solution and C max ¼ dzjKj e as defined
X in Proposition 2. Thus, there are two limits on the operating
cb kkb  Cmax , k 2 K, (17)
time of pickers: T and aC max , as imposed in (19) and (11),
b2Bþ
respectively. Setting a such that T ¼ aC max obviates (10).
kkb 2 f0, 1g, b 2 Bþ , k 2 K, (18) Therefore, after combining OBPS with the makespan object-
ive and applying
-constraint method, the final OBPS model
where the constraint set (16) imposes that each batch
that we tackle stays as (4)-(7).
selected in the solution of OB is assigned to one of the pick-
ers, which also ensures each order to be picked up by a Proposition 3. For the optimal objective value of OBPS, the
picker. This problem is similar to the parallel machine following holds:
scheduling problem in which the jobs to be processed are zOB
zOB  zOBPS  jKjda e:
the selected batches b 2 Bþ with processing time cb, and the jKj
machines are the counterparts of the picker set K. Parallel
machine scheduling with makespan objective is N P-hard Proof. Relaxing the operating time constraints (6), we obtain
(Graham et al. 1997), however, (15)-(18) can be solved the OB, which gives a lower bound on zOBPS. The upper
quickly for the size of instances considered in this article. bound results from the aggregation of the same constraints
Moreover, depending on the batch capacity, the cardinality (6), which yields
XX
of the selected batches Bþ is small compared with the num- cb kkb  jKjT,
ber of orders. The integrality of Cmax imposed in (14) can k2K b2B
be achieved by replacing C  max ¼ dC max e: Letting C max
where the right-hand side can be replaced by
denote a lower bound on makespan Cmax, the following


proposition sets bounds on Cmax :  OB   OB 
z z

jKjT ¼ jKjaC max ¼ jKja  jKj a :
Proposition 2. For the optimal objective value Cmax of jKj jKj
MOBPS, the following holds:
 OB  which ends the proof. w
z
 max :
C max ¼  Cmax C
jKj
Remark 1. The solution of OB establishes the bounds on both
zOBPS and Cmax

since zOB resides in the lower and upper
Proof. A lower bound of Cmax can be obtained by aggregat- OBPS
ing (10) as follows: bounds of z (Proposition 3), the lower bound C max of Cmax
X XX (Proposition 2), and the selected batches Bþ in the solution of
cb ^x b  cb kkb  jKjCmax OB are assigned to pickers yielding the upper bound C  max : As
b2B k2K b2B OBPS
stipulated in Proposition 3, the upper bound on z is slightly
P
where the first inequality follows due to b2B cb ^
x b ¼ zOB , larger than azOB due to the ceiling operator, i.e.,
which is the minimum total travel time to pick all orders  OB 
zOB z
obtained by the solution of OB. Therefore, Cmax  dzjKj OB
e jKja  jKj a :
jKj jKj
where the ceiling follows from (14). As for the upper bound,
 max is the makespan of the batches Bþ that minimize the
C Thus, a can be interpreted as the allowable deterioration in
total travel time required to collect all orders in the solution the OB objective zOB after the incorporation of the time
of OB so that it cannot be shorter than the minimum make- limit T. Though the larger values of a allows higher increase


span Cmax : w in zOBPS, they also relax the constraint set (6) through T in
the right-hand side, which potentially improves zOBPS.
6 İ. MUTER AND T. ÖNCAN

2.3. Overview of the proposed methodology for

with Cmax : This algorithm is inspired by the framework pro-
the OBPS posed in Muter (2020) that solves the PBPM. In their work,
Ardjmand et al. (2018) also adopted the makespan mini-
The lower and upper bounds on Cmax are of paramount
mization objective. Apart from the makespan minimization,
importance in finding the optimal value of makespan as
we aim to minimize the total travel time required to pick all
well as that of OBPS. The solution of OBPS is governed by orders at the minimum value of Cmax. As alluded to previ-
T, which is the operation time limit of the pickers, and ously, this approach is known as the lexicographic method
 max , which is the upper bound on the makespan. The
C in the multi-objective optimization literature, in which the
general overview of the proposed methodology to solve objectives are ranked in order of their priority. For a two-
OBPSP, which involves updating T and solving OBPS objective optimization problem, the optimal value for the
iteratively, is given in Algorithm 1. This algorithm com- primary objective is found, and the optimization problem
mences with the solution of OB (Line 1), which outputs a with the secondary objective is solved without deteriorat-
set of batches Bþ and the optimal objective value zOB that ing the value of the primary objective. This approach also
is used to derive enables one to observe how the total travel time required
 OB  to collect all orders changes for decreasing values
z
C max ¼ : of makespan.
jKj
Algorithm 2 lays out the proposed methodology for
Defining the set Bþ and the set of pickers K as the set of solving the MOBPS, which is similar in its early phases to
jobs and machines, respectively, we solve the parallel Algorithm 1. If the assignment of batches Bþ to the pick-
machine scheduling problem with makespan minimization, ers in the solution of (15)-(18) satisfies C  max ¼ C max ,
which is used to find an upper bound C  max on the make- these batches and the picker assignment are also optimal

span (Line 2). If T  C max , we basically find the optimal for the MOBPS. Otherwise, C 0
max is obtained through


solution for the OBPS, i.e., zOBPS ¼ zOB and optimal VNS, and the search for Cmax starts. The value of T is
batches Bþ and assignments ^k , b 2 Bþ and k 2 K:
k decreased by one in a loop as long as the optimal – not
b
only feasible – solution of OBPS is obtained, and T does
Otherwise, we solve OBPS with the value T, which will be
not fall below the lower bound C max , which otherwise
explained in the next section. If OBPS is infeasible for T,
attests to the optimality of Cmax

