Download as pdf or txt
Download as pdf or txt
You are on page 1of 50

Accepted Manuscript

A problem evolution algorithm with linear programming for the


dynamic facility layout problem— a general layout formulation

Yiyong Xiao , Yue Xie , Sadan Kulturel-Konak , Abdullah Konak

PII: S0305-0548(17)30164-8
DOI: 10.1016/j.cor.2017.06.025
Reference: CAOR 4280

To appear in: Computers and Operations Research

Received date: 19 September 2016


Revised date: 25 June 2017
Accepted date: 30 June 2017

Please cite this article as: Yiyong Xiao , Yue Xie , Sadan Kulturel-Konak , Abdullah Konak , A prob-
lem evolution algorithm with linear programming for the dynamic facility layout problem— a general
layout formulation, Computers and Operations Research (2017), doi: 10.1016/j.cor.2017.06.025

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT

Highlights:
 The Problem Evolution Algorithm (PEA) is developed to solve the Facility Layout Problem
(FLP).
 The PEA-LP works very well in solving various FLP benchmark problems.
 A polyhedral inner-approximation is proposed for the nonlinear department area constraints.
 Two symmetry-breaking constraints are introduced to increase the algorithmic efficiency.
 Relayout of department blocks in the context of the dynamic FLP was considered.

T
IP
CR
US
AN
M
ED
PT
CE
AC

1
ACCEPTED MANUSCRIPT

A problem evolution algorithm with linear programming for the dynamic facility layout

problem— a general layout formulation

Yiyong Xiao, School of Reliability and System Engineering, Beihang University, Beijing 100191, China email address:
xiaoyiyong@buaa.edu.cn

T
Yue Xie, School of Reliability and System Engineering, Beihang University, Beijing 100191, China email address:

IP
guanzhu218@163.com

CR
Sadan Kulturel-Konak (Corresponding Author) Management Information Systems, Penn State Berks, Reading, PA
19610, USA email address: sadan@psu.edu

Abdullah Konak

US
Information Sciences and Technology, Penn State Berks, Reading, PA 19610, USA email address: konak@psu.edu

Abstract: Facility layout problems (FLPs) are quite common and important in many industries. This paper presents a
AN
mixed integer linear programming (MILP) model for the dynamic facility layout problem, which is a generalization of

several special cases of FLPs studied in recent years. A new evolutionary meta-heuristic framework, named as the
M

Problem Evolution Algorithm (PEA), is developed as a general solution approach for FLPs. Computational

experiments show that the PEA combined with the linear programming (LP), called PEA-LP in short, performs well in
ED

various types of FLPs. In addition, a new polyhedral inner-approximation method is proposed based on secant lines

for the linearization of the non-linear constraint for department area requirements. This new method guarantees that the
PT

actual department area is always greater than or equal to the required area within a given maximum deviation error.

Furthermore, two new symmetry-breaking constraints which help to improve the computational efficiency of the MILP
CE

model are also introduced. Computational experiments on several well-known problem instances from the

literature are carried out to test the DFLP-FZ and the PEA-LP with promising results.
AC

Keywords: facility layout; evolution algorithm; mixed integer linear programming; hybrid optimization

1. Introduction

Nowadays manufacturing systems have been increasingly becoming more complex and flexible to respond to market

changes in product demands rapidly. The layout and re-layout of the manufacturing units (e.g., machine tools,

2
ACCEPTED MANUSCRIPT

manufacturing cells, work centers, departments, etc.) within a facility need to be timely adjusted so that the

manufacturing system can be transferred to fit for different production scenarios in a fast yet low-cost way. The

dynamic facility layout problem (DFLP), which was first introduced by Rosenblatt (1986) as an extension of the

facility layout problem (FLP), considers the optimization of the layout and re-layout of manufacturing units over

multiple planning periods to maintain the economical operation of the system under dynamic product types and

T
demands. Relocation of a manufacturing unit between two consecutive periods results in a re-layout cost although it

IP
may reduce material handling costs. Therefore, the objective of the DFLP is to minimize the total material handling

costs of all periods and the re-layout costs of modifying the facility layout between consecutive periods.

CR
In this paper, a general formulation of the DFLP is presented based on the zone-based FLP, which was first defined

US
by Montreuil et al. (2002, 2004). In the zone-based FLP, the facility is partitioned into several functional zones with

fixed relative positions, and then departments are assigned to and arranged within the zones to minimize the total
AN
facility layout cost. In the zone-based FLP, departments are not allowed to cross zone boundaries. Zone-based

facility layouts are very common in many industries because similar departments are often grouped together and be

placed in the same zone, while dissimilar ones may need to be separated in different zones due to various reasons.
M

Therefore, departments can be better organized and efficiently maintained in a zone-based layout. Furthermore, the
ED

boundaries of the zones provide the basis for designing aisles and material handling systems after a block layout. For

example, the Flexible Bay Structure (FBS), which is a version of the zone-based FLP, is frequently used in the literature

because it generates practical block layouts although it leads to higher costs (Meller, 1997; Chae and Peters, 2006;
PT

Konak et al., 2006). The zone-based layout model can support different space partitioning of a facility in order to
CE

better represent specific real-life industrial cases (Montreuil et al., 2004). In Fig.1, the actual layout diagrams of two

manufacturing plants are given, including one plant producing electronic products in Shanghai, China (Fig. 1A) and
AC

one plant producing mechanical products in Shenzhen, China (Fig.1 B). Both diagrams show that plant floors are

partitioned into several zones, such as office, warehouse, production, business, research and development (R&D), etc.

The arrangement of the departments in this facility follows a zone-based partition. The non-rectangular facility shape

in Fig. 1A is also quite common in real-life facilities and requires the zone-based layout design in order to implement it.

