Expert Systems with Applications 38 (2011) 2686–2700

Contents lists available at ScienceDirect

Expert Systems with Applications

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

Improving the productivity of order picking of a manual-pick and multi-level

rack distribution warehouse through the implementation of class-based storage
Felix T.S. Chan a, H.K. Chan b,⇑
a
Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hom, Hong Kong
b
Norwich Business School, University of East Anglia, Norwich, Norfolk, UK

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

Keywords: Determining the proper storage locations for potentially thousands of products is one major task that
Simulation faces the warehouse manager during the design of a new warehouse or the refurbishment of existing
Class-based storage facility. There are many factors affects the storage assignment. For example, order picking method, size
Order picking and layout of the storage system, material handling system, product characteristics, demand trends, turn-
Warehouse management
over rates and space requirements. Selecting appropriate storage assignment policies (i.e. random, ded-
icated or class-based) and routing methods (i.e. transversal, return or combined) with regards to above
factors is a possible solution to tackle this problem. However, the efficiency of these policies is interde-
pendent. This paper aims to present a simulation study of a real case regarding storage assignment prob-
lem of a manual-pick and multi-level rack warehouse. Performance of the case is measured in terms of
travel distance and order retrieval time. The results of the study indicate that the key to effective imple-
mentation of a storage assignment system is to match the types of warehouse storage system and the
variety of items in the customer order. Also, the use of key performance indicator should clearly reflect
the needs of the warehouse.
Ó 2010 Elsevier Ltd. All rights reserved.

1. Introduction ager during the design of a new warehouse or the refurbishment of

existing facility. There are many factors affects the storage assign-
Order picking, defined as the process of retrieving the required ment. For example, order picking method, size and layout of the
goods from storage according to the customer orders, has come un- storage system, material handling system, product characteristics,
der increase consideration as it has long been identified as the demand trends, turnover rates and space requirements. It is extre-
most labor-intensive and costly activity for almost every ware- mely difficult for a human to consider all the factors simulta-
house (De Koster, Le-Duc, & Roodbergen, 2006). In order to mini- neously and consistently and select the best location. Selecting
mize the throughput time of picking an order while maximize appropriate storage assignment policies (i.e. random, dedicated
the use of space, equipment and labor as well as the accessibility or class-based) and routing methods (i.e. transversal, return or
to all items, careful design and control of the order picking system combined) with regards to above factors is a possible solution to
is necessary. There are four typical tactical and operational deci- tackle this problem. However, the efficiency of these policies is
sion problems faced by the warehouse managers. They are (1) lay- interdependent. In different storage system (i.e. single or multi-le-
out design, (2) picking policies, (3) storage assignment policies and vel racks) and different levels of pick density, different combina-
(4) routing policies. Layout design concerns both the layout of the tion would results in different performance of picking. Therefore,
facility containing the order-picking system and the layout within simulation is the most appropriate tool to analyze the performance
the order-picking system; picking policies determine how orders of picking for each combination under different pick densities.
are to be grouped into picking tours; storage assignment policies This paper aims to improve the performance of storage assign-
find out where items are to be stored in the warehouse; and rout- ment problem of a manual-pick, multi-level rack warehouse of
ing policies indicate the sequence of items to be picked during any picking in terms of travel distance and order retrieval time through
tour. a simulation study. Data are obtained from a real case. The rest of
Determining the proper storage locations for potentially thou- the paper is organized as follows: Section 2 reviews relevant liter-
sands of products is one major task that faces the warehouse man- ature. Section 3 presents the background and relevant information
of the case. Section 4 explains the simulation study. Results and
⇑ Corresponding author. key findings are presented in Section 5. Section 6 concludes this
E-mail address: [email protected] (H.K. Chan). paper.

0957-4174/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2010.08.058
 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700 2687

2. Literature review dom, dedicated and class-based storage assignment policies. The
simulation results showed that dedicated storage helps to maxi-
Storage assignment is the most important factor affect the per- mize the system throughput while the random storage helps to
formance of picking. Many factors influence the location of prod- maximize the utilization of storage space. On the other hand,
ucts in the warehouse. For example, order picking method, size Ashayeri et al. (2002) presented a geometry-based analytical mod-
and layout of the storage system, material handling system, prod- el to compute the system cycle time in three class-based storage
uct characteristics, demand trends, turnover rates and space layouts as shown in Fig. 1. The results showed that the third stor-
requirements. Selecting appropriate storage assignment policy age layout has the shortest system process time while the second
helps to improve the performance of picking. layout has the longest system process time.
The storage assignment policies which assign products to stor- For the class-based storage assignment based on inventory
age locations are generally fall into three broad categories: (1) ran- turnover, Montulet et al. (1998) studied the problem of minimizing
dom storage, (2) dedicated storage and (3) class-based storage the peak load with single command cycles and dedicated storage
(Hausman, Schwarz, & Graves, 1976). Random and dedicated stor- policy. The peak load is the maximum value of the daily loads over
age are in fact extreme cases of the class-based storage policy. Ran- a fixed planning horizon. The daily load is the expected total time
dom storage has all products in a single class and dedicated storage of the command cycles. A mathematical model is formulated and a
has each of the products assigned to a separate class. The main idea branch-and-bound algorithm is developed to solve the problem.
of class-based storage is to divide products into classes. Each class Hsieh and Tsai (2001) presented a bill of material (BOM) oriented
is then assigned to a dedicated area of the warehouse. Storage class-based storage assignment method for AS/RS. This method
within an area is random. The advantage of this policy is that fast possesses not only the advantage of a class-based storage method,
moving products can be stored close to the depot while the flexibil- but also the feasibility to integrate an AS/RS into a Computer Inte-
ity and high storage space utilization of random storage are appli- grated Manufacturing (CIM) system. However, this type of assign-
cable. Generally, there are two kinds of class-based storage, (1) ment would only be used in the warehouses stored raw materials
dedicated purposes (e.g. Brynzér & Johansson, 1996; Liu, 1999); rather than finished goods.
and (2) ABC classification. (e.g. Ashayeri, Heuts, Valkenburg, Vera- For all types of order-picking system, an appropriate inventory
art, & Wilhelm, 2002; Montulet, Langevin, & Riopel, 1998). classification method is of vital importance to divided products
On the other hand, storage assignment in picker-to-parts system into classes that increase the efficiency of picking. In this regard,
is mainly divided into two aspects, the (1) products and (2) customer Guvenir and Erel (1998) and Partovi and Anandarajan (2002) pro-
orders. For products, Brynzér and Johansson (1996) pointed out that posed to classify inventory into classes based on multiple attri-
class-based storage according to product characteristics can in- butes. Guvenir and Erel (1998) compared the traditional ABC
crease the accuracy of picking and reduce order retrieval time. They analysis, Analytic Hierarchy Process (AHP) and Generic Algorithm
emphasized that people should consider order retrieval time and (GA) for their effectiveness on classifying the inventory. He criti-
travel distance as two different performance measures for order cized that the traditional ABC analysis only put the annual usage
picking. However, their research did not further investigate into value into consideration while AHP is too subjective that may
the relationship between order retrieval time and distance trav- causes errors. He found that GA can classify inventory that shows
elled. On the other hand, Vickson and Lu (1998) proved that the consistency to what decision maker desire. Partovi and Anandara-
class-based storage according to turnover of products can enhance jan (2002), on the other hand, proposed to classify the inventory of
the accessibility of fast moving items to the pickers. They used a pharmaceutical company by using Artificial Neural Network
Cube-per-Order Index (COI) analysis which is proposed by Heskeet (ANN). ANN is a mathematical model or computational model
(1963), Heskett (1964). The COI of a product is defined as the ratio of based on biological neural networks. They can be used to model
the product’s total required space to the number of trips required to complex relationships between inputs and outputs or to find pat-
satisfy its demand per period. The algorithm consists of locating the terns in data. He proved that ANN works better than AHP.As a
products with the lowest COI (i.e. highest sales rates or forecasted whole, there are some drawbacks of the inventory classification
demands) closest to the depot. As a result, the order retrieval time methods mentioned above. For example, number of input variables
is decreased. However, this research is not completely applicable is limited and they have difficulties in analyzing qualitative attri-
to multi-level rack warehouses where picking the products at higher butes. This is why professional judgment is not completely re-
level would results in longer time. placed by the use of the inventory classification methods.
For customer orders, Liu (1999) mentioned that studying the sta- To conclude, the usage of storage assignment policies depends on
tistical correlation between items can help to improve the efficiency the needs of the warehouse operators as well as some environmen-
of order picking. The items that are likely appeared at the same order tal constraints such as the size and layout of the storage system and
would be assigned to adjacent storage locations. The number of stor- the material handling system as mentioned. The literature review
age and retrieval operations can be reduced. However, there are for storage assignment is summarized in Table 1. It is found that
some drawbacks of this methodology. It requires the stability of or- most studies consider single-level rack warehouse, or focus on AS/
der contents, adhesiveness to particular storage assignment policy
and fixed sequence of items printed on the picking lists. Apart from
COI discussed above, another indicator, namely Entry-Item-Quan-
tity (EIQ) analysis (Li, 2009), is proved to be more superior in inven-
tory classification than other data mining techniques that results in
more reduction in travel distance.
In fact, most research on class-based storage has been per-
formed in the context of Automated Storage/Retrieval System
(AS/RSA). Items are divided into classes according to three aspects:
(1) system cycle time, (2) inventory turnover and (3) Bill of Mate-
rials (BOM). Ashayeri et al. (2002) and Mansuri (1997) researched
on the class-based storage according to system cycle time in AS/RS
and used the system throughput and process time for performance
measurement. Mansuri (1997) compared the effectiveness of ran- Fig. 1. Three class-based storage layout (Ashayeri et al., 2002).
2688 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700

