Version of Record: https://www.sciencedirect.

com/science/article/pii/S0377221718304818
Manuscript_b80fe2b2aff11f6acd545f3a6fbdfacb

A Model to Optimize Rack Layout in a Retail Store

Corinne H. Mowrey1, Pratik J. Parikh1*, and Kevin R. Gue2

1
Department of Biomedical, Industrial and Human Factors Engineering
Russ Engineering Center, 3640 Colonel Glenn Highway, Wright State University, Dayton, OH 45435

2
Department of Industrial Engineering
J B Speed Building, 2301 S 3rd St, University of Louisville, Louisville, KY 40292

*Corresponding author

Pratik J. Parikh, Ph.D.

pratik.parikh@wright.edu

© 2018 published by Elsevier. This manuscript is made available under the Elsevier user license
https://www.elsevier.com/open-access/userlicense/1.0/
 A Model to Optimize Rack Layout in a Retail Store

Abstract
The physical layout of a retail store is known to influence the attitude and behavior of shoppers and affect
store performance. An important role of the store's layout is to expose shoppers to merchandise in order to
facilitate consideration and ultimately purchasing of exposed products. This work focuses on determining
which rack layouts maximize exposure of products to a shopper. To do this, we introduce the retail rack
layout problem (RRLP), which identifies the optimal single or multi-column rack orientations in a
constrained space in order to maximize exposure to the shopper. To solve this problem, we propose a
mixed-integer non-linear optimization model, which we solve using a particle swarm optimization (PSO)
algorithm. Results indicate that an increase of exposure ranging from 213-226% (small head turns) and
17-18% (large head turns) over 90º-rack layouts can be achieved with angled-racks (acute or obtuse from
a shopper’s travel path) depending on required duration of exposure. The increase in exposure comes at
the expense of loss of display space and is sensitive to the shopper’s field of regard (angular limit and
depth of vision). We also show that multiple competitive layout designs with 1-, 2-, and even 3-columns
may exist for a given system configuration that offer similar exposure values.

Keywords
Facilities planning and design; retail layout; field of regard; product exposure; particle swarm
optimization

1 Introduction
In light of the increasing online presence of most major retailers, it is natural to perceive the diminishing
influence of retail stores. This is not necessarily the case. Physical stores offer advantages web stores do
not, such as the ability to feel and inspect products for quality and the opportunity to immediately take
possession of products and fulfill the customer's needs (Bloomberg, 2018). Research shows that for most
Americans, shopping online is an infrequent occurrence and upwards of 65% of online shoppers prefer to
buy from physical locations (Pew, 2016); Kohl's recently reported that 90% of their sales still occur
within a physical store (Mansell, 2018). Clearly, the retail store provides a critical interface between
retailers and shoppers; after all, the store is where sales happen (Dunne et al., 1995). Retailers recognize
that the store’s physical environment is a critical trigger point in the shopping cycle that can significantly
affect the attitude and behavior of shoppers (Shankar et al., 2011, Bloomberg, 2018). Successful design of
the store's layout can influence how shoppers navigate through the store and interact with products,

1
ultimately affecting their shopping experience and sales (Bitner, 1992; Turley and Millian, 2000; Burke,
2005; Mitchell, 2008; Lu and Seo, 2015).
While the facility layout problem (FLP) has been well-studied, fundamental differences exist
between the traditional FLP for manufacturing and warehousing facilities and the retail layout problem.
First, the objectives are often quite different; retailers tend to focus largely on maximizing sales instead of
minimizing travel time. Second, the travel path of the shopper is highly variable and unpredictable, as
opposed to the well-defined flow of materials in a manufacturing facility. Third, department adjacencies
are often purposely violated (e.g., bread and milk) to encourage shoppers to travel through most of the
store to increase impulse purchases. And finally, shopper demand for products can be influenced by the
layout, product placement, promotions, and stimuli. Given these substantial deviations from the
traditional FLP, it should be evident that traditional FLP metrics such as space utilization, flow of
material, material handling costs, and travel time (Meller and Gau, 1996; Singh and Sharma, 2006) maybe
limited or even inappropriate for retail facility layout problems.
Retail literature has identified several metrics pertaining to rack layouts such as sales
productivity, shrinkage, shopper time spent in the store, shopper satisfaction/frustration, number of
unplanned purchases, and exposure of products (Dunne et al., 1995; Turley and Milliman, 2000; Lu and
Seo, 2015). Among these, exposure of products (on the racks) has often been alluded to as an important
metric and as a sales stimulus (Cairns, 1962; Cairns, 1963; Anderson, 1979; Dreze et al., 1994; D'Astous,
2000). Given that approximately 80% of all purchase decisions are made in-store, including generally
planned, unplanned, and substitute purchases (POPAI, 2014), it seems intuitive that shoppers will only
buy what they see (Ebster and Garaus, 2015).
In our earlier work, we developed analytical and computational models to quantify the impact of
rack layout on exposure in a retail store (Mowrey et al., 2017). We defined exposure as the possibility
(not the probability) that a rack location could be seen by a shopper as they traveled past a store section.
Our goal was to develop a solution approach that generated a visual profile of a given rack layout, which
could then be used to assign products to rack shelves in order to maximize store revenue. We observed
that, while angled orientations could potentially improve exposure, it did so at the expense of increases in
required floor space. Under floor space constraints, this would result in a trade-off between exposure and
amount of available display on the racks. Empirical evidence reveals that retailers are not only willing to
make such a tradeoff, but have already done so. In our immediate geographical area, we observed two
mass merchandise retailers, a grocery retailer, and a gas station employing alternate rack layouts (Figure
1). Despite this evidence of use, we are not aware of any approaches in either the academic or trade
literature that addresses the tradeoff between exposure and display. Moreover, the question of how best to
layout racks so as to maximize exposure has yet to be addressed.

2
 . v v v v v v v v v v

Pharmacy

Baby and Kids

Men's Clothing
v

Shoes

Women's Clothing

Cash Registers

Subway Hair Vision Bath- Customer

Bank
Salon Center room Service

(a) (b)

Figure 1: Actual layouts of (a) a mass merchandise retailer and (b) a gas station in the Midwest US

In this paper, we focus on this very issue and subsequently propose an optimization model for the
retail rack layout problem (RRLP). The RRLP is a special case of the FLP that identifies the optimal
placement of rectangular racks within a constrained space in order to maximize the total rack exposure.
Given the complexity associated with solving the corresponding mixed-integer non-linear optimization
model, we propose a heuristic-based solution approach using particle swarm optimization (PSO). We also
conduct a sensitivity analysis of the near-optimal layouts generated by the PSO to variation in system and
shopper parameters. Based on this analysis, we derive several managerial insights. For instance,
orientation of racks in the column closest to the shopper should be less than the degree of horizontal head
movement exhibited by the shopper. Further, multiple competitive layout designs exist that offer similar
exposure values for a given system configuration, meaning designers have options when selecting a
layout that fits their unique preferences and needs.
In what follows, we first review existing literature in retail layout in Section 2. We then present
an optimization model for the RRLP in Section 3, and a particle swarm optimization based heuristic to
solve the model in Section 4. We evaluate the performance of this heuristic in Section 5 and discuss
results of our experimental study in Section 6. Lastly, we summarize our findings and contributions in
Section 7.

