Omega 36 (2008) 609 – 618

www.elsevier.com/locate/omega

Efficient formation of storage classes for warehouse storage location

assignment: A simulated annealing approach夡
Venkata Reddy Muppani (Muppant), Gajendra Kumar Adil∗
Shailesh J. Mehta School of Management, Indian Institute of Technology Bombay, Mumbai 400076, India

Received 30 March 2006; accepted 12 January 2007

Available online 18 May 2007

Abstract
Class-based storage policy distributes products among a number of classes and for each class it reserves a region within the
storage area. The procedures reported in the literature for formation of storage classes primarily consider order-picking cost
ignoring storage-space cost. Moreover, in these procedures items are ordered on the basis of their cube per order index (COI), and
items are then partitioned into classes maintaining this ordering. This excludes many possible product combinations in forming
classes which may result in inferior solutions. In this paper, a simulated annealing algorithm (SAA) is developed to solve an integer
programming model for class formation and storage assignment that considers all possible product combinations, storage-space
cost and order-picking cost. Computational experience on randomly generated data sets and an industrial case shows that SAA
gives superior results than the benchmark dynamic programming algorithm for class formation with COI ordering restriction.
䉷 2007 Elsevier Ltd. All rights reserved.

Keywords: Order picking; Storage-space cost; Class-based storage; Cube per order index; Simulated annealing

1. Introduction management are two important and closely knit activ-

ities that can have an impact on supply chain costs.
Organizations are increasingly adopting supply chain There is enormous pressure to simultaneously increase
management practices in the hope of reducing supply throughput rate and lower operating costs of ware-
chain costs as today’s competition is moving from houses. One means of improving throughput rate is by
“among organizations” to “between supply chains” using automatic storage and retrieval systems [2]. There
[1]. Effective supply chain management has become are many factors to be considered in order to keep
a potentially valuable way of securing competitive operating costs to a minimum in a warehouse. Storage
advantage and improving organizational performance. location assignment influences almost all key ware-
Thus, every activity and element in the supply chain house performance indicators including order-picking
is extremely important. Warehousing and inventory time and cost, productivity, shipping and inventory
accuracy, and storage density [3].
夡 This manuscript was processed by Guest Editors Angappa Hausman et al. [4] present three storage location as-
Gunasekaran and T.C. Edwin Cheng. signment policies: randomized storage, dedicated stor-
∗ Corresponding author. Tel.: +91 22 2576 7738;
fax: +91 22 2572 2872. age, and class-based storage. The randomized storage
E-mail addresses: [email protected] (V.R. Muppani policy allows items to be stored anywhere in the storage
(Muppant)), [email protected], [email protected] (G.K. Adil). area. Under the dedicated storage policy each region
0305-0483/$ - see front matter 䉷 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.omega.2007.01.006
610 V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 – 618