Table 1
Summary for research on storage assignment.

Author (Year) Storage assignment policy Performance indicator Order picking system
Brynzér and Johansson (1996) Class-based storage: product characteristics Order retrieval time Picker-to-parts
system
Mansuri (1997) Class-based storage: inventory turnover and system process time System throughput AS/RS
Daniels et al. (1998) Random storage: multiple storage locations for single item Warehouse costs Picker-to-parts
system
Vickson and Lu (1998) Class-based storage: inventory turnover System cycle time Picker-to-parts
system
Guvenir and Erel (1998) Class-based storage: multiple attributes Inventory management and
control
Montulet et al. (1998) Class-based storage: system cycle time System throughput AS/RS
Liu (1999) Class-based storage: statistical correlation between items in the same Travel distance Picker-to-parts
order system
Hsieh and Tsai (2001) Class-based storage: Bill of Materials Order retrieval time AS/RS
Ashayeri et al. (2002) Class-based storage: system process time Process time AS/RS
Partovi and Anandarajan Class-based storage: multiple attributes Inventory management and
(2002) control

RS. Few consider multi-level rack warehouse. Also, most of the stud- Table 2
ies focus on random and class-based storage. As summarized from Summary for research on configuration of classes and aisles.
warehousing literatures, random storage is widely used in ware- Author (Year) Experiment Experimental variables Performance
houses because it is simple to use. Every incoming pallet is assigned factor indicator
a location in the warehouse that is selected randomly from all eligi- Roodbergen Increment of Size of warehouse, Order
ble empty locations with equal probability (Petersen, 1997). It re- and De cross aisles number of orders and retrieval
sults in high space utilization at the expense of increased travel Koster width of aisles time
(2001)
distance (Choe and Sharp, 1991). It is commonly used in bulk storage
Jewkes et al. Configuration Assignment of products, Order
areas with a computerized inventory system, since present stock (2004) of classes position of order pickers retrieval
levels and locations need to be updated in real time. time
Configuration of classes is also an important research direction. Le-Duc and De Configuration Positioning of classes, Order
Jane and Laih (2005), Jewkes, Lee, and Vickson (2004), Le-Duc and Koster of classes routing policies retrieval
(2005) time
De Koster (2005) together with Heragu, Du, Mantel, and Schuur
Jane and Laih Configuration Product assignment, Order
(2005) explored the positioning, workload and size of each class. (2005) of classes workload of each class retrieval
Various possibilities exist for positioning of classes in the picker- time
to-parts system. There are two common ways which are called Heragu et al. Configuration Product assignment, Warehouse
(2005) of classes workload of each class costs
within-aisle storage and across-aisles storage. Based on a closed
form travel-time estimate for the return routing policy, Le-Duc
and De Koster (2005) optimize the storage-class positioning. They (number of classes, percentage of items per class, and percentage
claim that the across-aisle storage method is close to optimal. of the total pick volume per class) for picker-to-part systems. Also,
Alternatively, Jewkes et al. (2004) investigated the positioning of the optimal assignment depends on the routing policies, size of the
classes as well as the order pickers so as to minimize the order re- warehouse and number of SKUs per pick route. This leads to the
trieval time. However, the order retrieval time spent on vertical discussion of the next topic, how combination of different factors
movements is not considered. (storage assignment policies, routing policies and pick densities)
Jane and Laih (2005) criticized that class-based storage may affect the overall performance of order picking.
cause workload imbalance. By developing a clustering model, they Based on the above, it can be concluded that most research fo-
grouped items into discrete exclusive clusters. Items within a clus- cused on efficiency of picking in single-level rack warehouses or
ter will usually have similar demand patterns and handling meth- AS/RS systems. This paper attempts to solve the storage assign-
ods. The whole clustering process is based on the aggregate ment problem in a multi-level rack warehouse in a case study.
product demand rather than individual item demand. It results in Not only the order retrieval time spent on horizontal travels is con-
uniform demand patterns across all the clusters and consequently
cerned, but also the time spent on vertical movements. Relation-
balances the workload and minimizes the overall order retrieval ship between the order retrieval time and travel distance is
time. Moreover, Heragu et al. (2005) proposed a mathematical mod-
studied so as to find out the optimal storage assignment. This
el to investigate how the assignment of products and size of classes draws the major difference from the previous research. In addition,
affect the efficiency of picking. Three functional classes, namely re-
this paper also compares the performance of picking under three
served area, forward area and cross-docking area are defined. Based routing policies (i.e. transversal, return and combination). Simula-
on studying the physical flow of products, optimum size of each
tion models are built according to respective combination of poli-
functional class is determined and the cost of picking is reduced. cies as well as the basic parameters of the case company (e.g. Wu,
More studies concerning this area are summarized in Table 2.
Chen, Tsai, & Yang, 2010; Yoo, Cho, & Yücesan, 2010).
Undoubtedly, the efficiency of picking is closely related to the con-
figuration of classes and aisles. For example, the size of the classes
would affect the balance of workload. In real situation, the size of 3. Problem formulation
the warehouse is fixed. To decrease the number of aisles would in-
crease the size of each class or storage space. However, the impli- 3.1. Background of the case
cations on efficiency of picking should be evaluated. For example, it
may cause congestion within the aisles. To conclude, in the ware- In this study, a distribution warehouse for athletic products
housing literature, there is no firm rule to define a class partition (called ABC warehouse) is studied as a case to illustrate ways to
 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700 2689

Table 3
Pick process time tracking summary report of ABC warehouse.

