Anda di halaman 1dari 10

Omega 36 (2008) 609 618 www.elsevier.

com/locate/omega

Efcient formation of storage classes for warehouse storage location assignment: A simulated annealing approach
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. 2007 Elsevier Ltd. All rights reserved.
Keywords: Order picking; Storage-space cost; Class-based storage; Cube per order index; Simulated annealing

1. Introduction Organizations are increasingly adopting supply chain management practices in the hope of reducing supply chain costs as todays competition is moving from among organizations to between supply chains [1]. Effective supply chain management has become a potentially valuable way of securing competitive advantage and improving organizational performance. Thus, every activity and element in the supply chain is extremely important. Warehousing and inventory
This manuscript was processed by Guest Editors Angappa Gunasekaran and T.C. Edwin Cheng. Corresponding author. Tel.: +91 22 2576 7738; fax: +91 22 2572 2872. E-mail addresses: muppani@som.iitb.ac.in (V.R. Muppani (Muppant)), adil@iitb.ac.in, gajendra@som.iitb.ac.in (G.K. Adil). 2007 Elsevier Ltd. All rights reserved. 0305-0483/$ - see front matter doi:10.1016/j.omega.2007.01.006

management are two important and closely knit activities that can have an impact on supply chain costs. There is enormous pressure to simultaneously increase throughput rate and lower operating costs of warehouses. One means of improving throughput rate is by using automatic storage and retrieval systems [2]. There are many factors to be considered in order to keep operating costs to a minimum in a warehouse. Storage location assignment inuences almost all key warehouse performance indicators including order-picking time and cost, productivity, shipping and inventory accuracy, and storage density [3]. Hausman et al. [4] present three storage location assignment policies: randomized storage, dedicated storage, and class-based storage. The randomized storage policy allows items to be stored anywhere in the storage area. Under the dedicated storage policy each region

610

V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 618

may only be used for a specic item. The class-based storage policy distributes the items among a number of classes and for each class it reserves a region within the storage area. Randomized and dedicated storage are extreme cases of class-based storage policy: randomized storage considers a single class and dedicated storage considers one class for each item. If only order-picking cost is considered, dedicated policy may yield the lowest cost; on the other hand, if only space cost is considered, the completely random policy (one class) will yield the lowest cost solution. However, if we consider both the costs simultaneously, the best solution may be neither of these. The approaches reported in the literature for class formations primarily consider order-picking cost [5] even though class-based policies may offer reduction in cost through space savings in situations when space cost is signicant vis--vis picking cost. One criterion used to assign product classes to storage locations that captures item popularity and its storagespace requirement is cube-per-order index (COI) [6]. COI of an item is dened as the ratio of the items storage-space requirement (cube) to its popularity (number of storage/retrieval requests for the items). In the earlier works (e.g., [5,7]) items are ordered on the basis of their COI and then using this ordering, product classes are formed as shown in solutions 13 in Table 1. Items with lower COI values are put together in a class and located closer to I/O point than the items with higher COI values that were assigned to another class in order to reduce the order-picking cost. These restricted combinations in forming classes (i.e., class formation with COI ordering restriction) help overcoming the combinatorial nature of the problem by reducing the solution space [8]. However, this may not yield the least cost solution [9, pp. 284285] as illustrated in solution 4, Table 1. As a result of combining items into classes, storage area required reduces which translates into lower order-picking and storage-space costs. Grouping items having complementary planned inventory proles may require smaller region within the storage area than grouping based on COI values. However, for given classes, allocation ofstorage space based on COI of classes [10] minimizes order-picking cost as illustrated in solution 4, Table 1. In this paper, we develop a simulated annealing algorithm (SAA) to form storage classes efciently considering storage-space cost, order-picking cost, and class formation (part grouping) without any ordering restriction. The paper is organized as follows. In Section 2, pertinent literature is reviewed. We present problem formulation in Section 3 as in [10]. SAA is developed

