See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/315783295

A simulated annealing approach to optimal storing in a multi-level warehouse

Article · March 2017

CITATIONS READS

0 46

3 authors, including:

Carina Geldhauser
Technische Universität Dresden
7 PUBLICATIONS   10 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Carina Geldhauser on 19 February 2019.

The user has requested enhancement of the downloaded file.

 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A
MULTI-LEVEL WAREHOUSE

ALEXANDER ECKROT, CARINA GELDHAUSER, AND JAN JURCZYK

Abstract. We propose a simulated annealing algorithm specifically tailored to optimise to-

tal retrieval times in a multi-level warehouse under complex pre-batched picking constraints.
Experiments on real data from a picker-to-parts order picking process in the warehouse of
a European manufacturer show that optimal storage assignments do not necessarily display
arXiv:1704.01049v1 [cs.AI] 25 Mar 2017

features presumed in heuristics, such as clustering of positively correlated items or ordering

of items by picking frequency. In an experiment run on more than 4000 batched orders with
1 to 150 items per batch, the storage assignment suggested by the algorithm produces a 21%
reduction in the total retrieval time with respect to a frequency-based storage assignment.

1. Introduction
Warehouses play a key role in modern supply chains [21] and are a significant cost factor to a
company: according to the European Logistics Association/AT Kearney report [20], the capital
and operation costs of warehouses represent about 25% of the surveyed companies’ logistics costs
in 2003, while figures for the USA [33] indicate that warehousing contributed to the total logistics
costs with a share of 22%.
Order picking, generally defined as the process of retrieving products from storage in response
to a specific customer request, is the most labour-intensive operation in warehouses with manual
systems, and a very capital-intensive operation in warehouses with automated systems [25, 46].
Estimates of the percentage of order picking costs on the total warehouse costs range as high
as 55% in Drury [19] and Bartholdi et al[5] to 65% in Coyle et al. [12]. For these reasons,
warehousing professionals consider order picking as the highest priority area for productivity
improvements.
Over the last twenty years, many papers have studied order picking processes and optimal strate-
gies or heuristics to optimise subprocesses such as warehouse layout, storage assignment, order
batching, order release method and picker routing. However, there is little published research
on how to combine these subprocesses optimally: we mention here [13], which compares the
performance of S-shaped and Largest Gap routing heuristics for batches of 3 and 4 items, and
[1], which focuses on heuristics for order batching to improve the overall performance of order
picking systems.
The contribution of this work is to propose a combined optimisation of the storage assignments
and a routing heuristic for correlated batched orders of large size (up to 102 unique parts in our

Institut für Theoretische Physik, Universität Regensburg, Universitätsstrasse 31, 93053 Regens-
burg, Germany
Weierstrass Institute for Applied Analysis and Stochastics, Mohrenstrasse 39, 10117 Berlin, Ger-
many
Institut für Theoretische Physik, Universität Regensburg, Universitätsstrasse 31, 93053 Regens-
burg, Germany
E-mail addresses: alexander.eckrot@physik.uni-regensburg.de, geldhauser@wias-berlin.de,
jan.jurczyk@physik.uni-regensburg.de .
1
2 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE

tests). This is achieved via a hand-tailored algorithm based on a simulated annealing combina-
torial search, which incorporates variable parameters depending on the routing heuristics. The
algorithm is designed for multi-level warehouses, the routing heuristics is based on a warehouse
design with wide aisles, however, it can be adapted to narrow aisles.

2. Warehouse order picking

In human-operated warehouses, the most common system for order picking is the picker-to-
parts class, where the (human) order picker walks or drives along the aisles to pick items [14,
48]. Despite their ubiquity, pickers-to-part order-picking systems have not received comparable
attention from researchers, perhaps because of their variety and complexity [15].
An optimal strategy for a picker-to-parts picking process will minimize the time needed for picking
all orders in a given time frame. The precise quantity to minimize is indeed total retrieval time,
which is defined as the sum of pick and travel time and time due to delays. By using s-shaped
routing or other routing rules which make the aisles unidirectional, the delay part (which is
usually due to congestion) can be a priori removed from the model. The main components to
be minimized are therefore (1) the time needed for travelling the warehouse to collect the items
and (2) the time needed to perform one pick. If we assume that the microdesign of the storage
racks was already optimised and appropriate tools are used by the picker, the latter is basically
a constant, its size depending on the level in the warehouse where the item is located.
Minimizing the time needed for travelling the warehouse can be achieved by various ansatzes, the
most common ones involve solving the storage location assignment problem (SLAP), the routing
problem or allocating optimal batched orders for one tour of the picker.