Dept group Pick note type Total requested quantity Total picked quantity No. of pick notes Total man-hour spent Productivity (pcs/man-hr)
Accessories Allocation 8968 8968 259 14.30 626.92
Auto-replenishment 25256 25188 2453 153.11 164.51
Normal 33001 32915 1735 70.52 466.72
Accessories Total 67225 67071 4447 237.94 281.89

Apparel Allocation 410 409 30 2.25 181.76

Auto- replenishment 5541 5159 1944 129.52 39.83
Normal 20037 19222 4191 156.23 123.04
Apparel Total 25988 24790 6165 288.00 86.08

Total 93213 91861 10612 525.94 174.66

improve order picking productivity. The ABC warehouse is a man-

ual-pick, multi-level rack warehouse. In retail business of athletic
products, order picking is critical to the success of the company.
The shorter the order lead time, the higher the opportunity of sales
hit. The products are divided into 3 lines, i.e. men, women and kids
with 2 big categories namely apparel and accessories like caps and
water bottles. It processes around 500 to 1500 auto-replenishment
orders (i.e. orders based on point-of sales) per working day in nor-
mal seasons and delivers to about 200 local shops in tight time
windows.
The comparatively low picking productivity of auto-replenish- Fig. 2. Average time break down for pickers to complete an order.
ment orders is owing to their three characteristics: (1) small order
size with wide variety of items; (2) wide range of order volume due
to seasonality and promotions and (3) inconsistent number of level rack warehouses. The difference in picking speed/efficiency
delivery locations based on the point-of-sales data. Past experience in the moving distance between vertical and horizontal storage
shows that if batch picking is used, the reduction in travel times in a manual-pick multi-level rack warehouse should be considered.
cannot compensate the time spent on sorting the orders to about It indicates that the minimum moving distance does not equivalent
200 delivery locations. Therefore, discrete picking is used. Order to the minimum picking time. Objective of the research is to im-
pickers usually pick one order per journey and it results in long tra- prove picking performance of auto-replenishment orders in terms
vel distance and order retrieval time. of travel distance and order retrieval time of ABC warehouse by:
Generally, discrete picking (i.e. picking by order) and batch
picking (i.e. picking by article) are used in the ABC warehouse. Dis- (1) Classifying the items into 3 classes (A,B,C) by using COI and
crete picking is used when orders are fairly large whilst batch pick- EIQ method;
ing is used when orders are small. The order picking productivity is (2) Exploring the picking efficiency under different combina-
defined as pieces of SKU picked per man-hour. A computerized sys- tions of storage assignment policies, routing policies, and
tem is set to record the order retrieval time for individual pickers. compares their performances by simulation;
As shown in Table 3, the overall picking productivity for auto- (3) Performing sensitivity analysis based on different order vol-
replenishment orders is significantly lower than that of the alloca- ume and pick densities;
tion orders (i.e. medium-sized orders submitted by the salesperson (4) Studying the relationship between travel distance and order
based on the demand) and normal orders (i.e. indent order for new retrieval time in the manual-pick multi-level rack warehouse.
launched products).
Despite the characteristics of the customer orders, there is an- Altogether, three models of the same warehouse layout are built
other reason leads to low order picking productivity. Currently, with respective to different routing policies (i.e. transversal, return
the ABC warehouse randomly assigns the Stock Keeping Units and combined routings). Twenty-seven experiments are carried
(SKU) to storage locations according to their product category out to test the performance of picking under different combination
and codes alphabetically or numerically. Fast, medium and slow of factors. (i.e. 3 storage assignment policies  3 routing poli-
moving items are stored in same locations and same levels of racks. cies  3 pick densities). Computer simulation is employed to pur-
Hence, pickers have to travel long distance and spend more time sue the aforementioned objectives. ARENA was used to build the
on picking the hot items stored in high levels of rack. As you can model for analysis. The purpose of using COI index and EIQ analysis
see in Fig. 2, travel contributes to half of the total order picker’s is to classify products into classes. Computer simulation is to ob-
time. In order to fulfill the key performance indicator (KPI) of deliv- tain data for travel distance and order retrieval time of each com-
ery lead time of auto-replenishment orders, i.e. 1 day after the bination. And the statistical technique is used to analyze and
sales made, the order picking productivity is needed to be present the findings relevant to the objectives of the simulation
improved. study.

3.2. Objective 4. Methodology and implementation

In the research of improving order picking productivity, most In this research, only 120 types of products are selected to per-
studies focused on single-level rack warehouses, or considered form the ABC classification and simulation modeling. The selection
the AS/RS systems. Few of them consider manual-pick and multi- of products is based on historical picking list of auto-replenish-
2690 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700

(3) IQ: order quantity (Q) of each item I

(4) IK: frequency of order (K) of each item I

Generally, EIQ analysis consists of five major steps:

[Step 1:] Collect picking lists of customer orders for one or more
peak days.
[Step 2:] Construct EIQ table (Table 4 as an example) based on
the picking lists. The table shows particular number of items
appears in a particular order. It defines the number and quan-
tity of items that are going to investigate.
[Step 3:] Analyze the EQ, EN, IQ and IK obtained for the EIQ
Fig. 3. Number of auto-replenishment orders in December 08 & January 09. table as in Table 5 and (only IQ and EQ are illustrated in this
example).
[Step 4:] Arrange the EQ, EN, IQ and IK in descending order.
Divide the items into classes A, B and C based on percentage
ment orders on December 2008 and January 2009. As shown in
ratios of each class (i.e. class A: 60% of all items; class B: 30% of
Fig. 3, the number of auto-replenishment orders reaches to its peak
all items; class C: 10% of all items). This is illustrated in Table 6.
level on 13th December, 2008, which consisted of 617 orders. The
[Step 5:] Assign products to storage locations. A-items should
picked quantity is 6189 units with a wide variety of 120 items.
be stored in locations with high accessibility to order pickers.
4.1. ABC Classification
4.2. Warehouse design
In the ABC warehouse, products are grouped into three classes
in a way that the fastest moving class contributes to about 60% 4.2.1. Warehouse layout
of the turnover while the slowest moving class contributes to In the research, only 42 racks of the ABC warehouse are consid-
10% of the turnover. The assignment of products into classes is ered. The location of the depot is addressed in Fig. 4. Here are some
determined by the COI index or EIQ analysis. Fast moving items of the important configurations of the warehouse. Each rack has 4
are generally called A-items. The next fastest moving category of levels and 5 bays. Altogether there are 840 storage locations. The
products is called B-items, and so on. dimension (W  D  H) of a storage location is 2.75 m  0.9 m 
0.85 m which constitutes to 1.58 m3 (75% of the height). The total
4.1.1. COI index number of aisle is 22 with the width of 1 m.
The COI index of an item is the ratio of its total required space to
its demand per period. In the study, the COI value for 120 products 4.2.2. Storage assignment policies
is calculated. COI for each item can be found with three basic data: Currently, random storage is used in the ABC warehouse. In this
(1) size of item in cubic meter; (2) Expected number of orders per paper, class-based storage is proposed to improve the performance
year/ forecasted demands and (3) average inventory level. The of picking of auto-replenishment orders. Two positioning of classes
products with the lowest index value are located nearest the depot. is investigated. They are horizontal ABC class-based storage and
This method ensures that the greatest amount of product volume vertical ABC class-based storage. Altogether, three storage assign-
travels the minimum amount of distance. However, the type and ments are studied in the study.
quantity of the item retrieved and other items contained within
the order are not considered. 4.2.2.1. Random storage. Currently, random storage assignment
based on simple product categorization is used in the ABC ware-
4.1.2. EIQ analysis house. According to the restrictions on maximum loading of the
EIQ analysis is well adopted by researchers for logistics system storage racks, apparels are stored at level 1 to level 3 of 42 racks
or distribution planning purpose. It is an effective tool for ware- in the warehouse. Accessories, on the other hand, are stored sepa-
housing systems design, selection of material handling equipment, rately at level 4 of the 42 racks. Random storage is simple to use
inventory control and forecasting. Traditional ABC or POS data and performs better when the pick density is high. It is acted as
analysis is based on two factors: items (I) and quantity (Q). EIQ a base to find out the best combination of factors. An example is
data have three key factors (E, I, Q) which give more details on cus- shown in Fig. 5.
tomer orders. For example:
4.2.2.2. Horizontal ABC class-based storage. This horizontal ABC
(1) EQ: order quantity (Q) from each customer E class-based storage assignment is commonly used in single-level
(2) EN: number of items (N) ordered by each customer E rack warehouse. Firstly, classes A, B, C are determined by the COI