in Section 4. In Section 5, we discuss computational experience when SAA is applied to randomly generated problem instances and an industrial case. Further, the solutions generated are compared with those of dynamic programming algorithm (DPA) for class formation with COI ordering restriction. The paper is concluded in Section 6. 2. Literature review There are four methods used to reduce pick travel distances namely (1) determining good order-picking routes, (2) zoning the warehouse, (3) assigning orders to batches, and (4) assigning products to correct storage locations [11]. In this section, we review papers pertinent to storage location assignment. A detailed discussion of warehouse design literature can be found in [1215]. Harmatuck [16] has shown that the COI-based item allocation minimizes the expected order-picking cost for single-address stock location. Malmborg and Bhaskaran [17] studied the optimality of the COI-based assignment policy for the case in which a single storage transaction and a single retrieval transaction are interleaved in order to economize on the travel of the order-picking vehicle. Items allocation based on COI values gives optimal allocation in terms of order picking/storing time under dedicated storage policy for single command transaction [18]. Malmborg and Bhaskaran [19] provided the revised proof of COI rule and a tie-breaking rule based on Euclidean distances. Hausman et al. [4] observed that the travel time might be measured by the Chebyshev metric. Guenov and Raeside [20], in their experiments, estimated bounds for an optimum tour with respect to Chebyshev travel. Kanet and Ramirez [21] analyze picking decisions and propose an integer programming model that allows the inclusion of xed costs per pick and breakdown costs that occur when the quantity of items stored in a location is greater than the ordered quantity and therefore, it must be broken down into multiple lots. Hausman et al. [4] consider the problem of nding class regions for the class-based storage policy using grid search to determine optimal boundaries for three classes. The authors show that a class-based storage policy with relatively few classes yields mean travel times that are close to those obtained by a dedicated storage policy for continuous racks and a specic demand function. They assumed same area for dedicated policy as well as class-based policy. This is extended to establish class boundaries for any given number of classes in a square-in-time rack [22] and to any rectangular

V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 618


Table 1 An example illustrating storage classes formation
Policy Class formation With COI ordering restrictiona Products Inventory level
i i

611

Space requiredc (in units) i + i 25.00 9.00 40.00 30.40 40.00 25.00 45.40 213.14 66.78

Average demand
d

Costs Order pickingd 37.5 29.5 162 229 60.79 151.19 211.98 37.50 190.79 228.29 159.85 57.784 217.64 Storagee 25.00 9.00 40.00 74.00 30.40 40.00 70.40 25.00 45.40 70.40 53.28 9.00 62.28 Total 62.50 38.50 202.00 303.00 91.19 191.19 282.38 62.50 236.19 298.69 213.14 66.78 279.92

Solution 1 (dedicated)

P1 P2 P3 P 1/P 2 P3 P1 P 2/P 3

5 5 20 10 20 5 25 20 20

20 4 20 20.4 20 20 20.40 1.5 4.03

15 5 15 Total 20 15 Total 15 20 Total 35 59 Total

Solution 2 (class based) Solution 3 (class based) Solution 4 (class based)


a COI

Without COI ordering restrictionb

P2 P 1/P 3

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. Orderpicking 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 regions are not necessarily optimal when dual commands occur, but they argue that the mean dual command cycle time in an L-shaped class allocation will in general be no more than 3% above the optimum. They observe that the space requirements increase with the number of classes. Schwarz et al. [25] verify the analytical results in [4,24] using simulation. The simulations suggest that most results of the analytic models hold under stochastic conditions. Hackman and Rosenblatt [26] present a model for forward/reserve problem that considers both assignment (which item) and allocation (what amounts) for forward and reserve locations. The emphasis in the above models has been in reducing the time for storing/order picking. However, oor space utilization is also important for warehouses. There are papers dealing with joint inventory planning and warehouse sizing (see for instance, [27,28]). Policies such as cross docking can also help in reducing the oor space requirements. For a given inventory policy, shared storage policy gives a better oor space utilization. Petersen et al. [5] study the improvement in orderpicking time as a result of forming classes (based on grouping parts with similar COI values) over a random

