European Journal of Operational Research: Felix Weidinger, Nils Boysen, Michael Schneider

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 15

European Journal of Operational Research 274 (2019) 501–515

Contents lists available at ScienceDirect

European Journal of Operational Research


journal homepage: www.elsevier.com/locate/ejor

Discrete Optimization

Picker routing in the mixed-shelves warehouses of e-commerce


retailers
Felix Weidinger a, Nils Boysen a, Michael Schneider b,∗
a
Friedrich-Schiller-Universität Jena, Lehrstuhl für Operations Management, Carl-Zeiss-Str. 3, 07743 Jena, Germany
b
RWTH Aachen University, Deutsche Post – Chair Optimization of Distribution Networks, Kackertstr. 7, 52072 Aachen, Germany

a r t i c l e i n f o a b s t r a c t

Article history: E-commerce retailers face the challenge to assemble large numbers of time-critical picking orders, of
Received 29 August 2017 which each typically consists of just a few order lines and low order quantities. To efficiently solve this
Accepted 11 October 2018
task, many warehouses in this segment are organized according to the mixed-shelves paradigm. Incoming
Available online 16 October 2018
unit loads are isolated into single units, which are randomly spread all over the shelves of the warehouse.
Keywords: In such a setting, the probability that a picker always finds a demanded stock keeping unit (SKU) close-by
Facility logistics is high, irrespective of his/her current position in the warehouse. In spite of this organizational adaption,
Warehousing picker routing, i.e., the sequencing of shelf visits when retrieving a set of picking orders, is still an im-
Mixed-shelves portant optimization problem. In a mixed-shelves warehouse, picker routing is much more complex than
Picker routing in traditional environments: Multiple orders are concurrently assembled by each picker, many alternative
depots are available, and items of the same SKU are available in multiple shelves. This paper defines the
resulting picker-routing problem in mixed-shelves warehouses and provides efficient solution methods.
Furthermore, we use the developed methods to explore important managerial aspects. Specifically, we
benchmark mixed-shelves storage against traditional storage policies for scenarios with different ratios
between small-sized and large-sized orders. In this way, we investigate whether mixed-shelves storage is
also a suitable policy if an omni-channel sales strategy is pursued, and large-sized orders of brick-and-
mortar stores as well as small-sized online orders are to be jointly processed by the same warehouse.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction mortar stores (Brynjolfsson, Hu, & Smith, 2003). Consequently,


most online retailers offer a large assortment.
When setting up their warehouse operations, e-commerce re- • Varying workloads: Most online retailers have dynamically
tailers directly serving final customer demands in the business-to- grown over the past years (Laudon & Traver, 2007). Depending
consumer (B2C) segment typically face the following challenges: on the offered products, they face highly volatile demands, e.g.,
due to end-of-season sales. Thus, warehouses should be scal-
• Small orders: Private households tend to order rather few items able to flexibly react to varying workloads.
in low quantities, so that the typical picking order in the B2C • Tight delivery schedules: Next-day or even same-day deliver-
segment consists of just a few order lines, each demanding only ies have become an elementary component of many business
very few items. According to personal communications with a models and supply chains especially in the B2C segment (e.g.
German Amazon warehouse manager, their average order con- Yaman, Karasan, & Kara, 2012). This narrows the time window
tains only about 1.6 items. for order picking considerably and puts increasing stress on
• Large assortment: The term “long tail” describes the phe- warehouse operations.
nomenon that niche products account for a much larger
A warehousing setup specifically designed for these require-
proportion of sales in e-commerce than they do in brick-and-
ments is a mixed-shelves warehouse (e.g., Weidinger, 2018). Here,
unit loads are purposefully broken down into single items and ran-
domly spread all over the racks of the warehouse. This storage

Corresponding author.
policy is also denoted as scattered storage (Weidinger & Boysen,
E-mail addresses: felix.weidinger@uni-jena.de (F. Weidinger), nils.boysen@uni-
jena.de (N. Boysen), schneider@dpo.rwth-aachen.de (M. Schneider). 2018). With stock keeping units (SKUs) being stored all around,
URL: http://www.om.uni-jena.de/ (N. Boysen), http://www.dpo.rwth-aachen.de/ there is a good chance to always have one of the demanded
(M. Schneider) items close-by, irrespective of where the picker is currently

https://doi.org/10.1016/j.ejor.2018.10.021
0377-2217/© 2018 Elsevier B.V. All rights reserved.
502 F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515

Fig. 1. Basic elements of a mixed-shelves warehouse.

positioned in a warehouse. In this way, the unproductive walk- All around the warehouse, there are multiple access points to a
ing time of the pickers, which is the main drawback of traditional conveyor system, which moves completed bins towards the pack-
picker-to-parts systems (in which unit loads are kept together, e.g., ing stations of the shipping area. Thus, there is not a single central
de Koster, Le-Duc, & Roodbergen, 2007), is considerably reduced. depot but multiple ones spread all over the warehouse. Next to the
This helps to realize the tight delivery schedules of e-commerce. access points, there are stacks of empty bins, which can directly be
Moreover, storing a large assortment seems unproblematic as long loaded onto the cart for initiating the retrieval of subsequent pick-
as there is enough space for adding shelves to the warehouse. The ing orders.
standardized and low-tech equipment also facilitates the scalability In this way, the shelves are successively emptied by the pick-
in case of varying capacity situations; additional pickers and racks ers, so that refilling the shelves is another important task. Other
can simply be added to or removed from a warehouse. Finally, the than picking, which is executed under great time pressure due to
small order sizes reduce the chance that a picker has to visit mul- the tight delivery schedules, refill operations are not (that) time-
tiple storage positions to accumulate enough items of a SKU that is critical. At a distribution center we visited, refill operations have
ordered in a larger quantity. This would increase the walking time, been used to get new employees acquainted with the warehous-
so that care has to be taken that the SKU diversity of the racks fits ing processes and to employ redundant workforce during off-peak
the order sizes in a mixed-shelves warehouse. hours.
Mixed-shelves are applied in the European distribution centers In different real-world applications, the above processes may, of
of online retailer Amazon but also by many other warehouses in course, slightly vary. For instance, a pick-by-voice system can di-
the B2C segment (Weidinger & Boysen, 2018). rect the pickers and picking carts and mezzanine systems are not
a fixed requirement. Also, each bin filled by a picker is not nec-
essarily used to collect only a single order but can be assigned
1.1. Operations in a mixed-shelves warehouse to a batch of orders that are jointly picked. Traditionally, batch-
ing is particularly effective if the same SKU is needed for multiple
A typical warehouse setting of B2C online retailers operated un- orders (de Koster et al., 2007). In the packing and shipping area,
der the mixed-shelves policy applies the basic elements depicted however, these batches need to be separated into customer orders
in Fig. 1 and can be described as follows: again, e.g., by an automated sorter (Gallien & Weber, 2010; Johnson
The racks for storing items are headhigh to be conveniently ac- & Meller, 2002).
cessible by human pickers, who walk along the aisles while push-
ing small picking carts. A cart carries multiple bins in parallel, each 1.2. Decision problems and literature review
dedicated to a specific picking order. Due to the small order sizes,
the picking carts—although being small and agile—are still able to The main decision problems to be solved under a mixed-shelves
transport up to a handful of bins concurrently. To use space more policy are basically the same as in any other warehouse. Among
efficiently, the racks are often arranged on top of each other in the most elementary decisions are (e.g., Bartholdi & Hackman,
multi-story mezzanine systems. The SKUs are stored according to 2014):
the mixed-shelves policy, so that all items of a SKU are spread all • the storage assignment problem, which assigns SKUs to storage
over the warehouse. To route the workers during order picking, locations, and
they carry a small handheld scanner. The integrated display lists • the picker-routing problem, which determines the routes of the
which order line from which storage location is to be retrieved pickers through the warehouse when retrieving picking orders.
next. The optimization approach for picker routing that is devel-
oped in this paper can thus be applied to navigate the pickers via Instead of trying to summarize the vast body of literature which
their handhelds. Furthermore, the scanner is used to acknowledge has accumulated on these topics, we merely refer to the recent
the retrieval of an item from a shelf and its put-away in a bin. in-depth review papers, e.g., provided by de Koster et al. (2007)
In this way, the background information system also receives in- and Gu, Goetschalckx, and McGinnis (2007); Gu, Goetschalckx,
formation on empty storage positions where newly retrieved items McGinnis (2010). Our focus is on the peculiarities of mixed-shelves
can be restocked. warehouses. In real-world mixed-shelves warehouses, storage
F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515 503