2 Previous Research
It is well-known that the traditional FLP of finding the optimum placement of rectangular machines,
departments, or facilities in order to minimize distance or material flow is a challenging, non-linear
problem that is computationally intractable, in part due to its combinatorial nature. Various models and

3
approaches to solve FLP for manufacturing and warehousing facilities have been discussed in the
literature (Meller and Gau, 1996; Singh and Sharma, 2006; Gue and Meller, 2009; Kundu and Dan,
2012). However, literature on retail facility layout is fairly limited.
Retail layout involves two levels of decision making, physical layout of the store (departments,
aisles, and racks) and shelf space allocation. The shelf space allocation problem typically involves
deciding how much shelf space is allocated to a product category and can be viewed as a hierarchy of
decisions that first assigns product categories (already chosen as part of the assortment planning problem)
to departments or areas within the store, then to an area of a rack located in the department (i.e., middle of
rack or end) and finally allocating rack shelf space (defined by number of facings of a specific product) to
specific products and onto the racks themselves. This decision making problem has been the subject of
research in retailing and marketing literature for decades (Corstjens and Doyle, 1981; Botsali and Peters,
2005; Irion et al., 2012; Flamand et al., 2016). However, the shelf space allocation decision involves
aggregate estimates of the likelihood of a rack section being exposed to a shopper, which is only
determined after the physical layout of the store (especially, the orientation of racks) has been generated.
That is, the physical layout problem precedes the shelf space allocation problem. We now summarize
what little literature exists relevant to physical store layout before discussing the gaps that motivate our
study.
Peters et al. (2004) were the first to address the retail layout problem and considered three types
of retail layouts: aisle, hub-and-spoke, and serpentine. They developed a department location assignment
model to assigning departments (i.e., groups of products) to locations in order to maximize impulse
purchase revenue. They used a ‘visits’ measure to determine impulse purchases, which assumed a product
could become an impulse purchase if it was seen as the shopper passed through or visited the location
containing the product as they acquired their planned purchases. Peters et al. modeled the vision of a
travelling shopper as perpendicular sight; i.e., they assumed that if a shopper is next to a product, it must
be visible. While this is a reasonable approximation for visibility, the human eye’s field of view is
typically forward looking.
Yapicioglu and Smith (2012a) focused on optimally designing a departmental block layout
(departments and aisle space) that maximized the store's revenue, which involved determining the area
and location of departments within a racetrack structure. They too assumed that impulse purchases were
more likely to occur in departments that were frequently visited by shoppers. To account for this, a
mechanism was employed to locate departments with high impulse rates to areas with high shopper
traffic. Yapicioglu and Smith (2012b) reformulated this problem as a bi-objective problem where the first
objective maximized store revenue based on department layout and the second maximized the satisfaction

4
of departmental adjacencies. Visibility was defined broadly as the area within the vicinity of a shopper;
high traffic zones mean many shoppers will be in the area, therefore products in the area will be visible.
While the above approaches recognize the importance of visibility or exposure when designing a
retail layout, these measures are aggregate, and thus limited. Neither measure accounts for the shopper’s
field of view, the dynamics that ensue when a shopper walks past display racks, or possible obstructions
that could limit or eliminate product visibility. Good store layouts are those that place the greatest amount
of products within the shopper's field of vision for the longest amount of time (Dunne et al., 1995; Burke,
2005; Underhill, 2008).
To address some of these aspects, Sorensen (2008) presents a visibility-based attractiveness
measure that considers the size of marketing media in a store, its orientation to the shopper, and its
distance to the shopper in order to evaluate and compare the ability of the media to grab the shopper's
attention. Similarly, Lu and Seo (2015) present an ArcGIS-based computational approach to model
visibility access and exposure measures that consider the human field of vision in order to describe a
retail layout and compare it with others. Their measure is similar to the one we use, however, their
approach (which is strictly descriptive) is product specific, requiring knowledge of product location and
size, and does not consider bidirectional shopper traffic. Our recent work presented a much more refined,
analytical and computational, approach that did not require foreknowledge of products and human vision
(specifically, depth of vision and viewing angle) into the analysis rather than making generalized
assumptions about what is or can be seen. Our work also accounted for unidirectional and bidirectional
traffic, and obstructions due to other racks as the shopper travelled through the system (Mowrey et al.,
2017).
Although research exists on departmental layout and product placement on racks, we are unaware
of any work that bridges the gap between the two. Our contributions concentrate on the intermediate
layout problem, optimizing the rack layout within a department in order to maximize its exposure We use
the measure of visibility presented in our earlier work and refer to it as exposure. Exposure has been
shown to be a predictor of product engagement, making it immediately relevant to the retail world as a
precursor to product sales (Suher and Sorensen, 2010; Lu and Seo, 2015). We now present a
mathematical model and propose an efficient algorithm to solve this complex problem.

3 An Optimization Model for RRLP

The retail rack layout problem (RRLP) can be defined as the optimal arrangement of racks within one or
more columns in order to maximize the exposure of the rack facings to a traveling shopper. It is important
to distinguish between what a shopper is able to see from what a shopper actually does see. The latter is
beyond modeling, because it so strongly depends on the environment and the behavior of the individual

5
shopper in the moment. Therefore, we focus on the former—what could and could not have been seen by
a shopper. But even this objective requires careful definition. For example, if a specific location on a shelf
passes within the field of regard for a prototypical shopper for 1 millisecond, can we say that the location
was exposed? In what follows, we define an arbitrary minimum time of exposure, t, for a location to have
been exposed. The total exposure across all locations during a pathway is represented as Et. The best
value of this minimum time, which depends on what a retailer believes is necessary for the shopper to see,
perceive, and act, is beyond the scope of our research (see Section 6 where we experiment with two
values of t).

The RRLP is similar to the FLP in that it seeks to arrange rectangular objects (racks vs. departments
or machines) within a defined space without overlapping. The aisle structure in RRLP is simply the
spacing between the racks, so it can be sufficiently designed by including an additional spacing
requirement between racks. We make the following assumptions in developing our optimization model:
• the shopper walks in the middle of the main aisle and travels in a straight line.
• the shopper can turn their head and eyes combined no more than Φ in one direction (2Φ total)
from her direction of travel.
• the shopper can discern targets (symbols, logos, packaging style/color) up to a distance DOV
away.
• two, 2 ft. wide x 4 ft. long, rack segments are placed back-to-back forming a 4 ft. x 4 ft. segment
(which is in line with what we observed at most retail stores in our geographical region) and can
be attached end-to-end to form longer racks.
Figure 2 shows a schematic of a 2-column rack layout as an illustration of how we represent a
rack layout in our model; Tables 1 and 2 list the parameters and decision variables. A column of racks can
be broadly defined as an arrangement of racks that all align along the x or y-axis, have the same
orientation (θ) and are separated by a minimum spacing (Ac) to allow adequate room for shoppers to
travel. In this 2-column layout within a predefined space (L ft. long and W ft. wide), Column 1 has 5 racks
oriented at θ1 = 45°, each having 2-6 rack segments (or 8-24 ft long). Column 2 has 6 racks oriented at
θ2 = 90°, each 6 rack segments long. Columns 1 and 2 are separated by a cross-aisle that is Ca=5 ft wide.

6
 rikl

III xik" , yik"

IV

II
I

xik' , yik' xik' , yik' xik" , yik"

Figure 2: Example 2-column rack layout

Table 1: Parameters used in the model

Notation Definition
W width of the section (ft)
L length of the section (ft)
i, j index for rack in a column; i = 1, 2, ..., N, where N is the allowable number of racks in a
column
k index for column in the layout; k = 1, 2, ..., B, where B is allowable number of columns in
the layout
Ac width of the shopping aisle (ft) between racks in column k
Rs length of a rack segment (ft)
Rw width of a rack’s end cap display (ft)
Ca width of the cross-aisle between columns
Am width of the main aisle
Da perpendicular distance of closest corner to the shopper's directional path
α (1-α) percentage of traffic flowing in the primary (opposite) direction of travel
Φ shopper’s angular limit of vision corresponding to half of FoR
DOV shopper’s depth of focused vision, radius of FoR
P90 display (perimeter) of an equivalent 90°-rack layout for given system parameters (ft)
ω maximum allowable loss in display from the equivalent 90°-rack layout ( 0 ≤ ω ≤ 1 )
M large number

7
 Table 2: Decision variables used in the model
Notation Definition
length of a rack exposed for > t seconds on one side of the pathway for the α direction (ft)

length of a rack exposed for > t seconds on one side of the pathway for the 1-α direction (ft)

ck 1 indicates column k exists, 0 otherwise

wk width of column k (ft)
θk rack orientation measured clockwise from primary direction of travel for column k
vik 1 indicates rack i in column k exists, 0 otherwise
rikl number of rack segments in rack i (ft) in column k
x ik' x-coordinate of corner between face I and II
y ik' y-coordinate of corner between face I and II
"
x ik x-coordinate of corner between face II and III
"'
y ik y-coordinate of corner between face II and III
θ '
k adjusted rack orientation
tk 1 indicates θk is greater than 90°, 0 otherwise
tk' 0 indicates θk = 0°, 1 otherwise
tk" 0 indicates θk = 90°, 1 otherwise
x
zijk 1 indicates rack j is to the right of rack i in column k, 0 otherwise
y
zijk 1 indicates rack j is above rack i in column k, 0 otherwise

We propose the following optimization model to determine the optimal rack layout that
maximizes total exposure on one side of a pathway.