Implicitly or explicitly, the zone-based layout is widely applied in real-life facilities, and this is the main motivation in

this paper for adapting the zone-based layout to the DFLP as well. For example, a bay configuration where
3
ACCEPTED MANUSCRIPT

departments are arranged along a central aisle is frequently used for the layout of semiconductor manufacturing plants

(Peter and Yang, 1997). In flexible manufacturing systems, manufacturing cells are also located around a central

material handling system that interconnects all cells (Kusiak and Heragu, 1987). A zone-based block layout can be

adapted to represent such layout design skeletons in addition to other layout representations such as flexible bay. The

majority of the existing approaches to the DFLP address the version of the problem where department dimensions are

T
predetermined, or they use layout representation schemes where departments cannot be located freely in the facility

IP
(e.g., flexible bay). Only a few DFLP methods (Kulturel-Konak and Konak 2015; Lacksonen, 1997) can concurrently

determine the relative locations and dimensions of the departments over multiple planning periods using an unrestricted

CR
FLP model where departments can be located anywhere within the facility. While this unrestricted modeling approach

provides flexibility in terms of optimization, the resulting block layout can be difficult to transfer into a detailed layout

compromise between the unrestricted and restricted modeling approaches.


US
when aisles and input/output stations are incorporated into the design. On the other hand, the zone-based FLP is a

Departments can be located freely


AN
anywhere within zones which partition the facility based on the particular needs of the manufacturing system.

Therefore, practical block layouts can be achieved with the flexibility of determining department locations and
M

dimensions concurrently. This point is the main contribution of the proposed modeling framework in this paper.
ED

(Entry)
Packing &
Shipping

Rest Product Display Large Small


Area

Room Room Meeting Meeting Waiting Room


Non-usable zone Room Room

Business Room I
PT

Conference
Assembly
Area (III)

Inspection Receiving Staging Packaging R&D


Room Area Area Area Business Room II Room
Room
Finance Dep. Office
Staging Area for Acces. & SP Experimental
Manager
CE

Area
Painting

Staging Area for Large Parts


Room

Warehouse of
Warehouse Assembly Area (I) Product Work
Examine Assembly Assembly
Station Shell Screen
encapsu Aging
lation Area
shop 2
AC

Assembly Area (II) Rest Warehouse of


Room Materials Work Welding Rough Assembly Cleaning Aging
Staging Area shop 1 Station Testing Board &testing Area
for Pipe Grinder Balancing
Processing

Area Room
Area
Pipe

storage of
Office Tea &
Defective Product Aging
Utility Tea Grinder
Repair Examine Repair & Test
Service Rest In/Out Station Area Area
Room Room Room Area Line
Room (Lefter)

(A) (B)
Fig. 1 Examples of zone-based layout of two real-life plants

4
ACCEPTED MANUSCRIPT

In this paper, a general formulation of the DFLP with flexible zones (DFLP-FZ) is presented. In the DFLP-FZ, a

rectangular facility is first partitioned into several previously defined zones that are completely flexible in terms of their

shapes, sizes, boundaries, and relative positions. Then, the departments are placed properly within zones while

minimizing a cost function. In addition, the planning horizon has multiple periods with different patterns of material

flow. The cost function includes three general components: (1) the total material handling cost in all periods, (2) the

T
fixed and variable re-layout costs related to departmental location changes between two consecutive periods, and (3)

IP
the fixed cost of moving boundaries of zones between two consecutive periods. Compared to the traditional FLP, the

DFLP-FZ has two levels of optimization: (1) zone partition and (2) arrangements of departments within zones, which

CR
makes the formulated model a better fit for various real-life facility design problems but also harder to solve.

US
The contributions of this paper are in three folds: (1) the DFLP-FZ is formulated as a mixed-integer linear

programming (MILP) model; (2) the Problem Evolution Algorithm (PEA) is developed as a new solution approach for
AN
facility layout problems (FLPs). Combining the PEA with Linear Programming (LP), named the PEA-LP in short,

can efficiently solve the DFLP-FZ as well as other traditional FLPs with good quality solutions; and (3) a new

polyhedral inner-approximation is proposed to model non-linear constraints for ensuring minimum department area
M

requirements. The proposed polyhedral inner-approximation method ensures that the calculated department areas by
ED

the model are always greater than or equal to the minimum area requirements of the departments. On the other hand,

the existing polyhedral outer-approximation method may result in department areas smaller than the minimum area

requirement of the departments. In addition, the proposed polyhedral inner-approximation always uses the minimum
PT

number of ascent lines for a given approximation accuracy.


CE

The rest of the paper is organized as follows. In Section 2, a literature review is given, and the contributions of our

study are highlighted. In Section 3, the advantages of layout design with flexible zones are discussed, a new method
AC

for approximating the non-linear area constraint is introduced, the DFLP-FZ is formulated as an MILP, and two new

symmetry-breaking constraints are proposed. In Section 4, a heuristic solution approach named problem evolution

algorithm with linear programming (PEA-LP) is proposed. In Section 5, a case study is presented to demonstrate the

use of the zone-based DFLP approach. In Section 6, computational experiments are carried out to test the MILP

5
ACCEPTED MANUSCRIPT

model and the proposed PEA-LP based on well-known benchmark instances from the literature. Finally, in Section 7,

the conclusion is presented.

2. Literature Review

The FLP concerns with how to place departments in a rectangular facility area for an efficient operation of a

manufacturing system. A good facility layout will facilitate efficient material handling among the departments, which

T
is known to have a significant impact on the manufacturing costs, work-in-process, lead times, and productivity.

IP
Koopmans and Beckmann (1957) modeled the FLP as a quadratic assignment problem (QAP) that aims to find the

CR
optimal assignment of n departments to n fixed and discrete locations to minimize the total material handling cost

expressed as the product of material flow and travel distance. The QAP is NP-Complete (Sahni and Gonzalez, 1976).