may only be used for a specific item. The class-based in Section 4. In Section 5, we discuss computational
storage policy distributes the items among a number of experience when SAA is applied to randomly generated
classes and for each class it reserves a region within problem instances and an industrial case. Further, the
the storage area. Randomized and dedicated storage are solutions generated are compared with those of dynamic
extreme cases of class-based storage policy: randomized programming algorithm (DPA) for class formation with
storage considers a single class and dedicated storage COI ordering restriction. The paper is concluded in
considers one class for each item. Section 6.
If only order-picking cost is considered, dedi-
cated policy may yield the lowest cost; on the other 2. Literature review
hand, if only space cost is considered, the completely
random policy (one class) will yield the lowest cost There are four methods used to reduce pick travel
solution. However, if we consider both the costs simul- distances namely (1) determining good order-picking
taneously, the best solution may be neither of these. routes, (2) zoning the warehouse, (3) assigning orders
The approaches reported in the literature for class for- to batches, and (4) assigning products to correct stor-
mations primarily consider order-picking cost [5] even age locations [11]. In this section, we review papers
though class-based policies may offer reduction in cost pertinent to storage location assignment. A detailed dis-
through space savings in situations when space cost is cussion of warehouse design literature can be found in
significant vis-à-vis picking cost. [12–15].
One criterion used to assign product classes to storage Harmatuck [16] has shown that the COI-based item
locations that captures item popularity and its storage- allocation minimizes the expected order-picking cost for
space requirement is cube-per-order index (COI) [6]. single-address stock location. Malmborg and Bhaskaran
COI of an item is defined as the ratio of the item’s [17] studied the optimality of the COI-based assignment
storage-space requirement (cube) to its popularity (num- policy for the case in which a single storage transaction
ber of storage/retrieval requests for the items). In the and a single retrieval transaction are interleaved in order
earlier works (e.g., [5,7]) items are ordered on the ba- to economize on the travel of the order-picking vehi-
sis of their COI and then using this ordering, prod- cle. Items allocation based on COI values gives optimal
uct classes are formed as shown in solutions 1–3 in allocation in terms of order picking/storing time under
Table 1. Items with lower COI values are put together dedicated storage policy for single command transac-
in a class and located closer to I/O point than the items tion [18]. Malmborg and Bhaskaran [19] provided the
with higher COI values that were assigned to another revised proof of COI rule and a tie-breaking rule based
class in order to reduce the order-picking cost. These on Euclidean distances.
restricted combinations in forming classes (i.e., class Hausman et al. [4] observed that the travel time might
formation with COI ordering restriction) help overcom- be measured by the Chebyshev metric. Guenov and
ing the combinatorial nature of the problem by reduc- Raeside [20], in their experiments, estimated bounds
ing the solution space [8]. However, this may not yield for an optimum tour with respect to Chebyshev travel.
the least cost solution [9, pp. 284–285] as illustrated Kanet and Ramirez [21] analyze picking decisions and
in solution 4, Table 1. As a result of combining items propose an integer programming model that allows the
into classes, storage area required reduces which trans- inclusion of fixed costs per pick and breakdown costs
lates into lower order-picking and storage-space costs. that occur when the quantity of items stored in a loca-
Grouping items having complementary planned inven- tion is greater than the ordered quantity and therefore,
tory profiles may require smaller region within the stor- it must be broken down into multiple lots.
age area than grouping based on COI values. How- Hausman et al. [4] consider the problem of finding
ever, for given classes, allocation ofstorage space based class regions for the class-based storage policy using
on COI of classes [10] minimizes order-picking cost as grid search to determine optimal boundaries for three
illustrated in solution 4, Table 1. classes. The authors show that a class-based storage pol-
In this paper, we develop a simulated annealing algo- icy with relatively few classes yields mean travel times
rithm (SAA) to form storage classes efficiently consid- that are close to those obtained by a dedicated storage
ering storage-space cost, order-picking cost, and class policy for continuous racks and a specific demand func-
formation (part grouping) without any ordering restric- tion. They assumed same area for dedicated policy as
tion. The paper is organized as follows. In Section 2, well as class-based policy. This is extended to estab-
pertinent literature is reviewed. We present problem for- lish class boundaries for any given number of classes
mulation in Section 3 as in [10]. SAA is developed in a square-in-time rack [22] and to any rectangular
 V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 – 618 611

Table 1
An example illustrating storage classes formation

Policy Class Products Inventory level Space requiredc Average Costs

formation (in units) demand
i i i + i d Order pickingd Storagee Total

Solution 1 With COI P1 5 20 25.00 15 37.5 25.00 62.50

(dedicated) ordering P2 5 4 9.00 5 29.5 9.00 38.50
restrictiona P3 20 20 40.00 15 162 40.00 202.00
Total 229 74.00 303.00

Solution 2 P 1/P 2 10 20.4 30.40 20 60.79 30.40 91.19

(class based) P3 20 20 40.00 15 151.19 40.00 191.19
Total 211.98 70.40 282.38
Solution 3 P1 5 20 25.00 15 37.50 25.00 62.50
(class based) P 2/P 3 25 20.40 45.40 20 190.79 45.40 236.19
Total 228.29 70.40 298.69

Solution 4 Without COI P2 20 1.5 213.14 35 159.85 53.28 213.14

(class based) ordering P 1/P 3 20 4.03 66.78 59 57.784 9.00 66.78
restrictionb Total 217.64 62.28 279.92
a COI of P 1, P 2 and P 3 are 1.67, 1.80 and 2.67, respectively. Items are partitioned into classes maintaining this ordering. Successive classes
thus formed are allocated available storage locations closest to I/O point.
b Classes are formed without COI ordering restriction (i.e., P 1 is grouped with P 3) and are allocated sequentially from the I/O point in the

order of increasing COI of classes [10]. COI of classes {P 2} and {P 1, P 3} are 1.63 and 2.46, respectively, and are allocated storage locations
closest to I/O point in that order.
c The storage space allocated corresponds to one standard deviation above mean planned inventory level.
d Order-picking cost is computed assuming a rectangular layout of width 10 units and the I/O point at the bottom of the rectangle. Order-

picking distance for a class is the distance from the I/O point to the centroid of the space allocated to theclass.
e Storage-space cost = $1 per square unit and Order-picking cost = $1 per unit.