¼ C max : If no feasible
we need to update its value to find a feasible solution for
solution of OBPS exists for a given T, the minimum
which a becomes instrumental. As explained previously, a,
makespan is reached so that the algorithm terminates
which is initially equal to T=C max , corresponds to the

with the optimal value Cmax ¼ T þ 1: We choose to con-
maximum allowed percentage deterioration in zOB after 0
duct the search starting from T ¼ C max and decrementing
the incorporation of time limit T. We increase T in tan-
it by one due to the tightness of this upper bound in most
dem with a (Lines 7 and 8) until OBPS has a feasible, and
of the instances, which we will show in the computational
hence optimal, solution. Even though increasing a also
experiments.
increases the upper bound on zOBPS, the larger values of T
will yield better objective values as the limit on the operat-
ing time of pickers is relaxed. 3. Column generation for OBPS
In this section, we present the algorithm to solve the
Algorithm 1: Proposed Methodology for OBPS OBPS given in (4)-(7). This algorithm is called both in
1 Solve (1)-(3) (Bþ, zOB, C max ); Algorithm 1 and Algorithm 2 with a given value of T. In a
2 Solve (15)-(18) (C max ); nutshell, we first capitalize on the decomposable structure

3 if C max  T then of this problem that leads to a reformulation whose vari-
able set corresponds to the picker schedules, each of which
4 Stop, z OBPS ¼ z OB ;
corresponds to a set of order batches satisfying the oper-
5 Solve OBPS(T);
6 if Infeasible then ation time limit of the pickers T. The set of schedules
7 Update a; increases exponentially with the number of orders so that
OB the reformulation has a large set of columns, making it
8 T ¼ da zjKj e;
amenable to column generation algorithm (see L€ ubbecke
9 Return to Line 5; and Desrosiers (2005) for details). The main difficulty with
10 else this approach is that as the picker schedules are formed of
11 Stop; order batches, the generation of these schedules hinges
upon the set of batches to be pre-enumerated, which not
2.4. Overview of methodology for MOBPS only is arduous, but also has a detrimental impact on the
picker schedule generation subproblem. We overcome this
We now focus on the solution of MOBPS, which prioritizes hurdle by developing a subproblem that concurrently gen-
the makespan objective. To that end, we modify Algorithm erates batches and picker schedules. This column gener-
1 by embedding a search mechanism that solves OBPS itera- ation algorithm is reminiscent of that proposed in Muter
tively while decrementing the value of T until it coincides (2020) to solve PBPM.
 IISE TRANSACTIONS 7

of pickers. Since constraint set (6) does not satisfy the inte-
Algorithm 2: Proposed solution methodology for the grality property, the LP relaxation bound of the reformula-
MOBPS tion (24)-(27) is at least as tight as that of (4)-(7), as well as
that of (1)-(3) due to Proposition 2.
1 Solve (1)-(3) (Bþ, zOB, C max ); The application of column generation to MP works on a
2 Solve (15)-(18) (C  max ); restricted form of the LP Relaxation of MP (RMP) contain-
3 
if C max ¼ C max then ing a subset of schedules S, which is expanded with new
4 Stop, z OBPS ¼ z OB and Cmax

 max ;
¼C variables by solving the following Pricing SubProblem (PSP):
5 þ 
VNS(B ) ! (C max );0 X
0 ; min cs  ui ais  v (28)
6 T :¼ C max
s2S
i2I
7 while True do
8 T T  1; where ui, i 2 I and v are the dual variables associated with
9 if T < C max then (25) and (26), respectively, whose values are retrieved from
10

Stop, Cmax ¼ C max ; the solution of RMP. (28) is defined over the set of sched-
11 Solve OBPS(T); ules, whose complete description is not known a priori. This
12 if Infeasible then problem can be posed as an optimization problem that
13

Stop, Cmax ¼ T þ 1; forms the minimum reduced cost schedule based on revers-
ing the earlier substitutions as follows:
X X X
Model (4)-(7) has a special structure that can be exploited minimize ðcb  ui aib Þxb  v ¼ c b xb  v (29)
by Dantzig–Wolfe decomposition (Dantzig and Wolfe 1960). b2B i2I b2B
The constraint set (6) is separable for each picker k 2 K: Let X
S contain integer solutions in the polyhedron induced by subject to aib xb  1, i 2 I, (30)
s2S
the constraint set (6) for any k 2 K, which also corresponds X
to the set of all possible schedules that can be performed by cb xb  T, (31)
any picker. Note that since the pickers are identical, a b2B
unique set S independent of k is defined for the schedules.
xb 2 f0, 1g, b 2 B, (32)
For an integer solution s 2 S, kbs is a binary parameter that P
assumes value one only if s contains batch b. Replacing the where c b ¼ cb  i2I ui aib is the reduced cost of batch b 2
corresponding terms in (4)-(7) and defining a binary vari- B: The aim of this model is to select batches with minimum
able ys 2 S that indicates whether schedule s is selected, we total reduced cost such that each order is picked at most
obtain: once, and the total duration limit is not exceeded.
XX As in (1)-(3), the subproblem (29)-(32) is defined over all
minimize cb kbs ys (20) batches b 2 B, and a viable method is again column gener-
b2B s2S
XX ation. Solving a PSP by column generation is known as
subject to aib kbs ys ¼ 1, i 2 I, (21) nested column generation, and solving (29)-(32) by column
s2S b2B generation at every iteration can be very cumbersome. In
X order to cope with the complexity of this subproblem that
ys  jKj, (22)
generates picker schedules, we relax the constraint set (30)
s2S
and solve the following knapsack problem formed of the
ys 2 f0, 1g, s 2 S: (23) complete set of batches B:
X
The second set of constraints ensures that the number of ðKPSPÞ s ¼ maximize c b xb (33)
schedules selected is at most jKj, as each picker is not obli- b2B
gated to perform a schedule. For each s 2 S, the coefficients X
P cb xb  T, (34)
b2B aib k ¼ ais and
bs
in this model can be substituted as
P b2B
b2B cb k ¼ cs , which yields the following compact formu-
bs

lation for OBPS: xb 2 f0, 1g, b 2 B, (35)