assignments are typically not optimized but simply randomly de- variant determines cost-minimal routes that serve exactly one cus-
termined (Weidinger & Boysen, 2018). The paper on hand concen- tomer of each subset. All these VRP variants share some similari-
trates on the latter problem and aims to optimize picker routes in ties, but are not directly applicable to our picker routing problem.
mixed-shelves warehouses. Compared to traditional picker-routing It can be concluded that optimization approaches suited for
problems, where a single order is to be picked from a given set of picker routing in mixed-shelves warehouses are not yet available.
storage positions, and each tour starts and ends at a central depot, Note that this result seems especially unsatisfying because of the
the following peculiarities have to be considered: extreme time pressure (and the large potential of optimized picker
routes to reduce this pressure) existent in the warehouses of on-
• Items of the same SKU are available in multiple shelves, so that
line retailers.
an additional selection problem is to be solved. Moreover, or-
der lines requesting multiple items may need to be collected
by visiting multiple storage positions of the respective SKU. 1.3. Contribution and paper structure
• The picking cart is able to carry multiple bins concurrently, so
that multiple orders are processed in parallel. This paper treats a rich picker-routing problem, which is rele-
• There exist multiple access points to the conveyor system, so vant in the mixed-shelves warehouses of online retailers. Specif-
that we have many alternative depots where finished orders ically, the following peculiarities are considered: Multiple orders
can be handed over, and new orders can be initiated. are concurrently assembled by a picker, several alternative depots
are available, and items of the same SKU are available in multi-
Picker routing with alternative pick locations has been consid-
ple shelves. The resulting problem is defined (Section 2), and effi-
ered by Daniels, Rummel, and Schantz (1998). They solve the re-
cient solution procedures are introduced (Section 3) and tested in
sulting problem with facultative distance matrices by a tabu search
a comprehensive computational study (Section 4.3). Furthermore,
procedure. Weidinger (2018) addresses the same problem for rect-
we address important managerial aspects. A major drawback of
angular warehouses. These warehouses have specially structured
mixed-shelves warehouses is that multiple storage locations have
distance matrices, for which the (pure) routing problem is solvable
to be accessed whenever larger order quantities occur that have to
in polynomial time (Ratliff & Rosenthal, 1983). Integrating alterna-
be accumulated from multiple shelves. Therefore, we benchmark
tive pick locations, however, makes the problem strongly NP-hard
mixed-shelves storage against traditional storage policies for sce-
even for rectangular warehouses (Weidinger, 2018), so that simple
narios with different ratios between small-sized and large-sized
myopic selection rules coupled with the efficient routing procedure
orders. In this way, we investigate whether mixed-shelves stor-
of Ratliff and Rosenthal (1983) are applied. However, in both pa-
age is also a suitable policy if an omni-channel sales strategy is
pers the latter two peculiarities have not been considered. They
pursued, and large-sized orders of brick-and-mortar stores as well
assume a single central depot, exclude picking carts and a parallel
as small-sized online orders are to be jointly processed by the
processing of multiple orders.
same warehouse. This issue is addressed in Section 5 and, finally,
With a single central depot, a picking cart with capacity for
Section 6 concludes the paper.
multiple bins leads to a traditional batching problem (de Koster
et al., 2007 for an overview on the literature). In addition to the
routing problem, it has to be decided which orders should be 2. Problem definition
jointly retrieved in a tour. With multiple access points a picker
may pass some depot anyway, so that some completed order can Consider a mixed-shelves warehouse consisting of a set P of
be handed over en passant although other orders of the current storage positions, each initially storing γ p ≥ 0 items of one SKU
batch are not yet completed. Thus, a dynamic batching problem s ∈ S, where S denotes the total SKU set. Without loss of general-
arises, which (to the best of the authors’ knowledge) has not been ity, we assume that each storage position stores just a single type
treated in the warehousing literature. of SKU, which is defined by ηp . If a shelf carries items of multiple
Multiple depots are also rarely considered in the warehous- SKUs, then multiple storage positions having zero distance among
ing literature. In a related setting, conveyors run along the cross- each other can be introduced. Furthermore, we have a given set D
aisles, so that each tour can start and end anywhere in a cross of depots (also denoted as access points), where completed picking
aisle. De Koster and van der Poort (1998) extend the algorithm of orders are handed over to the conveyor system, and empty bins for
Ratliff and Rosenthal (1983) to derive optimal routes for rectan- new orders are retrieved. In total, we have V = P ∪ D positions in
gular warehouses in polynomial time for such a setting. We, how- the warehouse, and the picker’s walking distance among storage
ever, consider a discrete set of depots spread arbitrarily around the positions and/or depots v and v is denoted by dvv . We consider
warehouse. a single order picker, whose picking cart can carry up to C bins
Routing pickers in warehouses shares many similarities with concurrently for collecting picking orders. Initially, the picker is lo-
routing delivery tours of vehicles. Here, our three peculiarities may cated at an arbitrary depot. We have a given set of picking orders
be relevant too, so that we briefly summarize related problem set- O, and aos defines the number of items of SKU s demanded by or-
tings for extensions of the vehicle-routing problem (VRP) and its der o ∈ O.
derivatives. A truck has capacity for multiple items anyway and We aim at a solution defining the picker route through the
multiple depots are common in VRPs too (e.g., Montoya-Torres, warehouse such that all picking orders are retrieved. Specifically,
Franco, Isaza, Jiménez, & Herazo-Padilla, 2015). However, contrary a solution is represented by a sequence π defining the successive
to our setting, the depots generally mark the end point of a vehi- visits of the picker at positions in V. Each sequence position π i ,
cle route at which all orders are completely fulfilled. A more re- i = 1, . . . , m, is specified by a triple (v, δ, o). We have to distinguish
lated variant are VRPs with intermediate stops (Schneider & Hof, three cases: If the picker is located at a storage position, i.e., v ∈ P
2015), in which a certain resource can be replenished or unloaded holds, such a triple defines that during the ith stop the picker re-
at so-called intra-route facilities, and thus the capacity of the ve- trieves δ items of SKU ηv stored at storage position v for order
hicle with regard to that resource is increased. Examples include o. Note that if multiple successive triples all refer to an unaltered
stops for good replenishment at certain intermediate depots or for storage position v ∈ P, this means that the picker retrieves SKU ηv
waste disposal at intermediate collection sites. Finally, alternative stored at this position for multiple orders currently being active on
delivery locations are considered in the generalized VRP (Ghiani his/her cart. Stops do not only refer to storage positions but also to
& Improta, 20 0 0): given a partitioning of the customer set, this depots where finished orders are handed over and new orders are
504 F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515

started. At these positions, i.e., if v ∈ D holds, no items can be re- of expiry may need to be considered, so that, for instance, the
trieved, such that we have δ = 0 and o referring to a virtual order first-in-first-out rule has to be applied (Akkerman, Farahani, &
-1. Finally, triple (v, δ, o) at sequence position π 0 refers to the ini- Grunow, 2010). This can be easily considered by preselecting
tial position of the picker, so that v is fixed to the given starting these preferred (urgent) items. Our picker routing approach is
point of the picker. Given these definitions, we call a solution π then only allowed to retrieve inventory from these preselected
feasible if the following conditions hold: storage positions.
• We presuppose standardized bins for collecting orders, which
• An order o is started (handed over) at the depot preceding (suc-
all have equal size. This allows us to measure the picking cart
ceeding) the first (last) occurrence of o in the triples of π . Note
capacity C in number of bins. Typically, standardized bins are a
that there must always exist one optimal solution for which
prerequisite for a fail-safe transport on picking carts and con-
this property holds. Starting (finishing) orders at a depot ear-
veyors, so that this should rarely be a shortcoming.
lier (later) than required cannot improve the solution value be- • We restrict ourselves to rectangular warehouses, i.e., single-
cause, in this case, cart capacity is blocked longer than neces-
deep parallel racks along both sides of an aisle with cross aisles
sary. Therefore, in a feasible solution π , there has to be at least
at the top and bottom. This most basic setup is very common
one depot visit prior and after the first and last occurrence of
in the real world. For instance, at a German Amazon warehouse
each order, respectively. During the time between these two de-
we visited, the pickers were not allowed to change levels in
pot visits, we call this order active.
the multi-story mezzanine system, and in each level the setup
• In each sequence position i = 1, . . . , m of π where a storage po-
was rectangular. Moreover, the restriction to rectangular ware-
sition is visited, we have at most C active orders, so that the
houses allows us to apply the polynomial time routing algo-
given cart capacity is not exceeded.
rithm of Ratliff and Rosenthal (1983) for quickly solving an im-
• For each order o ∈ O and all SKUs s ∈ S, it must hold that the
portant subproblem to optimality. If more complex warehouse
number of items aos requested by order o is retrieved while it is
settings occur, this algorithm needs to be replaced either by
active. Thus, retrieved items δ summed over all triples (v, δ, o )
one of the extensions (de Koster & van der Poort, 1998; Rood-
in π with o = o and ηv = s must equal aos .
bergen & de Koster, 2001) still solving the subproblem to op-
• The available stock in each storage position can be depleted but
timality or by a heuristic (de Koster et al., 2007). However, we
may not become negative. Thus, for each storage position v ∈ P
leave an evaluation of these adaptions to future research.
retrieved items δ summed over all triples (v , δ, o) in π with • We follow the most common approach in the warehousing lit-
v = v may not exceed the given initial inventory γv .
erature and presuppose that distance minimization is the right
Among all feasible solutions π , we seek one that minimizes the performance criterion for measuring picking performance. As
length of the picker route, i.e., most other parts of the picker’s time, e.g., the time to retrieve
items from a shelf, are fixed once the pick list is determined,