rack [23]. Graves et al. [24] observe that L-shaped re- policy. This is possible as the reduction in total space
gions are not necessarily optimal when dual commands requirements in random storage can also reduce average
occur, but they argue that the mean dual command cy- retrieval costs since it creates an opportunity to locate
cle time in an L-shaped class allocation will in general items in closer regions [29].
be no more than 3% above the optimum. They observe van den Berg [7] presents a DPA that distributes items
that the space requirements increase with the number and locations among classes such that the mean single
of classes. command travel time is minimized considering reduc-
Schwarz et al. [25] verify the analytical results in tion in area as a result of grouping items. The algo-
[4,24] using simulation. The simulations suggest that rithm may be applied to a wide variety of warehousing
most results of the analytic models hold under stochas- systems such as warehouse layout, demand curve, and
tic conditions. Hackman and Rosenblatt [26] present a travel time metric. Applicability of this approach for
model for forward/reserve problem that considers both large problems poses computational difficulties. Mup-
assignment (which item) and allocation (what amounts) pani and Adil [10] developed a mathematical model to
for forward and reserve locations. capture space and order-picking costs and to form stor-
The emphasis in the above models has been in reduc- age classes with consideration to reduction in area when
ing the time for storing/order picking. However, floor using class-based storage in a warehouse. They pro-
space utilization is also important for warehouses. There posed a greedy heuristic procedure and DPA to solve
are papers dealing with joint inventory planning and the above model. They observe that there is a magni-
warehouse sizing (see for instance, [27,28]). Policies tude difference between the time taken by the greedy
such as cross docking can also help in reducing the heuristic and DPA and also demonstrate that at times
floor space requirements. For a given inventory policy, the greedy heuristic procedure provided better objective
shared storage policy gives a better floor space utiliza- function value.
tion. Petersen et al. [5] study the improvement in order- As can be seen from the literature review, only few
picking time as a result of forming classes (based on papers consider reduction in storage space (e.g., [7,10])
grouping parts with similar COI values) over a random and the relative costs of storage space and order picking
612 V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 – 618

[10] in class formation. Further, the procedures reported f space cost in $ per square feet for
in the literature for formation of classes are only on the planning horizon
the basis of COI of items which may be sub-optimal as fp footprint density (footprint
shown in Table 1. area required to store one unit
load of product p considering
3. Problem formulation the stacking height)
h order-picking cost in $ per foot per unit load
We assume storages and retrievals are performed in Ipt storage level in unit loads planned
single command cycles and all items are stored and for product p during period t
transported on identical storage media (e.g., pallets or Decision variables:
totes). The class-based storage policy is used. Each lo- 
1 if product p is assigned to class c,
cation is uniformly utilized and the assigned items are xpc =
distributed homogeneously in the space allocated for  0 otherwise
1 if location l is assigned to class c,
the class. This assumption implies that the geometric ylc =
center of the class is the same as the load center. We 0 otherwise
assume that inventory decision has been made inde-
pendent of storage decision and all times required in 3.2. Class formation and allocation model (CFAM)
the storage/retrieval process, except travel times, are in-
dependent of the storage allocation. We also assume Using the above notation and assumptions, the fol-
that there is no congestion between vehicles/cranes and lowing mathematical model can be formulated as
loads are not relocated. All the parameters of the mod- CFAM:
 
els are deterministically known and the travel time be- Minimize z = f · (al · ylc ) + 2h ·
tween any two positions is readily computed. Once the c c
storage locations are assigned they cannot be reshuffled  l   
l (al · dl · ylc )
during the planning horizon and we assume there is a ×  ∗ Dp · xpc
single input/output point. l (al · ylc ) p
The problem can be defined as: given P prod- (1)
ucts/items, their average demand and planned inventory subject to
levels for T periods and the layout of the storage area 
xpc =1 ∀p, (2)
divided into lattice of storage locations, the problem
c
is to establish classes of products and allocate them  
 