US
There are earlier optimal procedures for solving the QAP with a small number of machines, such as Lawer (1963) and

Burkar & Elshafer (1979). Montreuil and Ratliff (1989) proposed a design skeleton-based approach for the facility
AN
layout design with input/output stations, using the cut tree algorithm to generate flow network, and interacting with a

human designer to transpose the design skeleton into a facility layout. Montreuil et al. (1993) strengthened the design
M

skeleton-based approach by developing a linear programming model which efficiently generates a layout from a design

skeleton.
ED

The QAP version of the FLP requires the plant area to have the same number of pre-determined candidate sites to
PT

locate departments, without considering the area requirements of the departments. However, in most industrial cases,

departments have different area requirements, and they can also be continuously placed anywhere in the plant without
CE

pre-specified candidate sites, as long as they do not overlap. This version of layout problem is called the facility

layout problem with unequal area requirements (or the unequal area FLP). The material handling cost is usually
AC

estimated by the amount of material flow and the rectilinear distance between the centroids of each department pair.

The unequal area FLP is harder to solve than the QAP version even with a few departments. Regarding how the

shapes of the departments are modeled, the unequal area FLP has two major versions in the literature. The first one,

referred to as the unequal area FLP with fixed departmental shapes, assumes that all departments have given fixed

shapes. The layout arrangement considers only the optimization of the department coordinates and their vertical or

6
ACCEPTED MANUSCRIPT

horizontal orientations to minimize the material handling cost while satisfying that the departments do not overlap at

the same time. The literature includes many papers in this direction, including Lee and Lee (2002), Dunker et al.

(2005), Baykasoglu et al. (2006), Drira et al.(2007), Balakrishnan and Cheng (2009), McKendall and Hakobyan (2010),

and McKendall and Liu (2012).

The other version of the unequal area FLP, which is referred to as the unequal area FLP with flexible departmental

T
shapes, considers a more general and complex situation in which the side lengths of the departments in addition to their

IP
locations are all considered as decision variables. In other words, the side lengths of a department can have any value

CR
as long as the area requirement of the department is guaranteed (i.e., ai  li  li where ai, li and li denote the
x y x y

area, height and length requirement of department i, respectively) under a department shape constraint (e.g., the

aspect-ratio or minimum side-length constraints).


US
This version of the unequal area FLP is also categorized as the FLP

on the continuous plane (Montreuil, 1990; Kulturel-Konak and Konak, 2013) and harder to solve due to the
AN
linearization of the non-linear constraint ai  li  li for each department i. Relevent research on this venue of the
x y

unequal area FLP in literature includes Montreuil (1990), Bozer and Meller (1997), Meller et al. (1999), Sherali et al.
M

(2003), Konak et al. (2006), Ulutas and Kulturel-Konak (2012), Kulturel-Konak and Konak (2013, 2015), and

Gonçalves and Resende (2015).


ED

This paper also considers flexible departmental shapes. In the literature, the non-linear area constraint ai  li  li
x y
PT

is modeled in various ways to improve the tractability of the problem. Montreuil (1990) used a surrogate linear

constraint to substitute for the non-linear area constraint, but the approximation error could be as high as 44% for the
CE

aspect ratio of five. Meller et al. (1999) proposed a more accurate surrogate area constraint which reduced the

maximum error to 14.3% for the aspect ratio of five. Lacksonen (1997) used a piecewise linear approximation for the
AC

area constraint which was more accurate and resulted in the actual area between ai and 1.03ai. Sherali et al. (2003)

developed a polyhedral outer-approximation method based a set of tangential lines for supporting the non-linear

function, li  ai / li .
y x
This approach was also used by Kulturel-Konak and Konak (2013) and Gonçalves and Resende

(2015). Castillo and Westerlund (2005) developed a formula to estimate the number of tangential lines for the

outer-approximation method to guarantee the maximum error of area deviation under a given percentage (  % ).

7
ACCEPTED MANUSCRIPT

The dynamic nature of the FLP was first recognized by Hitchings (1970). Since manufacturing systems may have

to be updated over the time (e.g., adding/replacing equipment, new material handling tools, etc.), and product

demands/volumes may also be continuously changing according to market changes, there is a need for a dynamic

layout that can handle different production scenarios. The DFLP that integrates the optimization of the initial layout

and re-layouts of a facility over multiple periods was first introduced by Rosenblatt (1986) based on the QAP

T
formulation. The objective function includes the total material handling cost in all periods and the re-layout costs

IP
between the consecutive periods. Montreuil and Venkatadri (1991) presented a proactive methodology for the

dynamic layout optimization of the manufacturing system with multiple expansion (or decline) phases. Urban (1993)

CR
developed a heuristic called the steepest-descent pairwise-interchange procedure for the QAP version of the DFLP,

where for each period, the workflow from a number of consecutive periods are combined to determine the layout.

US
Lacksonen and Enscore (1993) compared five algorithms for the DFLP, including CRAFT, cutting planes, branch and

bound, dynamic programming, and cut trees, on a series of realistic test problems and found that the cutting plane
AN
algorithm performed the best. Lacksonen (1994, 1997) modeled the DFLP with unequal department areas in a

continuous plane that can optimally be solved by a MIP solver for small problems. Balakrishnan et al. (2003)
M

developed a hybrid genetic algorithm for the DFLP. Baykasoglu et al. (2006) studied the DFLP with budget

limitations on the reconfiguration of facilities. Balakrishnan and Cheng (2009) investigated the DFLP under rolling
ED

horizons and uncertainty. Kulturel-Konak and Konak (2015) studied the cyclic facility layout problem (CFLP), a

special case of the DFLP in which the production cycle repeats itself by going back to the first period after the last one.
PT

The state-of-the-art surveys of the DFLP can be found in Balakrishnan and Cheng (1998) and Kulturel-Konak (2007).