m
F (π ) = dvπi−1 ,vπi , picker walking is the main factor influencing picker time. Note
i=1 that in e-commerce often different priorities and varying dead-
lines have to be considered, which, e.g., vary whether or not
where vπi refers to position v ∈ V being visited during the ith visit
privileged delivery programs have been booked by a customer.
of π . Recall that vπ0 refers to the given start position of the picker.
However, assigning each order a specific weight and deadline
Our problem definition is based on some (simplifying) assump-
implies an unrealistically long planning horizon for picker rout-
tions, which we elaborate in the following:
ing, e.g., of several hours. We, however, assume a shorter plan-
• We only consider a single order picker and neglect all interde- ning horizon where the (urgent) orders to be processed next
pendencies to other pickers concurrently working in the ware- have already been determined on a preceding planning stage.
house. Thus, pickers are assumed to neither block each other in In such a setting, if a given order set is assembled faster by re-
front of shelves, in front of depots, or within narrow aisles (e.g., ducing the travel distances, then it becomes more likely that
Hong, Johnson, & Peters, 2012) nor to influence each other’s the orders’ tight delivery schedules are met. A more detailed
inventory levels. Large warehouses in the B2C segment use a discussion on this matter is provided by de Koster et al. (2007).
large number of pickers that work concurrently, so that it will
regularly occur that multiple pickers compete for items. As soon Given such a rich routing problem, it is anything but astounding
as a picker has retrieved some items, they are no longer avail- that (already a quite restricted version of) our mixed-shelves picker
able for other pickers, who have to sidestep to other shelves. routing (MSPR) turns out to be a complex matter (e.g., Weidinger,
The coordination of pickers is not part of our research, and we 2018).
assume that some preceding planning stage has already dis-
tributed the storage positions and orders among the workers.
Theorem 1. MSPR is strongly NP-hard even if only rectangular ware-
It seems a valid task for future research to integrate our (single
houses with a single depot are considered and only a single order is
picker) approach in a larger framework where it can be applied
picked.
to evaluate different distribution plans of shelves among multi-
ple pickers.
• We neglect the interdependency to replenishment operations The reduction is from the set cover problem, which belongs to
(Section 1.1). If stow workers replenish shelves in parallel to the Karp’s famous collection of 21 NP-complete problems (Karp, 1972):
picking operations, it may occur that additional items become Set cover: Given a set U of n elements and a collection S of sub-
available after a while. It seems risky to integrate items (and sets from U, is it possible to select exactly k subsets from S, such
their exact replenishment times) ahead of their actual place- that every element in U is contained in at least one of the selected
ment in a shelf; any delay would result in an invalid picker subsets.
tour. Therefore, we assume that items are only considered once
they are finally placed, and the picker is about to start the next Proof. In our transformation, we introduce n SKUs, |S| aisles, and
tour. a single picking order. An aisle is added to our warehouse for each
• All items of a SKU are assumed to be equally valid to sat- subset in S, and we store a single item of those SKUs corresponding
isfy a demand. In warehouses storing perishable goods, dates to the elements contained in the respective subset exactly in the
F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515 505

Table 1
Notation used in the MSPR-MIP.

S Set of SKUs (index: s)


O Set of orders (index: o)
P Set of storage positions
Ps Set of storage positions holding SKU s
D Set of depots
V Set of positions, with V = P ∪ D (indices: v, v )
m Maximum sequence position (index: i, i = 0, . . . , m)
C Capacity of picking cart
γv Number of items stored at storage position v ∈ P
dvv Distance between positions v and v
aos Quantity of items of SKU s demanded by order o
v0 Initial start position of the picker (with v0 ∈ V )
M Big integer (e.g., M = maxo∈O,s∈S {aos , |O|})
xvi Binary variables: 1, if v is visited at sequence position i; 0 otherwise
δvio Integer variables: quantity of items picked from storage position v at sequence position i for order o
ksio Binary variables: 1, if picking order o starts at sequence position i; 0 otherwise
keio Binary variables: 1, if picking order o ends at sequence position i; 0 otherwise
φvv i (auxiliary) binary variables: 1, if v is direct predecessor of v at sequence position i; 0 otherwise

middle of the respective aisle. Note that items can be stored in dif- one end point. Additionally, picking a specific order can only
ferent levels of a rack, such that the storage positions in the middle start and end at a depot (8), and SKUs can only be picked at
of an aisle can host multiple SKUs. The single order demands a sin- corresponding storage positions (9). Moreover, picking items of
gle unit of each of the n SKUs. Our single depot is located right un- an order is limited to the time span during which the order is
der the leftmost aisle, and if the length of the cross aisle to enter active (10). The capacity of the picker’s cart is considered by
the aisles has length m, then all our aisles have length l > 2 m. The constraints (11) and the auxiliary variable used to linearize the
question we ask is whether a solution with a tour length smaller objective function is set in formulas (12). Constraints (13) and
than (k + 1 ) · l exists. (14) state that the same position cannot be visited twice in a row
As the movement along the cross aisle is smaller than reaching and storage positions are only visited if demand is satisfied during
the middle of an aisle, we only have to choose aisles (subsets) that the visit. The start of the picker tour is set by (15), and, finally, the
altogether cover our order (set U). Thus, the one-to-one mapping domains of the variables are given by (16) and (17).
between the k subsets of set cover to be selected and the k aisles

m
to be visited in MSPR is readily available, and we abstain from a MSPR-MIP Minimize F = φvv i · dvv (1)
more detailed proof.  v∈V v ∈V i=1
subject to
3. Solving the mixed-shelves picker routing problem with  
multiple depots xvi ≤ xv(i−1) ∀i = 1 , . . . , m (2)
v∈V v∈V
In this section, we elaborate on how to solve MSPR. In

Section 3.1, we formulate a mixed-integer program, which allows xv0 ≤ 1 (3)
to solve MSPR with a standard solver. Then, we present heuristic v∈V
solution procedures extending former work on picker routing with
multiple storage locations per SKU. Daniels et al. (1998) were the 
m
first treating picker routing in a mixed-shelves warehouse. Their δvio ≤ γv ∀v ∈ P (4)
solution approach, however, presupposes a single depot and as- o∈O i=0
sumes that only a single order is picked at a time. We adapt their
nearest neighbor approach to our more complex problem setting in 
m

Section 3.2. Afterward (Section 3.3), we develop a more complex δvio = aos ∀o ∈ O; s ∈ S (5)
v∈Ps i=0
pool-based solution procedure extending the work of Weidinger
(2018), in which picker routing with multiple storage locations
per SKU in a rectangular warehouse is addressed, but, again, only 
m
ksio = 1 ∀o ∈ O (6)
a single depot and a single active picking order is assumed. In
i=0
Section 3.4, a local search heuristic is presented, which can be used
to further improve solutions of our construction heuristics. 
m
keio = 1 ∀o ∈ O (7)
3.1. Mixed-integer program i=0

 
Applying the notation summarized in Table 1 MSPR can be for- (ksio + keio ) ≤ M · xvi ∀i = 0, . . . , m (8)
mulated as a mixed-integer problem (called MSPR-MIP) consisting o∈O v∈D
of objective function (1) and constraints (2) to (17).
Objective function (1) minimizes the total length of the picker 
δvio ≤ M · xvi ∀v ∈ V ; i = 1, . . . , m (9)
tour. Constraints (2) and (3) ensure that position visits in the
o∈O
warehouse are properly sequenced. Inequalities (4) take care that
the available stock of items in a shelf is not exceeded during
 
i
order picking, and formulas (5) make sure that customer or- δvio ≤ M · (ksi o − kei o ) ∀o ∈ O, i = 1, . . . , m (10)
ders are satisfied. Since preemption is not allowed, constraints v∈P i =0
(6) and (7) determine that each order has only one start and
506 F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515