2.1. Storage assignment policies. Early scientific contributions to the storage location as-
signment problem in warehouses include a taxonomy of possible storage location assignment
policies, where the classification between dedicated storage, randomized storage and class-based
storage was introduced, see [22, 23, 29] and the references therein.
In our problem, we deal with a warehouse where the individual pieces used in the production of
certain products are stored. Therefore, the orders arrive already pre-batched, according to the
necessity of the assembly line. The order batches are quite large (on average 30 items per batch
in the data we used for testing our algorithm). We therefore seek a solution to the combined
SLAP for a warehouse with dedicated storage-policy and the routing problem for such pre-batched
orders.
The dedicated storage policy assigns a fixed location to each product, meaning that this location is
reserved even for products that are out of stock. While this may result in a low space utilisation,
which may imply higher maintenance costs and possibly longer routing times, dedicated storage
locations are an advantage in warehouses for production units, since every stored item is needed
regularly.
Moreover, dedicated storage locations help to increase the orientation of the order pickers, leading
to an increased routing velocity and fewer wrong picks. Dedicated storage policies allow for logical
grouping of storage items, which are often advantageous in retail warehouses [16] and in general
for items of very different weight, which can be stored in decreasing weight along the standard
picking route, implying a good stacking sequence.
However, as pointed out in the survey [15], analytical models for optimising dedicated storage
assignment in manual-pick order-picking systems are still lacking. Existing studies mainly focus
on random storage assignments.
The random storage policy (see for example [40]) reduces the total space required. This does
not automatically improve routing times, as travel distance might increase, see [10]. It needs a
computer-controlled environment to be efficient, as a lack of automation or technical equipment
 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE 3

assisting the picking process can lead to slow travelling times due to disorientation and higher
percentages of picking errors.
It was observed [47, 48] that a separation of the pick stock (forward area) from the bulk stock
(reserve area) can lead to a significant improvement of picking times: in the forward area, a
dedicated storage policy is applied, while the bulk area can follow a random storage policy. In
this way, the advantages of dedicated storage still hold and disadvantages are reduced. Indeed,
this policy is already adapted in many warehouses attached to production units, such as the one
in our reference case.
At the interface between research and industry, several papers, such as [6, 7, 9, 18, 35], describe
algorithms that solve the SLAP with real-life constraints. In particular the paper [9] also deals
with a multi-level warehouse situation, pointing out that the storage assignment systems need
to reflect the structure of the orders.

2.2. Routing policies. Storage assignment has an impact on the performance of the routing
method [41]. The problem of optimal routing for order picking classifies as a Steiner Traveling
Salesman Problem, which is in general not solvable in polynomial time. However, for a special
warehouse aisle configuration, Ratliff and Rosenthal [42] showed that there does exist an algo-
rithm that can solve the problem in running time linear in the number of aisles and the number
of pick locations. This algorithm was extended to other situations in [17, 44, 45].
However, algorithms are not yet available for every specific layout, and there remains the unsolved
problem of aisle congestion by pickers following different routes, and the fact that pickers may
deviate from routes that they deem illogical [24]. Because of this, the problem of routing order
pickers is mainly solved by using a heuristic, such as the s-shape method. In [30], a routing
strategy for a warehouse with U-shaped layout has been introduced and proven to be more
efficient under certain conditions.
Most heuristic methods for routing order pickers in single-block warehouses assume that the
aisles of the warehouse are narrow enough to allow the order picker to retrieve products from
both sides of the aisle without changing position [28, 40, 43]. A polynomial-time algorithm for
routing order pickers in wide aisles was proposed in [26].
One of the strengths of the algorithm we propose here is that it gives efficient heuristics both
for narrow and wide aisles: the routing strategy, the right-hand rule and the aisle length are
incorporated as variable parameters and can easily be changed, see formula (4.3).

2.3. Simulated annealing. The simulated annealing (SA) method is an optimisation algorithm
introduced by Kirkpatrick in 1983 [34] and separately by Černỳ in 1985 [8]. The idea behind
this method is to emulate the congelation of a crystal with many atoms. This process starts at
very high temperatures at which all atoms are free to move. By slowly decreasing the tempera-
ture atoms start to ”feel” the presence of others and arrange themselves into a structure, more
precisely: the interaction between atoms forces them to settle into a crystal shape, where the
overall interaction energy is minimized. The slow cooling process allows the system to explore
many configurations and enabling it to find a near optimal or optimal configuration.
Transferring this method to optimisation problems, one has to define an energy or cost function
f . In our case this cost function is the overall picking time in a warehouse. Changing the positions
of items (also called configuration) leads to an increase or decrease in the cost function. At each
iteration of this changing procedure (called update), the simulated annealing algorithm checks
whether the new configuration is better or worse than the one before and applies the Metropolis
criterion [38], which states the following: An improvement, in this case a change of configuration
which leads to a decrease of the cost function, will be accepted, while a worse configuration will
4 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE

only be adopted with a Boltzmann probability

∆f
(2.1) P(∆f, T ) = e− T

where ∆f is the difference of the values of the cost-function for the two configurations and T a
control parameter similar to a temperature.
When performing an optimisation, at first one chooses a large control parameter T and performs
a finite number of updates to the configuration. Then the parameter T is slightly decreased and
again a finite number of updates are realised. Repeating the last step transforms the update-
method from a random walk at large T to a local search update at small T . The last configuration
obtained is the one with the least energy.
Simulated annealing algorithms are nowadays widely used in economics, for example in packaging
problems [27], the production scheduling problem [36], the corridor allocation problem [2] [3] and
solving the SLAP [4, 39].
A simulated annealing heuristic to minimize total retrieval time involving order batching and
sequencing was introduced in [32]. Their algorithm uses a geometric cooling schedule [11] also
adapted in [37]. The algorithm proposed in [37] uses a single flexible heuristic based on random
moves in a structured manner, in comparison to multiple deterministic neighbourhood search
heuristics, as often found in the literature, and is comparably very fast.
Recent results [31] show that solutions obtained by simulated annealing, Iterated Local Search or
the Attribute-Based Hill Climber [49] may allow order picking systems to operate more efficiently
compared to those obtained with standard constructive heuristics such as the Earliest Due Date
rule.

3. Problem description and modelling

We consider a multi-level warehouse of a production site. Incoming orders consist of large batches
of individual items, which are picked manually into a bin using a picker-to-parts order picking
system and delivered to a couple of output locations. The first two levels of the warehouse are
easily accessible, while all levels above can only be accessed by a lifting device, resulting in a
higher picking time per item.
The items ordered in a batch are correlated to each other, as each batch corresponds to the parts
used in one step of the assembly line of the production site.
The difficulty of this SLAP is due to the high number of possible optional features among which
the customer can choose when ordering one of the (very few) main products: The number of
items which are needed for all options in a main product line is relatively low compared to the
number of items corresponding to one or more options. The correlations between individual
items to be picked are therefore quite complex and cannot be treated by a pure frequency-based
algorithm as proposed in [6].
Another level of complexity is added by different storage container classes used in the warehouse:
As they are of different size, they cannot be randomly exchanged. Moreover, the number of
different items stored per aisle depends on the composition of container classes in this aisle. Note
that standard simulated annealing algorithms are unable to deal with combinatorial constraints
as posed by storage containers of different size.

4. Solution approach
We build a simulated annealing algorithm which finds the configuration of warehouse items in a
multi-level, multi-container class warehouse minimizing the total retrieval time of batched orders
under a given routing in a reference time frame. For this, we need to specify (1) how we calculate
the retrieval time and (2) which update routine we implement in our SA algorithm to gradually
improve retrieval times.
 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE 5

4.1. Construction of retrieval time. As we consider pre-batched orders, meaning several

items collected into the same bin during one route, we have to calculate the retrieval times per
ordered bin. The total retrieval time t in the reference time frame is simply the sum over the
individual retrieval times tbin,i for each bin (labelled by i) ordered in the reference time frame:

X
N bin

(4.1) t= tbin,i
i=1
As mentioned above, our algorithm achieves this by both optimising the storage assignment and
using a heuristic to minimize routing times. The retrieval time per bin tbin is therefore split into
the pick time tp and the routing time tr . These are calculated independently from one another
and the retrieval time per bin then reads
(4.2) tbin = tp + tr .
4.1.1. Routing heuristic. The routing time is the time that a picker needs to physically move to
all locations in the warehouse where the goods for one bin are stored. The goal is to reduce the
overall routing time for all bins.
It is easy to see that for a warehouse with only parallel aisles, the s-shaped routing heuristic
can be applied both for wide aisles, enforcing a right-hand pick rule, and for narrow aisles (no
right-hand rule). However, while unidirectional narrow aisles have to be traveled completely, the
routing heuristic in wide aisles would be to pick with the right hand until the last item in the
batch for the right part of the aisle is reached, and then to turn around and pick from the “left”
part of the aisle, exiting at the same point where the aisle was entered.
We present a routing heuristic optimised for wide aisles. This formula can readily be adapted to
an s-shaped routing in unidirectional narrow aisles.
Dividing every aisle into subsections, the time needed to travel through one wide aisle while
applying a right-hand pick rule is given by the travel time per subsection τs multiplied by the
distance d of the subsection which is farthest away from the entrance of the aisle1. In case of
narrow aisles, the time needed to travel through one aisle is constant, namely just the number
of subsections per aisle multiplied τs . In other words, d is a constant and not a variable.
The routing time tr for one bin is then calculated as
(4.3) tr = Naisle · τaisle + d · τs
where Naisle is the number of aisles that have to be entered to collect all items in this batch,
τaisle the time needed to change from one aisle to the next and τs the time it takes to move one
subsection within one aisle.
One might ask why this heuristic should give a good result. In fact, considering an isolated
routing optimisation, as described in the literature review above, a simple s-shaped heuristic
does not always give the best results, even though it helps to avoid congestion. However, in our
case, it is precisely the mix of a solution to the SLAP with implicit routing heuristics which will
produce an optimal solution on the condition of this wide-aisle s-shaped routing heuristic.
4.1.2. Multi-level picking. The literature often distinguishes between low-level and high-level
picking. In low-level picking systems the picker can directly collect the items from the storage
racks, while high-level picking or “man-aboard order-picking” indicates the use of a lifting order-
pick truck or crane, see [15] for a detailed exposition. We design our algorithm in a way that
the pick times depend on the level where the item is stored. In the simplest case, when only a
1We call d the distance as it can be calculated by any norm which the user considers adapted. The easiest
notion of distance is twice the maximum norm, so that d = 2 max{xj − x}, where x is the entrance point of the
aisle and xj the j-th item to be picked in this aisle.
6 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE

distinction between low-level and high-level picking is made, we therefore work with two picking
times, namely τl for the lower level(s) and τu for “upper level” picking. The pick time per bin
can then be calculated as
(4.4) tp = Nl · τl + Nu · τu + Θ(Nu ) · τlift
where Nl is the number of items located in lower levels and Nu the number of items located
in the upper levels, respectively. The last term in equation (4.4) adds the time τlif t needed to
fetch or adjust the lifting device to the upper level(s). In the simplest situation of only one level
change, Θ(Nu ) function returns one if there are any elements to retrieve from the upper levels
and zero if the picker has no need to visit the upper levels.
Formula (4.4) is only the special case of the general multi-level picking time formula
X
L
(4.5) = t̃p = (L − 1) · τlift Nj · τj
j=1

where L is the number of level changes to be made by the picker and τj the picking time for an
item stored in level j.
By specifying different picking times for different levels or other special situations, the proposed
algorithm can be adapted to combined low- and high-level picking, and by adjusting τlift , τaisle and
τs , the algorithm applies also to differently shaped warehouses and variable routing heuristics.
4.2. Construction of moves. The total retrieval time constructed in (4.1) is the cost-function
which our simulated annealing algorithm has to minimize. As explained above, such an algorithm
changes again and again the location of the storage containers (in which the items are stored),
trying to find configurations with a lower total cost. These continuous changes are called moves.
It is of key importance for the performance of the algorithm to choose the correct type of moves.
The specific design our test warehouse adds an extra difficulty: the presence of different-size
storage containers in the warehouse translate into combinatorial compatibility conditions, i.e. is
has to be ensured that the algorithm does not exchange two storage containers of different size.
The crucial solution step here was to observe that the admissible combinations of container
classes form subsections in the aisle. Therefore, in the presented algorithm, two update routines
were implemented: the first routine exchanges two random boxes from the same size category, the
second routine swaps whole subsections in two randomly chosen levels. The first move is tasked
to cluster all items that are strongly correlated and minimizes the individual item picking time
for one bin. The second routine both ensures that a solution found is admissible and accelerates
the optimisation by searching for a more adequate location of a group of already clustered items,
e.g. a subsection consisting of items used to assemble a highly popular product option should be
placed at a part of the warehouse which is quickly reachable by the picker.
The combination of these two moves allows the simulated annealing algorithm to find warehouse
configurations that are minimized by the average picking time for each bin, while keeping the
overall performance picking time for the whole range of product options in mind.

5. Validation and results

5.1. Case description. We test our algorithm with real data from the production site of a
medium-size European company offering highly customizable products with a long lifespan.
When ordering a product, the customer chooses from a large number of possible options for
the product of his choice, which are assembled at the production site.
The individual parts for the product are prefabricated by subcontractors, shipped to the com-
pany and stored in the warehouse until needed, there is no just-in-time delivery. Orders to the
warehouse arrive as a batch, which is picked in a single journey. The picker stores the items
 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE 7

in the batch on a bin, which includes small trays for small pieces and a dedicated space for
heavy items, so there is no issue of considering heavy or delicate items when deciding the routing
strategy. When all items are collected, the picker delivers them to certain input points along the
assembly line.
The content of one batched order varies not only according to the output delivery point, which
is the input point of a specific step in the assembly line, but also highly depends on the end-
configuration of the product chosen by the customer. In other words, customised product options
lead to complex correlations between the items stored in the warehouse. Positively correlated
items are more likely to be found in one batched order arriving at the warehouse.

5.1.1. Sample warehouse design used in the algorithm. Our algorithm was designed to be general
enough to cope with several features appearing in the warehouse of the abovementioned com-
pany, which are variable aisle length, multiple storage levels, and different container types with
combinatorial restrictions due to their size.
As visualised in figure 1, the sample warehouse of the manufacturer has 7 aisles of variable length
(due to the constraints of the production site). Each aisle is 4 levels high and is divided into
(at most) 20 subsections. Individual items are stored in containers, of which two main container
type, namely large or regular size are used, and of these types, both high and low containers
are available. Each subsection of an aisle is wide enough to hold two large or three regular
containers, therefore up to six different items (six items meaning three regular size containers
of low height stacked on three other regular size containers of low height) can be stored at each
level of a subsection.
A total amount of 1268 different components are stored in the warehouse of the manufacturer
whose data we used. We considered only the primary pick location of these 1268 individual
items, as only this primary pick location is mentioned in the order sheet given to the picker.
The secondary location of this component is in the less frequented “cold” area of the warehouse,
where a random storage policy is exibited (surplus and refill storage).
An order batch contains between one and 150 items each. As mentioned, the bin has designated
storage options for different items. For example, small items go on the trays, so that no predefined
sequence of picking is needed. The delivery locations are designated spots at the edge of the
assembly line. As the retrieval time for one order batch is by orders of magnitude longer than
the travelling time to the delivery points along the assembly line, the influence of those delivery
points on the warehouse storage locations can be neglected.