policy. This is possible as the reduction in total space requirements in random storage can also reduce average retrieval costs since it creates an opportunity to locate items in closer regions [29]. van den Berg [7] presents a DPA that distributes items and locations among classes such that the mean single command travel time is minimized considering reduction in area as a result of grouping items. The algorithm may be applied to a wide variety of warehousing systems such as warehouse layout, demand curve, and travel time metric. Applicability of this approach for large problems poses computational difculties. Muppani and Adil [10] developed a mathematical model to capture space and order-picking costs and to form storage classes with consideration to reduction in area when using class-based storage in a warehouse. They proposed a greedy heuristic procedure and DPA to solve the above model. They observe that there is a magnitude difference between the time taken by the greedy heuristic and DPA and also demonstrate that at times the greedy heuristic procedure provided better objective function value. As can be seen from the literature review, only few papers consider reduction in storage space (e.g., [7,10]) 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 in the literature for formation of classes are only on the basis of COI of items which may be sub-optimal as shown in Table 1. 3. Problem formulation We assume storages and retrievals are performed in single command cycles and all items are stored and transported on identical storage media (e.g., pallets or totes). The class-based storage policy is used. Each location is uniformly utilized and the assigned items are distributed homogeneously in the space allocated for the class. This assumption implies that the geometric center of the class is the same as the load center. We assume that inventory decision has been made independent of storage decision and all times required in the storage/retrieval process, except travel times, are independent of the storage allocation. We also assume that there is no congestion between vehicles/cranes and loads are not relocated. All the parameters of the models are deterministically known and the travel time between any two positions is readily computed. Once the storage locations are assigned they cannot be reshufed during the planning horizon and we assume there is a single input/output point. The problem can be dened as: given P products/items, their average demand and planned inventory levels for T periods and the layout of the storage area divided into lattice of storage locations, the problem is to establish classes of products and allocate them to storage locations so that the total cost of order picking and storage space is minimized in a single command warehouse exploiting the reduction in area because of clubbing the parts into classes. We establish notation and then present formulation of a model as in [10]. 3.1. Notation We use the following notation in this paper. Indices: c l p t (c = 1, 2, 3, . . . , C = P ) for classes (l = 1, 2, 3, . . . , L) for storage locations (p = 1, 2, 3, . . . , P ) for products/items (t = 1, 2, 3, . . . , T ) for time periods Parameters: al footprint area of location l in square feet dl distance of location l from the input/output point Dp total number of picks (in unit loads) for product p in the planning period

f fp

h t Ip

space cost in $ per square feet for the planning horizon footprint density (footprint area required to store one unit load of product p considering the stacking height) order-picking cost in $ per foot per unit load storage level in unit loads planned for product p during period t 1 if product p is assigned to class c, 0 otherwise 1 if location l is assigned to class c, ylc = 0 otherwise xpc =

Decision variables:

3.2. Class formation and allocation model (CFAM) Using the above notation and assumptions, the following mathematical model can be formulated as CFAM: Minimize z=f
c l l (al

(al ylc ) + 2h
c

dl ylc ) (al ylc ) l

Dp xpc
p

(1) subject to xpc =1


c

p,

(2) (al ylc ) c,


l

Max
t p

t (Ip fp xpc )

(3) ylc
c

(4) (5)

xpc , ylc {0, 1} p, c, l.

The objective function (1) minimizes the sum of storage space cost and order-picking cost over the planning horizon. The storage-space cost for class c is f l (al ylc ). One way order-picking cost for class c is product of unit order-picking cost h, centroid distance of the allocated space from the input/output point l (al dl ylc )/ l (al ylc ) and total number of picks p Dp xpc . Constraint (2) ensures that each item is assigned to one and only one class. Constraint (3) ensures that there is adequate storage space to hold the items in a class c in each planning period t. A storage location, if at all assigned, is to be assigned to only one class. This is taken care of by constraint (4). Constraint (5) imposes binary restrictions on decision variables.