Maximize: t
E tot = α E αt + (1 − α )E 1t − α

α k (
Subject to: Et = f θ , Da, Ac, rl , r l , Rw, Φ, DOV, t
ik jk )
( )
(1)
E1t−α = f 180° − θ j , ( Am − Da), Ac, rikl , rjkl , Rw, Φ, DOV, t

b
 b 
∑k =1  k =1
∑
ck wk + Ca ck − 1 ≤ W

(2)

c k ≤ wk ∀k (3)

c k ≥ vik ∀i, k (4)

8
v ik M ≥ rikl ∀i, k (5)

x ik" = x ik' + rikl Rs sin θ k' ∀ i, k (6)

y ik" = y ik' − (2 t k − 1 )rikl Rs cos θ k' ∀ i, k

(7)

v ik (1 − t k )Rw cos θ k' ≤ x ik' ∀ i, k (8)

(1 − t k )x ik" (
+ t k x ik" + Rw cos θ k' ≤ w k ) ∀ i, k (9)

(1 − t k )y ik" + t k y ik' + Rw sin θ k' ≤ L ∀ i, k (10)

xik' +
(
tk" Rw + 1 − tk" Rsrikl + Ac )
≤ x 'jk + M 1 − zijk ( )
( ) ∀ k , i, j i ≠ j
x
(11)
cosθk + 1 − tk
' "

yik' +
(
tk' Rw + 1 − tk' Rsrikl + Ac )
≤ y'jk + M 1 − zijk ( )
( ) ∀ k , i, j i ≠ j
y
(12)
sinθk + 1 − tk
' '

x
zijk y
, zijk ≤ vik ∀ i , k (13)

x
z ijk + z xjik + z ijk
y
+ z yjik ≥ vik v jk ∀ k , i, j i≠ j (14)

θ k' ≤ 90 ∀k (15)

θ k' = t k (180 − θ k ) + (1 − t k )θ k ∀k (16)

(
θ k M 2 t k' − 1 ≥ t k' ) ∀k (17)

(90 ) (
− θ k' M 2 t k" − 1 ≥ t k" ) ∀k (18)

∑∑ v (Rsr )
b n
2 ik
l
ik + Rw ≥ (1 − ω ) P90 (19)
k =1 i =1

0 ≤ θ k < 180 °
wk , xik' , xik" , yik' , yik" ≥ 0
{
rikl ∈ 0, Z + } (20)
x
ck , vik , zijk , zijky , t k , t k' , t k" ∈ {0,1}

9
 The main decision (or system design) variables in this model include the width of column k (wk),
l
the orientation of racks in column k (θk), the number of fixed-length rack sections for each rack ( rik ), and
the location of each rack’s lower, interior corner ( x ik' , y ik' ). The objective of the model is to maximize the
total exposure of a section ( ) that is located on one side of the pathway while considering both
directions of travel. Constraints (1) help estimate and . Given that the exposure estimate for a
general rack layout is not in a closed form, we use the algorithm presented in Mowrey et al. (2017) to
estimate the exposure values. This algorithm accounts for the human’s field of view (both DOV and Φ),
rack layout and inherent obstructions, and length of time of the exposed location. As shoppers travel
bidirectionally along the aisle, we estimate exposure of the left or right side of the pathway from each
direction separately and then take the weighted average based on the traffic flow parameter, α. Our
approach is applicable to two-way travel because it considers exposure experienced by the shopper from
both directions when selecting the layout that maximizes total exposure. It is important to note though,
that the exposure experienced in the α direction will not necessarily equal the exposure experienced by
the 1-α direction.
Embedded in this optimization model is a design problem that is jointly solved in order to
estimate exposure. This design problem identifies the number of rack-columns (and their associated
widths) that exists within the store sub-section, and then determines the maximum number of racks that
can be feasibly placed.
Constraint (2) ensures that the available width of the section is not exceeded and is utilized as
either available space for racks within each column or as a cross-aisle between columns. Constraint (3)
requires a column’s width to equal 0 when it does not exist, while Constraints (4) and (5) prevent racks
from existing in a column that does not exist. Constraints (6) and (7) calculate the x and y coordinates of
the lower, outside corner point (corner between Faces II and III) based on the coordinates of the lower,
interior corner point (located between Faces I and II) (see Figure 2). Constraints (8), (9), and (10) ensure
that rack i is defined within the bounds of column j, while Constraints (11) - (14) ensure that rack j is
sufficiently east (right) or north (above) of rack i so that minimum shopping aisle distance is maintained.
Constraints (15) - (18) allow the model to adapt to all rack orientations including 0° and 90° orientations.
Constraint (19) limits the loss in display of a layout to ω of P90, ω∈(0, 1); P90 is simply the total display of
the equivalent 90°-rack layout, which is straightforward to obtain and can be estimated a priori. Finally,
Constraints (20) indicate bounds on the decision variables.
Our model is inspired by the MIP model for the traditional FLP in the continuous space
(Montreuil, 1990; Tompkins et al., 2010). Montreuil’s MIP is known to be a challenging combinatorial
optimization problem and only problems with less than 15 departments have been solved optimally

10
(Meller and Gau, 1996; Liu and Meller, 2007; Kulturel-Konak and Konak, 2013). Our RRLP, beyond
inheriting the complex combinatorial challenges from Montreuil’s MIP, also includes additional features
such as nonlinear constraints and an exposure evaluator that is not a closed-form expression. Though a
linearized form of our problem could, in theory, be formulated, the large number of binary variables
required to solve it would likely lead to computational intractability. We, therefore, pursue a
metaheuristic approach to obtain high quality solutions to the RRLP, similar to how FLPs are typically
solved (Montreuil et al., 2004; Liu and Meller, 2007; Bozer and Wang, 2012; Jolai et al., 2012; Kulturel-
Konak and Konak, 2013).
Our proposed solution approach uses particle swarm optimization (PSO) to solve the RRLP. The
primary advantage of PSO is its natural ability to search within a continuous space, eliminating the need
to devise a method of mapping the continuous FLP to a discrete representation as done in sequence-pair
(Meller et al., 2007) and graph-pair representation (Bozer and Wang, 2012). Further, PSO is robust, has a
simple structure, is easy to implement, and has been shown to obtain high-quality solutions to nonlinear,
non-differentiable, multi-modal problems in a relatively quick time (Shi and Eberhart, 1998; Bansal et al.,
2011; Jolai et al., 2012). We next present our PSO and details of implementation.