5.1.2. Routing heuristics in the company. Our algorithm is designed to adapt to several routing
situations. The routing heuristics used in our sample warehouse is as follows: pickers start with
picking from the lower two levels of a wide aisle. They follow a right-hand rule, which means
that they pick only on their right while traveling along the aisle. Once the last item to their
right is reached, they turn around and pick the other side of the aisles until they arrive back
to the entrance of this aisle and change to the next aisle. After completing the routing in the
lower levels, a lifting order-pick truck is fetched and the picker starts picking the first upper level.
After all items of the batch stored in the first upper level are picked, the height of the order-pick
truck is adjusted and the picker continues the same routing on the second upper storage level.
Using this routing heuristics, the company arranged their storage allocation based on the picking
frequency.

5.1.3. Individual pick times. Note that the time required for each pick (denoted by τj in formula
(4.5)) changes depending on the level from which items are picked. In our sample warehouse,
we consider the simple situation of only two different picking times, τl for the lower levels, and
τu for the upper levels, see formula (4.4). The time required for each pick in the upper levels is
8 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE

7.5 7.5
a) b)
6.0 6.0

4.5 4.5

3.0 3.0

1.5 1.5

0.0 0.0

7.5 7.5
c) d)
6.0 6.0

4.5 4.5

3.0 3.0

1.5 1.5

0.0 0.0

Figure 1. Warehouse design divided into the different levels: a) ground level,
b) first level, c) second level, d) third level. Each pixel represents a designated
storage space for one item. The colours indicate the logarithm of the total
picking rate of an item. Red items are much more frequently picked than blue
items. Pathways within the aisle are white. Dark blue indicates empty or non-
exchangeable space due to real world constraints.

significantly longer than for the lower levels, as more careful steering is needed and the picker
is less mobile. The choice of pick times and the time for aisle change/ level change used in the
experiments of our algorithm are listed below. Please recall equations (4.3) and (4.4) for the
definition. It is worth mentioning that the solution quality does not depend on the exact times
as long as they are in reasonable proportions to each other.
Variable Time in seconds
τaisle 30
τs 2 per subsection
τl 15
τu 30
τlif t 120

5.2. Validation of method. The experiments were carried out with 4192 pre-batched orders,
representing the products assembled in the reference time frame. All experiments were run single-
threaded on Intel(R) Core(TM) i7-4790 CPU running at 3.60GHz and with 8 GB of memory.
We used a geometric cooling schedule starting at T = 107 and stopping when convergence is
 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE 9

7.5 7.5
a) b)
6.0 6.0

4.5 4.5

3.0 3.0

1.5 1.5

0.0 0.0

7.5 7.5
c) d)
6.0 6.0

4.5 4.5

3.0 3.0

1.5 1.5

0.0 0.0

Figure 2. Optimised warehouse design divided into the different levels: a)

ground level, b) first level, c) second level, d) third level. The colours indicate
the logarithm of the total picking rate of an item. Red items are much more
frequently picked than blue items. Pathways within the aisle are white. Dark
blue indicates empty or non-exchangeable space due to real world constraints.

reached with a cooling constant α = 0.95, resulting in about 2 · 106 iterations. The total running
time for one annealing was about 5 hours.

5.3. Results. After optimisation of storage locations by our algorithm, the total retrieval time
for all 4 · 104 pre-batched orders in the reference time frame was calculated to be approximately
38% lower than for a random item distribution and 21% lower compared to the total retrieval
time with the pre-otimised storage location configuration used by the company.
This is a significant improvement, in particular as the initial storage allocation used a frequency-
based heuristics similar to [6]: Fixing the routing heuristics, the company’s logistics division had
already re-arranged the items used more often to the first aisles and those used less were put
in the back aisles. However, as our results show, this simple heuristics has its drawback when
handling large batched orders.

5.3.1. Detailed analysis - global distribution. Figure 2 shows the optimised warehouse design.
Red items have the highest picking frequency, blue items the lowest. In comparison to the
starting situation (Figure 1), the lower levels shown in a) and b) are better stocked. They
contain the more frequently picked items (red, orange, yellow/green pixels), while the upper
levels c) and d)) contain more blue/green items. In fact, the algorithm removed all orange/red
10 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE

items from the upper level, which is to be expected since the higher picking time in the top floors
and the need for a lifting device have a significant impact on the optimisation.
The canonical entrance to the warehouse is on the right-bottom corner of the picture. The first
aisle from the right, which is the first aisle visited, initially had a lot of very frequently picked
items on the first two levels, as this was the preferred storage location in the heuristics used by
the company. After optimisation by our algorithm, most of these items were moved, as the aisle
is very short one and therefore inefficient to visit from a global optimisation point of view, as
taken by our algorithm.