X where the objective is stated as maximizing the negation of
ðMPÞ minimize c s ys (24)
s2S
the total reduced costs of the batches. There are efficient
X algorithms to solve large-scale instances of this problem
subject to ais ys ¼ 1, i 2 I, (25) (Martello and Toth 1990). The batches P selected by KPSP
s2S aib 
x b > 1 for
s2S may cover orders more than once, i.e.,
X
ys  jKj, (26) an order i. This requires constraint set (25) to be relaxed to
s2S a greater-than-or-equal type. On the other hand, in the opti-
mal LP relaxation solution of this problem, this constraint
ys 2 f0, 1g, s 2 S: (27)
set is satisfied with equality since covering an order more
MP is a set partitioning problem with an extra constraint than once can never decrease the objective to total travel
that limits the number of selected schedules to the number time to retrieve all orders.
8 İ. MUTER AND T. ÖNCAN

The overview of the algorithm to find the optimal solu- enumeration of B– can be time-consuming and spawn many
tion of the MP given in (24)-(27) is outlined in Algorithm batches that will yield a large knapsack problem. Therefore,
3. We first initialize the set of batches in KPSP with B,  we postpone the complete enumeration by conducting a
which comprises the batches generated in the course of col- heuristic enumeration that produces only a subset of B–,
umn generation to solve the OB (Algorithm 1, Line 1), and denoted by B  (Line 9). Regardless of whether exact or
then the RMP is formed with S by assigning a distinct batch heuristic enumeration is called, the new batches are added
 to each s 2 S (Line 1). We first solve KPSP only with
in B to B  (Lines 10 and 17) before KPSP is solved (Lines 11
 until it fails to generate a picker
currently existing batches B and 18).
schedule, which follows Lines 3 and 4 of Algorithm 3. If s > v, a new schedule s is added to S (Lines 13, 29),
When it fails to generate a schedule, an algorithm must be and the column generation algorithm continues. If s  v
designed to generate schedules s 2 SnS that contain batches and / ¼ 0, the algorithm either stops at the optimal LP
 Since KPSP is a knapsack problem, it is well- solution or detects infeasibility. The former initializes efforts
missing in B:
to find the optimal integer solution in the solution of both
known for this problem that batches b 2 B with c b > 0 can
OBPSP (Algorithm 1) and MOBPSP (Algorithm 2). The LP
be eliminated without deteriorating the optimal value s.
infeasibility of OBPS prompts the increase of a along with T
Hence, the complete set of batches with negative reduced
in OBPSP (Algorithm 1, Line 7) whereas it certifies optimal-
cost, denoted by B ¼ fb 2 B : c b < 0g, are generated, and
ity for MOBPSP (Algorithm 2, Line 13). Note that re-invok-
then, KPSP formed of batch set B– is solved. This approach
ing column generation by setting T to a larger value does
to solve the KPSP is known as the two-phase approach,
not impact the currently existing schedules in S: On the
where a subset of variables in the PSP is generated through
other hand, decrementing T, as implemented in Algorithm 2
implicit enumeration in the first phase, and the PSP is
(Line 8), removes those schedules with total duration larger
solved with these variables in the second. It has been applied than T – 1 when the column generation is recalled.
to solve subproblems arising in problems such as the PBPM Finally, in order to find the integral optimal solution,
and the Multi-depot Vehicle Routing with Inter-depot rather than branch-and-price, we employ an enumeration
Routes (Muter et al. 2014). strategy, which has been applied to VRP (Baldacci et al.
2011) and its extensions, including MTVRP (Mingozzi et al.
Algorithm 3: Column Generation to solve (26)-(29) 2013), and OBP (Muter and Oncan € 2015). Given an upper
bound and a lower bound along with a set of dual variables
1 Initialize the RMP and PSP (S, B),  / ¼ 0;
associated with the constraints for an integer program, this
2 while True do method involves enumerating variables with a reduced cost
3 Solve RMP; smaller than or equal to the gap between these bounds. The
4 Solve KPSP(B); 
solution of the reduced integer problem consisting of these
5 if s > v then enumerated variables is also optimal for the original integer
6 S S [ s;
program (see Nemhauser and Wolsey (1998), page 389,
7 else Proposition 2.1). Accordingly, we generate the batches that
8 if / ¼ 1 then can be selected in the optimal solution of OBPS using the
9 Generate B ; upper bound (z OBPS ) and the lower bound (z OBPS ), where
10 Update B;  z OBPS comes from C  0 or solving the final RMP as an inte-
11 Solve KPSP(B)  ! fsg; max
ger problem by a MILP solver and z OBPS results from solv-
12 if s > v then ing MP by column generation. We denote these batches by
13 S S [ s;
B0 ¼ fb 2 B : c b  z OBPS  z OBPS g (Algorithm 3, Line 25),
14 Go to Line 3; where c b is calculated by using the dual variables retrieved
15 / ¼ 0; at the termination of the column generation algorithm. If
16 Generate B ¼ fb 2 B : c < 0g; the cardinality of B0 is no larger than a prespecified thresh-
17 Update B; 
old X, then the reduced OBPS ((4)-(7)) formed of B0 is
18 Solve KPSP(B)  ! fsg;
solved using an off-the-shelf MILP solver, which finds the
19 if s > v then optimal solution to the OBPS.
S S [ s;
20 For the generation of sets B– and B0 , we employ the
21 / ¼ 1; labeling algorithm designed for the OB, which was proposed
22 else in Muter and Oncan€ (2015). Briefly, this algorithm itera-
23 Stop; tively treats the orders in the order of their indices by
24 if Feasible then extending the labels accumulated on them to the later ones.
25 Generate B0 ¼ fb 2 B : c b  z OBPS  z OBPS g; A label Xbi residing on order i 2 I is associated with batch b
26 if jB0 j  X then
and holds the information on the orders it contains (RiP )
27 Solve (4)-(7);
and their total number of items (RiL ), total reduced cost (RiC )
P
In the early stage of column generation, in which the and total cost (RiB ), where RiC ¼ RiB  j2Ri uj : Extension of
P
dual variables are known to be erratic, the complete a label from order i to j > i involves checking whether the
 IISE TRANSACTIONS 9