to storage locations so that the total cost of order Max (Ipt ·fp · xpc )  (al ·ylc ) ∀c,
picking and storage space is minimized in a single t
p l
command warehouse exploiting the reduction in area (3)
because of clubbing the parts into classes. We establish 
ylc 1 ∀l (4)
notation and then present formulation of a model as
c
in [10].
xpc , ylc ∈ {0, 1} ∀p, c, l. (5)
3.1. Notation The objective function (1) minimizes the sum of stor-
age space cost and order-picking cost over the plan-
We use the following notation in this paper.  horizon. The storage-space cost for class c is f ·
ning
Indices: l (al · ylc ). One way order-picking cost for class c
c (c = 1, 2, 3, . . . , C = P ) for classes is product of unit order-picking cost h, centroid dis-
l (l = 1, 2, 3, . . . , L) for storage locations tance
 of the allocated  space from the input/output point
p (p = 1, 2, 3, . . . , P ) for products/items l (a l · d l · ylc )/ l (al · ylc ) and total number of picks
t (t = 1, 2, 3, . . . , T ) for time periods p D p · x pc . Constraint (2) ensures that each item is
assigned to one and only one class. Constraint (3) en-
Parameters: sures that there is adequate storage space to hold the
al footprint area of location l in square feet items in a class c in each planning period t. A storage
dl distance of location l from the input/output point location, if at all assigned, is to be assigned to only one
Dp total number of picks (in unit loads) class. This is taken care of by constraint (4). Constraint
for product p in the planning period (5) imposes binary restrictions on decision variables.
 V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 – 618 613

4. Solution approach The proposed SAA exploits the structure of the

CFAM and it searches in space, xpc , while determin-
The CFAM involves nonlinear objective function and ing values of ylc in evaluation step by solving AM as
integer variables posing difficulties in optimally solving described before. Thus, solution (SOL) to the CFAM
large problems. Simulated annealing (SA) is a meta- consists of class formation, CF and allocation of space
heuristic for solving such combinatorial optimization to classes formed, AS.
problems [30]. Koulamas et al. [31] provided a survey
on applications of simulated annealing in layout prob- 4.1.1. Storage classes in initial solution
lems. We develop an SAA to solve CFAM. Solution to To obtain initial classes, CF0 , create C (=P ) classes
CFAM comprises of class formation (xpc variables) and having one item each with item 1 in class 1, item 2
space allocation to the classes (ylc variables). If classes in class 2 and so on. Then, compute COI for each
are given, i.e., the value of xpc variables fixed (at say class c (COIc ) using expression (9) and arrange the
Xpc ), the storage locations, ylc , can be optimally deter- classes in increasing order of COIc . Solve the AM
mined by solving the following allocation model (AM). for CF0 to obtain space allocation, AS0 , and objective
Allocation model (AM): value Z0 . CF0 and AS0 together form initial solution,

Minimize z2 = Alc .ylc , (6) SOL0 .
c l
subject to 4.1.2. Generation of storage classes in neighborhood

al · ylc Bc ∀c, (7) solution
l Generation of neighborhood classes and allocation to
ylc ∈ {0, 1} ∀c, l, (8) storage spaces are described below.
Neighborhood storage classes: Randomly select two
where
  classes, k and l such that l  = k from the current solution,
 CFi . Randomly select a part, pi from within the parts
Bc = Max (Ipt · fp · Xpc ) and in the class k and move the part pi from class k to class
t
p
l to get CFi+1 .
 
· Xpc Allocation of storage space: Solve the AM for CFi+1
p Dp
Alc = f · al + 2h(al · dl ) · . to obtain space allocation and objective value Zi+1 to
Bc
get solution SOLi+1 .
It can be proved that the optimum solution for AM can
be obtained using the following three steps. The classes 4.1.3. Simulated annealing algorithm (SAA)
are ranked in the increasing order of their COI (COIc ), SA initialization:
which is computed using the following equation as
in [10]. • Read input data: warehouse layout information,
Area allocated to class c number of products (P), number of periods (T),
COIc = demand (Dp ), foot print density (fp ) and storage
Demand rate of products in class c
 levels (Ipt ) for all the items.
Maxt p (Ipt · fp · xpc ) • Define the annealing parameters: initial tempera-
=  . (9)
p (Dp /T ) · xpc
ture, T0 ; maximum transitions at each temperature,
Lmax ; maximum accepted transition at each tem-
The storage locations are ranked in increasing order perature, ATmax ; temperature decrementing factor,
of their rectilinear distances from I/O point. If there is ; maximum number of iterations, imax ; and final
a clash, it is resolved by ranking based on Euclidean acceptance ratio, Rf .
distance of the locations from I/O point [19]. The classes
are allocated required storage space by picking storage SA Loop:
locations in the ranked order of storage locations. Step 0: Initialize iteration counter: i = 0, temperature
4.1. Simulated annealing Ti = T0 . Generate an initial solution as in Section 4.1.1.
Let the best solution be represented by SOLbest and the
In this section, we describe procedure to develop an best objective value by Zbest . Assign initial operation
initial solution, procedure to generate a neighborhood assignment SOL0 to SOLbest and initial objective value
solution given any solution and SAA. Z0 to Zbest .
 614 V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 – 618