5.3.2. Detailed analysis - batch-induced correlations. As already discussed above, the picking
process of large batched orders results in complex correlations between the individual items in
the warehouse. The correlations are not necessarily related to the frequency of picks of a single
item: some items are very basic and are used for every single product that this company is
producing, independently of the product option chosen by the customer, while other items are
specific to one product option and appear in exactly one batch if and only if this product option
was ordered.
One might conjecture that clustering the items according to their correlation to each other leads
to a lower total retrieval time. By clustering we mean that highly correlated items are stored in
neighbouring storage containers.
To check if our algorithm does indeed cluster correlated items, we visualise in figure 3 the
changes in correlation between neighbouring items. The visualisation is done via the average
jaccard-similarity coefficient of a batch which contains item i and a batch which contains a
direct neighbour of item i,which we call j. Roughly speaking, a high jaccard-similarity coefficient
means that item i and its neighbour j are positively correlated in the picking process, i.e. a large
percentage of batches which contain i also contain j.
The calculation of the jaccard-similarity coefficient goes as follows: Denote {Bi } a batch in
which i occurs and {Bj } a batch in which j occurs. The jaccard-similarity coefficient measures
the “similarity” between two finite sample sets {Bi } and {Bj } and is defined by
T
|{Bi } {Bj }|
(5.1) sim ({Bi } , {Bj }) = S .
|{Bi } {Bj }|
Moreover, define the set of neighbours of item i as the set of items j, which are in the same or
adjacent subsection of an aisle. We require that j has to be stored in the same level category
as i, meaning that if i is stored in a lower level, then also j has to be stored in a lower level to
qualify as a neighbour.
Figure 3 shows that the heuristic approach taken by the company (plot a) of figure 3) shows a
high similarity between a few items, i.e. objects which are often picked to the same batch are
stored in neighbouring containers. However, the optimised configuration(plot b) of figure 3) does
not display areas of a high similarity.
Consequently, in contrast to intuition, clustering of correlated items does not necessarily lead to
a reduction in retrieval times. This phenomenon can be explained with the relatively low impact
of item distance to the total retrieval time of a large batch, in relation to the sum of the picking
times of the individual items. This behaviour is also present in the other warehouse levels.
To conclude, the analysis of correlations of neighbouring items shows that the examined ware-
house cannot be simply divided into “hot areas” and a “cold areas” with respect to batched
orders: the complex structure of the batches leads to highly non-trivial optimised configurations.

5.3.3. Detailed analysis - individual picking times. Another important question is whether an
optimal total retrieval time for a large number of batches might go to the expense of very long
retrieval times for a few batches. The last step in our analysis is therefore to check the individual
 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE 11

1.0 1.0

a) b)
0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0.0 0.0

Figure 3. The average jaccard-similarity for each item in the lowest level a)
before b) after optimisation

1000
Original
Optimised
800

600
Frequency

400

200

0
0.0 0.2 0.4 0.6 0.8 1.0
Normed batch picking time

Figure 4. Histogram of individual batch picking times normed to the maximum

of the original picking time (blue) before and (green) after optimisation.

picking times per item. For better visibility, the displayed picking times are normed to the
maximum batch picking time in the original configuration. The histogram of individual picking
times are shown in figure 4: Before the optimisation, the distribution of individual picking times
is centred around a value of 0.16 with a right handed fat tail. The algorithm is able to shift the
12 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE

distribution to shorter times and eliminates parts of the fat tail. This proves that the optimal
storage assignment obtained by our algorithm bears a lower probability of extremely long retrieval
times for arbitrary batches.