Table 4
EIQ table.

Items (unites = pieces)

Entry of orders I1 I2 I3 I4 I5 I6 EQ EN
E1 3 5 0 1 2 3 14 5
E2 2 0 4 6 7 0 19 4
E3 4 0 0 0 0 8 12 2
E4 2 8 0 3 5 2 20 5
EQ Total
IQ 11 13 4 10 14 13 IQ Total 65 EN Total
IK 4 2 1 3 3 3 IK Total 16
 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700 2691

Table 5 the efficiency of three common routing heuristics. They are: (1)
Analysis of EIQ. transversal, (2) return and (3) combined.
IQ Cumulative IQ % of cumulative IQ
(a) Analysis of IQ 4.2.3.1. Transversal/S-shape heuristic. One of the simplest heuristics
I1 11 11 16.9 for routing order pickers is the transversal (or S-shape) heuristic.
I2 13 24 36.9 Routing order pickers by using the transversal method means that
I3 4 28 43.1 any aisle containing at least one pick is traversed entirely. Aisles
I4 10 38 58.5
I5 14 52 80.0
without picks are not entered. From the last visited aisle, the order
I6 13 65 100 picker returns to the depot. This type of routing is often used when
TOTAL 65
the pick density is high. Order pickers can visit more pick locations
in a trip.
EQ Cumulative EQ % of cumulative EQ
(b) Analysis of EQ 4.2.3.2. Return heuristic. Return method means an order picker en-
E1 14 14 21.5 ters and leaves each aisle from the same end. Only aisles with picks
E2 19 33 50.8 are visited. This type of routing is commonly used in class-based
E3 12 45 69.2
storage when the pick density is low.
E4 20 65 100.0
TOTAL 65
4.2.4. Combined/composite heuristic
The composite heuristic combines the best features of the re-
turn and transversal heuristics. Aisles with picks are either entirely
Table 6
traversed or entered and left at the same end. It minimizes the tra-
ABC classification based on IQ.
vel distance between the farthest picks in two adjacent aisles.
No. SKU Quantity of % of Cumulative Cumulative ABC
# an item for total IQ % of total classification
4.2.5. Pick density
all orders IQ IQ
(IQ) Pick density is defined as the variety of items in a customer or-
der which affects the performance of picking. It is calculated by:
1 108 2300 23.3 2300 23 A
2 131 2000 20.2 4300 43.5 A no: of items in a customer order
3 191 1800 18.2 6100 61.8 B Pick density ¼ :
4 114 1000 10.1 7100 71.9 B
total no: of items in the warehouse
5 106 800 8.1 7900 80.0 B For example, there are 24 items in a customer order while the total
6 119 450 4.6 8350 84.5 B
number of items is 120, the pick density is 24/120 = 20%. Based on
7 128 300 3.0 8650 87.6 B
8 132 280 2.8 8930 90.4 C
the historical picking list of the ABC warehouse, three common pick
9 134 250 2.5 9180 92.9 C densities are found: 30% (i.e. number of items in an auto-replenish-
10 101 188 1.9 9368 94.8 C ment order = 36 items), 15% (i.e. number of items in an auto-replen-
11 110 80 0.8 9448 95.6 C ishment order = 18 items) and 5 % (i.e. number of items in an auto-
12 112 75 0.8 9523 96.4 C
replenishment order = 6 items). In order to find out the optimal
13 120 70 0.7 9593 97.1 C
14 107 65 0.7 9658 97.8 C combination of storage assignment and routing policies, sensitivity
15 127 60 0.6 9718 98.4 C analysis is performed under different pick densities.
16 122 50 0.5 9768 98.9 C
17 105 45 0.5 9813 99.3 C
4.2.6. Time and distance measurement
18 124 35 0.4 9848 99.7 C
19 103 20 0.2 9868 99.9 C
4.2.6.1. Order retrieval time measurement. The order retrieval time
20 113 10 0.1 9878 100.0 C should be divided into three components: (1) time for horizontal
travels, (2) time for vertical movements and (3) actual picking
time.
In the ABC warehouse, individual items such as apparel and
index and EIQ analysis. The fast moving items (i.e. A-items) are accessory are picked in pieces. An order picker picks one order at
then assigned to store closer to the depot. The travel distance a time following a route up and down each aisle until the entire or-
can be reduced. See Fig. 6 for an example. der is picked. The order picker will usually use some type of pick-
ing cart. For items stored at the first and second levels of the racks,
after walking to the pick locations, the order picker used to open
4.2.2.3. Vertical ABC class-based storage. Different from horizontal the carton box directly and select the right item of right quantity
ABC class-based storage, fast moving items (i.e. A-items) are stored and right size from the box and close the carton box. The order re-
at lower levels of the racks (i.e. level one and level two) in the ver- trieval time should include the horizontal travels and actual pick-
tical ABC class-based storage. The chance for order pickers to pick ing time.
the items stored at higher level of the racks with a ladder is re- For items stored at the third and fourth levels of the racks, the
duced. It therefore reduces the order retrieval time. It is found order picker has to reach the carton boxes with a ladder. The order
more suitable to be used in multi-level rack warehouse. Fig. 7 de- retrieval time should include the horizontal travels, vertical move-
picts this scheme. ments as well as the picking time. The vertical movements of the
order picker refer to climbing up and down the ladder. Without
4.2.3. Routing policies doubt, the time spent on picking at higher levels of the rack must
When an order picker starts at the depot, where he receives a be longer than that of at lower levels.
picking list, has to visit all the pick locations and finally has to re- As a result, the order retrieval time is calculated based on the
turn to the depot. In practice, the problem of routing order pickers operation process and the productivity tracking report:
in a warehouse is mainly solved by using heuristics. The heuristics
help to sequence the items on the picking list to ensure a good Time spent on horizontal travel = horizontal distance/50 m per
route through the warehouse. This study focuses on analyzing minute
2692 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700

Fig. 4. Layout plan of the ABC warehouse.