Table 1. A list of symbols used in Algorithm 4 determining a fixed sequence, we prefer to consider a ran-
Symbol Definition dom ordering of neighborhood search structures. Finally,
B~ : an OBP solution found during the run of the VNS note that we have applied the sequential neighborhood
B~
: the best known feasible OBP solution change strategy rather than the nested one.
A~
: the best known feasible OBPSP solution A formal outline of the proposed VNS algorithm for both
C 0 : objective value, i.e. makespan, of the best known feasible OBPSP
max the OBP and OBPSP is depicted in Algorithm 4. The VNS
~z OB : objective value of the best found OBP
A : an OBPSP solution found during the run of the VNS
algorithm is performed given a picker routing policy such as
~
f ðBÞ : ~
objective function value of the OBP solution B the return, traversal and midpoint and a set of neighborhood
zðAÞ : objective function value of the OBPSP solution A search structures namely, N 1 , N 2 and N 3 : The output of
g : iteration limit the VNS algorithm consists of both OBP and OBPSP solu-
tions B~
and A~
with objective values ~z OB and C 0 ,
max
total number of items violates the capacity constraints respectively. For the sake of clarity, the notation used in
(RiL þ dj  C) and whether the lower bound on its total Algorithm 4 is presented in Table 1.
reduced cost upon completion satisfy the conditions for B– The VNS iterates g times. The parameter g plays a vital
or B0 : The heuristic enumeration of B role in the efficiency of both the VNS and Algorithm 2.
 eliminates some of
the batches through dominance rules, in which the total Hence, in order to find the most suitable value for this par-
numbers of items and reduced costs of pairs of batches on ameter we have run the VNS when g 2 ½jIj; 3  jIj: We
the same order are compared. Namely, a batch b1 associated have observed that fixing g ¼ 3  jIj seems to yield satisfac-
with Xbi 1 on i dominates another batch b2 associated with tory performance for most of the cases. Therefore, in our
computational experiments, we set g ¼ 3  jIj: Here j j
Xbi 2 only if RbC1  RbC2 and RbL1  RbL2 :
denotes the cardinality of a set. Given an initial OBP solu-
tion Bþ the VNS first runs the Perturbation procedure,
4. VNS algorithm which will probably yield an infeasible OBP solution. Then,
the FeasibilityChase procedure is performed with the aim of
In this section, we introduce a VNS algorithm specially restoring the feasibility. Unless a feasible OBP solution is
devised for the OBPSP. Although the main focus is to find a reached, the VNS continues with the next iteration where
promising OBPSP solution, the proposed VNS also explores the process will start from scratch with the Perturbation pro-
various OBP solutions. The VNS algorithm is a well-known cedure. For the other case, namely when a feasible OBP
metaheuristic algorithm used to solve several combinatorial solution is obtained then the VNS proceeds with the local
and global optimization problems (Hansen et al. 2019). As search process by first determining a random sequence of three
recent advances on the VNS we can mention a continuous neighborhood search operators N 1 , N 2 and N 3 : The neigh-
VNS algorithm for solving systems of nonlinear equations borhood search operators are performed one after the other
(Pei, Drazic, et al. 2019) and a novel hybridization of the
and the best known OBP solution, i.e., ~z OB , is updated when-
VNS with another metaheuristic (Pei, Liu, et al. 2019). The
ever it is necessary. Furthermore, after the run of each neigh-
VNS algorithm basically consists of the following three
borhood search operator the BatchesToPickersAssignment
steps: systematic change of the neighborhood, local search
procedure is performed to attain an approximate OBPSP solu-
and perturbation which are alternately performed until a
tion. Then, we compute the makespan of the OBPSP solution
stopping criterion is satisfied (Hansen and Mladenovic
obtained with the BatchesToPickersAssignment procedure. If
2001). Perturbation or shaking step is employed to jump
the makespan value of the incumbent OBPSP solution is less
from the local optimum solutions and start exploring distant
than the objective value of the best known feasible OBPSP
neighboring solutions. Local search consists of performing  0 , then we update C  0 : Note that when we
solution, i.e., C max max
one or many neighborhood search operators.
In the context of both the OBP and OBPSP, the neigh- do not update the best known feasible OBP solution then we
borhood search scheme includes three straightforward struc- consider the next neighborhood search operator in the ran-
tures, i.e., N 1 , N 2 and N 3 : The first structure, i.e., N 1 , is a domly predefined sequence. Furthermore, in the case where
move operator that chooses an order from a batch and none of the neighborhood search structures improves the best
inserts it into another one. The second one, namely N 2 , is known feasible OBP solution then we continue with the next
a swap operator that exchanges two orders belonging to two iteration of the VNS. The steps of the VNS algorithm are per-
different batches and the third structure, N 3 , is a cyclical formed until the iteration limit g is reached.
exchange operator that selects three orders and changes their Within the proposed VNS algorithm we perform three pro-
batches in cyclic order. cedures. They are Perturbation, BatchesToPickersAssignment
Note that, the sequence of neighborhood search struc- and FeasibilityChase, where each of them employs the match-
tures may also affect the overall performance of the VNS ing neighborhood proposed in Oncan € et al. (2008) to develop
(Huber and Geiger 2017). According to our preliminary very large-scale neighborhood search algorithms for the parti-
experiments, we have observed that none of the neighbor- tioning problems. We should also remark that the matching
hood search structure orders, including sorting them from neighborhood has been successively implemented within the
the smallest and fastest to explore, to the largest one and Iterated Local Search Algorithm with Tabu Thresholding for
slowest to explore, is the winner. Therefore, rather than €
the OBP in Oncan (2015). For the details of the neighborhood
10 İ. MUTER AND T. ÖNCAN