6. Discussion/Conclusion
This paper presents a simulated annealing algorithm to reduce picking times of potentially large
sets of batched orders by a combined optimisation of the storage assignments in a multi-level
warehouse and an implicit s-shape routing heuristic.
Our main contribution is a general solution method for the solution of the storage location as-
signment problem under adaptable routing methods, which is flexible enough to be used in a
multi-level warehouse setting. While the current algorithm was designed for single-block ware-
house layouts, extensions to more general warehouse settings can easily be done by adapting
formula (4.3). Notably, this algorithm can also be applied to parts of the warehouse without
resulting in phantom constraints or other unnatural solutions to this combined optimisation
problem.
To our knowledge, our article is the first to investigate the impact of large batched orders on the
optimal storage assignment. Contrary to intuition, we give evidence that, even for batched orders
of very large size (e.g. over 100 items in one batch), clustering of correlated items in specific
parts of the warehouse does not necessarily lead to a reduction in retrieval times. Moreover, we
show that clustering heuristics are not even beneficial to reduce the probability of extremely long
retrieval times for rarely occurring batches.
We tested our algorithm on real data, optimising the storage assignments in a four-level ware-
house of a manufacturer with pre-batched orders of 1-150 items per order. The optimal storage
assignment suggested by the algorithm reduces the total retrieval time by 21% compared to
heuristics based on the frequency of picking for individual pieces. Assuming a random item dis-
tribution in the warehouse, this simulated annealing algorithm reduces the overall picking time
approximately by 38%. These savings are achieved without changing the routing heuristics and
without splitting large pre-batched orders. The simultaneous delivery of all items, even of a very
large batch, is crucial to maintaining efficiency of the assembly line. Note that the algorithm is
very fast, as all parts of the algorithm are designed and optimised for multi-level warehouses.
No black-box packages are used.
The main novelty of our method is to provide a structured approach for the storage location
assignment problem with very general pre-batching constraints in multi-level warehouses settings.
In particular, the algorithm is able to deal with complex correlated batches and able to find highly
non-trivial optimised configurations.

Acknowledgements
We gratefully acknowledge the financial, data and feedback contribution from our partner com-
pany. A.E. and C.G. thank Stiftung der Deutschen Wirtschaft for partial funding of this project.
C.G.’s research is supported by ERC grant no. 277749 “EPSILON”.

References
[1] New batch construction heuristics to optimise the performance of order picking systems. International Journal
of Production Economics, 131(2):618 – 630, 2011.
[2] H. Ahonen, A. G. de Alvarenga, and A. Amaral. Simulated annealing and tabu search approaches for the
corridor allocation problem. European Journal of Operational Research, 232(1):221–233, 2014.
[3] A. R. Amaral. The corridor allocation problem. Computers & Operations Research, 39(12):3325–3330, 2012.
[4] E. Atmaca and A. Ozturk. Defining order picking policy: A storage assignment model and a simulated
annealing solution in as/rs systems. Applied Mathematical Modelling, 37(7):5069–5079, 2013.
 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE 13

[5] J. J. Bartholdi III and S. T. Hackman. Warehouse & distribution science: release 0.92. Atlanta, GA, The
Supply Chain and Logistics Institute, School of Industrial and Systems Engineering, Georgia Institute of
Technology, 2011.
[6] C. Battista, A. Fumi, F. Giordano, and M. Schiraldi. Storage location assignment problem: implementation in
a warehouse design optimization tool. In Proceedings of the Conference Breaking Down the Barriers between
Research and Industry, 2011.
[7] H. Brynzer and M. Johansson. Storage location assignment: Using the product structure to reduce order
picking times. International Journal of Production Economics, 46:595 – 603, 1996.
[8] V. Černỳ. Thermodynamical approach to the traveling salesman problem: An efficient simulation algorithm.
Journal of optimization theory and applications, 45(1):41–51, 1985.
[9] F. T. Chan and H. K. Chan. Improving the productivity of order picking of a manual-pick and multi-level rack
distribution warehouse through the implementation of class-based storage. Expert Systems with Applications,
38(3):2686–2700, 2011.
[10] K. Choe and G. Sharp. Small parts order picking: design and operation. Online tutorial, 1991.
[11] H. Cohn and M. Fielding. Simulated annealing: searching for an optimal temperature schedule. SIAM Journal
on Optimization, 9(3):779–802, 1999.
[12] J. J. Coyle, E. J. Bardi, C. J. Langley, et al. The management of business logistics, volume 6. West Publishing
Company St Paul, MN, 1996.
[13] M. De Koster, E. S. Van der Poort, and M. Wolters. Efficient orderbatching methods in warehouses. Inter-
national Journal of Production Research, 37(7):1479–1504, 1999.
[14] R. De Koster. How to assess a warehouse operation in a single tour. Technology Report. Erasmus University,
Netherlands, 2004.
[15] R. De Koster, T. Le-Duc, and K. J. Roodbergen. Design and control of warehouse order picking: A literature
review. European Journal of Operational Research, 182(2):481–501, 2007.
[16] R. De Koster and A. Neuteboom. The logistics of supermarket chains: a comparison of seven chains in the
Netherlands. Elsevier Business Information, 2001.
[17] R. De Koster and E. Van der Poort. Routing order pickers in a warehouse: a comparison between optimal
and heuristic solutions. IIE transactions, 30(5):469–480, 1998.
[18] R. Dekker, M. De Koster, K. J. Roodbergen, and H. Van Kalleveen. Improving order-picking response time
at Ankor’s warehouse. Interfaces, 34(4):303–313, 2004.
[19] J. Drury. Towards more efficient order picking. IMM monograph, 1, 1988.
[20] European Logistics Association and AT Kearney Management Consultants. Differentiation for performance
excellence in logistics, 2004.
[21] E. Frazelle. Supply chain strategy: the logistics of supply chain management. McGrraw Hill, 2002.
[22] E. Frazelle and G. P. Sharp. Correlated assignment strategy can improve any order-picking operation. In-
dustrial Engineering, 21(4):33–37, 1989.
[23] E. H. Frazelle. Stock location assignment and order picking productivity. Georgia Tech Theses and Disser-
tations, 1989.
[24] N. Gademann and S. Velde. Order batching to minimize total travel time in a parallel-aisle warehouse. IIE
transactions, 37(1):63–75, 2005.
[25] M. Goetschalckx and J. Ashayeri. Classification and design of order picking. Logistics World, 2(2):99–106,
1989.
[26] M. Goetschalckx and H. D. Ratliff. An efficient algorithm to cluster order picking items in a wide aisle.
Engineering Costs and Production Economics, 13(4):263–271, 1988.
[27] A. M. Gomes and J. F. Oliveira. Solving irregular strip packing problems by hybridising simulated annealing
and linear programming. European Journal of Operational Research, 171(3):811–829, 2006.
[28] R. W. Hall. Distance approximations for routing manual pickers in a warehouse. IIE transactions, 25(4):76–
87, 1993.
[29] W. H. Hausman, L. B. Schwarz, and S. C. Graves. Optimal storage assignment in automatic warehousing
systems. Management Science, 22(6):629–638, 1976.
[30] S. Henn, S. Koch, H. Gerking, and G. Wäscher. A u-shaped layout for manual order-picking systems. Logistics
Research, 6(4):245–261, 2013.
[31] S. Henn and V. Schmid. Metaheuristics for order batching and sequencing in manual order picking systems.
Computers & Industrial Engineering, 66(2):338–351, 2013.
[32] S. Hong, A. L. Johnson, and B. A. Peters. Batch picking in narrow-aisle order picking systems with consid-
eration for picker blocking. European Journal of Operational Research, 221(3):557–570, 2012.
[33] Establish. Inc and Herbert W. Davis &. Co. Logistic cost and service, 2005.
[34] S. Kirkpatrick. Optimization by simulated annealing: Quantitative studies. Journal of statistical physics,
34(5-6):975–986, 1984.
 14 A SIMULATED ANNEALING APPROACH TO OPTIMAL STORING IN A MULTI-LEVEL WAREHOUSE