i  out as a pretty bad choice if started from another access point. We
(ksi o − kei o ) ≤ C ∀i = 1, . . . , m (11) tested several simple decision rules for order batching, but none
i =0 o∈O of them turned out significantly better than our most basic index-
based approach.
φvv i + 1 ≥ xv(i−1) + xv i ∀v, v ∈ V, i = 1, . . . , m (12)
3.3. Pool-based construction heuristic
xv(i−1) + xvi ≤ 1 ∀v ∈ V, i = 1, . . . , m (13)
Real-world warehouses of online retailers contain thousands of
SKUs (Section 1), so that we aim at a fast procedure being able to
 handle a vast assortment of SKUs. Consequently, we apply a con-
xvi ≤ δvio ∀v ∈ P, i = 0, . . . , m (14)
struction heuristic, which extends partial solutions in a stepwise
o∈O
manner. To gain promising solution values with small optimality
gaps, we generate many partial solutions in parallel in each step
xv0 0 = 1 (15) and only add the most promising among them to a central pool of
solutions. Furthermore, we apply the efficient algorithm of Ratliff
and Rosenthal (1983) as a subprocedure to determine the optimal
δvio ≥ 0 ∀v ∈ V ; o ∈ O; i = 0, . . . , m (16)
circular picker tour in a rectangular warehouse once the orders to
be processed and the storage positions to be visited in between
two successive visits at a depot are selected. In the following, we
xvi , ksio, keio, φvv i ∈ {0, 1} ∀v, v ∈ V ; o ∈ O; i = 0, . . . , m (17)
sketch the basic idea of our heuristic, which mainly consists of two
Note that our model deviates from our problem definition in hierarchical steps, before we elaborate each step in detail:
Section 2, where picking at a single storage position for multi-
ple orders is modeled by multiple successive (virtual) visits at the • On the upper level, a pool of multiple partial solutions is man-
same position. In the model, picks for several orders during the aged. Starting with an empty tour where the picker is located
same visit are possible due to the δ -variables. Consequently, suc- in his/her initial position, partial solutions are selected from
cessive visits of the same position are prohibited by constraints the pool and extended in a stepwise manner. In each step of
(13) of our model. Even if constraints (13) and (14) are optional the lower level, we start at the picker’s current depot, select
and the same solutions are obtained without them, they reduce a small subset of yet unscheduled orders, schedule them, and
the search space considerably. During preliminary tests employ- end at the succeeding depot where finished orders are handed
ing our small dataset (Section 4.1), these additional constraints in- over. This is done λ · μ times by randomly selecting and extend-
creased the fraction of instances solved to proven optimality from ing a partial solution from the pool, where λ and μ are two
25% to 99% when applying Gurobi with a time limit of 30 minutes. parameters defining the search effort. Among all existing and
Further note that the binary x-variables can easily be eliminated new partial solutions, the λ most promising ones are kept in
and replaced by the φ -variables. However, the resulting model, al- the pool. Once all partial solutions are completed and all or-
though being much more handy, increases the computational time ders are scheduled, the best complete solution of the pool is
of the presented model by a factor of more than seven for our returned.
small dataset. Thus, we abstain from eliminating the x-variables. • The extension of partial solutions is handled on the lower level
of our heuristic. We receive a partial solution of positions al-
3.2. Nearest neighbor heuristic ready visited and orders already processed. This partial solu-
tion is extended up to a specific number of subsequent orders,
Our nearest neighbor (NN) heuristic is a straightforward ap- which is also set by the upper level. To do so, we first select
proach to find a first feasible solution for MSPR. First, picking or- the (small) subset of orders to be scheduled next. Note that
ders are sequenced according to increasing index, and the first C some orders may not be completed, so that the picking cart
orders according to this sequence are activated. As long as not all need not be completely empty when extending a partial so-
active demands are satisfied, the nearest storage position provid- lution. For the active orders, we select the respective storage
ing items of a demanded SKU is visited. When having collected all positions to satisfy the demanded SKUs and determine the best
active demands, the nearest depot is visited and processing of the picker tour along these selected positions by applying the dy-
next batch of C successive orders starts, until all orders are com- namic programming procedure of Ratliff and Rosenthal (1983).
pleted. A more formal definition of NN is given in Algorithm 1 . Note that this procedure derives an optimal circular tour and,
thus, always returns the picker to the initial depot. In our prob-
lem, however, this needs not be the best choice, so that we
Algorithm 1: Nearest neighbor heuristic. do not necessarily take over the complete solution but check
1 for (i = 1; i ≤ |O|; i += C) do whether a premature shortcut to another depot seems advis-
2 determine demand of active orders able. Based on the work of Weidinger (2018), we apply three
i, i + 1, . . . , min{i + C − 1, |O|}; different priority rules for selecting the storage positions to be
3 while (not all demand satisfied) do visited, so that multiple partial solutions are generated among
4 visit nearest storage position containing a not yet which one is returned to the upper level, chosen with a proba-
satisfied demand; bility based on the solution quality.
5 visit nearest depot to activate new picking orders;
6 return the generated tour; Note that by varying the steering parameters λ and μ, our
pool-based heuristic can easily be adjusted to problem instances
of varying size. In warehouses of the largest size, the dynamic pro-
Note that batching, i.e., deciding which orders should be con- gramming approach of Ratliff and Rosenthal (1983) applied as a
currently processed on the cart, is a complex matter especially in subprocedure can be substituted by a faster heuristic, e.g., a near-
a multi-depot warehouse. Picking two orders jointly in a batch may est neighbor approach or some other routing heuristic (de Koster
be advantageous when starting from a specific depot but may turn et al., 2007), to still get acceptable computational times. Further
F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515 507

note that detailed tests on tuning the steering parameters λ and μ


Algorithm 3: Extending a partial solution on the lower level
are reported in Section 4.2.
of the pool-based procedure.
In the following, we describe both levels in more detail and
start with Algorithm 2 of the upper level. In this step, a pool 1 c = randomly selected integer between 1 and
min{|O˜ |, C − |Ǒ|};
2 Oˆ = select c orders of O˜ ;
Algorithm 2: Pool management on the upper level of the 3 determine (remaining) demand of Oˆ and Ǒ;
pool-based procedure. 4 maxLength = 0; List l = new List(); // the new subtours are
1 determine UBF by solving the nearest neighbor heuristic; stored in l
2 define pool, pool  ; // used to save (new) partial 5 foreach (depot vd ) do
solutions 6 foreach (penalty rule ) do
3 pool.Add(new PartialSolution(null)); // picker at start 7 r = randomly drawn double value between 0.0 and 1.0;
position 8 if (vd is picker position ∨ r < |D2 | ) then
4 for (i = C; i ≤ |O|; i += max{1, min{C, |O| − i}}) do 9 π̆ = GenerateSubtour(demand, vd , );
5 pool  .Add(NearestNeighbor(i)); // partial nearest 10 maxLength = max{maxLength, π̆ .Length};
neighbor solution 11 l.Add(π̆ );
6 for ( j = 1; j < λ · μ; j++) do 12 foreach (π̆ in l) do
7 select partial solution π˜ from pool randomly; 13 compute probability for selecting subtour π̆ by
while (π˜ .OrdersP rocessed < i) do
 maxLength−π̆ .Length +1
8
P (π̆ ) =  ;
9 π˜ += new PartialSolution(π˜ ); // generate new π̆  ∈l (maxLength−π̆ .Length+1 )
partial solution 14 select subtour π̆ based on probabilities;

10 if (π˜ .T ourLength < UBF ) then 15 determine lowest position v∗ of π̆ ∗ at which an order is
11 pool  .Add(π˜ ); completely picked; // consider both possible directions
12 foreach (PartialSolution π˜ in pool ∪ pool  ) do of the tour and select based on distances
π˜ .TourLength r = randomly drawn double value between 0.0 and 1.0;
(tie-breaker: ππ˜˜ .ItemsPicked
.TourLength 16
13 rate π˜ by π˜ .OrdersC );
omplet ed 17 if (r < 0.5) then
14 store exclusively the λ partial solutions with lowest rates 18 stop subtour at storage position where order is completed
from pool ∪ pool  in pool; (v∗ unchanged);
15 return best solution in pool; 19 else
20 proceed subtour until the picker comes along a depot
(update v∗ );
of partial solutions is managed. By randomly selecting partial so- 21 select depot to be visited after subtour
lutions from the pool and handing them over to the lower level, maxv ∈D {dist[v∗ ,v
d
]}−dist[v∗ ,vd ]+1
λ · μ new partial solutions are generated and among all existing ( P ( vd ) =  d
∗  ∗  );
v ∈D maxv ∈D {dist[v ,vd ]}−dist[v ,vd ]+1
and new partial solutions, the λ most promising ones are kept in d d
22 update π̆ ∗ , P˜, O˜ , and Ǒ;
the pool. Pool management is defined in detail by Algorithm 2.
23 return π̆ ∗ ;
At first, an upper bound UBF is determined by executing the
nearest neighbor heuristic (Section 3.2; line 1). Afterward, the pool
is initialized with an empty partial solution, where the picker has
not picked any orders yet and is located at the starting position of
the tour. Additionally, pool , storing newly generated partial solu-
tions, is defined (lines 2–3). At first, we determine how many orders (c + |Ǒ| ) are to be
In each of the following iterations, λ · μ new partial solutions picked in parallel during the extension of the partial solution. We
are generated, which differ from the partial solutions generated choose c with equal probabilities from integers 1 to min{|O˜ |, C −
in the previous iteration by C less orders not yet processed. One |Ǒ|}. In doing so, the number of partly satisfied orders has to
of these partial solutions is generated by partially executing the be considered because they reduce the remaining cart capacity.
nearest neighbor approach. The remaining (λ · μ ) − 1 solutions are Based on this first decision, a set Oˆ of orders to be picked with
constructed by extending solutions from the pool (lines 5–9). A |Oˆ | = c and Oˆ ⊆ O˜ is generated by randomly selecting orders of O˜
new partial solution obtained from the lower level is added to the with equal probabilities. Having these two sets, the (remaining) de-
temporary pool pool , if the length of the partial tour is lower than mand of orders Ǒ ∪ Oˆ can be determined (lines 1–3). This demand
upper bound UBF (lines 10–11). is addressed partly or fully during the next subtour. In this con-
In the second step of each iteration, all partial solutions text a subtour is defined as a sequence of stopovers starting and
(pool ∪ pool ) are ranked by the length of the partial solution di- ending at a depot and visiting pick positions in between solely.
vided by the number of orders already fulfilled. As orders can be Making use of the notation employed in Section 2, a partial so-
processed partially, the distance divided by the number of items lution π̆ = ( (v0 , δ0 , o0 ), (v1 , δ1 , o1 ), . . . , (vm , δm , om ) ) is a subtour if
picked acts as a tie-breaker. The λ partial solutions with the low- v0 , vm ∈ D and vi ∈ P ∀i = 1, . . . , (m − 1 ) holds.
est ratios are kept in the pool and provide the basis for the next To determine a set of subtours, then, storage positions to be
iterations (lines 12–14). In the last iteration, partial solutions are visited are selected. Afterwards, the shortest circular tour between
completed. Among all complete and feasible solutions the one with these storage positions is determined by using the efficient proce-
the lowest objective value is returned (line 15). dure of Ratliff and Rosenthal (1983). Note that this procedure de-
Next, we detail how partial solutions are extended on the lower rives the optimal circular tour and, thus, always returns the picker
level by Algorithm 3 . The algorithm receives the set O˜ of orders to the initial depot. In our problem, however, this may not the best
that are not picked yet and the set P˜ of storage positions, whose choice, so that we do not necessarily take over the complete so-
initial inventory is reduced according to the already picked items. lution but check whether a premature shortcut to a depot seems
Additionally, set Ǒ contains orders, which are still active but not advisable (see below). We apply three different penalty values
yet completed. Their demands are partly satisfied. to select the storage positions, which base on the approximated
508 F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515