McKendall and Liu (2012) and Arostegui et al. (2006) summarize and compare several meta-heuristics for the DFLP.
CE

Recent research on meta-heuristics for the DFLP can be found in Hosseini-Nasab and Emami (2013), Chen (2013),

Ulutas and Islier (2015).


AC

The facility layout problem in multi-bay environments, another extension of the FLP, referred to as FLP with flexible

bay structure (FLP-FBS), is concerned with determining the most efficient assignment of departments to parallel and

flexible bays (Meller, 1997; Chae and Peters, 2006; Konak et al., 2006). The FLP-FBS is common in several

industries, such as heavy manufacturing, semiconductor, and retail, where the facility is first divided into multiple

parallel bays with varying widths, and then the facilities/departments are arranged in the bays without overlapping.
8
ACCEPTED MANUSCRIPT

Konak et al. (2006) first formulated a linear MIP model for the unequal area FLP-FBS that can be optimally solved by

MIP solvers up to twelve departments. Several heuristic algorithms for larger problems of the unequal area FLP-FBS

can be found in the literature, such as the Ant System (AS) algorithm by Wong and Komarudin (2010) and Komarudin

and Wong (2010), the artificial immune system (AIS) based algorithm by Ulutas and Kulturel-Konak (2012), and the

probabilistic tabu search with linear programming by Kulturel-Konak (2012).

T
There are also several other variants of the FLP in literature for some other special cases of industries, such as the

IP
single row facility layout problem which arranges the departments on a straight line (Solimanpur et al., 2005;

CR
Hungerländer and Rendl, 2013), the grid-based facility layout problem in which the facility is divided into basic

squares or rectangles having a unit area (Armour and Buffa, 1963), the multi-floor FLP which fulfills the layout

US
optimization in a facility that has been separated into multiple sub-floors by inner structural walls, passages, or fixed

split lines (Meller and Bozer, 1997; Lee et al., 2005), the multi-objective FLP which considers multiple objective
AN
functions in the facility layout design, such as the material flow, the number of material handling devices, and the

average work-in-progress (Saraswat et al., 2015).


M

3. Problem formulation

3.1 Layout design with flexible zones


ED

In many industries, it is quite common to partition the facility into several functional zones based on the
PT

requirements of the manufacturing system first and then arrange departments within these zones. The zone-based

layout improves the organization of manufacturing resources through grouping departments with similar functions and
CE

interactive works together into the same zone. Thereby, the operational efficiency of the plant can be achieved

through better maintenance, information exchange, plant management, etc.


AC

In this paper, we consider the layout of departments in a facility with flexible zones. Zone boundaries can be

anywhere in the facility, zone shapes/sizes (widths and lengths) are variables, and their relative positions are not fixed

to one another. We also assume that each department belongs to a zone, and each zone has at least one department

(excluding non-usable zones). An MILP model for the DFLP-FZ is given in Section 3.3.

9
ACCEPTED MANUSCRIPT

The flexibility of zones in our MILP model has other advantages. By pre-specifying some properties of zones in

the DFLP-FZ model, we can model many practical FLP cases as follows:

1) By fixing relative positions of some zones in the model, such as a zone must be on the right (or top) of another one,

the DFLP-FZ model then reduces to the zone-based unequal area FLP studied by Montreuil (2002, 2004), where the

relative positions of zones are assumed to be fixed in advance.

T
2) By restricting relative positions of zones along an axis direction (horizontal or vertical) such that all zones are

IP
arranged in multiple parallel arrays, the DFLP-FZ model reduces to the unequal-area FLP with FBS (Meller, 1997;

CR
Chae and Peters, 2006; Konak et al., 2006; Kulturel-Konak, 2012).

3) By defining some non-useable zones with fixed locations and shapes, the DFLP-FZ model applies to some

US
non-rectangular facilities, such as the plant of a mechanical product in Fig. 1A. Other typical examples of

non-rectangular facility building floor are given in Fig. 2(A) while in Fig. 2(B) they are transformed into a rectangular
AN
facility with special non-useable zones marked as X. To model such cases, departments are not allowed to be assigned

to non-useable zones.
M
ED
PT

X1
CE

X X1
X
X2
X2
AC

B
Fig. 2 Transforming non-rectangle facilities to rectangle ones with fixed non-useable zones

4) By defining some zones as fixed bars or lines, the DFLP-FZ model applies to the case of the multi-floor FLP that

consider inner structure walls and un-occupiable passages in the facility (Meller and Bozer, 1997; Lee et al., 2005).

10
ACCEPTED MANUSCRIPT

Technically, a narrow non-usable zone can be defined as a fixed bar or line that represents an un-occupiable passage or

inner structure walls.

3.2 A polyhedral inner-approximation of depart area requirements

In the DFLP-FZ, the locations and shapes of departments are modeled on the continuous space. Each department i

T
x y
has four continuous decision variables, its centroid location denoted as ( ci , ci ) and its length and width denoted as

IP
Each department i has an area requirement of ai, and li  li  ai must be satisfied.
x y x y
( li , li ). However, this

CR
nonlinear relationship makes the FLP difficult to solve optimally. To approximate the non-linear area constraint,

Montreuil (1990) used a surrogate linear constraint called bounded perimeter constraint, which is defined as

US
2 ai  lix  liy  ai (1  i ) / ai , i where  i is the aspect ratio of department i (i.e., the ratio of the longer side

length to the shorter one). Meller et al. (1999) pointed out that even if the bounded perimeter constraint is satisfied,
AN
the actual area can be much smaller than ai for a large value of  i (e.g., the error can be as high as 44% for  i =5).