[35] R. Kutzelnigg. Optimal allocation of goods in a warehouse: Minimizing the order picking costs under real-life
constraints. In 3rd IEEE International Symposium on Logistics and Industrial Informatics, pages 65–70,
Aug 2011.
[36] T. Loukil, J. Teghem, and P. Fortemps. A multi-objective production scheduling case study solved by simu-
lated annealing. European journal of operational research, 179(3):709–722, 2007.
[37] M. Matusiak, R. de Koster, L. Kroon, and J. Saarinen. A fast simulated annealing method for batch-
ing precedence-constrained customer orders in a warehouse. European Journal of Operational Research,
236(3):968–977, 2014.
[38] N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller. Equation of state calculations
by fast computing machines. The journal of chemical physics, 21(6):1087–1092, 1953.
[39] V. R. Muppani and G. K. Adil. Efficient formation of storage classes for warehouse storage location assign-
ment: a simulated annealing approach. Omega, 36(4):609–618, 2008.
[40] C. G. Petersen. An evaluation of order picking routeing policies. International Journal of Operations &
Production Management, 17(11):1098–1111, 1997.
[41] C. G. Petersen. The impact of routing and storage policies on warehouse efficiency. International Journal of
Operations & Production Management, 19(10):1053–1064, 1999.
[42] H. D. Ratliff and A. S. Rosenthal. Order-picking in a rectangular warehouse: a solvable case of the traveling
salesman problem. Operations Research, 31(3):507–521, 1983.
[43] K.-J. Roodbergen. Layout and routing methods for warehouses. Erasmus University Rotterdam, 2001.
[44] K. J. Roodbergen and R. De Koster. Routing order pickers in a warehouse with a middle aisle. European
Journal of Operational Research, 133(1):32–43, 2001.
[45] K. J. Roodbergen and R. Koster. Routing methods for warehouses with multiple cross aisles. International
Journal of Production Research, 39(9):1865–1883, 2001.
[46] J. A. Tompkins, J. A. White, Y. A. Bozer, and J. M. A. Tanchoco. Facilities planning (third ed). John Wiley
& Sons, 2003.
[47] J. P. Van den Berg, G. P. Sharp, A. N. Gademann, and Y. Pochet. Forward-reserve allocation in a warehouse
with unit-load replenishments. European Journal of Operational Research, 111(1):98–113, 1998.
[48] J. P. van den Berg and W. Zijm. Models for warehouse management: Classification and examples. Interna-
tional Journal of Production Economics, 59(1):519–528, 1999.
[49] I. M. Whittley and G. D. Smith. The attribute based hill climber. Journal of Mathematical Modelling and
Algorithms, 3(2):167–178, 2004.

View publication stats