4 A Particle Swarm Optimization (PSO) Based Heuristic for the RRLP

PSO is a nature-inspired evolutionary computation method introduced by Kennedy and Eberhart (1995)
that optimizes continuous nonlinear functions. PSO utilizes a population of particles that explore the
solution space by simulating social behavior that resembles a flock of flying birds or a school of
swimming fish. Each particle represents a potential solution to the problem and evolves through
cooperation and competition among the particles themselves through generations (Shi and Eberhart,
1998). As the swarm searches over time, individuals are drawn towards each other's successes, resulting
in convergence in optimal regions of the search space (Clerc and Kennedy, 2002).
Our proposed PSO implementation for the RRLP accepts a set of initial feasible solutions
generated using a straightforward greedy approach before improving them. It uses two key subroutines
for every particle at every iteration: (i) layout design, which builds a feasible arrangement of racks in each
column for a given solution (first used to generate initial feasible solutions and then in improving these
solutions) and (ii) exposure evaluation, which estimates the exposure for this feasible rack layout by
simulating a shopper walking through the system. We then use personal and neighborhood best solutions
to update the particles after each iteration. The high level structure of the PSO is as follows:

11
 Initialize population of particles with random positions and velocities
Do
For each particle:
Evaluate feasibility of the encoded solution
If Feasible:
Convert encoding to rack layout (Rack design subroutine)
Evaluate Constraint
If satisfies constraint:
Evaluate fitness function (Exposure subroutine)
If fitness value is greater than global best:
Set current solution as global best
If fitness value is greater than neighborhood best:
Set current solution as neighborhood best
If fitness value is greater than particle best:
Set current solution as particle best
Else:
Reject solution
Else:
Reject solution
End
For each particle:
Update particle velocity
Update particle position
End
Until termination criterion is met

Below we first discuss how we represent a solution for the RRLP in PSO and then summarize the
key steps in each subroutine.

4.1 RRLP Solution Representation in PSO

Our proposed PSO-based solution approach searches the multi-dimensional solution space to determine
the number of columns (k), their associated widths (wk), and the corresponding rack orientations (θk). We
used a swarm of 20 particles in line with Clerc and Kennedy (2002). In our experiments, each particle
represents a solution within the search space and is treated as a point in multi-dimensional space where
each dimension represents a different variable or parameter in the solution.
We represent a solution as a string of real numbers, which consists of 2K-1 elements, where K is
the number of possible columns of racks. The first K elements represent the rack angle (θk) for each
possible column (k) and the following K-1 elements represent the column width ratios (WR), defined as
the ratio of the width of column k (wk) to available space. The number of WR is one less than the number
of columns; i.e., for a 2-column layout, there is only one WR, but for a 3-column, there exist two, WR1
and WR2. To illustrate this encoding, assume that the available width of the store section where the racks
would be laid out is W=40ft and the spacing between all columns is Ca=4ft. If K=3 columns, the solution

12
representation would include 5 elements. A possible solution representation for a particle could be {θ1, θ2,
θ3, w1/W, w2/(W-w1-Ca)} = {45,90,30,0.5,0.5}. From the width ratios, we can calculate the column widths
as w1=WR*(available space)=0.5*40=20ft, w2=0.5*(40-20-4)=8ft, and w3=40-20-4-8-4=4ft. Notice that
the available space reduces for each subsequent column as it must account for the space required by the
cross-aisle that is located between 2 columns. Feasible solutions will satisfy 0° ≤ θk < 180° and 0 ≤ WR ≤
1 for each column. Solutions that are not feasible are not evaluated and not considered as either personal
or local best (Kennedy, 2007).

4.2 The Subroutines: Layout Design and Exposure Evaluation

The layout design subroutine interprets the encoded solution into a physical layout. This routine first
determines the number of columns, k, and their associated widths, wk, in the layout. Next, it fits as many
racks at θk orientation as is feasible into each column. One approach to do this is to assume that the
starting position of the first rack in a column is always a corner of the available space. We use the lower
left-hand corner as the starting point for acute orientations and the upper left-hand corner as the starting
point for obtuse orientations. Each rack is always designed for its maximum allowable length without
violating the boundaries of the column. Once the layout is designed from the solution encoding, it is
evaluated with respect to the model constraints. We calculate the total amount of display for each layout
to see if it satisfies the minimum requirement. Solutions that meet the display requirement are passed on
for exposure evaluation. If the minimum display constraint is not met, the solution is assigned an
exposure value of 0, ensuring that it will not be selected as either a personal or neighborhood best
solution.
The second subroutine, exposure evaluation, accepts each feasible layout solution and evaluates it
in accordance with the specified fitness function. In our case, our fitness function, E, is not a closed form
solution. To determine the amount of exposure a layout produces for a given shopper’s Φ and DOV, we
use the slicing algorithm (Mowrey et al., 2017).

4.3 Personal and Neighborhood Best Solutions

Each particle (corresponding to a feasible rack layout) retains the memory of its personal best solution
(pid) and its neighborhood best solution (gid), and uses this information to update its trajectory within the
search space and determine its solution in the next iteration. Obviously, pid is the current best value of
dimension d of the ith particle; gid is the best value among all the particles in the neighborhood.
A popular neighborhood structure used in PSO is the star, or global best topology, which is known
to produce rapid swarm convergence, and so is frequently used throughout literature (Yapicioglu et al.,
2007). Unfortunately, when all particles receive the same information, they all tend to travel to the same

13
location in space. More often than not, the region of the search space that produces the first good result
becomes an attractor for the entire swarm and often leads to early convergence (Kennedy, 2007). To
avoid this pitfall, we utilize a ring, or nearest neighbor topology (referred to here as local best) using a
neighborhood size of 3 particles (Kennedy, 2007; Yapicioglu et al., 2007). In our PSO, proximity was
determined based on the particle's sequence in its initial generation. The 3-particle neighborhood was
defined as particle n-1, particle n, and particle n+1. The first and last particle were treated as neighbors. In
this neighborhood structure, each particle only communicates with its two nearest neighbors. The local
best topology is good for functions with local optima because it slows the spread of solutions through the
swarm and increases the probability of finding the global optimum. After updating the personal and local
best solutions, the algorithm updates each particle's velocity and position.

4.4 Solution Updating

A particle's position within the search space at iteration (t) is based on its previous position and its
updated velocity as shown in (21). The particle velocity, specific to each dimension of the solution, is
given in (22) and is composed of three primary parts; velocity from the previous iteration, cognitive or
selfish influence, which uses the particle's personal best to simulate the private thinking of the individual
particle, and social influence, which represents collaboration among the particles in the swarm using the
local best (Shi and Eberhart, 1998; Jolai et al., 2012; Ozturkoglu et al., 2014).

xidt = xidt−1 + vidt (21)

( ( ) (
vidt = K vidt −1 + c1r1 pid − xidt −1 + c2 r2 gid − xidt −1 )) (22)

The movement toward personal or local best is determined by r1 and r2, which are both random
numbers, uniformly distributed [0,1]. To prevent an explosion of the particle’s velocity towards infinity
due to this random weighting, Eberhart and Shi (2000) proposed the use of a liberal velocity limit, Vmax. In
θ
accordance with their recommendation, we set -180° ≤ Vmax ≤ 180° for orientation components (θ) and
WR
-1≤ Vmax ≤ 1 for width ratio components (WR). We set acceleration constants c1=c2=2.05 and constriction

coefficient K=0.7298 as suggested by Clerc and Kennedy (2002). The PSO terminated after 100 iterations
with no improvement in the global best solution. Particle convergence was observed.