distances to be additionally covered if the respective position


Algorithm 4: Heuristic improvement procedure.
is chosen:
v 1 for (i = 0; i < κ ; i++) do
SinglePosition : SPd (v ) = dvvd (18)
2 reduce inventory γ according to tour π ;
3 π = π;
MinMax : max (v ) = max{dvv }, starting with V˜ = {vd } (19) 4 foreach subtour π̆ in π  in increasing order do
v ∈V˜
5 increase inventory γ according to subtour π̆ ;
MinMin : min (v ) = min{dvv }, starting with V˜ = {vd } (20) 6 initialize new subtour π̆  ;
v ∈V˜ 7 foreach SKU s demanded in π̆ in randomized sequence
For all three rules, SKUs are processed in non-descending or- do
der by the quantity of storage positions available per SKU. Among 8 while demand of s not satisfied do
all storage positions not depleted yet and belonging to the cur- 9 determine additional distance for each storage
rently considered SKU, the one with the lowest penalty value is position v ∈ Ps at each potential position within
selected until the demand of the considered SKU is satisfied. To π̆  ;
avoid visiting more positions than required, the selected positions 10 insert best storage position at best position into
are processed in non-ascending order of their penalty values, in a π̆  and reduce inventory and demand;
second step, and deselected if the demand can still be satisfied af- 11 while capacity C not fully utilized during subtour π̆  do
terwards. The three approaches differ only by the computation of 12 foreach order o in randomized sequence do
the penalty values. The SinglePosition rule uses the distance start- 13 if any demand of o satisfiable without additional
ing from a given depot vd to the considered storage position v effort then
to approximate the additional distance to be covered when visit- 14 plan new picking job substituting last
ing v. The two other rules MinMax and MinMin employ the maxi- possible job of order o;
mum/minimum distance to all storage positions selected until now. 15 replace π̆ by π̆  in π  ;
The set of selected storage positions is initialized containing the 16 if (π  .Length < π .Length) then
predefined depot vd in both rules. 17 π = π ;
All three rules depend on a given start depot. In a first step, 18 return improved solution;
we create subtours based on the depot the picker is currently
located at by applying all three rules. All resulting subtours are
added to set l of potential subtours. However, it can be advan-
tageous to change the depot of the picker first before continuing First, the inventory of each storage position is reduced by the
the picking process. To enable our procedure to do so, each of the total number of items picked from that very position during the
remaining depots is used as the start point of an additional sub- initial tour (line 2). In this way, improving a subtour does not
tour. Each of these additional subtours is added to l with a prob- result in a shortage of items during a downstream picking stop.
ability of min{1, |D2 | } for each rule. Since we have three rules, the The new solution π  is then generated on the basis of the cur-
mean value of subtours generated is 3 · (1 + (|D| − 1 ) · |D2 | ) ≈ 9 (for rent solution π (line 3). During the improvement phase, sub-
large |D|). Among these subtours, we choose one with a probabil- tours are unfixed and improved in a subsequent manner. Note that
ity based on the tour length needed to satisfy the total demand. the definition of a subtour in Section 3.3 still holds in this con-
The longer a subtour, the lower the chances to pick it (line 13). text. The improvement procedure considers subtours in increas-
Having selected a subtour, we check in both directions at which ing order (line 4). The inventory of the warehouse is increased
point of the tour at least one order is completely picked and according to the picking jobs during the unfixed subtour π̆ (line
choose the direction with a shorter distance to this point. In a sec- 5), and a greedy insertion logic is applied to replan the unfixed
ond step, we cut the subtour at that very position with a probabil- subtour in a successive manner. Starting with a subtour π̆  con-
ity of 50%. In the other half of the cases, we cut the tour when the sisting only of the fixed start point vstart and end point vend
picker comes along a depot anyway. Now that we have the pick (π̆  = {(vstart , 0, −1 ), (vend , 0, −1 )}), all currently demanded SKUs
positions of the subtour, we choose the depot to be visited after- are processed in a randomly determined sequence during the re-
ward with probabilities based on the distance to be covered from pair phase (lines 6–7). For each SKU, all combinations of storage
the last storage position visited to reach the respective depot in the positions and sequence positions within the newly generated sub-
same way used to select the subtour (line 21). Finally the parame- tour π̆  are compared, and the combination leading to the short-
ters P˜, O˜ , and Ǒ are updated and the generated subtour is returned est additional distance is added to the subtour as long as the de-
(lines 22–23). mand of the current SKU is not satisfied, before the next SKU is
processed (lines 8–10).
3.4. Improvement heuristic Afterwards, it is checked if full picker capacity is utilized during
the currently unfixed subtour (line 11). If not, the orders starting
Once a feasible solution is generated, whether by the nearest right after the currently unfixed subtour are evaluated in a ran-
neighbor heuristic or our pool-based approach, an iterative search domized sequence, and new picking jobs are added to subtour π̆ 
procedure can be employed to further improve the solution. The if the corresponding position could be visited without additional
procedure repeatedly unfixes parts of the tour and tries to connect effort (lines 12–14). Newly planned picking jobs of preponed or-
both loose ends of the tour by a shorter subtour still satisfying the ders substitute the latest picking job of the same SKU of that very
predefined demand. In doing so, the subtour is generated making order. In this way, there is a high probability that the order can
use of a greedy insertion algorithm based on a randomly generated both be started and completed earlier. After the preponement step,
sequence of demanded SKUs. Additionally, the tour is tried to be subtour π̆ is substituted by π̆  in the new tour π  (line 15). Note
shortened by preponing order processing intervals. Note that pre- that even if the newly generated subtour is longer than π̆ , a better
vious and subsequent pick positions have to be considered when solution can be obtained (e.g., if an order was preponed). For this
generating a new subtour to obtain a feasible solution. In the fol- reason, π̆  is accepted for sure, whereas an evaluation is executed
lowing, we will detail the single steps of the iterative improvement once all subtours are processed. The new tour π  , obtained by the
heuristic (Algorithm 4). single improvement runs, is accepted as the current solution, if it
F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515 509

Table 2
Parameter values for instance generation.

Symbol Description Small Large

w Number of aisles 2 40
h Length of aisles [in #storage positions] 5 150
|S| Number of SKUs {3, 6, 9} {20 0 0, 350 0, 50 0 0}
|D | Number of depots {2, 4} {26, 80}
|O| Number of orders 3 10
omax Number of items per order {1, 2} {1, 2}
γ max Max. number of items stored per storage position {1, 2} {1, 4}
C Capacity of picking cart 2 4

is the new best known solution (lines 16–17). This procedure is re- per order ranges between 1 and 2 items (Section 1), so that pa-
peated for κ iterations, such that κ is a parameter determining the rameter omax is realistic for all mixed-shelves warehouses where
search effort of the greedy improvement heuristic (line 1). Finally, no batching is applied.
the best solution found is returned (line 18). We repeat this procedure 20 times for each parameter set-
ting, so that we receive 480 small instances and another 480 large
4. Performance of solution procedures instances.