scheme, i.e., the Perturbation, BatchesToPickersAssignment and have also been employed by Muter and Oncan € (2015). In
FeasibilityChase procedures, we refer to the online supplement. this study, we have utilized these instances with maximum
capacity of the batches Q ¼ 24 and by varying the number
Algorithm 4: Variable Neighborhood Search of pickers, i.e., jKj ¼ f2, 3, 4g, as in Ardjmand et al. (2018).
In our test bed, the number of orders jNj ranges between 20
1 Initialization: Given an initial OBP solution Bþ compute and 100 in increments of 10 with 10 instances generated for
f ðBþ Þ, i.e. the cost function of the initial OBP solution; each order size, and the locations of the items in the orders
2 Set ~z OB f ðBþ Þ and C 0 1; are randomly generated within the layout depicted with
max
3 while t < g do Figure 1, in which L ¼ 10, w ¼ 2.4. Consequently, we con-
4 B ~ Perturbation(Bþ); \
Random move
\ sider in total 270 test instances for each routing policy. All
5 ~
if Infeasible(B) then test instances are available upon request from the corre-
6 ~
B FeasibilityChase(B); ~ sponding author.
~
7 if Feasible(B) then The key parameter for OBPS that drives the trade-off
8 \
Local search begins
\ between the total travel time to retrieve all orders and make-
9 Determine a random sequence pðiÞ for i ¼ 1, :::, 3 span is a, which is initialized as T=C max : In real-life instan-
10 k 0 \
Sequential neighborhood change step
\ ces of OBPSP, T is known, however, in generating instances,
11 while k  3 do we set it to
~ < ~z OB then  OB 
12 if f ðBÞ z
T¼ a
13 B~
~ ~z OB
B; ~ k
f ðBÞ; 0; jKj
14 A BatchesToPickersAssignment(B); ~
and a to 1.05. If OBPS is infeasible for a ¼ 1:05, we
15 
if zðAÞ < C max then
0
increase it by 0.05. This means that with the initial setting
16 0
C max zðAÞ; A~
A; a ¼ 1:05, the optimal objective of OBPS is imposed to be
\
Update the best feasible OBPS solution
\ no worse than approximately 5% of that of the OB as
17 k k þ 1; explained in Remark 1, and likewise, the makespan of the
18 B~ ~
N pðkÞ ðBÞ; OBPS solution is also limited to be within 5% of the lower
19 \
Local search ends
\ bound on makespan. A similar approach has been taken in
20 t t þ 1; the generation of the MTVRP instances in which T is set to
21 Output: B~
, A~
, ~z OB , C  0 and stop; 

max z 1:05
T¼ ,
jKj
5. Computational experiments where z
is the best bound for VRP and ½z returns the inte-
In this section, we report the results of the computational ger value closest to z. On the other hand, the value of a has
experiments on the randomly generated instances. We tested been limited to 1.05 and 1.1 in the literature, which leads to
Algorithm 1 and 2 that are proposed to solve OBPSP and a feasible solution in most of the instances with at least 50
MOBPSP using three routing policies, namely traversal, customers. Even though this is also the case in our experi-
return and midpoint, and compared their performances on ments, in several instances with small number of orders, we
the instances. In the literature, the only exact method that is show that larger a values may be needed to achieve a feas-
based on solving a model using an off-the-shelf MILP solver ible/optimal solution. We remark that this approach can be
is proposed in Ardjmand et al. (2018) where makespan modified to generate the Pareto front by increasing a in
minimization is the sole objective. The authors tested their small increments, starting from one. Increasing the value of
model on instances with up to 25 orders and small picker a after the model finds a feasible solution may decrease the
capacity, and the solver finds the optimal solution in a time objective of OBPS. Finally, we set X ¼ 100, 000, that is, the
limit of 1800 seconds for only instances with five orders. MILP solver to find the optimal solution of OBPS is called
The MOBPSP minimizes not only the makespan as the pri- only when B0  X:
mary objective, but also the total travel time to collect all Recall that Algorithms 1 and 2 comprise two phases, each
orders as the secondary one. We will show that we solve associated with the solution of one optimization problem.
instances of both OBPSP and MOBPSP with up to 100 The first one involves the solution of OB, and the second
orders to optimality. one takes the solution of OB as an input and solves the
The experiments were conducted on the nodes of com- OBPS, possibly several times for various values of T. We
puting cluster Balena hosted at the University of Bath, each allocate 1800 seconds to each phase, including the column
with two Intel E5-2650 v2 CPUs (16 cores in total) running generation, enumeration of columns for the reduced MILPs
at 2.60 GHz. The LPs and MILPs solved in the course of the and solution of these MILPs. We did not incorporate the
column generation for OB and MP are solved by CPLEX running time of the VNS into the time limits due to its
12.9. As the test bed, we have adopted the instances which short duration compared with the other components. If the
€
had been generated by Oncan (2015) following the proced- OB is not solved to optimality due to time limit or B0 > X,
ure proposed in Hwang and Kim (2005). These instances Algorithms 1 and 2 terminate with the unsolved label.
 IISE TRANSACTIONS 11