Fig. 5. Rack design of random storage.

Time spent on vertical movement = (vertical distance  2)/50 m height of reached each level (Vz) where z = 1–4. In fact, the mea-
per minute surement of travel distance can be applied in three routing policies,
Time spent on picking at 1st and 2nd levels of the rack = i.e. transversal, return and combined routing policies.
0.27 min
Time spent on picking at 3rd and 4th levels of the rack =
2.13 min 4.3. Simulation modeling

4.2.6.2. Travel distance measurement. In the ABC warehouse, there The objective of the simulation study is to determine which
are 42 racks and 22 aisles. As shown in Fig. 8, the horizontal dis- combination of factors provides the shortest travel distance and or-
tance is the sum of three components: the depth of the travelled der retrieval time for pickers to complete all the auto-replenish-
bays (Hx), the width of travelled bays (Hy) and the width of the vis- ment orders in a day. Three models of the same warehouse
ited aisles (Hm), where x = 1–42, y = 1–5 and m = 1–22. For vertical layout are built with respective to different routing policies (i.e.
distance, there are 4 levels in each rack. The height of each level transversal, return and combined routings). Table 7 summarizes
equals to 0.85 m. Therefore, vertical distance is the sum of the some common parameters of the three models.
 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700 2693

Fig. 6. Rack design of horizontal ABC class-based storage.

Fig. 7. Rack design of vertical ABC class-based storage.

Table 7
Parameters used in the simulation model.

Parameter Value Explanation

Total storage 840 42 racks  4 levels  5
locations bays
Size of a storage 2.75 m  0.9 m  0.85 m Width  Depth  Height
location
No. of aisles 22
Width of aisle 1m
Picking time 2.13 m Level 3–4 of the rack
0.27 m Level 1–2 of the rack
Travel speed 50 m/min
Travel time Horizontal distance/ Horizontal travels
speed
(Vertical distance  2)/ Vertical movements
speed
Fig. 8. Configuration of the rack and aisles for measurement of travel distance.
2694 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700

Fig. 9. General layout of the warehouse model.

In addition, the following assumptions are made for the simula- pick densities (i.e. 5%, 15% and 30%). The order data is then read
tion study: from a database file. For each order arrival, the start time is re-
corded for measurement of total order retrieval time. When order
(1) All items are stored in standard size of cartons in one loca- picking is completed, the end time is also recorded. Then, a deci-
tion only sion module is used to ensure the pick density used for the exper-
(2) No shortage occurs in the warehouse iment is right. After checking, the item number of each item in the
(3) Strict order picking is used for auto-replenishment orders, same order is extracted from the spreadsheet. The systems will as-
an order picker is responsible to pick one order at a time sign a unique storage location number for each item according to
(4) Time and distance measurement is focused on the retrieval
of items from the storage locations and deliver the picked
items to the depot. Some sequential activities of order pick-
ing are excluded. For example, pack the items according to
the orders, update the warehouse management system by
barcode scanning and prepare for shipping.

Based on above, a general layout for the three warehouse mod-

els is shown in Fig. 9. With the purpose to build a model close to
the actual environment of the warehouse ABC, 42 racks with 22
aisles are built in real dimensions (in feet). The depot is located
at the center of the racks. In order to measure the total travel dis-
tance, stations are set at the depot, two ends of the aisle as well as
each bay of the racks. The stations are connected by segments of
lines set with different distances. Each distance set consists of a
distance set identifier and either a station range and associated dis-
tance matrix or a list of station pairs and individual distance. This
facilitates the distance setting and enhances the measurement of
the total travel distance.
In additional, in order to guide the order pickers by routing heu-
ristics, the routes are connected by intersections. It is used to de-
fine specific routing in three models with the predefined logic,
sequence and decision making points. For example, specific entity
attributes such as item number or index are used to define the se-
quences of item picking. If the item number is matched with a spe-
cific station number, the order pickers can be guided to the desired
sequence set number value.
For the simulation of the order retrieval process, the logic for or-
der arrival and dispatch is the same for three models. The major
difference is in the logic for routing. First, the logic for order arrival
is illustrated in the following example.
When simulation starts, auto-replenishment order arrival is
created. Auto-replenishment orders are classified into different Fig. 10. Flow of the simulation process.
 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700 2695

Table 8
Ten products with the lowest COI index.

No. SKU Dept Season Size (Cu. Expected no. of orders (Per Average inventory Average no. of daily Required storage space COI
no. Ft) quarter.) (Units) orders (Cu. Ft.)
1 I90 Acc LR 0.05 3000 1800 35 88.99 2.52
2 I103 Acc LR 0.04 1800 1500 21 60.00 2.83
3 I96 Acc LR 0.05 1800 1250 21 61.80 2.92
4 I114 Acc LR 0.05 2500 1800 29 88.99 3.03
5 I97 Acc LR 0.05 850 700 10 34.61 3.46
6 I47 App 2009Q1 0.06 2500 1700 29 102.00 3.47
7 I85 Acc LR 0.05 1000 850 12 42.02 3.57
8 I111 Acc LR 0.05 900 800 11 39.55 3.74
9 I119 Acc 2009Q1 0.05 1800 1600 21 80.00 3.78
10 I95 Acc LR 0.05 1200 1100 14 54.39 3.85

the item number. The assignment is based on the use of storage 5.1.2. EIQ analysis
assignment policy. The storage location number specifies the data The EIQ table is constructed according to the picking lists of auto-
such as aisle number, rack number, station number and level num- replenishment orders with high demand rates. Since the size of the
ber. If the items are stored in the same bay of the same rack, they EIQ table is too big, key information is summarized in Table 9 only.
will have the same station number. The station number is a key to Altogether, 617 auto-replenishment orders (E1–E617) are collected.
determine the distance travelled by the order picker. After that, the As mentioned, the size of the auto-replenishment order is small
station number of each item in the same order is arranged in with wide variety of items. From the EIQ table, it is found that the
ascending order. This sequence impacts the logic of the routing average order quantity is 10.03 items. And the maximum number
heuristics. Then, the resource, order picker, is requested to perform of items by each customer is 31 with a variety of 120 items.
the sequential order picking from the depot of the warehouse. And In order to divide 120 products into appropriate classes, four
the measurement for travel distance is begun. types of analysis are performed (i.e. EQ, EN, IQ and IK). It is found
After building the models, twenty-seven experiments are car- that the IQ analysis (i.e. order quantity of each item) and IK analy-
ried out to test the performance of picking under different combi- sis (i.e. frequency of order of each item) can better reflect the
nation of factors. (i.e. 3 storage assignment policies  3 routing inventory turnovers of the ABC warehouse. As shown in Fig. 11,
policies  3 pick densities). First, one of the pick densities (i.e. small number of items has a significant impact on overall inven-
30%, 15%, 5%) and storage locations of the 120 products based on tory turnover. Some of these items have the high values of IQ
a storage assignment policy (i.e. random, vertical and horizontal and IK at the same time. For example, item I190 consistently be-
ABC class-based storage assignments) are set in the model. When come the item with highest order frequency and quantity.
an auto-replenishment order arrives in the system, the system From above analysis, fast moving products are identified. To
would assign the pick location of respective items. And the order determine the number of items in each class, percentage ratios
picker will complete the picking process according to the prede- are defined according to the total turnover. For instance, A-items
fined routing policy (i.e. transversal, return and combined). The will typically contain items that account for 60% of the total turn-
system will update the database once the picking of the order is over. B-items will account for 30 % while C-items will make up the
completed. After running the simulation, the respective total order remaining 10%. Table 10 shows the results obtained from different
retrieval time and travel distance are obtained. Fig. 10 shows the analysis. After comparing the results of the above analysis with the
flow of the simulation process. COI index, IK analysis is found the best for product classification.
The A-items classified by IK analysis has a comparatively low COI
5. Results and statistic analysis index. As a result, the accuracy of the classification is ensured.
Appendix B shows the detailed ABC classification of the 120 se-
5.1. ABC Classification lected products.