In this section, the performance of our solution procedures is


tested. As no established testbed is available for our routing prob- 4.2. Parameter tuning for the heuristic procedures
lem, instance generation is detailed in Section 4.1. Afterward, pa-
rameter tuning (Section 4.2) and performance results of our solu- This section explores the impact of the steering parameters of
tion procedures are presented (Section 4.3). both, the pool-based procedure (Section 3.3) as well as the im-
All computations are executed on a 64-bit PC with an Intel Core provement heuristic (Section 3.4). Recall that parameters λ and
i7-6700K CPU (4 x 4.0 gigahertz), 64 gigabytes main memory, and μ of the pool-based procedure define the number of partial so-
Windows 7 Enterprise. The procedures are implemented using C# lutions held in the pool and the relative number of new solutions
(Visual Studio 2017), and the off-the-shelf solver Gurobi (version generated, respectively. Specifically, we combine parameter values
7.5.1) is applied for solving the MIP models. λ ∈ {10, 15, 20, 30} and μ ∈ {1, 3, 5, 10} in a full-factorial manner
and solve all small and large data instances to evaluate the impact
4.1. Instance generation of these parameters. The results are summarized in Fig. 2, where
we relate objective value (‘F ’) and solution time (‘sec’) to the vary-
We generate two different sizes of test instances. The ones in ing values of the steering parameters.
the small dataset are still solvable to optimality by a standard The results reveal that an increasing λ value leads to a more
solver, while the ones labeled large represent instances of real- extensive search and, therefore, lower objective values and a
world size. The parameters handed over to our data generator are higher computation time. This holds true for both datasets. The
presented in Table 2. The procedure of instance generation is sum- analogous effect is recorded for μ in most cases. However, an
marized in the following. interesting effect can be observed for the small dataset. Choosing
Warehouse layout: The warehouse is equipped with parallel a μ value of 10, up to 300 new partial solutions are generated in
racks arranged along w aisles, which lead into the cross aisles at each iteration. As the size of the instances is restricted, however,
the front and rear of the warehouse. Storage positions within a many of these partial solutions are similar. In consequence, the
rack are quadratic and the depth of a rack is assumed to equal the pool becomes degenerated and further search is mainly based on
width of an aisle (Fig. 4). Therefore, we can measure distances via the same partial solutions. This leads to a nearly constant solution
the resulting grid of squares. Two depots are located at both cross quality irrespective of the λ value. The same effect, although at a
aisles at the end of each (normal) aisle. Whenever |D| < 2 · w, not much larger μ value, is assumed for the large dataset.
any aisle can house depots at both ends. They are skipped such The search effort of the improvement heuristic introduced in
that depots are located in an equidistant manner along the cross Section 3.4 is specified by steering parameter κ . To get insight into
aisles. the impact of this parameter, we vary parameter values κ ∈ {0, 1,
Storage assignment: As storage turnover often follows the Pareto 2, 3, 5, 10, 15, 20, 30, 40, 50, 100, 150}, where the start solution
principle (e.g., de Koster et al., 2007), we subdivide SKUs into 20% is determined making use of the nearest neighbor procedure (NN)
A-, 30% B-, and 50% C-products. First, between 1 and γ max items of and the pool-based procedure (using steering parameters μ = 5,
each SKU are assigned randomly to one storage position per SKU λ = 20 and μ = 10, λ = 30), respectively. Results are presented in
to assure that each product can be found in the warehouse at least Fig. 3.
once. The remaining storage positions are assigned with a chance The local search procedure is able to improve average solution
of 80% to an A-product, with a chance of 15% to a B-product, and values of both procedures and all parameter combinations inves-
with a chance of 5% to a C-product. Once the class is selected, the tigated. Naturally, a higher number of iterations leads to a bet-
SKU to be assigned is drawn from the set of SKUs belonging to that ter solution quality. The potential of the improvement procedure,
class with equal probabilities. Finally, between 1 and γ max items of however, is depending on the solution quality of the start solu-
the chosen SKU are assigned to the respective storage position. tion. The solutions of the pool-based procedure can only be im-
Generation of pick lists: Pick lists are generated by randomly se- proved marginally for the small dataset, while the nearest neigh-
lecting |O| · omax items stored in the warehouse. In this way, we bor heuristic provides plenty potential for further improvement.
can ensure feasible problem instances because the quantity of de- The time consumption of the improvement procedure is hardly
manded items cannot exceed the number of those available. Note measurable. Only for the large dataset a slight linear increase ex-
that the cart capacity C of order bins within the mixed-shelves ists. Even when executing κ = 150 iterations, the total computa-
warehouses we visited varied between 2 and 4 bins, so that these tion time rises only by about 0.15 seconds, irrespective of the start
values are realistic. Further note that the average number of items solution.
510 F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515

Fig. 2. Influence of parameters μ and λ on solution quality (a, b) and time consumption (c, d) for small (a, c) and large (b, d) instances.

4.3. Computational performance in 96.6% of all instances for which the optimal objective value is
available. The average gap to the optimal solutions is merely about
This section is dedicated to the performance of our solution 0.55%, and in spite of this very convincing performance, the aver-
methods. Specifically, we compare the performance of standard age solution time ranges around one-hundredth of a second only.
solver Gurobi solving MSPR-MIP, nearest neighbor heuristic (NN), Compared to the pool-based procedure without improvement, the
pool-based procedure (pool), and both of the latter two heuris- time consumption does not vary in a measurable way, the solution
tics executed with the additional improvement procedure via local quality, however, increases slightly due to the improvement heuris-
search (NN+ , pool+ ). Note that Gurobi solving MSPR-MIP is given tic applying local search. The nearest neighbor approach is even
a maximum solution time of 30 minutes. First, we report the so- faster; its solution time is barely measurable. The solution quality,
lution performance for the small dataset. The results are summa- though, is not that good. The average gap to the optimum amounts
rized in Table 3. We report the number of optimal solution found to 26.14%. Even when the solution is further improved making use
(‘#opt’), the average relative gap to the optimal solutions (‘gap’), of the local search procedure, the average optimality gap still adds
and the solution time (‘sec’) in CPU-seconds for each procedure. up to 22.0%.
Our small instances have only 20 storage positions and between The results for the large dataset are reported in Table 4. Un-
3 and 6 items to be picked in total. Nonetheless, standard solver fortunately, standard solver Gurobi is not able to find feasible
Gurobi is not able to find all optimal solutions, i.e., 477 of 480 solutions for any of the instances. In spite of 64 gigabytes of
instances are solved to proven optimality. The main impact factor main memory, it always returned a buffer overflow. Therefore, we
on Gurobi’s performance is the length of the pick list. If each or- can only compare the results of our four heuristics. Again, the
der demands just omax = 1 item, Gurobi finds optimal solutions for pool-based(+ ) approach clearly outperforms the NN(+ ) heuristic.
all instances in about 1.92 seconds on average. However, if the or- The pool+ procedure is able to find the best known solution for
der size increases and omax = 2 items are demanded, the average 391 instances (81.5%). However, in 89 cases, NN+ finds a solution
computation time increases to 193.62 seconds and 98.8% of the in- of higher quality than pool+ . The solutions of NN+ , though, are
stances are solved to proven optimality in the given time frame. only slightly better, resulting in a mean gap to the best solution of
Our improved pool-based heuristic pool+ (executed with parame- about 1% for the pool+ approach. The gap of NN+ amounts to more
ter values λ = 30, μ = 3, and κ = 150) finds the proven optimum than 11%. Solving the large dataset, the additional improvement
F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515 511

Fig. 3. Influence of parameter κ on solution quality (a, b) and time consumption (c, d) for small (a, c) and large (b, d) instances.

Table 3
Solution performance for the small dataset.

omax |S| γ max |D | Gurobi pool+ pool NN+ NN


(λ = 30, μ = 3, κ = 5) (λ = 30, μ = 3) (κ = 15)
#opt/sec #opt/gap/sec #opt/gap/sec #opt/gap/sec #opt/gap/sec

1 3 1 1 20/1.32 20/0.00%/0.01 20/0.00%/0.01 14/9.05%/ < 0.001 14/9.05%/ < 0.001


2 20/0.99 20/0.00%/0.01 20/0.00%/0.01 14/8.75%/ < 0.001 13/9.75%/ < 0.001
2 1 20/1.84 20/0.00%/0.01 20/0.00%/0.01 16/7.62%/ < 0.001 15/8.53%/ < 0.001
2 20/1.15 20/0.00%/0.01 20/0.00%/0.01 14/8.75%/ < 0.001 13/9.46%/ < 0.001
6 1 1 20/2.66 20/0.00%/0.01 20/0.00%/0.01 8/18.93%/ < 0.001 6/23.74%/ < 0.001
2 20/1.57 20/0.00%/0.01 20/0.00%/0.01 13/8.51%/ < 0.001 11/11.49%/ < 0.001
2 1 20/2.89 19/1.25%/0.01 19/1.25%/0.01 12/11.22%/ < 0.001 10/13.04%/ < 0.001
2 20/1.49 20/0.00%/0.01 20/0.00%/0.01 14/14.40%/ < 0.001 13/16.55%/ < 0.001
9 1 1 20/2.85 19/0.91%/0.01 19/0.91%/0.01 7/19.15%/ < 0.001 2/35.73%/ < 0.001
2 20/1.55 18/1.96%/0.01 18/1.96%/0.01 8/20.63%/ < 0.001 6/23.31%/ < 0.001
2 1 20/3.15 19/1.25%/0.01 19/1.25%/0.01 9/14.56%/ < 0.001 9/16.14%/ < 0.001
2 20/1.54 20/0.00%/0.01 20/0.00%/0.01 10/19.10%/ < 0.001 8/24.93%/ < 0.001
2 3 1 1 20/62.89 20/0.00%/0.01 20/0.00%/0.01 0/34.46%/ < 0.001 0/38.10%/ < 0.001
2 20/45.21 20/0.00%/0.01 20/0.00%/0.01 3/33.86%/ < 0.001 2/36.63%/ < 0.001
2 1 20/56.61 20/0.00%/0.01 20/0.00%/0.01 5/23.69%/ < 0.001 5/23.69%/ < 0.001
2 20/46.34 20/0.00%/0.01 20/0.00%/0.01 3/27.74%/ < 0.001 2/29.17%/ < 0.001
6 1 1 20/349.39 17/2.83%/0.01 17/2.83%/0.01 1/34.23%/ < 0.001 1/39.27%/ < 0.001
2 20/170.34 18/1.34%/0.01 18/1.34%/0.01 2/31.84%/ < 0.001 2/34.70%/ < 0.001
2 1 20/267.18 19/0.91%/0.01 18/1.68%/0.01 3/31.13%/ < 0.001 3/39.52%/ < 0.001
2 20/195.19 20/0.00%/0.01 20/0.00%/0.01 6/23.20%/ < 0.001 4/27.64%/ < 0.001
9 1 1 19/274.14 16/1.68%/0.01 16/1.68%/0.01 4/34.88%/ < 0.001 4/42.76%/ < 0.001
2 20/193.66 19/0.56%/0.01 18/1.18%/0.01 4/28.87%/ < 0.001 3/35.21%/ < 0.001
2 1 18/352.64 18/0.00%/0.01 17/1.11%/0.01 1/32.50%/ < 0.001 1/42.60%/ < 0.001
2 20/309.85 19/0.50%/0.01 19/0.50%/0.01 4/32.68%/ < 0.001 2/38.75%/ < 0.001
aggregated: 477/97.77 461/0.55%/0.01 458/0.65%/0.01 175/22.00%/ < 0.001 149/26.14%/ < 0.001
512 F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515

Fig. 4. Warehouse setup for (a) dedicated, (b) mixed-shelves, and (c) split storage.