Meller et al. (1999) proposed a more accurate surrogate area constraint, which is 2(lix  liy )  3 ai  f  limax , i , where
M

limax  max{lix , liy } . For f=0.95, this approach can achieve a closer approximation to ai than the bounded perimeter
ED

constraint (e.g., a maximum error of 14.3% for  i =5). In Sherali et al. (2003), the experimental value for parameter f

is suggested as f =1.9. Lacksonen (1997) used a piecewise linear approximation for the area constraint which results
PT

in the actual area between ai and 1.03ai. However, this piecewise linearization increases the problem complexity due

to additional two binary variables for each department (Meller et al., 1999, Konak et al., 2006). Sherali et al. (2003)
CE

proposed a polyhedral outer-approximation that uses a number of tangential lines to support the non-linear function

liy  ai / lix . Generally, increasing the number of tangential support lines results in a more accurate approximation to
AC

ai. However, the number of the support lines required to reduce the area approximation error to a target level is not

 ln(limax / limin ) 
defined in Sherali et al. (2003). Castillo and Westerlund (2005) provided a formula, Ci   ,
 ln((1   ) / (1   )) 

to estimate the number of tangential support lines to guarantee that the final area of department i is within an % error

range of the required area.

11
ACCEPTED MANUSCRIPT

All above methods based on the polyhedral outer-approximation usually result in smaller department areas than the

required ones. On the other hand, a larger actual area will often be a safer design in many industrial cases. In this

paper, we propose a polyhedral inner-approximation method to model department areas, which guarantees that the

actual area is always greater than or equal to the required one. In addition, this method can restrict the maximum

aactual  ai
approximation error within a given error ratio, i.e.,  i , for any department i. The polyhedral
ai

T
inner-approximation method uses a number of secant lines (instead of tangential lines) to approximate the non-linear

IP
relationship ai  x  y internally. As shown in Fig. 3, the secant lines are determined by starting from point (x1, y1),

CR
which corresponds to the minimum permissible width and the maximum permissible height of department i. In our

approach, the ending point of a secant line is calculated based on the given error ratio  i . In other words, we can

US
control the error of the area approximation by locating the starting and ending points of the secant lines. On a secant

line expressed as y  kx  b, (k  0, b  0, x  0, y  0) , each point (x, y) corresponds to an area value, i.e., ai  x  y , and
AN
the maximum area on the line can be derived as a  x  y  b / (4k ) . Therefore, given an error ratio  i and a
* * * 2

starting point on the area function, i.e., (x1, y1) in Fig. 3, we can calculate the next point (x2, y2) to determine a secant
M

( x  y )  ai
line (x1, y1)-(x2, y2) that satisfies  i for any point (x', y') on the line. In this way, we can continue to
ai
ED

max
determine the next point (x3, y3) starting from (x2, y2), and so on, until the point ( xi , yimin ) is covered by the last secant
PT

line.
CE
AC

12
ACCEPTED MANUSCRIPT

y ai=xy
max
y (x1, y1)

(x*, y*)

(x2, y2)

T
(x3, y3)

IP
(x4, y4)
min
y

CR
x
min
x
max
x

US
Fig.3 An inner-approximation method for the non-linear area linearization

Eq.(1) gives the formula for calculating the minimum number of secant lines to approximate a department area ai

with the maximum allowed approximation error ratio  i of the actual area to the required area as follows:
AN
ln ximax  ln ximin
ni   1, i , (1)
ln(1  2i  2 i  i 2 )
M

max min
where xi and xi are the maximum and minimum permissible side lengths of department i in the x axis direction,
ED

respectively. The detailed derivation of Eq. (1) is given in the Appendix. In the symmetric case of xi
max
 yimax , Eq.

(1) can be rewritten as Eq. (2) in the following manner:


PT

ln i
ni   1, i (2)
ln 
CE

where   1  2i  2 i  i 2 and  i is the maximum aspect ratio for department i. Note that the number of

secant lines is dependent on only  i and  i , but not on ai. In Table 1, we present the minimum number of required
AC

secant lines for different values of  i and  i for an intuitive view.

Table 1 Minimum required number of secant lines with respect to different  i and  i
i  i =1.5  i =2  i =3  i =4  i =5
3.00% 3 4 5 6 6
2.00% 3 4 5 6 7
1.00% 4 5 7 8 10
0.50% 4 6 9 11 13

13
ACCEPTED MANUSCRIPT

0.10% 8 12 19 23 27
0.05% 11 17 26 33 37

In our approach, each department i is associated with a set of ni secant lines, i.e., y  Kip x  Bip , p  1,2,..., ni , that

can be used for the non-linear area approximation in the MIP model formulated in the next section. Like Sherali et

al.’s (2003) polyhedral outer-approximation, our method does not require binary variables. Unlike Sherali et al.’s

T
(2003) method which uses uniformly spaced tangential support lines, the end-points of secant lines are calculated to

IP
achieve a target maximum area approximation error in our method. Thereby, the proposed method always uses the

CR
minimum number of secant lines for a given maximum area approximation error. Eq. (3) provides the equations for

calculating the slope Kip and intercept Bip of the secant lines of department i.

Kip  

i
2 p 1
 1
, Bip  p yimax ,

US p  1, 2,..., ni , i , (3)
AN
max
where yi is maximum side length of department i in y axis direction. The detailed derivation of Eq. (3) can be

found in the Appendix.


M

In order to make the area linearization symmetric in the x and y axes directions, the set of secant lines can also be
ED

max
determined by starting from the middle point ( ai , ai ) of the curve toward two directions ( xi , yimin ) and

( ximin , yimax ). In this case, the following Eq. (4) can be used to determine the slope Kip and intercepts Bip for two
PT

directions:
CE

 1  1
 Kip    2 p 1 , Bip   p ai , p  1, 2,..., nix
 ( ) ( )
 i , (4)
K   1  1
, Bip   p ai , p  1, 2,..., n y

AC

( )
 
( )
ip 2 p 1 i

ln ximax  ln ai ln yimax  ln ai
where   1  2  2    ,   1  2  2    ,
 
nix  and niy 
2 2
. If
ln  ln 

ln i
ximax  yimax , we will have nix  nix  . It should be noted that in Eq. (4), the slope Kip depends only on  i
2ln 

14
ACCEPTED MANUSCRIPT

while intercept Bip depends on  i and is proportional to ai . Some values of Kip and Bip that can be used directly

for general area approximation are provided in Table 2 where column bip/ai0.5 means that the value in this column needs

to be multiplied by ai before using it as bip. For example, to approximate area ai=4 and aspect ratio  i =3 with

the maximum area approximation error of 1.0%, six secant lines are needed to be used, which are (-0.819, 3.638004),

(-1.221, 4.441996), (-0.54936, 2.979536), (-1.82031, 5.423664), (-0.36849, 2.440246), and (-2.71378, 6.622282).