5 Performance of PSO-based Heuristic

To evaluate the performance of the proposed PSO-based heuristic we compared the solutions generated
by the PSO to a grid search algorithm that conducts an exhaustive search of a manually-specified subset

14
of the solution space. We evaluated the solution quality and run times over three different store section
sizes (22 ft x 40 ft, 40 ft x 40 ft, and 58 ft x 40 ft), and three different Φ values (45º, 60º, and 90º)
(Table 3). The grid search algorithm assumed a maximum of 2 rack columns within the section and
evaluated every integer combination of θ1 and θ2 (in 5° rack angle increments) for width-ratio 0.1≤WR≤1
at each 0.1 ft increment. In the PSO we allowed θ1 and θ2 to assume values to the 0.1° rack angle
increments; WR was allowed to assume any decimal value. The CPU times are based on a personal
computer with an Intel(R) Core(TM) i7-3770 CPU 3.4 GHz processor and 16 GB RAM; the PSO was
coded in Python.
As you can see from Table 3, our PSO found solutions as good or better than those discovered by
the course grid search within a fraction of the grid search's time. Note that while we let the PSO evaluate
much finer rack angles and various WR values (possibly finding better solutions than the grid search
algorithm), the fact that it found such solutions so quickly provides some evidence that it is an effective
solution to this fairly complex RRLP. We also verified that the final solution generated by the PSO
was robust to the choice of initial feasible particles on several configurations (results not shown for
brevity). We, hence, use this heuristic in our experimental study to derive managerial insights.

Table 3: Performance of PSO against Grid Search ( E tot0 = amount of rack exposed for >0 seconds)

22' x 40' 40' x 40' 58' x 40'

0 0 0
Etot Time % of Grid Etot Time % of Grid Etot Time % of Grid
(min) Time (min) Time (min) Time
(ft) (ft) (ft)
Grid 38.55 898.88 68.60 1793.98 95.30 2758.50
Φ=45°

PSO 39.75 32.58 3.62% 70.88 53.10 2.96% 103.38 100.06 3.63%

Grid 54.25 867.20 97.40 1795.75 134.85 2641.12

Φ=60°

PSO 55.38 29.20 3.37% 98.88 74.11 4.13% 147.25 120.02 4.54%

Grid 72.00 899.81 136.25 1869.45 204.35 3259.58

Φ=90°

PSO 72.00 29.95 3.33% 136.62 62.22 3.33% 204.75 109.45 3.36%

6 Experimental Study
In our experiments, we used 20 particles in the PSO. The first 19 particles were randomly generated;
orientations (θ) were randomly selected between the values of 0 and 180º, and width ratios (WR) were
randomly selected between the values of 0 and 1. In the instances where a minimum display constraint
was enforced, randomly-generated seeds that produced layouts with insufficient display were discarded
and new, random solutions were generated until the display constraint was satisfied. To ensure the PSO

15
always found a layout that was as good or better than the 90º equivalent, the last (20th) particle was
specified to be a one-column, 90º layout. Termination criteria was specified as 100 iterations with no
improvement in the global solution. Depending on the specified intensity and display constraints,
infeasible solutions were encountered approximately 20-27% of the search time.
In the remainder of this section, we analyze the best rack layout designs generated by the PSO-
based heuristic. We consider a 44 ft x 44 ft section and assume bidirectional shopper traffic (α=0.5) with
shoppers traveling 5 ft away from the nearest point on a rack. We allow up to three columns of racks
(B=3) assuming shopping aisle and cross-aisle widths are 4 ft each. Considering that the degree with
which a shopper scans the environment in front of them depends on many factors such as mood, task,
stimuli, and familiarity with the store, we evaluate two levels of head and eye rotation: small movement,
Φ = 45º (one side from the forward line of sight), and large movement, Φ = 90º. The depth of focused
vision (DOV) was set to 25 ft and 65 ft, and the pathway exposure was estimated from 10 ft prior to the
first rack (similar to entering a store section from an intersection).
For each shopper-system combination, the PSO generated thousands of feasible layouts for the
specified store sub-section that is to be laid out. Each layout generated by the PSO is associated with its
corresponding display (total linear feet of rack perimeter that products may occupy) and exposure.
Because the 90°-rack angle is the most common orientation in retail stores, we report %-change in
exposure and %-loss in display compared to an equivalent one-column 90°-rack layout – a one-column
layout of 90º racks that would fill the specified design space while adhering to cross-aisle (Ca) and
shopping aisle (Ac) spacing – for a given system configuration.
Figure 3 shows all feasible layouts generated for Φ=45º and Φ=90º when DOV=65ft across
multiple values of 100ω% (Constraint 19 in the model). The dotted line indicates the maximum positive
%-change in exposure over 90º-rack layout that could be achieved across various values of loss of
display. Having derived this relationship, we then explore the impact of the shopper’s FoR.

16
 (a) Change in exposure over 90º
(b)
Figure 3: Loss of display and change in exposure, E tot compared to 90º layout
0

for all feasible layouts at (a) Φ = 45º and (b) Φ = 90º when DOV = 65 ft

6.1 Impact of Shopper’s FoR Visual Parameters: DOV and Φ

Figure 4: Penetration of the FoR into the
44' x 44' design section (with example
90° racks) for Φ=45º (dotted line) and
90º (solid line) and DOV=25 ft (black
line) and 65 ft (grey line)
We model two scenarios with FoR, one that partially
penetrates the 44 ft x 44 ft design section (i.e., DOV = 25 ft)
and the other that fully (or almost fully) penetrates it (i.e.,
DOV = 65 ft), for two shopper head movements, Φ = 45°
and 90° (Figure 4).
We first ran the 4 combinations of Φ and DOV, and
then analyzed all solutions whose total exposure was within
1% of the maximum exposure found. Figures 5-7 illustrate
our findings for the case when ω=1 (Constraint 19 in the
RRLP model). Based on these results, we observe the
following:

Obs. 1: For a given DOV, when shoppers make small head turns, angled rack layouts can offer
substantial increase in exposure.
Figure 5a and 5b illustrate solutions generated with small shopper head movements (Φ = 45º). We see
that with small head turns, gains in exposure can be quite substantial. Solutions generated with small head
movements result in the highest increase in E tott ; 226.5% with DOV = 65 ft and t > 0 sec (126.6% for E tot2

17
), and 67.3% with DOV = 25 ft (65.2% for E tot2 ). This is for two reasons: first, angled-rack layouts align
better with the shopper's head movements and second, for small head movements, only a small percentage
of racks are exposed when racks are oriented at 90º (i.e., the baseline exposure at 90° itself is fairly low).
The joint effect is that E tott increases substantially. In contrast, large head movements limited the

increase in exposure of angled racks over 90°; ~21% for t > 0 sec and ~61% for t > 2 sec.

(a) θ1 = 44.3º (b) θ1 = 30.4º

for Φ = 45° and t > 0 sec for Φ = 45° and t > 2 sec

(c) (θ1, θ2, θ3) = (72º, 90º, 90º)

for Φ = 90° and t > 0 sec

Figure 5: Example designs for DOV = 65 ft

With large head movements, exposure of the 90º racks is much higher. So even though angled rack
layouts offer a further increase in E tott , the %-increase from 90º is limited. For this reason, alternate rack
layouts may be more desirable in settings associated with small head movements. For instance, marketing
literature has identified, but not quantified in terms of head movement, several motivational and
situational factors that affect shopper behavior (Ng, 2003). It is conceivable that a utilitarian shopper

18
familiar with the store layout and under time constraints may exhibit small head movements, while a
leisure shopper who is unfamiliar with the layout and not under any time constraints may exhibit large
head movements.