Step 1: Execute outer loop, i.e., steps (1.1–1.7) until Problem instances each with 40 items were gen-
conditions in step 1.7 are met. erated considering three key factors that influence
order-picking and space costs. These are pick density
1.1. Initialize inner loop counter l=0, and accepted num- of products measured as reciprocal of COI [3], inven-
ber of transitions AT = 0. tory and the relative cost of order picking and storage
1.2. Initialize solution for inner loop, SOL0i =SOLi , Zi0 = space. Five inventory patterns with low variability and
Zi . another five inventory patterns with high variability
1.3. Execute inner loop, i.e., steps (1.3.1–1.3.4) until having coefficient of variation (COV) of inventory av-
conditions in step 1.3.4 are met. eraged for all items varying in the range of 4.97–45.30
1.3.1. Set l = l + 1. and 110.77–293.21, respectively, were used. Similarly,
Generate a neighborhood solution as in Sec- five pick density patterns with low variability and an-
tion 4.1.2 (get SOLli , Zil ). other five pick density patterns with high variability
1.3.2.  = Zil − Zil−1 . having COV of pick density of items varying in the
1.3.3. If  0 or random (0, 1) e−/Ti then range of 0.10–0.22 and 0.30–0.50, respectively, were
• SOLli and Zli are accepted. used. Twenty pick density-inventory combinations:
• Update accepted transition AT to AT + 1. Low–Low, Low–High, High–Low, and High–High are
Else created, i.e., five data points of pick density with low
• Solution is rejected, SOLli = SOLl−1 variability combining the five data points of inventory
i , Zi =
l
l−1 with low variability yielding five combinations. The
Zi .
storage space allocated corresponds to one standard
1.3.4. If one of the following conditions holds true:
deviation above mean planned inventory level.
AT ATmax or l Lmax , then assign l to Li
Eighty problem instances were generated by com-
(length of Markov chain), terminate the inner
bining four levels of space cost per square foot (f) for
loop and go to 1.4; else continue the inner loop
the duration of the planning period: $0, $0.42, $0.8 and
and go to 1.3.1.
$1.1 with the above 20 pick density-inventory combi-
nations. We assume order-picking cost per foot (h) to
1.4. Set: i = i + 1.
be $0.0025 and T to be 7 periods for all the problem in-
1.5. Update SOLi = SOLl−1 i−1 , Zi = Zi−1 . If Zi < Zbest
l−1
stances. We compared the SAA with a benchmark DPA
then SOLbest = SOLi and Zbest = Zi .
for class formation with COI ordering restriction [10].
1.6. Reduce the cooling temperature: Ti =  ∗ Ti−1 .
Percentage benefit of class-based solution obtained
1.7. If one of the following conditions holds true:
by DPA and SAA over the dedicated solution for the
i imax , or the acceptance ratio (defined as
80 problem instances described above is shown against
AT/Li )Rf , then terminate the outer loop and go
COV of inventory and COV of pick density of all items
to 2, else continue the outer loop and go to 1.1.
in Table 2. It can be observed that SAA performs bet-
Step 2: Print the best solution obtained and terminate
ter than DPA in 64 of 80 problem instances tested with
the procedure.
the difference in objective values as high as 7.34%. A
4.2. SA parameter setting paired t-test rejected the equality of the DPA and SAA
solutions (t-statistic = 7.615, p = 0.000 at 95% confi-
To determine good annealing parameters for a given dence level). Thus, SAA performs better than DPA as
problem is often a hard task and needs experimentation conjectured and illustrated in Table 1 for the SAA pa-
[32]. The following values for the SA parameters were rameters stated in Section 4.2 for the problem instances
selected after experimentation: tested.
In order to understand the influence of inventory
=0.98, T0 =7500.0, Lmax =200, ATmax =100, and pick density characteristics on the performance
imax = 2000. of class-based solutions, the problem instances are
grouped based on COV of inventory and COV of pick
5. Computational experience density as shown in Table 3. Average percentage bene-
fit of SAA and DPA over dedicated solution for all the
In order to assess the quality of solution generated by four groups is shown. It can be observed that when the
SAA, a rectangular storage area of 60 × 48 ft 2 divided average COV of inventory of items is high, class-based
into unit cells of 1 ft × 1 ft and I/O point located on the solutions offer greater benefit over dedicated solution.
periphery at the center of the width is assumed. This is due to the fact that, when items are grouped into
 V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 – 618 615