T
Table 2 The slopes and intercepts of the secant lines with respect to different  i and  i

IP
i p kp bip/ai0.5 i i p kip bip/ai0.5 i
3.0% 1+ -0.70843 1.708432 2.0 0.5% 5+ -0.28035 1.061597 4.1

CR
1- -1.41157 2.411568 2.0 5- -3.56703 3.786749 4.1
2+ -0.35555 1.210308 4.0 6+ -0.21133 0.921705 5.5
2- -2.81258 3.404092 4.0 6- -4.73197 4.361481 5.5
3+ -0.17844 0.857421 7.9 0.1% 1+ -0.93872 1.938723 1.1
3- -5.60414 4.805107 7.9 1- -1.06528 2.065277 1.1
2.0% 1+
1-
2+
2-
3+
-0.75434
-1.32566
-0.42925
-2.32967
-0.24426
1.754343
2.325657
1.323376
3.083024
0.998279
1.8
1.8
3.1
3.1
5.4
US 2+
2-
3+
3-
4+
-0.8272
-1.20889
-0.72893
-1.37187
-0.64234
1.819923
2.200093
1.708404
2.343708
1.603717
1.3
1.3
1.5
1.5
1.7
AN
3 -4.09408 4.087033 5.4 4- -1.55682 2.496699 1.7
1.0% 1+ -0.819 1.819002 1.5 5+ -0.56603 1.505446 1.9
1- -1.221 2.220998 1.5 5- -1.7667 2.659677 1.9
2+ -0.54936 1.489768 2.2 6+ -0.49878 1.413197 2.1
2- -1.82031 2.711832 2.2 6- -2.00488 2.833293 2.1
M

3+ -0.36849 1.220123 3.3 7+ -0.43953 1.3266 2.4


3- -2.71378 3.311141 3.3 7- -2.27517 3.018242 2.4
4+ -0.24717 0.999284 4.9 8+ -0.38731 1.24531 2.8
4- -4.04579 4.042895 4.9 8- -2.5819 3.215264 2.8
ED

5+ -0.16579 0.818416 7.4 9+ -0.3413 1.169001 3.1


5- -6.03161 4.936364 7.4 9- -2.92998 3.425148 3.1
0.5% 1+ -0.86823 1.868226 1.3 10+ -0.30075 1.097368 3.5
1- -1.15177 2.151774 1.3 10- -3.32499 3.648732 3.5
2+ -0.65448 1.622041 1.8 11+ -0.26502 1.030124 4.0
PT

2- -1.52793 2.478359 1.8 11- -3.77325 3.88691 4.0


3+ -0.49336 1.408297 2.3 12+ -0.23354 0.967001 4.6
3- -2.02692 2.854511 2.3 12- -4.28194 4.140637 4.6
4+ -0.3719 1.22272 3.1 13+ -0.2058 0.907746 5.2
CE

4- -2.68888 3.287752 3.1 13- -4.85921 4.410926 5.2

A department area ai may have minimum and maximum side length requirements on both x and y directions, which
AC

min max min max


can be noted as xi , xi , yi , and yi , respectively. It may also have a maximum permissible aspect ratio

requirement, denoted as  i , and subject to the restriction from the length/width (noted as Lx , Ly ) of the facility as well.

x x y y
The lower and upper bounds on the side lengths of department i, noted as ( li , li ) and ( li , li ), can be calculated by

Eq. (5) as follows:

15
ACCEPTED MANUSCRIPT

 li x  min{Lx , max{ ai / i , ximin }}




 li
x
 min{Lx , ai / i , ximax }
 y i (5)
 li  min{Ly , max{ ai / i , yimin }}
 y

 li  min{Ly , ai / i , yimax }

It should be noted that since the inner-approximation always results in an actual area greater than or equal to the

required area, the best solution obtained with inner-approximation might not be the optimal one for the corresponding

T
FLP.

IP
Lemma 1. For an unequal area FLP, the objective function value of the true optimal solution is between the objective

CR
function values of the best solutions found using the outer and inner approximation methods for the linearization of

department areas.

US
Proof. Let s and s be the optimal solutions found using the outer-approximation and inner-approximation methods for

an unequal area FLP. Now assume that s* is the optimal solution for the true department areas.
AN
We can create a

solution s from solution s* by reducing the side lengths of some departments in either of the x-axis or y-axis directions

or both axes directions such that the department area constraints imposed by the outer-approximation method are
M

satisfied and the relative department positions in s* are maintained. By definition, this new solution s must have a
ED

greater than or equal to the total cost (TC) of solution s. Otherwise, solution s will be the optimal solution obtained

by the outer-approximation method. Hence, TC(s)TC(s) TC(s*) is always satisfied. Now let us consider solution
PT

s. Since some departments in solution s might have areas greater than or equal to the required ones, we can always

create a new solution s from solution s by reducing the side lengths of some departments in either of the x-axis or
CE

y-axis directions or both axes directions such that the minimum area requirment of each department is satified, and the

relative department positions are maintained. Therefore, the total cost of the new solution s will be less than or equal
AC

to that of solution s and greater than or equal to that of solution s*. Hence, TC(s*)TC(s) TC(s). In conclusion,

TC(s)TC(s*) TC(s) is always true.