Obs. 2: Rack layouts are sensitive to the amount of FoR penetration (i.e., DOV of 25 ft vs. 65 ft).
We know that before a location can be exposed, it must first exist and, if so, then it must fall within the
FoR. Consequently, best layouts are those that strategically place as many racks as possible within the
FoR in order to maximize total exposure for at least t seconds ( ). With a DOV = 65ft, the entire
design section falls within the shopper's FoR as they travel (i.e., full penetration) and contributes to the
overall exposure. Our experiments showed that when shoppers made small head turns (Φ = 45º), the best
layouts increased total exposure substantially (by 227% for by 120% for considering a walking
speed of 2.5 fps) over the corresponding 90º-layout and resulted predominantly in 1-column layouts. The
rack angles for these 1-column layouts ranged between 30.4°-44.8° depending on the value of t; see
Figure 5 (a) and (b). Layouts corresponding to large head turns (Φ = 90º) increased total exposure by 20%
for and resulted in 3-column layouts (Figure 5c); increased by 60% and resulted in 1-column
layouts with racks oriented at 89º.
These results indicate that a large FoR provides more opportunities for racks to be exposed (both
in the ability to see around an obstructing corner and depth of vision) over the course of the shopper's
path, making multiple column solutions possible. For multiple-column solutions that were prominent
when t > 0 seconds, we noticed that while the racks in the first column were non-orthogonal, the racks in
the 2nd and 3rd columns were often orthogonal, potentially allowing the recapture of display reduction in
the 1st column. The staggering of racks in the 1st column with respect to the 2nd column, allows exposure
of the racks in the latter (Face I and II as defined in Figure 2) as the shopper travels between racks in the
first column. With a DOV = 25 ft, the FoR only partially penetrates the 44 ft x 44 ft section. In that case, it
would make sense to orient the racks in the sections in which the FoR penetrates; the rest of the area in
which the FoR does not fall will not contribute to total exposure and so could have racks placed at any
angle as long as the total space and minimum display constraints are not violated. For such a situation,
our experiments demonstrate the existence of many competitive designs with varying number of columns.
0
The best layouts for small head turns (Φ = 45º) increased total exposure (by 67% with Etot and 28% with
2 0 2
Etot ) over 90º layout, while the increase in exposure was limited (18% with Etot and 0% with Etot ; 1-

column 90° layout was still the best) with large head turns (Φ=90º).
Figure 6 shows 1-, 2- and 3- column solutions for DOV = 25 ft that are all equivalent in terms of
total exposure. In the 2- and 3-column designs (Figure 6 (b) and (c)), both θ2 and θ3 assumed values

19
between 0º and 180º. We illustrate example layouts from these results in Figure 7. Note that if a pair of
racks (back-to-back) to be oriented at a specific angle cannot fit within the space, then it may result in an
empty space (e.g., top section of column 1 in Figure 7 (a), (b) and (c)).

θ1 θ1 θ2

t > 0 sec: o Φ = 45° Φ = 90°

t > 2 sec: o Φ = 45° Φ = 90°

(a) 1-column solutions

θ3 θ3
180
165
150
135
120
105
θ2
θ2

90
75 θ2 θ1
60
45
30
15 t > 0 sec: o Φ = 45° Φ = 90°
0 t > 2 sec: o Φ = 45° Φ = 90°
0 15 30 45 60 75 90
θθ1
1
(c) 3-column solutions
t > 0 sec: o Φ = 45° Φ = 90°
t > 2 sec: o Φ = 45° Φ = 90°

(b) 2-column solutions

Figure 6: Best rack angles for DOV=25 ft solutions

Obs. 3: Rack angles for the first column (θ1) tend to be at angles less than Φ, independent of DOV.
At both DOV values, the rack angles in the first column that produce the highest amounts of exposure are
less than or equal to their respective Φ. Notice that a range of values <45º seem to dominate designs for
small head movements (Φ = 45º) and a range of values <90º dominate designs for large head movements
(Φ = 90º). Similar results were seen at t > 0 for Ф = 22.5º and 67.5º where rack angles in the first column
were observed to range 0º to 20.3º and 62.1º to 66.7º, respectively. When we impose the constraint that
exposed locations must be seen for a minimum amount of time (say, t > 2 seconds vs. t > 0), ‘fringe’

20
locations with low intensity no longer count toward exposure and the best range of rack angles shifts from
43.5°-44.8° for 0
E tot to 24.5°-27.3° for 2
E tot when Φ = 45º.
These findings are different than what was observed in a model with unconstrained space where
30º racks produced the most exposure over 90º (Mowrey et al., 2017). Intuitively, exposure should
increase as we allow the racks to be more angular. But this angularity increases loss in display, which
then reduces the %-increase in t
E tot , eventually making the design worse than the equivalent 90º-layout.
Clearly, the shopper's head and eye movements (horizontal scanning behavior) strongly influence the rack
layout (at least those closest to the shopper).

(a) (θ1, θ2) = (44.8º, 90º) (b) (θ1, θ2, θ3) = (85.3º, 90.0º, 0º)
for Φ = 45° and t > 0 sec for Φ=90° and t > 0 sec

(c) (θ1, θ2, θ3) = (26.1º, 90º, 90º) (d) θ1 = 90.0º

for Φ = 45° and t > 2 sec for Φ = 90° and t > 2 sec

Figure 7: Example designs for DOV = 25 ft

The best layouts appear to be relatively robust to changes in α (see Table 4); Observations 1-3
appear to hold. Angled racks still offer substantial increases in exposure when shoppers make small head
turns (Observation 1). The best layouts associated with full penetration of the design space (DOV = 65 ft)

21
favor one-column designs, while single and multiple column designs are competitive for partial
penetration (DOV = 25 ft) (Observation 2). The orientation of the racks in the column closest to the aisle
is still less than or equal to Ф (Observation 3), although there does appear to be a slight tendency toward
more acute orientations in θ1 when α = 1 (unidirectional traffic) and the FoR fully penetrates the design
space (DOV = 65 ft).

Table 4: Best solutions and their increase in exposure (E) over the 90º-rack layout
for multiple α values when t > 0 sec.

α=1 α = 0.75 α = 0.5

Range No. of Range No. of Range No. of
DOV % ΔE % ΔE % ΔE
of θ1 Columns of θ1 Columns of θ1 Columns
42.7º- 42.7º- 43.5º-
25 ft 160.7% 1, 2, 3 115.5% 1, 2, 3 67.1% 1, 2, 3
Ф = 45º

44.8º 42.9º 44.8º

38.8º- 42.7º- 44.1º-
65 ft 450.5% 1 343.0% 1 226.3% 1
44.9º 44.3º 44.4º
78.1º- 77.8º-
25 ft 52.8% 90º 1 35.0% 1, 2, 3 18.1% 1, 2, 3
Ф = 90º

86.4º 89.9º
72.2º- 78.2º- 71.0º-
65 ft 55.6% 1 30.4% 1 17.8% 2, 3
77.7º 88.1 89.6º

6.2 Impact of Allowable Loss of Display

While the above observations were for situations where we set ω=1, it is likely that retail managers would
prefer to limit the loss in the display space compared to their current layout (often 90° rack angles) in
exchange for increased exposure. Therefore, we explore the impact display has on designs generated for
the same Φ and DOV combinations previously discussed. For each combination, we plotted every feasible
solution generated by the PSO, similar to Figure 3 and then identified the maximum exposure increase
achieved for each corresponding loss of display. Recall that loss of display is the difference in linear feet
of display between the alternative layout and the traditional 90º layout that would be associated with a
given system. Figure 8 illustrates the relationship between %-loss in display and %-change in t
E tot

compared to equivalent 90° racks across all Φ and DOV combinations.

22
 250% (DOV, > t sec) 250% (DOV, > t sec)
(65ft, 0 sec) (65ft, 0 sec)
Change in exposure over 90º

Change in exposure over 90º

200% 200%
(65ft, 2 sec) (65ft, 2 sec)
150% 150%
(25ft, 0 sec) (25ft, 0 sec)
100% (25ft, 2 sec) 100% (25ft, 2 sec)

50% 50%

0% 0%
0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100%
-50% -50%

-100% -100%
Loss of Display Loss of Display

(a) Φ = 45° (b) Φ = 90°

Figure 8: %-change in t
with respect to 90° orientation for various %-loss in display (dotted line
E tot
represents t > 2 sec and solid line represents t > 0 sec)

Obs. 4: Percentage change in exposure of an angled layout over 90º layout is a function of display loss.
While exposure is often quite high with some angled racks, the angularity increases loss in display. This
reduction in display limits the increase in exposure, eventually making certain rack layouts worse than the
equivalent 90º-layout. This observation seems to hold regardless of the duration of exposure (t); see
Figure 8. Interestingly, in scenarios where the FoR is large (Φ = 90° and DOV = 65 ft), the reduction in
exposure for the 90º-layout is larger than the reduction in exposure for angled rack layouts when t > 2 sec;
both angled rack layouts and 90º-layouts see a reduction in t
E tot as low intensity locations are excluded.
Consequently, we observe a large positive %-change in exposure over 90º for t > 2 sec as compared to t >
0 sec. Under certain circumstances, such as when the FoR partially penetrates the store sub-section (i.e.,
DOV=25 ft), the %-change in exposure is the highest over a range of values for %-display loss. In our
experiments, this range was observed to be 13-75% of loss, depending on system parameters.