V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 618

613

4. Solution approach The CFAM involves nonlinear objective function and integer variables posing difculties in optimally solving large problems. Simulated annealing (SA) is a metaheuristic for solving such combinatorial optimization problems [30]. Koulamas et al. [31] provided a survey on applications of simulated annealing in layout problems. We develop an SAA to solve CFAM. Solution to CFAM comprises of class formation (xpc variables) and space allocation to the classes (ylc variables). If classes are given, i.e., the value of xpc variables xed (at say Xpc ), the storage locations, ylc , can be optimally determined by solving the following allocation model (AM). Allocation model (AM): Minimize subject to al ylc ylc {0, 1} where Bc = Max
t p p Dp t (Ip fp Xpc ) l

The proposed SAA exploits the structure of the CFAM and it searches in space, xpc , while determining values of ylc in evaluation step by solving AM as described before. Thus, solution (SOL) to the CFAM consists of class formation, CF and allocation of space to classes formed, AS. 4.1.1. Storage classes in initial solution To obtain initial classes, CF0 , create C (=P ) classes having one item each with item 1 in class 1, item 2 in class 2 and so on. Then, compute COI for each class c (COIc ) using expression (9) and arrange the classes in increasing order of COIc . Solve the AM for CF0 to obtain space allocation, AS0 , and objective value Z0 . CF0 and AS0 together form initial solution, SOL0 . 4.1.2. Generation of storage classes in neighborhood solution Generation of neighborhood classes and allocation to storage spaces are described below. Neighborhood storage classes: Randomly select two classes, k and l such that l = k from the current solution, CFi . Randomly select a part, pi from within the parts in the class k and move the part pi from class k to class l to get CFi+1 . Allocation of storage space: Solve the AM for CFi+1 to obtain space allocation and objective value Zi+1 to get solution SOLi+1 . 4.1.3. Simulated annealing algorithm (SAA) SA initialization: Read input data: warehouse layout information, number of products (P), number of periods (T), demand (Dp ), foot print density (fp ) and storage t levels (Ip ) for all the items. Dene the annealing parameters: initial temperature, T0 ; maximum transitions at each temperature, Lmax ; maximum accepted transition at each temperature, ATmax ; temperature decrementing factor, ; maximum number of iterations, imax ; and nal acceptance ratio, Rf . SA Loop: Step 0: Initialize iteration counter: i = 0, temperature Ti = T0 . Generate an initial solution as in Section 4.1.1. Let the best solution be represented by SOLbest and the best objective value by Zbest . Assign initial operation assignment SOL0 to SOLbest and initial objective value Z0 to Zbest .

z2 =
c l

Alc .ylc ,

(6)

Bc

c,

(7) (8)

c, l,

and Xpc

Alc = f al + 2h(al dl )

Bc

It can be proved that the optimum solution for AM can be obtained using the following three steps. The classes are ranked in the increasing order of their COI (COIc ), which is computed using the following equation as in [10]. COIc = = Area allocated to class c Demand rate of products in class c Maxt
t p (Ip

fp xpc )

p (Dp /T ) xpc

(9)

The storage locations are ranked in increasing order of their rectilinear distances from I/O point. If there is a clash, it is resolved by ranking based on Euclidean distance of the locations from I/O point [19]. The classes are allocated required storage space by picking storage locations in the ranked order of storage locations. 4.1. Simulated annealing In this section, we describe procedure to develop an initial solution, procedure to generate a neighborhood solution given any solution and SAA.