Table 2
Benefit of class based (SAA and DPA) over dedicated solution

Problem instances

1 2 3 4 5 6 7 8 9 10

f =0
%Benefit over dedicated
SAA 1.83 4.49 6.74 8.69 10.59 1.41 3.94 5.71 7.60 9.20
DPA 1.87 4.44 5.94 7.92 9.52 1.56 3.95 5.28 7.13 8.60
Average COV of inventory 4.97 14.90 24.81 34.61 45.30 4.97 14.90 24.81 34.61 45.30
COV of pick density of all SKUs 0.10 0.13 0.16 0.19 0.22 0.30 0.35 0.40 0.45 0.50

11 12 13 14 15 16 17 18 19 20

%Benefit over dedicated

SAA 20.89 22.73 27.06 24.68 26.06 17.49 19.20 23.94 21.17 22.61
DPA 16.05 17.38 25.05 18.71 21.10 14.18 15.49 22.93 16.96 19.00
Average COV of inventory 110.77 151.17 190.00 257.46 293.21 110.77 151.17 190.00 257.46 293.21
COV of pick density of all SKUs 0.10 0.13 0.16 0.19 0.22 0.30 0.35 0.40 0.45 0.50

f = 0.42 1 2 3 4 5 6 7 8 9 10

%Benefit over dedicated

SAA 2.47 5.82 10.06 11.90 14.55 1.68 4.39 6.66 8.70 10.52
DPA 2.48 6.09 9.79 11.63 14.10 1.73 4.46 6.33 8.35 10.07
Average COV of inventory 4.97 14.90 24.81 34.61 45.30 4.97 14.90 24.81 34.61 45.30
COV of pick density of all SKUs 0.10 0.13 0.16 0.19 0.22 0.30 0.35 0.40 0.45 0.50

11 12 13 14 15 16 17 18 19 20

%Benefit over dedicated

SAA 23.92 26.79 34.51 28.23 30.92 19.42 21.03 27.01 22.75 24.14
DPA 20.98 23.73 34.44 23.94 27.51 16.70 17.91 26.08 19.01 21.19
Average COV of inventory 110.77 151.17 190.00 257.46 293.21 110.77 151.17 190.00 257.46 293.21
COV of pick density of all SKUs 0.10 0.13 0.16 0.19 0.22 0.30 0.35 0.40 0.45 0.50

f = 0.8 1 2 3 4 5 6 7 8 9 10

%Benefit over dedicated

SAA 2.75 6.79 11.37 13.07 16.38 1.81 4.55 7.32 9.57 11.67
DPA 2.75 6.99 11.05 13.28 16.05 1.86 4.85 7.07 9.23 11.22
Average COV of inventory 4.97 14.90 24.81 34.61 45.30 4.97 14.90 24.81 34.61 45.30
COV of pick density of all SKUs 0.10 0.13 0.16 0.19 0.22 0.30 0.35 0.40 0.45 0.50

11 12 13 14 15 16 17 18 19 20

%Benefit over dedicated

SAA 25.58 28.92 38.54 30.24 33.26 20.90 22.20 28.65 23.57 25.57
DPA 23.75 26.99 38.54 27.16 31.28 18.55 19.80 28.45 20.48 22.87
Average COV of inventory 110.77 151.17 190.00 257.46 293.21 110.77 151.17 190.00 257.46 293.21
COV of pick density of all SKUs 0.10 0.13 0.16 0.19 0.22 0.30 0.35 0.40 0.45 0.50

f = 1.1 1 2 3 4 5 6 7 8 9 10
%Benefit over dedicated
SAA 2.90 7.32 12.05 14.27 17.27 1.94 5.07 7.84 10.10 12.40
DPA 2.90 7.44 11.81 14.09 16.99 1.96 5.12 7.58 9.89 11.99
Average COV of inventory 4.97 14.90 24.81 34.61 45.30 4.97 14.90 24.81 34.61 45.30
COV of pick density of all SKUs 0.10 0.13 0.16 0.19 0.22 0.30 0.35 0.40 0.45 0.50
11 12 13 14 15 16 17 18 19 20
%Benefit over dedicated
SAA 26.55 29.86 40.53 31.20 34.51 21.76 23.50 28.78 24.64 26.56
DPA 25.23 28.63 40.53 29.05 33.34 19.71 21.01 29.47 21.64 24.14
Average COV of inventory 110.77 151.17 190.00 257.46 293.21 110.77 151.17 190.00 257.46 293.21
COV of pick density of all SKUs 0.10 0.13 0.16 0.19 0.22 0.30 0.35 0.40 0.45 0.50
616 V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 – 618