Table 4
Solution performance for the large dataset.

omax |S| γ max |D | pool+ pool NN+ NN


(λ = 30, μ = 10, κ = 150) (λ = 30, μ = 10) (κ = 150)
#best/gap to best/sec #best/gap to best/sec #best/gap to best/sec #best/gap to best/sec

1 20 0 0 1 1 18/0.77%/19.60 3/4.15%/19.53 8/9.42%/0.08 2/18.98%/ < 0.001


3 19/0.53%/14.78 4/5.52%/14.71 1/21.68%/0.08 0/34.61%/ < 0.001
4 1 13/2.28%/19.39 2/6.67%/19.32 9/9.37%/0.08 0/22.27%/ < 0.001
3 19/0.17%/14.57 1/5.02%/14.50 2/19.57%/0.07 0/30.03%/ < 0.001
3500 1 1 18/0.38%/21.68 5/3.22%/21.54 7/8.27%/0.11 1/14.59%/ < 0.001
3 18/0.10%/16.62 5/3.18%/16.49 2/21.77%/0.11 0/30.62%/ < 0.001
4 1 20/0.00%/21.91 1/4.32%/21.77 3/13.53%/0.11 0/24.11%/ < 0.001
3 19/0.07%/16.12 7/3.60%/15.99 2/22.60%/0.11 0/30.56%/ < 0.001
50 0 0 1 1 18/0.46%/23.18 5/3.26%/22.98 6/10.03%/0.15 0/21.37%/ < 0.001
3 20/0.00%/17.69 5/2.88%/17.51 0/20.12%/0.15 0/25.37%/ < 0.001
4 1 18/0.25%/24.31 5/2.22%/24.12 3/14.45%/0.15 1/18.77%/ < 0.001
3 20/0.00%/17.56 5/3.05%/17.38 0/26.74%/0.15 0/33.87%/ < 0.001
2 20 0 0 1 1 11/4.61%/20.74 2/10.00%/20.64 13/3.14%/0.07 0/22.86%/ < 0.001
3 17/1.03%/16.09 0/8.71%/16.01 3/10.34%/0.07 0/32.24%/ < 0.001
4 1 15/1.91%/20.60 0/9.94%/20.50 7/7.33%/0.08 0/35.83%/ < 0.001
3 14/1.58%/16.07 0/11.32%/15.99 6/9.95%/0.08 0/38.14%/ < 0.001
3500 1 1 11/2.86%/23.70 1/7.18%/23.51 10/4.35%/0.12 0/20.64%/ < 0.001
3 15/1.31%/17.58 1/7.58%/17.41 6/5.81%/0.12 0/24.99%/ < 0.001
4 1 11/1.86%/23.59 0/5.71%/23.39 10/5.90%/0.12 0/24.71%/ < 0.001
3 17/0.53%/17.54 1/5.80%/17.37 3/11.91%/0.12 0/34.64%/ < 0.001
50 0 0 1 1 14/2.23%/25.35 0/5.24%/25.06 8/5.41%/0.15 0/20.79%/ < 0.001
3 16/1.94%/19.28 2/5.19%/19.07 4/5.84%/0.15 0/24.32%/ < 0.001
4 1 13/1.99%/26.17 1/6.30%/25.91 11/4.23%/0.15 0/16.36%/ < 0.001
3 17/0.53%/18.63 3/4.05%/18.40 4/10.36%/0.15 0/26.58%/ < 0.001
aggregated: 391/1.14%/19.70 59/5.59%/19.54 128/11.76%/0.11 4/26.14%/ < 0.001
F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515 513

step shows a higher effect, so that both (unimproved) construction feasible instances. Within our test instances, we vary the appor-
heuristics show significant gaps compared to their improved tionment between online and batch orders. When picking a total of
version. While the computational time of NN(+ ) is still well below 120 items, up to 6 batch orders are possible. We vary the fraction
one fifth of a second, the CPU time of our pool-based(+ ) approach of items belonging to online orders as follows: 0%, 17%, 33%, 50%,
amounts to approximately 20 seconds. This still seems acceptable 67%, 83%, and 100%. Depot settings as shown in Table 2 are used
in practical scenarios, so that we conclude that our pool-based and we generate 240 instances per parameter setting, so that in to-
heuristic with improvement (pool+ ) is well suited for solving tal 3360 instances, i.e., 480 instances per data point (item fraction
MSPR of real-world size. of online orders), are generated. All resulting instances are solved
with our improved pool-based heuristic (executed with parameter
values λ = 30, μ = 10, and κ = 150).
5. Managerial aspects
Irrespective of the storage assignment strategy, the setting of
the warehouse is identical to the one described in Section 4.1. De-
Compared to traditional storage policies, where unit loads are
pots are spread all over the warehouse, and two depots can always
kept together and are jointly stored in specific shelves, mixed-
be found on both ends of the same aisle. However, we additionally
shelves warehouses have two main drawbacks, which give rise to
have to process batch orders now. As those orders are often packed
the three research questions to be answered in this section:
on separate stations in business practice, we introduce a special
• To spread items of the same SKU all over the shelves, the unit batch order depot on the upper left corner of the warehouse, where
loads that the SKUs arrive on need to be broken down and lo- each batch order picker tour starts and ends. Due to the higher vol-
gistics workers have to bring the individual items to diverse ume of these orders, picker capacity amounts to C = 1 for this kind
storage positions. Thus, to harvest the advantages of mixed of order. Picker capacity for online orders still amounts to C = 4.
shelves during order picking, additional effort when replen- Having this basic setting, the orders are to be satisfied by three al-
ishing the racks arises (Weidinger & Boysen, 2018). Therefore, ternative warehouse settings, which are schematically depicted (in
we are interested in a performance comparison of two differ- reduced size) by Fig. 4 and explained in the following:
ent warehouse settings: (a) a mixed-shelves warehouse, and
(a) Dedicated storage: All items of the same SKU are stored at clus-
(b) a traditional warehouse applying dedicated storage, where
tered positions within the warehouse. A-class SKUs are stored
all items of a SKU are assigned to a joint and unique stor-
in the aisles closest to the depot; C-class SKUs farthest away.
age position. We generate picking orders and pick the same
(b) Mixed-shelves storage: The items of the SKUs are randomly
set of orders from both settings. By comparing both results,
spread all over the warehouse.
we can quantify the gains of scattered storage in terms of re-
(c) Split storage: The warehouse is subdivided into two parts. On
duced picker travel. Practitioners can then weigh up these gains
the left side, there is the batch area with the depot in the
against the additional effort to be spent during the put-away
upper left corner and dedicated storage applied. In the remain-
process of mixed-shelves storage.
ing part of the warehouse, mixed-shelves storage is applied.
• Mixed-shelves storage is especially suited for small-sized or-
The size of the scattered section as well as the relative quan-
ders demanding just a few items per order line. As soon as an
tity of items per SKU stored there corresponds to the relative
occasional larger order with multiple items occurs, the stock
amount of items picked for online orders. Note that anticipating
in a single shelf may not be sufficient, and the picker has
the fraction of online orders is an aggregated forecast, which is
to visit multiple storage positions in order to gather enough
assumed to be possible with the required precision.
items. Thus, large-sized orders increase the picking effort un-
der mixed-shelves storage. Therefore, it is interesting to know The results of these tests are depicted in Fig. 5. Specifically, we
at what apportionment of large- and small-sized orders the report the total picking distance (left) and the relative increase of
turning point between dedicated and mixed-shelves storage is picking distance over split storage (right) for all three warehouse
reached. settings and relate these results to the fraction of items demanded
• Today, many retailers aim at an omni-channel sales strategy, by online orders. The following answers to our research questions
so that often (large-sized) orders of physical brick-and-mortar can be derived from these graphs:
stores and (small-sized) customer orders placed via a webshop
need to be fulfilled from the same warehouse (Agatz, Fleis- • If we have only online orders, i.e., the percentage of items de-
chmann, & Van Nunen, 2008). In this case, depending on the manded by online orders is 100%, mixed-shelves storage leads
apportionment of large- and small-sized orders, either dedi- to much better picking performance than dedicated storage;
cated or mixed-shelves storage will show advantageous. How- the total travel distance of the picker more than halves. This
ever, there is also a third alternative, which is to subdivide considerable advantage seems to outweigh the additional ef-
the warehouse into a dedicated storage area, where items of fort of the put-away process in order to realize mixed-shelves
the same SKU are kept together and batch orders of brick-and- storage. This assessment is supported by the fact that only or-
mortar stores are processed, and a mixed-shelves section ded- der picking is the time-critical process determining the cus-
icated to small-sized online orders. Therefore, another aim of tomers’ waiting times, whereas replenishment has no direct
this section is to benchmark a split-warehouse strategy against impact on customer satisfaction and is, thus, not (that) time-
the other two alternatives. critical (Section 1.1).
• The turning point between mixed-shelves storage and ded-
To answer these research questions, our computational exper- icated storage is already reached once 33% of all items are
iment is designed as follows. We generate large instances with demanded by online orders. At a larger (smaller) fraction
20 0 0 SKUs. We assume that, in total, 120 items, i.e., 1% of the total mixed-shelves (dedicated) storage becomes preferable. Evi-
stock, have to be picked in each instance. The items are demanded dently, dedicated storage considerably suffers when having to
by two different kinds of orders. There are small-sized online or- process small-sized online orders, so that the turning point
ders each demanding two items of randomly determined SKUs is reached astonishingly early. Dedicated storage requires long
and large-sized batch orders requiring 10 items of two SKUs each. walks between the storage positions of online orders, whereas
These SKUs are randomly chosen among all those SKUs having a mixed-shelves enlarge the probability of close-by storage loca-
minimum stock level of 20 items, so that we are sure to generate tions of requested items. On the other hand, our results clearly
514 F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515