3.3 A mixed integer linear programming (MILP) model

In this section, we present a mixed-integer linear programming (MILP) model for the DFLP-FZ. First, the

parameters and decision variables used in the model are given in Tables 3 and 4, respectively.

16
ACCEPTED MANUSCRIPT

Table 3. Parameters

General parameters
T set of planning periods
N set of departments
Z set of zones
t index of periods, t  T
i, j index of departments, i, j  N
index of zones, k , z  Z

T
z, k
x, y axis directions

IP
e an axis direction, e {x, y}
Parameters of the facility and zones

CR
Lix , Liy side lengths of the facility floor in the x and y axis directions in period t
 x
t
set of zone pairs (z, k) such that zone z must be to the west of zone k in period t
 ty set of zone pairs (z, k) such that zone z must be to the south of zone k in period t
Ozt
t
US
fixed cost of moving boundaries of zone z between periods t-1 and t
set of zones with fixed shape and position in period t, i.e., for element ( z, xz1 , yz1 , xz 2 , yz 2 ) t ,
AN
(xz1, yz1) and (xz2, yz2) are the fixed south-west and north-east corners of zone z in period t
 zt 1/0 parameter indicating if whether zone z is non-usable in period t
Parameters on department shape and material flows
ait minimum area requirement of department i in period t
M

 it maximum permissible aspect ratio of department i in period t (the ratio of the longer side to the
shorter side)
lit x , lit y
ED

minimum permissible side lengths of department i in the x and y axis directions in period t
lit x , lit y maximum permissible side lengths of department i in the x and y axis directions in period t
fijt material flow cost per unit distance between departments i and j in period t
PT

Ft set of department pairs with positive flow costs in period t, Ft  {(i, j ) | i  j, fijt  0}, t T .

Rizt 0/1 parameter denoting if department i must be assigned to zone z in period t


CE

Parameters on the cost of department rearrangement


U it variable rearrangement cost per unit distance of department i between periods t-1 and t
Vit fixed rearrangement cost of department i between periods t-1 and t
AC

Qit fixed cost of moving department i from its current zone to another one between periods t-1 and t
Parameters on department area approximation by secant lines
i maximum error allowed for department area approximation
nit minimum number of secant lines that are required to guarantee the actual area of department i in
period t not exceeding the required area by i%
Kitp slope of the pth secant line of department i in period t
Bitp intercept of the pth secant line of department i in period t

17
ACCEPTED MANUSCRIPT

Table 4. Decision Variables

Decision variables on zones’ locations and shapes


west and east x axis boundaries of zone z in period t, respectively
w1ztx , wzt2 x

south and north y axis boundaries of zone z in period t, respectively


w1zty , wzt2 y

relative locations of zones z and k in the x and y axis directions in period t such that
 zktx ,  zkty

T
 zktx  1 (or  zkty  1 ) if zone z is enforced to precede zone k in the x (or y) axis

IP
direction,  zkt  0 (or  zkt  0 ) otherwise
x y

CR
Decision variables on departments’ locations and shapes
rizt department-zone assignment such that rizt=1 if department i is located in zone z in
period t, rizt=0 otherwise
x y
cit , cit

litx , lity
US
centroid coordinates of department i in the x and y axis directions in period t

side lengths of department i in the x and y axis directions in period t


AN
rectilinear distances between the centroids of departments i and j in the x and y axis
d ijtx , dijty
directions in period t
the relative locations of departments i and j in the x and y axis directions in period t
M

sijtx , sijty
such that sijtx =1 (or sijty =1) if department i is enforced to precede department j in the x
ED

(or y) axis direction, sijtx =0 (or sijty =0) otherwise

Decision variables on zones’ rearrangements between two consecutive periods


1x 2x
the relocation of zone z in x axis between periods t-1 and t such that ozt =1 and ozt
PT

1x 2x
ozt , ozt =1 if zone z is relocated between periods t-1 and t in the west and east x axis
1x 2x
boundaries, respectively, ozt =0 and ozt =0 otherwise
1y 2y
1y 2y
the relocation of zone z in x axis between periods t-1 and t such that ozt =1 and ozt
ozt , ozt
CE

=1 if zone z is relocated between periods t-1 and t in the south and north y axis
1y 2y
boundaries, respectively, ozt =0 and ozt =0 otherwise
Decision variables on departments’ rearrangements between two consecutive periods
relocation distances of the centroid of department i in the x and y axis directions from
AC

uitx , uity period t-1 to period t


vit department rearrangements such that vit=1 if department i is relocated between
periods t-1 and t, vit=0 otherwise
qit department-zone reassignment from period t-1 to t such that qit=1 if department i is
assigned to a different zone between periods t-1 and t, qit=0 otherwise

18
ACCEPTED MANUSCRIPT

Objective function
The objective function of the problem is to minimize the total material handling and re-layout costs. The re-layout

cost includes four components: (i) the fixed re-layout cost incurred if a department is relocated between two

consecutive periods, (ii) the variable re-layout cost estimated based on how much a department is moved between two

consecutive periods, (iii) the fixed re-layout cost incurred if a department is moved from its current zone to another one

T
between two consecutive periods, and (iv) the fixed re-layout cost incurred if the boundaries of a zone are moved

IP
between two consecutive periods.

CR
Below, we present the mathematical formulation for Problem DFLP-FZ. Note that the linearization of constraints

with the absolute value function is not shown for brevity in the model presentation, and the notation e={x, y} is used to

indicate both x and y axis directions.