Table 2. Result of the proposed methodology with return policy

MOBPS OBPS
Cmax


C max zMOBPS ðC max C 0max Þ z OBPS
zOBPS Cmax Cmax
jIj  jKj # MOBPS C max Cmax

zOB C max Time z OB zOB Cmax

C max # OBP # OBPS Tot. Time
20-2 9 3.2 2.1 2.4 0.9 433.7 2.2 1.9 0.3 3.6 10 10 56.1
20-3 10 6.2 5.4 5.3 3.2 158.6 6.5 4.4 0.4 6.6 10 10 73.6
20-4 10 17.3 13.5 14.7 7.3 430.1 11.4 12.1 1.2 18.7 10 10 130.1
30-2 3 0.2 0.0 0.0 – 2063.9 0.5 0.1 0.0 1.6 10 9 225.9
30-3 1 0.0 0.2 0.0 1.4 1830.0 2.4 0.0 0.2 5.1 10 4 960.5
30-4 5 3.3 5.2 2.2 0.5 1329.5 1.0 1.2 1.7 4.5 10 6 262.7
40-2 7 0.1 0.0 0.0 – 601.9 – 0.0 0.0 0.1 10 10 189.4
40-3 1 0.2 1.9 0.3 – 584.1 0.7 0.0 1.9 2.9 10 6 202.2
40-4 0 – – – – 1417.0 1.1 0.1 – 5.6 10 4 121.5
50-2 5 0.1 0.3 0.0 – 657.3 – 0.0 0.3 0.3 7 7 299.1
50-3 2 0.1 0.0 0.0 – 1363.0 0.2 0.0 0.0 0.9 7 7 339.1
50-4 0 – – – – 1149.4 0.4 0.0 – 2.6 7 5 348.4
60-2 7 0.0 0.0 0.0 – 55.0 – 0.0 0.0 0.0 7 7 30.9
60-3 2 0.0 0.0 0.0 – 1705.5 – 0.0 0.0 0.7 7 7 86.3
60-4 2 0.1 0.0 0.0 – 513.6 – 0.0 0.0 1.4 7 6 91.8
70-2 6 0.0 0.0 0.0 – 140.7 – 0.0 0.0 0.0 6 6 155.8
70-3 5 0.0 0.0 0.0 – 333.7 – 0.0 0.0 0.0 6 6 244.3
70-4 2 0.0 0.0 0.0 – 611.0 – 0.0 0.0 0.9 6 6 233.2
80-2 6 0.0 0.0 0.0 – 242.2 – 0.0 0.0 0.0 6 6 80.3
80-3 6 0.0 0.0 0.0 – 79.0 – 0.0 0.0 0.0 6 6 78.0
80-4 5 0.0 0.0 0.0 – 537.2 – 0.0 0.0 0.0 6 6 163.5
90-2 5 0.0 0.0 0.0 – 143.6 – 0.0 0.0 0.0 5 5 134.8
90-3 5 0.0 0.0 0.0 – 157.0 – 0.0 0.0 0.0 5 5 159.0
90-4 4 0.0 0.0 0.0 – 506.2 – 0.0 0.0 0.0 5 5 76.4
100-2 5 0.0 0.0 0.0 – 782.3 – 0.0 0.0 0.0 6 6 283.8
100-3 6 0.0 0.0 0.0 – 171.0 – 0.0 0.0 0.0 6 6 169.0
100-4 5 0.0 0.0 0.0 – 69.3 – 0.0 0.0 0.1 6 6 78.2

The results of our proposed methodology on the ran- z OBPS attains tighter bounds than z OB in all tables. The
domly generated instances using return, midpoint and tra- second column shows the ratio of the optimal solutions for
versal routing policies are reported in Table 2, Table 3 and OBPS
OBPS and OB (zzOB ). The next column in these tables,
Table 4, where the latter two are given in the online supple- 
 max and
namely C max , summarizes the relative gap between C
Cmax
ment. The instances are identified with their order and
picker sizes, jIj  jKj, in the first column, where each row Cmax which is the makespan outputted by Algorithm 1. The
reports the results of the solution of 10 instances. The ratios figures under “# OBP” and “# OBPS” are the number of
represent the percentage gaps, i.e., OBP and OBPS instances, respectively, that are solved to
ðU1 U2 Þ100
optimality. The last column gives the solution time of
U1
U2 reports U2 : Algorithm 1 for OBPS. We point out that all percentage
The table is split into two sets of columns: the columns gaps and times are averaged over instances that are solved
associated with the solution of MOBPSP and those associ- to optimality.
ated with the solution of OBPSP. In the former, we report The major finding is that the upper bound C  max is very
the number of instances solved to optimality (# MOBPS), tight for instances with 50 or more orders. In fact, it coin-
the average percentage gap between the optimal makespan cides with the optimal makespan in many instances.

C
Concretely, it means that solving the OB problem suffices to
Cmax and the lower bound of makespan C max (C max ), the
max find the optimal for the OBPS and MOBPS in that the opti-

average percentage gap between the upper bound C max and mal batch set of Bþ lends itself to a perfect partition in the



Cmax (CC
max ) and the average percentage gap between zMOBPS solution of the parallel machine scheduling problem (15)-
max
MOBPS
and zOB (z zOB ), which shows the deterioration in the min- (18). On the other hand, in small instances, especially those
with 20 orders, the gap between the upper bound and the
imum total travel time to collect all orders after the make-
optimal makespan is large for all routing policies. For
span is minimized. In the last two columns, we report the
 max instances with 20 and 30 orders, VNS improves the upper
percentage gap between C  max and C
0 C
max (C
 0 ), which shows bound C  max , which then enhances the performance of
max
the improvement in upper bound achieved by VNS, and the Algorithm 2. Moreover, in these instances, the general pat-
solution time of Algorithm 2. In the second set of columns tern is that the gap between the upper bound C  max and the
allotted to OBPSP, we first report the average percentage lower bound C max increases as the number of pickers
ratio of the lower bounds of column generation applied to increases. A similar pattern occurs for the gap between the
MP and OB, denoted by z OBPS and z OB , respectively. When optimal makespan Cmax

and the lower bound C max , and the
€
calculating z OB , Muter and Oncan (2015) utilized valid 
upper bound C max and the optimal makespan Cmax

: This
inequalities to enhance the LP relaxation bound of column increase in gaps with the number of pickers manifests itself
generation, which we did not employ in our proposed col- as fewer numbers of instances solved to optimality in all
umn generation for MP. Except for a couple of instances, routing policies, i.e., the largest number of instances solved
12 İ. MUTER AND T. ÖNCAN

to optimality is when jKj ¼ 2: The gap between C  max and collect all orders and the makespan of the orders that arises

in the wave picking systems. These two objectives are inher-
Cmax is smaller in virtually all instances than that between

ently linked, which we show by analyzing their bounds. We
Cmax and C max , that is, the optimal makespan is closer to
the upper bound than the lower bound. In the results of devised two models, each of which prioritize one of these
OBPS
OBPS, note that CCmax
max
bounds zzOB from above, as pointed out objectives and developed exact algorithms to find their opti-