Fig. 5. Performance benchmark for the three warehouse settings depending on the fraction of items demanded by online orders.

indicate that applying mixed-shelves storage when less than plied in parallel in separate areas. We benchmark these warehouse
33% of items stem from online orders is not recommendable. If settings by varying the apportionment between small-sized on-
no online orders occur, dedicated storage reduces picker travel line orders and large-sized orders, e.g., posed by brick-and-mortar
by a remarkable 189%. stores, additionally serviced under an omni-channel sales strategy.
• Finally, we benchmark mixed-shelves storage and dedi- Our results show which storage policy is superior under which cir-
cated storage with the compromise strategy denoted as cumstances.
split-warehouse storage. The maximum advantage of split- This paper considers only the single-picker case. Future re-
warehouse storage, i.e., 39.7%, is reached at the turning point search should extend our picker-routing problem to a multiple-
where 33% of all items are ordered via the online channel. We picker environment. In real-world warehouses, multiple pickers are
have to keep in mind, however, that the price for this advantage employed in parallel, who influence each other with regard to the
is organizational overhead for managing and operating two sep- shelves they access and the inventory they retrieve from there. We
arate warehouse areas. Therefore, switching to split-warehouse assume that the coordination of pickers has already been accom-
storage will not pay if the gains in terms of reduced picker plished on a superordinate planning level. How exactly our single-
travel are too small, so that the corridor where split-warehouse picker case can be applied in a larger solution framework that also
storage should be considered as a serious alternative is com- includes picker coordination remains up to future research.
paratively small. For instance, the corridor where picker travel
is reduced by about 25% just ranges somewhere between 25% Appendix A. Impact of cart capacity on picker travel
and 50% of all items ordered online.
This appendix explores the impact of the capacity of the pick-
Note that another managerial aspect is considered in the ap- ing cart on the picking performance. Specifically, we vary param-
pendix, where we explore the impact of an increasing capacity of eter C ∈ {1, 2, 4, 6, 8}, which defines the number of orders that
the picking cart on the picker travel. The results presented there can concurrently be assembled by the picker. We solve each of the
show that the positive effect of additional cart capacity quickly di- large instances (Section 4.1) with any of the aforementioned capac-
minishes. Small and agile carts with a capacity for up to a handful ities by applying our improved pool-based heuristic (with parame-
of orders lead to considerable reductions of picker travel. Larger ter values λ = 30, μ = 10, and κ = 150). The resulting average total
and bulky carts that would considerably increase the ergonomic picking distance for any of these capacities is plotted in Fig. 6.
burden for the pickers only lead to small marginal reductions and,
thus, need not be considered.

6. Conclusion
1,000
This paper defines the picker-routing problem for the mixed-
shelves warehouses applied by many online retailers. In such ware-
800
houses, unit loads are purposefully broken down, and single items
of the same SKU are spread all over the shelves. During picker
routing, this gives the additional flexibility of having alternative 600
shelves that a specific SKU can be retrieved from. Furthermore, our
picker-routing problem considers (a) a cart pushed by the picker
400
that allows to assemble multiple orders concurrently, and (b) mul-
tiple access points to the central conveyor system where com-
pleted orders are handed over. The resulting optimization prob- 200
lem is defined, computational complexity is proven, and a suited
heuristic solution procedure is developed. This procedure is able to
0
handle the thousands of SKUs relevant in real-world warehouses. 1 2 3 4 5 6 7 8
Furthermore, we address managerial aspects in our computational
studies. Specifically, we compare mixed-shelves storage with ded-
icated storage and a split warehouse where both policies are ap- Fig. 6. Impact of cart capacity C on picking performance.
F. Weidinger et al. / European Journal of Operational Research 274 (2019) 501–515 515

The results show that the positive effect of additional Gu, J., Goetschalckx, M., & McGinnis, L. F. (2010). Research on warehouse design
cart capacity quickly diminishes. Increasing the capacity from a and performance evaluation: A comprehensive review. European Journal of Op-
erational Research, 203, 539–549.
single order to two and four concurrent orders leads to consider- Hong, S., Johnson, A. L., & Peters, B. A. (2012). Batch picking in narrow-aisle order
able reductions of the picker travel. From then on, only moderate picking systems with consideration for picker blocking. European Journal of Op-
additional gains can be realized. This is good news for the practi- erational Research, 221, 557–570.
Johnson, E., & Meller, R. D. (2002). Performance analysis of split-case sorting sys-
tioner. Already relatively small and agile picking carts with a ca- tems. Manufacturing & Service Operations Management, 4, 258–274.
pacity for just a few orders lead to a good performance, so that Karp, R. M. (1972). Reducibility among combinatorial problems. In R. E. Miller, &
larger carts that would considerably increase the ergonomic bur- J. W. Thatcher (Eds.), Complexity of computer computations (pp. 85–103). New
York: Plenum Press.
den for the pickers need not be considered.
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-
References search, 182, 481–501.
de Koster, R., & van der Poort, E. S. (1998). Routing orderpickers in a warehouse:
Agatz, N. A., Fleischmann, M., & Van Nunen, J. A. (2008). E-fulfillment and multi- A comparison between optimal and heuristic solutions. IIE Transactions, 30,
-channel distribution - a review. European Journal of Operational Research, 187, 469–480.
339–356. Laudon, K. C., & Traver, C. G. (2007). E-commerce. Pearson/Addison Wesley.
Akkerman, R., Farahani, P., & Grunow, M. (2010). Quality, safety and sustainability in Montoya-Torres, J. R., Franco, J. L., Isaza, S. N., Jiménez, H. F., & Herazo–
food distribution: A review of quantitative operations management approaches Padilla, N. (2015). A literature review on the vehicle routing problem with mul-
and challenges. OR Spectrum, 32, 863–904. tiple depots. Computers & Industrial Engineering, 79, 115–129.
Bartholdi, J. J., III, & Hackman, S. T. (2014). Warehouse & distribution science. release Ratliff, H. D., & Rosenthal, A. S. (1983). Order-picking in a rectangular warehouse:
0.96. Supply Chain and Logistics Institute. A solvable case of the traveling salesman problem. Operations Research, 31,
Brynjolfsson, E., Hu, Y. J., & Smith, M. D. (2003). Consumer surplus in the digital 507–521.
economy: Estimating the value of increased product variety at online book- Roodbergen, K. J., & de Koster, R. (2001). Routing order pickers in a warehouse with
sellers. Management Science, 49, 1580–1596. a middle aisle. European Journal of Operational Research, 133, 32–43.
Daniels, R. L., Rummel, J. L., & Schantz, R. (1998). A model for warehouse order Schneider, M., Schneider, A., & Hof, J. (2015). An adaptive VNS algorithm for vehicle
picking. European Journal of Operational Research, 105, 1–17. routing problems with intermediate stops. OR Spectrum, 37, 353–387.
Gallien, J., & Weber, T. (2010). To wave or not to wave? Order release policies for Weidinger, F. (2018). Picker routing in rectangular mixed shelves warehouses. Com-
warehouses with an automated sorter. Manufacturing & Service Operations Man- puters & Operations Research, 95, 139–150.
agement, 12, 642–662. Weidinger, F., & Boysen, N. (2018). Scattered storage: How to distribute stock keep-
Ghiani, G., & Improta, G. (20 0 0). An efficient transformation of the generalized ve- ing units all around a mixed-shelves warehouse. Transportation Science (In
hicle routing problem. European Journal of Operational Research, 122, 11–17. Press). doi: 10.1287/trsc.2017.0779.
Gu, J., Goetschalckx, M., & McGinnis, L. F. (2007). Research on warehouse operation: Yaman, H., Karasan, O. E., & Kara, B. Y. (2012). Release time scheduling and hub
A comprehensive review. European Journal of Operational Research, 177, 1–21. location for next-day delivery. Operations Research, 60, 906–917.

You might also like