614

V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 618

Step 1: Execute outer loop, i.e., steps (1.11.7) until conditions in step 1.7 are met. 1.1. Initialize inner loop counter l=0, and accepted number of transitions AT = 0. 1.2. Initialize solution for inner loop, SOL0 =SOLi , Zi0 = i Zi . 1.3. Execute inner loop, i.e., steps (1.3.11.3.4) until conditions in step 1.3.4 are met. 1.3.1. Set l = l + 1. Generate a neighborhood solution as in Section 4.1.2 (get SOLl , Zil ). i 1.3.2. = Zil Zil1 . 1.3.3. If 0 or random (0, 1) e /Ti then l and Zl are accepted. SOLi i Update accepted transition AT to AT + 1. Else Solution is rejected, SOLl = SOLl1 , Zil = i i Zil1 . 1.3.4. If one of the following conditions holds true: AT ATmax or l Lmax , then assign l to Li (length of Markov chain), terminate the inner loop and go to 1.4; else continue the inner loop and go to 1.3.1. 1.4. Set: i = i + 1. l1 1.5. Update SOLi = SOLl1 , Zi = Zi1 . If Zi < Zbest i1 then SOLbest = SOLi and Zbest = Zi . 1.6. Reduce the cooling temperature: Ti = Ti1 . 1.7. If one of the following conditions holds true: i imax , or the acceptance ratio (dened as AT/Li ) Rf , then terminate the outer loop and go to 2, else continue the outer loop and go to 1.1. Step 2: Print the best solution obtained and terminate the procedure. 4.2. SA parameter setting To determine good annealing parameters for a given problem is often a hard task and needs experimentation [32]. The following values for the SA parameters were selected after experimentation: =0.98, T0 =7500.0, imax = 2000. Lmax =200, ATmax =100,

5. Computational experience In order to assess the quality of solution generated by SAA, a rectangular storage area of 60 48 ft 2 divided into unit cells of 1 ft 1 ft and I/O point located on the periphery at the center of the width is assumed.

Problem instances each with 40 items were generated considering three key factors that inuence order-picking and space costs. These are pick density of products measured as reciprocal of COI [3], inventory and the relative cost of order picking and storage space. Five inventory patterns with low variability and another ve inventory patterns with high variability having coefcient of variation (COV) of inventory averaged for all items varying in the range of 4.9745.30 and 110.77293.21, respectively, were used. Similarly, ve pick density patterns with low variability and another ve pick density patterns with high variability having COV of pick density of items varying in the range of 0.100.22 and 0.300.50, respectively, were used. Twenty pick density-inventory combinations: LowLow, LowHigh, HighLow, and HighHigh are created, i.e., ve data points of pick density with low variability combining the ve data points of inventory with low variability yielding ve combinations. The storage space allocated corresponds to one standard deviation above mean planned inventory level. Eighty problem instances were generated by combining four levels of space cost per square foot (f) for the duration of the planning period: $0, $0.42, $0.8 and $1.1 with the above 20 pick density-inventory combinations. We assume order-picking cost per foot (h) to be $0.0025 and T to be 7 periods for all the problem instances. We compared the SAA with a benchmark DPA for class formation with COI ordering restriction [10]. Percentage benet of class-based solution obtained by DPA and SAA over the dedicated solution for the 80 problem instances described above is shown against COV of inventory and COV of pick density of all items in Table 2. It can be observed that SAA performs better than DPA in 64 of 80 problem instances tested with the difference in objective values as high as 7.34%. A paired t-test rejected the equality of the DPA and SAA solutions (t-statistic = 7.615, p = 0.000 at 95% condence level). Thus, SAA performs better than DPA as conjectured and illustrated in Table 1 for the SAA parameters stated in Section 4.2 for the problem instances tested. In order to understand the inuence of inventory and pick density characteristics on the performance of class-based solutions, the problem instances are grouped based on COV of inventory and COV of pick density as shown in Table 3. Average percentage benet of SAA and DPA over dedicated solution for all the four groups is shown. It can be observed that when the average COV of inventory of items is high, class-based solutions offer greater benet over dedicated solution. This is due to the fact that, when items are grouped into

V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 618


Table 2 Benet of class based (SAA and DPA) over dedicated solution
Problem instances 1 f =0 %Benet over dedicated SAA DPA Average COV of inventory COV of pick density of all SKUs 2 3 4 5 6 7 8 9 10