Table 3
Average percentage benefit of algorithms over dedicated solution

Problem characteristic Number of problem Average percentage benefit of algorithms

instances considered over dedicated solution

Average COV of in- COV of pick density SAA DPA

ventory of items of all items
Objective value Objective value

Low Low 20 9.07 8.86

High Low 20 29.25 26.67
Low High 20 6.60 6.41
High High 20 23.25 20.78

Table 4
Average CPU times of algorithms
Problem characteristic Number of problem Average CPU times of
instances considered algorithms (s)

Average COV of inventory of items COV of pick density of all items SAA DPA

Low Low 20 8.17 25.58

High Low 20 7.55 5.29
Low High 20 7.57 25.50
High High 20 5.91 5.17

Aisle width
= 2 ft

Aisles

Aisle width
= 4 ft Input/
Output
(4 ft X 4 ft)

Aisle width
= 4 ft

Fig. 1. Layout of the warehouse (industrial case).

classes, the space saving potential is higher with higher took slightly more time than DPA. Further, the CPU
COV of inventory. It can also be seen that SAA domi- time required by DPA is sensitive and is found to in-
nates over DPA in all the four groups but is relatively crease with average COV of inventory levels of items.
more when average COV of inventory of items is high.
It can be observed from Table 4 that when the av- 5.1. Application to an industrial problem
erage COV of inventory of items is low, SAA requires
considerably less CPU time than DPA for the problem A distribution warehouse, which stores food and
instances tested on the other hand when it is high, SAA beverage products, having 45 items is considered. The
 V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 – 618 617

Table 5
Product classes formed by DPA and SAA

Class number Product(s) assigned by algorithm

DPA SAA

1 P 42, P 44, P 8, P 39, P 1, P 40, P 21, P 3, P 25, P 13

P 41, P 30, P 18, P 43, P 6, P 7, P 34, P 35,
P 26, P 33, P 16, P 45, P 28, P 20, P 24,
P 31, P 29, P 23, P 11, P 10, P 4, P 15,
P 19, P 17, P 22, P 32, P 27, P 37, P 36,
P 5, P 14, P 9, P 38
2 P 12 P 2, P 12
3 P 25, P 13, P 2 P 42, P 44, P 8, P 39, P 1, P 40, P 21, P 3, P 41
4 P 30
5 P 18, P 43, P 6, P 7
6 P 34, P 35, P 26, P 33, P 16, P 45, P 28, P 20,
P 24, P 31, P 29, P 23, P 11, P 10, P 4, P 15,
P 19, P 17, P 22, P 32, P 27, P 37, P 36, P 5,
P 14, P 9, P 38

historical demand data for each item over an 11-month when the variability of inventory of items is high. Com-
period were collected. Inventory and total demand were putational time required by DPA is found to be more
in the range of 10–8640 units per month and 19–16 843 sensitive with problem data than SAA.
units, respectively. We assume the order-picking cost
per foot is $0.0025 and space cost per square foot for References
the planning horizon is $1.5. Based on firm’s policy, 13
stack levels are assumed. The layout of the warehouse [1] Li S, Ragu-Nathan B, Ragu-Nathan TS, Subba Rao S.
with dimensions 104 ft × 32 ft is considered and shown The impact of supply chain management practices on
competitive advantage and organizational performance. Omega
in Fig. 1. 2006;34(2):107–24.
The cost of allocation for dedicated storage is [2] Wen U-P, Chang D-T. Picking rules for a carousel conveyor in
$20533.9. In Table 5, we show the three product an automated warehouse. Omega 1988;16(2):145–51.
classes and five product classes formed by DPA and [3] Frazelle E. World-class warehousing and material handling.
New York, NY: McGraw-Hill; 2002.
SAA, respectively. The cost of allocation using DPA
[4] Hausman WH, Schwarz LB, Graves SC. Optimal storage
is $18544.79 and it took 83.89 s of CPU time. SAA assignment in automatic warehousing systems. Management
took 96 s of CPU time with $17949.52 as the cost of Science 1976;22(6):629–38.
allocation. SAA resulted in a better solution than DPA. [5] Petersen CG, Aase GR, Heiser DR. Improving order-picking
In this particular case, SAA has taken more CPU time performance through the implementation of class-based storage.
International Journal of Physical Distribution and Logistics
than DPA. Management 2004;34(7):534–44.
[6] Heskett JL. Cube-per-order index—a key to warehouse
6. Conclusion stock location. Transportation and Distribution Management
1963;4:27–31.
[7] van den Berg JP. Class-based storage allocation in a single
In this paper, a simulated annealing algorithm command warehouse with space requirement constraints.
(SAA) to form storage classes is developed consider- International Journal of Industrial Engineering 1996;3(1):21–8.
ing storage-space cost, order-picking cost, and product [8] Hassan MMD. Toward re-engineering models and algorithms
grouping without COI ordering restriction. Computa- of facility layout. Omega 2000;28(6):711–23.
[9] Francis RL, McGinnis LF, White JA. Facility layout and
tional experience on randomly generated data sets and location: an analytical approach. 2nd ed., Englewood Cliffs,
an industrial case shows that SAA on an average per- NJ, USA: Prentice-Hall; 1992.
formed better than dynamic programming algorithm [10] Muppani VR, Adil GK. Storage classes formation in the
(DPA) for class formation with COI ordering restric- presence of space cost for warehouse planning. International
tion. This paper demonstrated that forming classes Journal of Services Operations and Informatics 2006;1(3):
286–303.
based on COI ordering of items might not yield opti- [11] Chen M-C, Wu H-P. An association based clustering approach to
mal solution. Further, it was observed that class-based order batching considering customer demand patterns. Omega
solutions offer greater benefit over dedicated solution 2005;33(4):333–43.
618 V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 – 618