Cmax mal solutions. Both of these algorithms were based on col-
in Remark 2. The same interpretation also applies to C max umn generation, which generates picker schedules formed of
MOBPS
and z zOB : In both bounds for OBPS and MOBPS, the aver- order batches. The challenge we faced in forging a solution
age deterioration in zOB with the assignment of batches to algorithm for the subproblem of column generation was the
pickers is approximately half of that in the lower bound dependence of it on a set of batches that is large in size.
of makespan. Our approach was to simplify this subproblem to a knapsack
Algorithms 1 and 2 demonstrate similar results on all of problem by relaxation and implicitly enumerating the neces-
the routing heuristics in terms of finding the optimal solu- sary batches before its solution.
tions for OBPS and MOBPS, respectively. Out of 270 instan- We have conducted extensive computational experi-
ces, the number of instances solved to optimality for ments, in which the proposed algorithms are used to solve
MOBPS using return, midpoint and traversal are 124, 93 and randomly generated instances from order batching litera-
72, respectively. For the OB (OBPS), the number of instan- ture. We showed that for the large instances, the picker
ces solved to optimality are 201ð177Þ, 171ð143Þ and scheduling aspect of the problem becomes trivial as the
147ð126Þ using return, midpoint and traversal, respectively. order batching solution gives a very tight bound for both
We point out that the larger the number of instances solved objectives considered in this article. This enabled us to
to optimality for OB, the more instances for OBPS and solve instances with up to 100 orders to optimality. We
MOBPS are successfully solved. Overall, the return policy
have also discussed similarity of this problem with the
features more solved instances than the others, due to the
problems from different realms, such as vehicle routing
fact that its lower and upper bounds are tighter than the
and machine scheduling, which will be examined in future
other policies. The solution times of the problems are in
research for unification.
line with the number of solved instances, that is, the solu-
tion time of the return routing policy is the shortest among
three policies. The solution time does not increase substan-
tially in tandem with the number of orders. On the contrary, Acknowledgments
some of the instances with 20 orders have long solution This research was done prior to the first author’s employment at
times, due to the value of a being updated a number of Amazon. Thanks are due to anonymous referees for their valuable
times in the course of Algorithm 1, which invokes comments which have helped improve the paper.
Algorithm 3 as many times.
Last, but not least, we suggest the first exact algorithms
for both OBPSP and MOBPSP. In order to evaluate the
Funding
efficiency of the proposed algorithms we consider solving The second author acknowledges the partial support by the Scientific
the MOBPSP via the multi-objective solver of CPLEX. To and Technical Research Council of Turkey - TUB € ITAK
_ (Grant
this end, we have developed a multi-objective MILP model No: 217M477).
of the MOBPS. Although, we only present a multi-objective
MILP model of the MOBPS with return policy in the sup- Notes on contributors
plemental online material, the multi-objective MILP models
_
Ibrahim Muter is a Senior Research Scientist at Amazon Web Services
of the MOBPS with other routing policies can be devised
(AWS). He received his PhD in Industrial Engineering from Sabancı
similarly. We have performed some preliminary experi- University in 2011. Before joining AWS, he worked at Bahçeşehir
ments by trying to solve the MOBPS for small-sized instan- University as assistant professor and then at University of Bath as
ces with 20 orders. However we could not obtain optimum senior lecturer. His research interests include integer programming,
solutions within a CPU time limit of 14 hours and for only decomposition methods and their applications.
three instances CPLEX could report solutions with €
Temel Oncan is a Professor of Industrial Engineering at Galatasaray
unknown status. In addition, for instances with size larger University. He received his PhD in Industrial Engineering from
than 20 orders, the CPLEX’s multi-objective solver could _
Bogaziçi University, Istanbul, in 2004. His research interests include
not report a solution within a CPU time limit of 14 hours. combinatorial optimization, integer programming, network design,
Hence, for the sake of brevity we do not report the details facility location and routing problems. He has published papers in
of our experiments with the multi-objective solver Computer & Operations Research, European Journal of Operational
Research, IIE Transactions, Annals of Operations Research and Journal
of CPLEX.
of the Operational Research Society.

6. Conclusions ORCID
In this article, we have tackled the integrated OBPSP, and _
Ibrahim Muter http://orcid.org/0000-0003-2202-4751
considered two objectives, namely the total travel time to €
Temel Oncan http://orcid.org/0000-0003-0595-0673
 IISE TRANSACTIONS 13