615

1.83 1.87 4.97 0.10 11

4.49 4.44 14.90 0.13 12 22.73 17.38 151.17 0.13 2 5.82 6.09 14.90 0.13 12 26.79 23.73 151.17 0.13 2 6.79 6.99 14.90 0.13 12 28.92 26.99 151.17 0.13 2 7.32 7.44 14.90 0.13 12 29.86 28.63 151.17 0.13

6.74 5.94 24.81 0.16 13 27.06 25.05 190.00 0.16 3 10.06 9.79 24.81 0.16 13 34.51 34.44 190.00 0.16 3 11.37 11.05 24.81 0.16 13 38.54 38.54 190.00 0.16 3 12.05 11.81 24.81 0.16 13 40.53 40.53 190.00 0.16

8.69 7.92 34.61 0.19 14 24.68 18.71 257.46 0.19 4 11.90 11.63 34.61 0.19 14 28.23 23.94 257.46 0.19 4 13.07 13.28 34.61 0.19 14 30.24 27.16 257.46 0.19 4 14.27 14.09 34.61 0.19 14 31.20 29.05 257.46 0.19

10.59 9.52 45.30 0.22 15 26.06 21.10 293.21 0.22 5 14.55 14.10 45.30 0.22 15 30.92 27.51 293.21 0.22 5 16.38 16.05 45.30 0.22 15 33.26 31.28 293.21 0.22 5 17.27 16.99 45.30 0.22 15 34.51 33.34 293.21 0.22

1.41 1.56 4.97 0.30 16 17.49 14.18 110.77 0.30 6 1.68 1.73 4.97 0.30 16 19.42 16.70 110.77 0.30 6 1.81 1.86 4.97 0.30 16 20.90 18.55 110.77 0.30 6 1.94 1.96 4.97 0.30 16 21.76 19.71 110.77 0.30

3.94 3.95 14.90 0.35 17 19.20 15.49 151.17 0.35 7 4.39 4.46 14.90 0.35 17 21.03 17.91 151.17 0.35 7 4.55 4.85 14.90 0.35 17 22.20 19.80 151.17 0.35 7 5.07 5.12 14.90 0.35 17 23.50 21.01 151.17 0.35

5.71 5.28 24.81 0.40 18 23.94 22.93 190.00 0.40 8 6.66 6.33 24.81 0.40 18 27.01 26.08 190.00 0.40 8 7.32 7.07 24.81 0.40 18 28.65 28.45 190.00 0.40 8 7.84 7.58 24.81 0.40 18 28.78 29.47 190.00 0.40

7.60 7.13 34.61 0.45 19 21.17 16.96 257.46 0.45 9 8.70 8.35 34.61 0.45 19 22.75 19.01 257.46 0.45 9 9.57 9.23 34.61 0.45 19 23.57 20.48 257.46 0.45 9 10.10 9.89 34.61 0.45 19 24.64 21.64 257.46 0.45

9.20 8.60 45.30 0.50 20 22.61 19.00 293.21 0.50 10 10.52 10.07 45.30 0.50 20 24.14 21.19 293.21 0.50 10 11.67 11.22 45.30 0.50 20 25.57 22.87 293.21 0.50 10 12.40 11.99 45.30 0.50 20 26.56 24.14 293.21 0.50

%Benet over dedicated SAA DPA Average COV of inventory COV of pick density of all SKUs f = 0.42 %Benet over dedicated SAA DPA Average COV of inventory COV of pick density of all SKUs

20.89 16.05 110.77 0.10 1 2.47 2.48 4.97 0.10 11

%Benet over dedicated SAA DPA Average COV of inventory COV of pick density of all SKUs f = 0.8 %Benet over dedicated SAA DPA Average COV of inventory COV of pick density of all SKUs

23.92 20.98 110.77 0.10 1 2.75 2.75 4.97 0.10 11

%Benet over dedicated SAA DPA Average COV of inventory COV of pick density of all SKUs f = 1.1 %Benet over dedicated SAA DPA Average COV of inventory COV of pick density of all SKUs

25.58 23.75 110.77 0.10 1 2.90 2.90 4.97 0.10 11

%Benet over dedicated SAA DPA Average COV of inventory COV of pick density of all SKUs