Before the simulation, 120 products are classified into classes A, 5.2. Simulation results and statistic analysis
B and C based on their turnover rates. As a whole, class A inventory
contributes to 60% of the total turnover of the selected products. Based on the predefined assumptions and parameters of the or-
Class B inventory account for 30% of the turnover and class C der-picking system, three simulation models are built by using
inventory should make up the remaining 10% of the turnover. ARENA with respect to three routing policies. Twenty-seven com-
The classification of products is determined by COI index and EIQ binations of factors (i.e. 3 storage assignment policies  3 routing
analysis.

Table 9
5.1.1. COI index
Summary for the EIQ table.
The COI index takes the turnover rate and required space of
individual items into consideration. Table 8 shows ten items with Analysis factors Quantity
the lowest COI index. These items can be classified as A-items. No. of customer order (E) 617
The advantage of using COI index over the EIQ analysis is that 38 No. of item (I) 120
the new items can be classified into a particular class-based on Total order quantity from all customers (Total EQ = Total IQ) 6189
Average order quantity from each customer (Average EQ) 10.03
the forecasted sales level. In the study, 120 products is selected Average order quantity of each item (Average IQ) 51.58
from past auto-replenishment orders. Classification of products Range of EQ 1 to 126
can be extended by studying the correlation of products. If items Range of IQ 0 to 705
within the same class are found to have higher frequency to be ap- Range of EN 1 to 31
Range of IK 0 to 73
peared in the same order, they should be stored closer to each
Total no. of items ordered by each customers (Total EN = Total IK) 1229
other.
2696 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700

and second levels of the racks minimizes vertical movements of

the order picker. Also, routing the order picker by combined heu-
ristics allows higher flexibility in visiting the pick locations.
In contrast, the use of return routing in horizontal ABC class-
based storage would results in longest order retrieval time.
Although A-items are stored near to the depot, the accessibility
to the order picker decreases as the level of their storage locations
increases. Order picker has to spend more time on vertical move-
ments and picking. In additional, return routing requires the order
picker to enter and leave each aisle from the same end. That means
the order picker has to visit the aisles with picks twice in time.
Hence, the total order retrieval time is increased by duplicated
travels.
The mean and standard deviation of total order retrieval time
for each experimental factor is calculated and shown in Table 12.
Moreover, one-way ANOVA test is used to analysis the variance
of the results with the level of significance equals to 0.05. Here
are some of the important findings:

5.2.1.1. Storage assignment policy. The average total order retrieval

Fig. 11. IQ and IK analysis. time for the vertical ABC class-based storage is the shortest among
three storage assignment policies. It is slightly shorter than that of
the random storage. In comparison, the difference becomes larger
when comparing the average total order retrieval time between
Table 10 the vertical ABC class-based storage and that of the horizontal
Classification of 120 items by different analysis. ABC class-based storage. Under the F-test, the result also shows
that average total order retrieval time of the vertical ABC class-
Class IQ analysis (% of IK analysis (% of IQ x IK analysis (%
(Percentage 120 items) 120 items) of 120 items) based storage is similar to that of the random storage. However,
turnover) the average total order retrieval time of horizontal ABC class-based
Class A (60) 16 (13) 22 (18) 23 (19) storage is considerably longer than that of the other two storage
Class B (30) 20 (17) 28 (23) 13 (11) assignment policies.
Class C (10) 84 (70) 70 (58) 84 (70) Base on above statistic analysis, the use of vertical ABC class-
based storage and random storage policies in multi-level rack
warehouse results in similar performance of picking in terms of or-
der retrieval time. On the other hand, horizontal ABC class-based
Table 11 storage, which is commonly used in single-level rack warehouse,
Abbreviation for different experimental factors. is found less efficient in multi-level rack warehouse. As mentioned,
Policies for order picking Experimental factors Abbreviation the order retrieval time increases as the frequency of high-level or-
der picking increases.
Storage assignment Horizontal ABC class-based S1
policy storage
Vertical ABC class-based storage S2 5.2.1.2. Routing policy. Combined routing has the shortest order re-
Random S3 trieval time among all routing policies. It is allows order picker to
Routing policy Transversal R1 switch between the transversal and return policies so as to mini-
Return R2
Combined R3
mizes the travel distance between the farthest picks in two adja-
Pick density 5% D1 cent aisles. Hence, order picker can complete the order picking in
15% D2 the shortest route. On the other hand, transversal and return poli-
30% D3 cies have the similar total order retrieval time. Owing to the dupli-
cated travels in the same aisle, return policy leads to the longest
order retrieval time. By performing pair-wise comparison of F-test,
policies  3 pick densities) are simulated. After running the simu- it is found that the difference in total order retrieval time between
lation, results are obtained. They are divided into three parts: (1) combined routing and any of the transversal or return routings is
order retrieval time, (2) travel distance and (3) overall perfor- large. However, the difference becomes minor between the trans-
mance. To better illustrate the results, abbreviations for different versal and return routings.
experimental factors are used as shown in Table 11.
5.2.1.3. Pick density. The results show a normal behavior that the
total order retrieval time increases in direct proportion with the
5.2.1. Order retrieval time pick density. Since the variety of items in the auto-replenishment
By using simulation, total order retrieval time for each combi- orders increases, more time is spent on both travelling and picking.
nation of factors is obtained. It represents the time required for Also, the results of F-test indicate that the changes in order retrie-
an order picker to complete all auto-replenishment orders. As val time are obvious for different pick densities (i.e. 5%, 15% and
shown in Fig. 12, the shortest and the longest order retrieval time 30%).
appear in the vertical ABC class-based storage and the horizontal
ABC class-based storage respectively. Indeed, the use of combined 5.2.2. Travel distance
routing in vertical ABC class-based storage under different pick In multi-level rack warehouse, the minimum order retrieval
densities shows a consistency success in reducing the total order time is not equivalent to the minimum travel distance. To study
retrieval time. The main reason is the storage of A-items at first the relationship in between the time and distance, the total order
 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700 2697

Total Order Retrieval Time

350000
S1_R1
300000
S1_R2
250000
S1_R3
200000
S2_R1
150000
S2_R2
100000
S2_R3
50000
S3_R1
0
D1 D2 D3 S3_R2

S1 S2 S3
R1 R2 R3 R1 R2 R3 R1 R2 R3
D1 111096 117989 97912 81517 86691 72017 88632 87983 77671
D2 180277 189655 164098 151514 160873 136222 155887 166755 142307
D3 321886 331808 305373 240293 250085 223879 247674 256377 230426
Average 204420 213151 189128 157775 165883 144039 164064 179372 150135

Fig. 12. Total order retrieval time for all combination of factors (seconds).

Table 12
Total order retrieval time for all experimental factors (seconds).