[12] Cormier G, Gunn EA. A Review of warehouse models. [22] Rosenblatt MJ, Eynan A. Deriving the optimal boundaries for
European Journal of Operational Research 1992;58:3–13. class-based automatic storage/retrieval systems. Management
[13] van den Berg JP. A literature survey on planning and Science 1989;35:1519–24.
control of warehousing systems. IIE Transactions 1999;31: [23] Eynan A, Rosenblatt MJ. Establishing zones in single-command
751–62. class-based rectangular AS/RS. IIE Transactions 1994;26(1):
[14] Rouwenhorst B, Reuter B, Stockrahm V, van Houtum GJ, 38–46.
Mantel RJ, Zijm WHM. Warehouse design and control: [24] Graves SC, Hausman WH, Schwarz LB. Storage-retrieval
framework and literatures review. European Journal of interleaving in automatic warehousing systems. Management
Operational Research 2000;122:515–33. Science 1977;23(9):935–45.
[15] Gu J, Goetschalckx M, McGinnis LF. Research on warehouse [25] Schwarz LB, Graves SC, Hausman WH. Scheduling policies
operation: a comprehensive review. European Journal of for automatic warehousing systems: simulation results. AIIE
Operational Research 2007;117(1):1–21. Transactions 1978;10(3):260–70.
[16] Harmatuck DJ. A comparison of two approaches to stock [26] Hackman ST, Rosenblatt MJ. Allocating items to an automated
location. Logistics and Transportation Review 1976;12(4): storage and retrieval system. IIE Transactions 1990;22(1):7–14.
282–5. [27] Goyal SK. A note on ‘multi-product inventory situations with
[17] Malmborg CJ, Bhaskaran K. On the optimality of the cube per one restriction’. Journal of the Operational Research Society
order index for conventional warehouses with dual command 1974;29:269–71.
cycles. Material Flow 1987;4:169–75. [28] Hariga MA, Jackson PL. The warehouse scheduling problem:
[18] Goetschalckx M, Ratliff HD. Shared storage policies based formulation and algorithms. IIE Transactions 1996;28(2):
on the duration stay of unit loads. Management Science 115–27.
1990;34(9):1120–32. [29] Malmborg CJ. Storage assignment policy tradeoff. International
[19] Malmborg CJ, Bhaskaran K. A revised proof of optimality Journal of Production Research 1996;34(2):363–78.
for cube-per-order index rule for stored item location. Applied [30] Chen J-F. A hybrid heuristic for the uncapacitated single
Mathematical Modelling 1990;14:87–95. allocation hub location problem. Omega 2007;35(2):211–20.
[20] Guenov M, Raeside R. Real time optimization of man on [31] Koulamas C, Antony SR, Jaen R. A survey of simulated
board order-picking. In: Proceedings of the 10th international annealing applications to operations research problems. Omega
conference on automation in warehousing; 1989. p. 89–94. 1994;22(1):41–56.
[21] Kanet JJ, Ramirez RG. Optimal stock picking decisions in [32] Laarhoven PJMV, Aarts EHL. Simulated annealing: theory and
automatic storage and retrieval systems. Omega 1986;14(3): applications. Norwell, MA, USA: Kluwer Academic Publishers;
239–44. 1987.