26.55 25.23 110.77 0.10

616

V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 618

Table 3 Average percentage benet of algorithms over dedicated solution


Problem characteristic Average COV of inventory of items COV of pick density of all items Number of problem instances considered Average percentage benet of algorithms over dedicated solution SAA Objective value Low High Low High Low Low High High 20 20 20 20 9.07 29.25 6.60 23.25 DPA Objective value 8.86 26.67 6.41 20.78

Table 4 Average CPU times of algorithms


Problem characteristic Number of problem instances considered COV of pick density of all items Low Low High High 20 20 20 20 Average CPU times of algorithms (s) SAA 8.17 7.55 7.57 5.91 DPA 25.58 5.29 25.50 5.17

Average COV of inventory of items Low High Low High

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 COV of inventory. It can also be seen that SAA dominates over DPA in all the four groups but is relatively more when average COV of inventory of items is high. It can be observed from Table 4 that when the average COV of inventory of items is low, SAA requires considerably less CPU time than DPA for the problem instances tested on the other hand when it is high, SAA

took slightly more time than DPA. Further, the CPU time required by DPA is sensitive and is found to increase with average COV of inventory levels of items. 5.1. Application to an industrial problem A distribution warehouse, which stores food and beverage products, having 45 items is considered. The

V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 618


Table 5 Product classes formed by DPA and SAA
Class number Product(s) assigned by algorithm DPA 1 P 42, P 44, P 8, P 39, P 1, P 40, P 21, P 3, 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 P 12 P 25, P 13, P 2 SAA P 25, P 13

617

2 3 4 5 6

P 2, P 12 P 42, P 44, P 8, P 39, P 1, P 40, P 21, P 3, 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

historical demand data for each item over an 11-month period were collected. Inventory and total demand were in the range of 108640 units per month and 1916 843 units, respectively. We assume the order-picking cost per foot is $0.0025 and space cost per square foot for the planning horizon is $1.5. Based on rms policy, 13 stack levels are assumed. The layout of the warehouse with dimensions 104 ft 32 ft is considered and shown in Fig. 1. The cost of allocation for dedicated storage is $20533.9. In Table 5, we show the three product classes and ve product classes formed by DPA and SAA, respectively. The cost of allocation using DPA is $18544.79 and it took 83.89 s of CPU time. SAA took 96 s of CPU time with $17949.52 as the cost of allocation. SAA resulted in a better solution than DPA. In this particular case, SAA has taken more CPU time than DPA. 6. Conclusion In this paper, a simulated annealing algorithm (SAA) to form storage classes is developed considering storage-space cost, order-picking cost, and product grouping without COI ordering restriction. Computational experience on randomly generated data sets and an industrial case shows that SAA on an average performed better than dynamic programming algorithm (DPA) for class formation with COI ordering restriction. This paper demonstrated that forming classes based on COI ordering of items might not yield optimal solution. Further, it was observed that class-based solutions offer greater benet over dedicated solution

when the variability of inventory of items is high. Computational time required by DPA is found to be more sensitive with problem data than SAA. References
[1] Li S, Ragu-Nathan B, Ragu-Nathan TS, Subba Rao S. The impact of supply chain management practices on competitive advantage and organizational performance. Omega 2006;34(2):10724. [2] Wen U-P, Chang D-T. Picking rules for a carousel conveyor in an automated warehouse. Omega 1988;16(2):14551. [3] Frazelle E. World-class warehousing and material handling. New York, NY: McGraw-Hill; 2002. [4] Hausman WH, Schwarz LB, Graves SC. Optimal storage assignment in automatic warehousing systems. Management Science 1976;22(6):62938. [5] Petersen CG, Aase GR, Heiser DR. Improving order-picking performance through the implementation of class-based storage. International Journal of Physical Distribution and Logistics Management 2004;34(7):53444. [6] Heskett JL. Cube-per-order indexa key to warehouse stock location. Transportation and Distribution Management 1963;4:2731. [7] van den Berg JP. Class-based storage allocation in a single command warehouse with space requirement constraints. International Journal of Industrial Engineering 1996;3(1):218. [8] Hassan MMD. Toward re-engineering models and algorithms of facility layout. Omega 2000;28(6):71123. [9] Francis RL, McGinnis LF, White JA. Facility layout and location: an analytical approach. 2nd ed., Englewood Cliffs, NJ, USA: Prentice-Hall; 1992. [10] Muppani VR, Adil GK. Storage classes formation in the presence of space cost for warehouse planning. International Journal of Services Operations and Informatics 2006;1(3): 286303. [11] Chen M-C, Wu H-P. An association based clustering approach to order batching considering customer demand patterns. Omega 2005;33(4):33343.