References Huber, S. and Geiger, M.J. (2017) Order matters - a variable neighbor-
hood search for the swap-body vehicle routing problem. European
Ardjmand, E., Shakeri, H., Singh, M. and Bajgiran, O.S. (2018) Journal of Operational Research, 263, 419–445.
Minimizing order picking makespan with multiple pickers in a wave Hwang, H. and Kim, D.G. (2005) Order-batching heuristics based on
picking warehouse. International Journal of Production Economics, cluster analysis in low-level picker-to-part warehousing system.
206, 169–283. International Journal of Production Research, 43(17), 3657–3670.
Baldacci, R., Mingozzi. A. and Roberti, R. (2011) New route relaxation L€
ubbecke, M.E. and Desrosiers, J. (2005) Selected topics in column
and pricing strategies for the vehicle routing problem. Operations generation. Operations Research, 53(6), 1007–1023.
Research, 5(5), 1269–1283. Martello, S. and Toth, P. (1990) Knapsack Problems: Algorithms and
Chabot, T., Lahyani, R., Coelho, L.C. and Renaud, J. (2017) Order Computer Implementations, John Wiley & Sons, New York, NY.
picking problems under weight, fragility and category constraints. Matusiak, M., de Koster, R. and Saarinen, J. (2017) Utilizing individual
International Journal of Production Research, 55(21), 6361–6379. picker skills to improve order batching in a warehouse. European
Chen, T-L., Cheng, C-Y., Chen, Y-Y. and Chan, L-K. (2015) An effi- Journal of Operational Research, 263, 888–899.
cient hybrid algorithm for integrated order batching, sequencing Mingozzi, A., Roberti, R. and Toth, P. (2013) An exact algorithm for
and routing problem. International Journal of Production Economics, the multi-trip vehicle routing problem. INFORMS Journal on
159, 158–167. Computing, 25, 193–207.
Dantzig, G.B. and Wolfe, P. (1960) Decomposition principle for linear Muter, I. _ (2020) Optimization of single and parallel batch processing
programs. Operations Research, 8(1), 101–111. machines. European Journal of Operational Research, 285(2),
de Koster, R., Le-Duc, T. and Roodbergen, K.J. (2007) Design and con- 470–483.
trol of warehouse order picking: A literature review. European Muter, I.,_ Cordeau, J-F. and Laporte, G. (2014) A branch-and-price
Journal of Operational Research, 182, 481–501. algorithm for the multidepot vehicle routing problem with interde-
de Koster R., Roodbergen, K. and van Voorden, R. (1999) Reduction of
pot routes. Transportation Science, 48(3), 425–441.
walking time in the distribution center of De Bijenkorf, in New Muter, I. €
_ and Oncan, T. (2015) An exact solution approach for the
Trends in Distribution Logistics, Springer, Berlin, pp. 215–234.
order batching problem. IIE Transactions, 47(7), 728–738.
Ehrgott, M. (2005) Multicriteria Optimization. Springer-Verlag, Berlin,
Nemhauser, G.L. and Wolsey, L.A. (1988) Integer and Combinatorial
Heidelberg.
Optimization, John Wiley & Sons, New York, NY.
Faber, N., de Koster, R.B.M. and Smidts, A. (2018) Survival of the fit- €
Oncan, T. (2015) MILP formulations and an iterated local search algo-
test: The impact of fit between warehouse management structure
rithm with tabu thresholding for the order batching problem.
and warehouse context on warehouse performance. International
European Journal of Operational Research, 243(1), 142–155.
Journal of Production Research, 56(1-2), 120–139. €
Oncan, T., Kabadi, N.S., Nair, K.P.K. and Punnen, A.P. (2008) VLSN
Frazelle, E. (2001) Supply Chain Strategy, McGraw-Hill, New York,
search algorithms for partitioning problems using matching neigh-
NY.
bourhoods. Journal of the Operational Research Society, 59(3),
Gademann, N. and van de Velde, S. (2005) Order batching to minimize
total travel time in a parallel-aisle warehouse. IIE Transactions, 388–398.
37(1), 63–75. Pei, J., Drazic, Z., Drazic, M., Mladenovic, N. and Pardalos, P.M.
Gademann, N., van Den Berg, J.P. and van Der Hoff, H.H. (2001) An (2019) Continuous variable neighborhood search (C-VNS) for solv-
order batching algorithm for wave picking in a parallel-aisle ware- ing systems of nonlinear equations. INFORMS Journal on
house. IIE Transactions, 33, 385–398. Computing, 31(2), 235–250.
Graham, R., Lawler, E.L., Lenstra, J.K. and Rinnooy Kan, A.H.G. Pei, J., Liu, X., Fan, W., Pardalos, P.M. and Lu, S. (2019) A hybrid BA-
(1977) Optimization and approximation in deterministic sequencing VNS algorithm for coordinated serial-batching scheduling with dete-
and scheduling: A survey. Annals of Discrete Mathematics, 51, riorating jobs, financial budget, and resource constraint in multiple
187–326. manufacturers. Omega, 82, 55–69.
Hansen, P. and Mladenovic, N. (2001) Variable neighborhood search: Petersen, C.G. (1997) An evaluation of order picking routeing policies.
Principles and applications. European Journal of Operational International Journal of Operations and Production Management,
Research, 130, 449–467. 17(11), 1098–1111.
Hansen, P., Mladenovic, N., Brimberg, J. and Moreno Perez, J.A. Petersen, C.G. and Schmenner, R.W. (1999) An evaluation of routing
(2019) Variable neighborhood search, in Handbook of and volume-based storage policies in an order picking operation.
Metaheuristics, Springer, New York, NY, pp. 57–97. Decision Sciences, 30(2), 481–501.
Henn, S. (2012) Algorithms for on-line order batching in an order Pinto, A.R.F., Crepaldi, A.F. and Nagano, M.S. (2018) A genetic algo-
picking warehouse. Computers and Operations Research, 39, rithm applied to pick sequencing for billing. Journal of Intelligent
2549–2563. Manufacturing, 29, 405–422.
Henn, S. (2015) Order batching and sequencing for the minimization Pinto, A.R.F. and Nagano, M.S. (2020) Genetic algorithms applied to
of the total tardiness in picker-to-part warehouses. Flexible Services integration and optimization of billing and picking processes.
and Manufacturing Journal, 27, 86–114. Journal of Intelligent Manufacturing, 31, 641–659.
Henn, S. and Schmid, V. (2013) Metaheuristics for order batching and Scholz, A., Schubert, D. and W€asher, G. (2017) Order picking with
sequencing in manual order picking systems. Computers and multiple pickers and due dates - simultaneous solution of order
Industrial Engineering, 66(2), 338–351. batching, batch assignment and sequencing, and picker routing
Henn, S. and W€asher, G. (2012) Tabu search heuristics for the order problems. European Journal of Operational Research, 263, 461–478.
batching problem in manual order picking systems. European Tompkins, J.A., White, J.A., Bozer, Y.A. and Tanchoco, J.M.A. (2010)
Journal of Operational Research, 222, 484–494. Facilities Planning, John Wiley & Sons, Hoboken, NJ.