6.3 Impact of Shape

We next investigate how the shape of a design section impacts the best layouts when t > 0sec. We define
the shape of our section in terms of aspect ratio (AR), where AR=L/W. We consider three shapes; (i) the
original 44 ft x 44 ft (AR=1); (ii) 20 ft x 88 ft (AR=0.23); and (iii) 84 ft x 25 ft (AR=3.36), the last two are
shown in Figure 9. The equivalent 90° rack layout for each of these 3 aspect ratios has a total of 132, 4-ft,
rack sections to enable a fair comparison with angled-rack layouts. We assume that shoppers begin their
travels 10 ft away from the first rack in all scenarios.

23
 (a) AR=0.23 (b) AR=3.36

Figure9: Penetration of the FoR into the (a) AR<1 and (b) AR>1 design section (with example 90° racks)
for Φ = 45º (dotted line) and 90º (solid line) and DOV = 25 ft (black line) and 65 ft (grey line)

Table 5 summarizes the results for all combinations of AR, Φ and DOV; where E90 is the exposure
(in ft) of the 90º-rack layout, θ1 is the angle of the racks in the first column of the best solution, E* and D*
are exposure and display associated with the best solution (both in ft), and %ΔE and %ΔD are the %
changes in E and D with respect to the 90º-rack layout. While these results support the previous
observations 1-3, they also reveal additional insights.

Table 5: Exposure (E) and display (D), both in linear ft, associated with the 90º-rack layout and the best
(*) solutions for each scenario when t > 0 sec. (Note: The best solution for AR = 0.23, Φ = 90º, and
DOV = 25 ft was the 90º-rack layout.)
Φ=45º Φ=90º
Space DOV 0*
(AR) (ft) E90
θ1
0*
E tot
% ΔE
D*
% ΔD
E90
θ1 E tot % ΔE
D*
% ΔD
(ft) (ft) (ft) (ft) (ft)
(ft)
20'x88'
25 44.7º 35.9 43.5% 440 -20.3% 72 90.0º 72.0 0.0% 552 0.0%
(0.23)
44'x44'
25 49 43.8º 81.9 67.1% 448 -22.2% 144 88.7º 170.0 18.1% 480 -16.7%
(1.0)
84'x25'
89 43.6º 154.8 73.9% 456 -26.0% 264 88.4º 340.0 28.8% 560 -9.1%
(3.36)
20'x88'
25 44.7º 44.4 77.5% 440 -20.3% 192 84.0º 221.5 15.4% 512 -7.3%
(0.23)
44'x44'
65 49 44.1º 159.9 226.3% 448 -22.2% 288 71.2º 339.1 17.8% 488 -15.3%
(1.0)
84'x25'
89 43.6º 265.8 198.6% 512 -16.9% 308 89.2º 403.8 31.1% 560 -9.1%
(3.36)

24
Obs. 5: Long and skinny store sections result in higher exposure with competitive loss of display.
Table 5 shows that the E 0
tot
increases as aspect ratio (AR) increases for each Φ and DOV combination

(for both 90° and angled rack layouts). This is intuitive because as the width of the design space decreases
and length increases (for a given area), racks that previously fell outside the FoR now fall within the FoR
increasing the overall exposure. Further, racks that were previously obstructed now move closer to the
pathway making them more likely to be exposed (see Figure 9).
However, this exposure increase comes with larger losses of display in some cases. Considering
the loss of display (% ΔD) associated with each combination of AR, Φ, and DOV in Table 4, we noticed
that in many cases, the best layouts for odd shapes (AR ≠ 1) result in smaller losses than AR = 1. Further,
for all combinations, the long and narrow shape (AR=3.36) physically produced the highest total exposure
and total display (D) with respect to the square shape (AR = 1) and the short and wide shape (AR = 0.23).
Further, the best layouts (in terms of total exposure) produced by AR ≠ 1 generally consist of
multiple short racks (see Figure 11) and can produce more display through their increased numbers of
endcaps. When we take into account that AR > 1 appears to produce the most total exposure and relatively
lower display loss than AR ≤ 1, racks sections with an AR > 1 may benefit substantially from angled
racks.

(θ1, θ2, θ3) = (44.7º, 90º, 90º)

Figure 10: Example 3-column design for AR=0.23, Φ = 45°,

and DOV = 65 ft

θ1 = 43.6º

Figure 11: Example 1-column

design for AR=3.36, Φ = 45°,
and DOV = 65ft
25
7 Conclusion
The physical layout of a retail store affects the attitude and behavior of shoppers as well as the
performance of the store. An important role of the layout is to expose customers to merchandise in order
to facilitate consideration and ultimately to increase the purchase of products. Key shelf locations known
to be highly visible to shoppers are extremely valuable to retailers and are sought after by manufacturers
because they are known to increase sales and satisfy customers. Observations of actual, in-store layouts at
many top retailers in our area have shown that retailers do consider placing racks at an angle as an
acceptable and appropriate layout alternative in specific store sections; e.g., the entrance of the store, and
areas with high impulse products or novel products that the retailer wants the shopper to see. Not every
section in the store may need to have angled racks, though, due to considerable lost in display space.
The main conclusion of our work is that angled-rack layouts can produce higher levels of
exposure to a traveling shopper when compared to 90º-rack layouts, but at a cost of some loss of display
space. Our experimental study indicates that a substantial increase in exposure could be achieved with a
20-30% loss in display. In many cases, these benefits come from non-intuitive layouts using multiple
columns and racks at angles of 26°, 45°, or 71°. It is just such an observation that suggests a need for
analytical models when considering such design questions.
How to resolve the fundamental tradeoff between rack exposure and rack display will vary from
retailer to retailer, according to the nature of the products placed on these racks, demand rates, profit
margins, and so on. It is our hope that insights presented in this paper would inform such decisions at the
least, and perhaps, in time, would be incorporated into software tools to assist in decision making and
store planning. Future research in this area could extend this work to non-rectangular rack designs that
have recently been introduced by leading rack manufacturers. Considering the vertical head movement of
shoppers, especially when the rack heights are varying, would be intriguing.

Acknowledgements
This research was supported in part by the National Science Foundation under grants CMMI #1548394
and #1548404.

26
References

Anderson, E. E. (1979). An analysis of retail display space: Theory and methods. Journal of Business,
52(1), 103-118.

Bansal, J. C., Singh, P. K., Saraswat, M., Verma, A., Jadon, S. S., & Abraham, A. (2011). Inertia weight
strategies in particle swarm optimization. 2011 Third World Congress on Nature & Biologically
Inspired Computing, 633-640.

Bitner, M. J. (1992). Servicescapes: The impact of physical surroundings on customers and employees.
Journal of Marketing, 56, 57-71.

Bloomberg. (2018). What Drew Amazon and Alibaba to Bricks-and-Mortar: Q&A. Bloomberg
QuickTake. Retrieved from https://www.bloomberg.com/news/articles/2018-01-30/what-drew-
amazon-and-alibaba-to-brick-and-mortar-quicktake-q-a

Botsali, A. R., & Peters, B. A. (2005, May 2005). A Network Based Layout Design Model for Retail
Stores. Paper presented at the Industrial Engineering Conference, Atlanta, GA.