618

V.R. Muppani (Muppant), G.K. Adil / Omega 36 (2008) 609 618 [22] Rosenblatt MJ, Eynan A. Deriving the optimal boundaries for class-based automatic storage/retrieval systems. Management Science 1989;35:151924. [23] Eynan A, Rosenblatt MJ. Establishing zones in single-command class-based rectangular AS/RS. IIE Transactions 1994;26(1): 3846. [24] Graves SC, Hausman WH, Schwarz LB. Storage-retrieval interleaving in automatic warehousing systems. Management Science 1977;23(9):93545. [25] Schwarz LB, Graves SC, Hausman WH. Scheduling policies for automatic warehousing systems: simulation results. AIIE Transactions 1978;10(3):26070. [26] Hackman ST, Rosenblatt MJ. Allocating items to an automated storage and retrieval system. IIE Transactions 1990;22(1):714. [27] Goyal SK. A note on multi-product inventory situations with one restriction. Journal of the Operational Research Society 1974;29:26971. [28] Hariga MA, Jackson PL. The warehouse scheduling problem: formulation and algorithms. IIE Transactions 1996;28(2): 11527. [29] Malmborg CJ. Storage assignment policy tradeoff. International Journal of Production Research 1996;34(2):36378. [30] Chen J-F. A hybrid heuristic for the uncapacitated single allocation hub location problem. Omega 2007;35(2):21120. [31] Koulamas C, Antony SR, Jaen R. A survey of simulated annealing applications to operations research problems. Omega 1994;22(1):4156. [32] Laarhoven PJMV, Aarts EHL. Simulated annealing: theory and applications. Norwell, MA, USA: Kluwer Academic Publishers; 1987.

[12] Cormier G, Gunn EA. A Review of warehouse models. European Journal of Operational Research 1992;58:313. [13] van den Berg JP. A literature survey on planning and control of warehousing systems. IIE Transactions 1999;31: 75162. [14] Rouwenhorst B, Reuter B, Stockrahm V, van Houtum GJ, Mantel RJ, Zijm WHM. Warehouse design and control: framework and literatures review. European Journal of Operational Research 2000;122:51533. [15] Gu J, Goetschalckx M, McGinnis LF. Research on warehouse operation: a comprehensive review. European Journal of Operational Research 2007;117(1):121. [16] Harmatuck DJ. A comparison of two approaches to stock location. Logistics and Transportation Review 1976;12(4): 2825. [17] Malmborg CJ, Bhaskaran K. On the optimality of the cube per order index for conventional warehouses with dual command cycles. Material Flow 1987;4:16975. [18] Goetschalckx M, Ratliff HD. Shared storage policies based on the duration stay of unit loads. Management Science 1990;34(9):112032. [19] Malmborg CJ, Bhaskaran K. A revised proof of optimality for cube-per-order index rule for stored item location. Applied Mathematical Modelling 1990;14:8795. [20] Guenov M, Raeside R. Real time optimization of man on board order-picking. In: Proceedings of the 10th international conference on automation in warehousing; 1989. p. 8994. [21] Kanet JJ, Ramirez RG. Optimal stock picking decisions in automatic storage and retrieval systems. Omega 1986;14(3): 23944.

Anda mungkin juga menyukai