S1 S2 S3 R1 R2 R3 D1 D2 D3
Mean 202233 155899 161524 175420 183135 161101 91279 160843 267533
Standard deviation 915 931 935 1121 1124 1117 1109 1111 1107

retrieval time is obtained by using simulation. Total travel distance When the pick density is low, the shortest and the longest travel
refers to the required travel distance for the order picker to com- distance are appeared in the horizontal ABC class-based storage
plete all the auto-replenishment orders in a day. Fig. 13 shows and the vertical ABC class-based storage respectively. The combi-
the total travel distance of each of the twenty-seven combination nation with the shortest travel distance is the usage of combined
of factors. routing in horizontal ABC class-based storage. The main reason is
Generally, the average total travel distance increases as the pick that fast moving items are stored near the depot. Unnecessary
density increases. However, it is found that a storage assignment travels are avoided. Therefore, the travel distance is minimized.
policy that has the longest total travel distance in one pick density In contrast, the use of return routing in vertical ABC class-based
can become the one that has the shortest total travel distance in storage is the combination with the longest travel distance. It is be-
another pick density. Therefore, the discussion on the results of to- cause fast moving items are distributed at the lower level storage
tal travel distance is divided into two parts, the low and high pick locations around the warehouse. The travel distance is increased
density. when the variety of items is narrow. If return routing is applied,

25000
S1_R1
20000 S1_R2

15000 S1_R3
S2_R1
10000
S2_R2
5000 S2_R3
S3_R1
0
D1 D2 D3 S3_R2

S1 S2 S3 Average
R1 R2 R3 R1 R2 R3 R1 R2 R3
D1 11675 14645 9365 13345 16909 11952 11952 14923 10962 12881
D2 14586 17556 12276 15008 18572 15001 15001 17972 14012 15423
D3 18744 21714 16434 17503 21067 19576 19576 22546 18586 19165
Average 15002 17972 12692 15285 18849 15510 15510 18480 14520 15823

Fig. 13. Total travel distance for all combination of factors (feet).
2698 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700

Table 13
Total travel distance for all experimental factors (feet).

S1 S2 S3 R1 R2 R3 D1 D2 D3
Mean 15222 16077 16170 15266 18434 13770 12881 15423 19165
Standard deviation 102 101 103 121 120 198 115 116 114

the performance of picking become worse as duplicate travels may icies results in different varied performance in terms of total order
occur. retrieval time and travel distance. The major influence factor is the
When the pick density is high, the results are different. The use of storage assignment policies. Under low pick density, hori-
shortest and longest total travel distance appeared in the vertical zontal ABC class-based storage policy assigns the fast moving
ABC class-based storage and the random storage respectively. If items near the depot. The performance is the best in terms of travel
the combined routing is used in the vertical ABC class-based stor- distance instead of order retrieval time. For high pick density, ver-
age, the total travel distance is the shortest. Since the variety of tical ABC class-based storage policy distributes fast moving items
pick items increases, the order picker needs to visit more storage at lower level of the racks around the warehouse and results in less
locations. In the vertical ABC class-based storage, fast moving vertical movements. It reduces the order retrieval time and the tra-
items are stored at lower level of the racks. Order picker can pick vel distance at the same time. Table 14 summarizes the average
the items more conveniently when walking around the storage values of total order retriever time and travel distance for different
locations with combined routing. The order retrieval becomes experimental factors from the previous session. Here are some of
more efficient as vertical movements are minimized. In contrast, the major findings:
if return routing is applied in random storage, the travel distance
is increased significantly. The reason is that fast moving items 5.2.3.1. Storage assignment policy. Vertical ABC class-based storage
may stored at the back of the warehouse or at higher level of the has the best performance with regards to total order retrieval time
racks while return routing would results in duplicate travels as but the worst performance in terms of total travel distance. Hori-
mentioned. zontal ABC class-based storage has the opposite performance as
The mean and standard deviation of total travel distance for compared to the vertical ABC class-based storage. Moreover, ran-
each experimental factor is computed and shown in Table 13. dom storage has a better impact on reducing the total order retrie-
Moreover, one-way ANOVA test is used to analysis the variance val time than reducing the total travel distance.
of the results with the level of significance equals to 0.05. Here
are some of the important findings:
5.2.3.2. Routing policy. Combined routing achieved the best perfor-
mance in both order retrieval time and travel distance. Moreover,
5.2.2.1. Storage assignment policy. In the F-test, the total travel dis- transversal and return routings has similar performance in two
tance for horizontal ABC class-based storage is significantly shorter indicators. And return routing is found to be the worst routing
than that of vertical ABC class-based and random storages. In fact, strategy.
the result is similar for vertical ABC class-based storage and ran-
dom storage. This indicates the use of horizontal ABC class-based
5.2.3.3. Pick density. The performance of picking decreases as the
storage in multi-level rack warehouse can reduce the total travel
pick density increases. The major finding is that the total order re-
distance. However, its drawback is the long order retrieval time
trieval time increases in direct proportion to the pick density while
owing to the higher frequency of high-level picking.
the total travel distance is not.
From above findings, the most appropriate combination of stor-
5.2.2.2. Routing policy. Without doubt, the flexible combined rout- age assignment and routing polices with respect to the order re-
ing consistently performs more superior than transversal and re- trieval time and travel distance is identified according to
turn routing in terms of total order retrieval time as well as different levels of pick density. The results are showed in Table 15.
travel distance. Transversal routing results in shorter travel dis- Regarding to the total order retrieval time, regardless to the le-
tance than that of return routing since it avoids duplicated travels vel of pick density, the use of combined routing in vertical ABC
at the same aisle. It is consistent to the results of total order retrie- class-based storage is the most appropriate combination. For total
val time. The result of F-test shows the use of routing policies re- travel distance, pick density is a key factor to determine the use of
sults in significant difference in travel distance. storage assignment policy. Under low pick density, the use of com-
bined routing in horizontal ABC class-based storage policy is the
5.2.2.3. Pick density. Similar to the results in total order retrieval
time, total travel distance increases as the pick density increases.
However, it is not in direct proportion. The setting of pick densities Table 14
Average of total order retrieval time and travel distance for different experimental
results in considerate difference in travel distance as demonstrated
factors.
in the F-test.
Policies for order Experimental Total order Total travel
picking factors retrieval time distance
5.2.3. Overall performance of picking
From the previous session, it is found that various combinations Storage assignment S1 202233 15222
policy S2 155899 16077
of storage assignment policies, routing policies and pick densities S3 161524 16170
result in different performance of picking in terms of order retrie- Routing policy R1 175420 15266
val time and travel distance. In order to analyze the overall perfor- R2 183135 18434
mance of each combination, these two performance indicators are R3 161101 13770
Pick density D1 91279 12881
considered at the same time.
D2 160843 15423
The consistency of the results on two performance indicators D3 267533 19165
only occurs in the high pick density. When the pick density is
Total average 173218 15823
low, different combination of storage assignment and routing pol-
 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700 2699