Bozer, Y. A., & Wang, C.-T. (2012). A graph-pair representation and MIP-model-based heuristic for the
unequal-area facility layout problem. European Journal of Operational Research, 218, 382-391.

Burke, R. (2005). Retail Shoppability: A Measure of the World's Best Stores Future Retail Now: 40 of
the World's Best Stores (pp. 206-219). Washington, DC: Retail Industry Leaders Association.

Cairns, J. P. (1962). Suppliers, retailers, and shelf space. Journal of Marketing, 26, 34-36.

Cairns, J. P. (1963). Allocate space for maximum profits. Journal of Retailing, 39(2), 41-45.

Clerc, M., & Kennedy, J. (2002). The particle swarm - Explosion, stability, and convergence in a
multidimensional complex space. IEEE Transactions on Evolutionary Computation, 6(1), 58-73.

Corstjens, M., & Doyle, P. (1981). A model for optimizing retail space allocations. Management
Science, 27(7), 822-833.

d'Astous, A. (2000). Irritating aspects of the shopping environment. Journal of Business Research, 49,
149-156.

Dreze, X., Hoch, S. J., & Purk, M. E. (1994). Shelf Management and Space Elasticity. Journal of
Retailing, 70(4), 301-326.

Dunne, P. M., Lusch, R. F., & Gable, M. (1995). Retailing (2nd ed.). Cincinnati, Ohio: South-Western
College Pub.

27
Ebster, C., & Garaus, M. (2015). Store Design and Visual Merchandising: Creating Store Space That
Encourages Buying (2nd ed.). New York, NY: Business Expert Press.

Eberhart, R. C., & Shi, Y. (2000). Comparing inertia weights and constriction factors in particle swarm
optimization. Proceedings of the 2000 Congress on Evolutionary Computation, 1, 84-88.

Flamand, T., Ghoniem, A., & Maddah, B. (2016). Promoting impulse buying by allocating retail shelf
space to grouped product categories. Journal of Operational Research Society, 67, 953-969.

Gue, K. R., & Meller, R. D. (2009). Aisle configurations for unit-load warehouses. IIE Transactions,
41(3), 171-182.

Harwood, T., & Jones, M. (2014). Mobile eye-tracking in retail research. In M. Horsley, N. Toon, B.
Knight, & R. Reilly(Eds.), Current Trends in Eye Tracking Research (pp. 183-199).
Switzerland: Springer International Publishing.

Irion, J., Lu, J.-C., Al-Khayyal, F., & Tsao, Y.-C. (2012). A piecewise linearization framework for retail
shelf space management models. European Journal of Operational Research, 222, 122-136.

Jolai, F., Tavakkoli-Moghaddam, R., & Taghipour, M. (2012). A multi-objective particle swarm
optimisation algorithm for unequal sized dynamic facility layout problem with pickup/drop-off
locations. International Journal of Production Research, 50(15), 4279-4293.

Kennedy, J., & Eberhart, R. (1995). Particle swarm optimization. Proceedings of ICNN'95 - International
Conference on Neural Networks(4), 1942.

Kennedy, J. (2007). Some Issues and Practices for Particle Swarms. 2007 IEEE Swarm Intelligence
Symposium, 162.

Kulturel-Konak, S., & Konak, A. (2013). Linear Programming Based Genetic Algorithm for the Unequal
Area Facility Layout Problem. International Journal of Production Research, 51(14), 4302-
4324.

Kundu, A., & Dan, P. (2012). Metaheuristic in facility layout problems: Current trend and future
direction. International Journal of Industrial and Systems Engineering, 10(2), 238-253.

Liu, Q., & Meller, R. D. (2007). A sequence-pair representation and MIP-model-based heuristic for the
facility layout problem with rectangular departments. IIE Transactions, 39, 377-394.

Lu, Y., & Seo, H.-B. (2015). Developing visibility analysis for a retail store: a pilot study in a bookstore.
Environment and Planning B: Planning and Design, 42, 95-109.

28
Mansell, K. (2018, January). Putting Customers First. Keynote address presented at the 2018 ICR
Conference, Orlando, FL.

Meller, R. D., Chen, W. P., & Sherali, H. D. (2007). Applying the sequence-pair representation optimal
facility layout designs. Operations Research Letters, 35(5), 651-659.

Meller, R. D., & Gau, K. Y. (1996). The facility layout problem: Recent and emerging trends and
perspectives. Journal of Manufacturing Systems, 15(5), 351-366.

Mitchell, P. (2008). Discovery-Based Retail. Minneapolis, MN: Bascom Hill Publishing Group.

Montreuil, B. (1990). A Modelling Framework for Integrating Layout Design and flow Network Design.
Proceedings of the Material Handling Research Colloquium (pp. 43-58). Hebron, KY: Material
Handling Institute.

Montreuil, B., Ouazzani, N., Brotherton, E., & Nourelfath, M. (2004). Coupling zone-based layout
optimization, ant colony system and domain knowledge Proceedings of the 8th International
Material Handling Research Colloquium. Graz, Austria: Material Handling Institute of America.

Mowrey, C. H., Parikh, P. J., & Gue, K. R. (2017). The impact of rack layout on visual experience in a
retail store. INFOR: Information Systems and Operational Research, 1-24.

Ng, C. F. (2003). Satisfying shoppers' psychological needs: From public market to cyber-mall. Journal of
Environmental Psychology, 23, 439-455.

Ozturkoglu, O., Gue, K. R., & Meller, R. D. (2014). A constructive aisle design model for unit-load
warehouses with multiple pickup and deposit points. European Journal of Operational
Research, 236(1), 382-394.

Peters, B. A., Klutke, G.-A., & Botsali, A. R. (2004). Research Issues in Retail Facility Layout Design
Progress in Material Handling Research: 2004 (pp. 399-414).

Pew Research Center. (2016). Online Shopping and E-Commerce. Retrieved from pewresearch.org:
http://www.pewinternet.org/2016/12/19/online-shopping-and-e-commerce/

POPAI. (2014). Mass Merchant Shopper Engagement Study Executive Summary Report. Retrieved from
www.shopassociation.org: http://www.shopassociation.org/a-r-e-popai-mass-merchant-shopper-
study/

Shankar, V., Inman, J. J., Mantrala, M., Kelley, E., & Rizley, R. (2011). Innovations in shopper
marketing: Current insights and future research issues. Journal of Retailing, 87(Supplement 1),
S29-S42.

29
Shi, Y., & Eberhart, R. (1998). A modified particle swarm optimizer. The 1998 IEEE International
Conference on Evolutionary Computation Proceedings, 69-73.

Singh, S. P., & Sharma, R. K. (2006). A Review of different approaches to the facility layout problems.
International Journal of Advanced Manufacturing Technology, 30(5/6), 425-433.

Sorensen, H. (2008). Long tail media in the store. Journal of Advertising Research (September), 329-
338.

Suher, J., & Sorensen, H. (2010). The power of atlas: Why in-store shopping behavior matters. Journal
of Advertising Research, 50(1), 21-29.

Tompkins, J. A. (2010). Facilities planning (4th ed.). Hoboken, NJ

Turley, L. W., & Milliman, R. E. (2000). Atmospheric effects on shopping behavior: A review of the
experimental evidence. Journal of Business Research, 49(2), 193-211.

Underhill, P. (1999). Why we buy : the science of shopping: New York : Simon & Schuster, 1999.

Yapicioglu, H., & Smith, A. E. (2012a). Retail Space Design Considering Revenue and Adjacencies
Using a Racetrack Aisle Network. IIE Transactions, 44(6), 446-458.

Yapicioglu, H., & Smith, A. E. (2012b). A Bi-Objective Model for the Retail Spatial Design Problem.
Engineering Optimization, 44(3), 243-266.

Yapicioglu, H., Smith, A. E., & Dozier, G. (2007). Solving the semi-desirable facility location problem
using bi-objective particle swarm. European Journal of Operational Research, 177(2), 733-749.

30