Table 15 characteristics of the warehouse into consideration. For the

Appropriate combination of factors for different customer orders. ABC warehouse, the storage space is large and the size of
Order Appropriate combination of polices auto-replenishment order is small with wide variety of
Pick density Storage assignment Routing items. The results of simulation proved that the use of ran-
dom storage is just a little bit worse than that of vertical
(a) Based on order retrieval time
ABC class-based storage assignment policy. However, it is
D1 (5%) S2 (Vertical ABC class-based storage) R3 (Combined)
greatly superior to the horizontal ABC class-based storage
D2 (15%) S2 (Vertical ABC class-based storage) R3 (Combined)
D3 (30%) S2 (Vertical ABC class-based storage) R3 (Combined)
assignment policy.
(b) Based on travel distance
D1 (5%) S1 (Horizontal ABC class-based storage) R3 (Combined)
It is found that further research can be made in following areas.
D2 (15%) S1 (Horizontal ABC class-based storage) R3 (Combined)
D3 (30%) S2 (Vertical ABC class-based storage) R3 (Combined) (1) More combination of factors can be included in the study of
improving the performance of order picking. Since the
research scope of the warehouse example used in this study
most appropriate combination. It is inconsistent to what has been is limited and the research time is insufficient, other factors
mentioned in the total order retrieval time. For high pick density, that affect the performance of picking are not included in the
vertical ABC class-based storage policy with combined routing study. It suggested that further research can be carried out
should be used. It is the same combination that mentioned in the on the design of the warehouse layout and the congestion
total order retrieval time. Based on above analysis, warehouse problems if more order pickers pick the items at the same
operators can select the most appropriate combination according pick wave. This would make the research on improving the
to the characteristics of the customer orders. performance of picking become more comprehensive and
more adapt to the diversity of the warehouse operation.
6. Conclusions (2) Also, further research can be extended to all the sequential
process of order retrieval. For example, order receiving,
This study takes a multi-level rack warehouse as an example to inventory replenishment and distribution operations. It
improve the performance of picking by performing twenty-seven increases the integration of the warehouse activities. By sim-
experiments on different combinations of storage assignment pol- ulation with actual data, overall logistics performance and
icies, routing policies and pick densities. The performance of pick- customer service level can be improved.
ing is measured in terms of order retrieval time and travel distance.
By simulation with ARENA, the results show that different combi-
nation of factors has different performance under different perfor-
References
mance indicators. Here are some of the important findings:
Ashayeri, J., Heuts, R. M., Valkenburg, M. W. T., Veraart, H. C., & Wilhelm, M. R.
(1) Use of storage assignment policy (2002). A geometrical approach to computing expected cycle times for
The use of storage assignment policy is greatly depends on zonebased storage layouts in AS/RS. International Journal of Production
Research, 40(17), 4467–4483.
the type of the storage system. For multi-level rack ware- Brynzér, H., & Johansson, M. I. (1996). Storage location assignment: Using the
house, the use of vertical ABC class-based storage assign- product structure to reduce order picking times. International Journal of
ment policy can improve the performance of picking in Production Economics, 46–47, 595–603.
Daniels, R. L., Rummel, J. L., & Schantz, R. (1998). A model for warehouse order
terms of total order retrieval time. On the other hand, the picking. European Journal of Operational Research, 105(1), 1–17.
use of horizontal ABC class-based storage assignment policy De Koster, R., Le-Duc, T., & Roodbergen, K. J. (2006). Design and control of
in single-level rack warehouse can results in shorter travel warehouse order picking: A literature review. European Journal of Operational
Research, 182(2), 481–501.
distance. Guvenir, H. A., & Erel, E. (1998). Multicriteria inventory classification using a genetic
(2) Use of combination of factors and selection of performance indi- algorithm. European Journal of Operational Research, 105(1), 29–37.
cator Hausman, W. H., Schwarz, L. B., & Graves, S. C. (1976). Optimal storage assignment
in automatic warehousing systems. Management Science, 22(6), 629–638.
Under different performance indicators such as order retrie-
Heragu, S. S., Du, L., Mantel, R. J., & Schuur, P. C. (2005). Mathematical model for
val time and travel distance, the performance of each combi- warehouse design and product allocation. International Journal of Production
nation of factors can be different. The one performs better in Research, 43(2), 327–338.
terms of order retrieval time can results in longer travel dis- Heskeet, J. L. (1963). Cube-per-order index – A key to warehouse stock location.
Transportation and Distribution Management, 3, 27–31.
tance. For example, the results of this study shows that hor- Heskett, J. L. (1964). Putting the cube-per-order index to work in warehouse layout.
izontal ABC class-based storage has opposite performance Transport and Distribution Management, 4, 23–30.
under different performance indicators. Hsieh, S., & Tsai, K.-C. (2001). A BOM oriented class-based storage assignment in an
automated storage/retrieval system. The International Journal of Advanced
(3) Use of routing policy Manufacturing Technology, 17(9), 683–691.
In this study, combined routing always has superior perfor- Jane, C. C., & Laih, Y. W. (2005). A clustering algorithm for item assignment in a
mance in performance indicators. It is due to its flexible use synchronized zone order picking system. European Journal of Operational
Research, 166(2), 489–496.
of transversal and return routings. Jewkes, E., Lee, C., & Vickson, R. (2004). Product location, allocation and server home
(4) Performance on travel distance changes under high pick density base location for an order picking line with multiple servers. Computers and
With regards to the total travel distance, the performance of Operations Research, 31(4), 623–636.
Le-Duc, T., & De Koster, R. (2005). Travel distance estimation and storage zone
storage assignment policy changes as pick density increases. optimisation in a 2-block class-based storage strategy warehouse. International
In the multi-level rack warehouse, if the variety of pick items Journal of Production Research, 43(17), 3561–3581.
increases, the advantage of using the horizontal ABC class- Li, M.-L. (2009). Goods classification based on distribution center environmental
factors. International Journal of Production Economics, 119(2), 240–246.
based storage under low pick density will lose. Since the fre-
Liu, C. M. (1999). Clustering techniques for stock location and order-picking in a
quency of high-level picking increases and results in longer distribution center. Computers and Operations Research, 26(10–11), 989–1002.
travel distance. Mansuri, M. (1997). Cycle-time computation and dedicated storage assignment for
(5) Use of random storage AS. RS systems. Computers and Industrial Engineering, 33(1–2), 307–310.
Montulet, P., Langevin, A., & Riopel, D. (1998). Minimizing the peak load: an
In fact, the use of random storage is not the worst choice in alternate objective for dedicated storage policies. International Journal of
the multi-level rack warehouse. However, it should take the Production Research, 36(5), 1369–1385.
2700 F.T.S. Chan, H.K. Chan / Expert Systems with Applications 38 (2011) 2686–2700

Partovi, F. Y., & Anandarajan, M. (2002). Classifying inventory using an artificial Wu, H.-H., Chen, C.-P., Tsai, C.-H., & Yang, C.-J. (2010). Simulation and scheduling
neural network approach. Computers and Industrial Engineering, 41(4), 389– implementation study of TFT-LCD Cell plants using Drum-Buffer-Rope system.
404. Expert Systems with Applications, 37(12), 8127–8133.
Petersen, C. G. (1997). An evaluation of order picking routing policies. International Yoo, T., Cho, H., & Yücesan, E. (2010). Hybrid algorithm for discrete event simulation
Journal of Operations & Production Management, 17(11), 1098–1111. based supply chain optimization. Expert Systems with Applications, 37(3),
Roodbergen, K. J., & De Koster, R. (2001). Routing methods for warehouses with multiple 2354–2361.
cross aisles. International Journal of Production Research, 39(9), 1865–1883.
Vickson, R. G., & Lu, X. (1998). Optimal product and server locations in one-
dimensional storage racks. European Journal of Operational Research, 105(1),
18–28.