Problem DFLP-FZ:
US
AN
min TC    fijt (dijtx  d ijty )
tT ( i , j )Pt

   (V v it it  Qit qit  U it uitx  U it uity ) (6)


M

tT ,t 1 iN

  O
tT ,t 1 zZ
zt (o1ztx  ozt2 x  o1zty  ozt2 y )
ED

Constraints for zone partitions in the facility:

(1)  zktx   kztx   zkty  kzty  1 z  k , t


PT

(2) wzt2e  wkt1e  M (1   zkt


e
) z  k , t, e

wzt2e  Let z, t, e


CE

(3)

(4) wzt2e  w1kte ( z, k ) te , t , e


AC

(5) w1ztx  x1 , wzt2 x  x2 , w1zty  y1, wzt2 y  y2 ( z, x1, x2 , y1, y2 ) t , t

Constraints for department arrangements in zones:

(6) rizt  Rizt i, t, z

(7) r
zZ
izt 1 i, t

19
ACCEPTED MANUSCRIPT

(8) 1   zt   rizt  M (1   zt ) z, t


iN

(9) w1zte  cite  0.5lite  M (1  rizt ) i, z, t, e

(10) wzt2e  cite  0.5lite  M (1  rizt ) i, z, t, e

(11) sijte   zkt


e
 M (2  rizt  rikt ) i  j, k  z, t , e

T
Constraint for department layout:

IP
(12) sijtx  s xjit  sijty  s yjit  1 i  j, t

CR
(13) cejt  0.5l ejt  cite  0.5lite  M (1  sijte ) i  j, t, e

(14) lite  lite i, t, e

(15) lite  lite US


i, t, e
AN
(16) lity  Kitplitx  Bitp i  p  nit , i, t

Constraint for bounding the rectilinear distance of a department pair:


M

(17) dijte | cite  cejt | e,(i, j )  Ft , t

Constraints for departments’ re-arrangement:


ED

(18) uite | cite  cie,t 1 | i, t  1, e

Mvit | lite  lie,t 1 | i, t  1, e


PT

(19)

(20) Mvit | cite  cie,t 1 | i, t  1, e


CE

(21) qit | rizt  ri , z ,t 1 | i, z, t  1

Constraints for zones’ re-arrangement:


AC

(22) M  zt1e | w1zte  w1ze,t 1 | z, t  1, e

(23) M  zt2e | wzt2e  wz2,et 1 | z, t  1, e

Constraint for decision variables’ domains:

20
ACCEPTED MANUSCRIPT

cite , lite , d ijte , uite , w1zte , wzt2e  0


(24)  i, j, z, k , t, e
rizt , vit , qit , sijt ,  izkt , ozt , ozt {0,1}
e e 1e 2e

In the model above, Constraints (1)-(4) are used to partition the facility into non-overlapping zones. Constraints (1)

and (2) ensure that zones do not overlap. Constraint (3) requires that all zones must be within the boundaries of the

facility. Constraint (4) is for enforcing the requirements on zones' relative positions as specified in  t . Note that if
e

T
te   , all zones can be freely located in the facility. Constraint (5) ensures that some zones must have fixed

IP
positions as specified in  t . This constraint is only needed if the problem data specify some zones with predefined

CR
locations. In addition, although the relative positions of zones are determined by decision variables  zkt , their
e

relationships must satisfy the pre-specified requirements given by  t and  t . Therefore, if  t and  t are
e e

e 1e 2e US
defined for a problem, some of the  zkt and (wzt , wzt ) variables can be fixed accordingly before solving the model in
AN
order to reduce computational complexity.

Constraints (6)-(11) define the relationships among zones and departments. Constraint (6) enforces pre-defined
M

department to zone assignments if there exists any. Constraint (7) requires that each department must be assigned to a

zone. Constraint (8) requires that each usable zone has at least one department and prevents assigning departments to
ED

non-usable zones. Constraints (9) and (10) require each department to be located within the boundaries of the zone to

which it is assigned. Constraint (11) makes the relative position relationships among departments consistent with
PT

those among zones.

Constraints (12)-(16) are used to locate departments within the facility in a consistent manner. Constraints (12) and
CE

(13) prevent overlapping of departments. Constraints (14) and (15) require that the side lengths of departments must

satisfy the maximum and minimum side length limits, respectively. Constraint (16) is the linearization of the
AC

department area requirement using a set of secant lines of which the slopes and intercepts are calculated based on Eq.

(2) or Eq. (4). Constraint (17) bounds the rectilinear distance between departments i and j.

Constraints (18)-(21) are used to determine re-arrangements of departments in consecutive periods. Constraint (18)

is used to calculate the rectilinear distance that the center of department i has moved from period t-1 to t. Constraints

(19) and (20) set variable vit to 1 if department i has different side lengths or different center coordinates in periods t-1

21
ACCEPTED MANUSCRIPT

and t. Constraint (21) ensures qit to be 1 if department i is assigned to different zones in periods t-1 and t.

Constraints (22) and (23) bound the zone re-arrangement variables which indicate whether the boundaries of zone z

have been changed from period t-1 to t or not. Constraint (24) defines the domains of all decision variables.

Particularly, the valid inequality Constraint (25) given below requires that if department j is on the right/top of

department i and department j' is on the right/top of department j then department j' must be on the right/top of

T
department i, which is not needed for Problem DFLP-FZ, but improves the computational efficiency of the MIP solvers

IP
(reducing about 1/3 of the computing time in our experiments).

CR
(25) sijte  s ejjt  1  sijet i  j, j  j , t , e

3.4 Symmetry breaking constraints

US
The solution space of the FLP in a facility with a rectangle floor is known to be symmetric due to the mirror effect of
AN
the axes. Symmetric solutions have different values for the decision variables but identical interdepartmental

distances. Heuristics or exact approaches for the FLP can be improved by eliminating the symmetricity of the

solution space. We first briefly review the existing symmetry-breaking constraints and then introduce two new types
M

of constraints for breaking diagonal and parallel symmetries in the FLP.


ED

Meller et al. (1999) proposed a symmetry-breaking constraint, called p-position strategy, that restrict the location of

a specified department p to one side of the central axis. The p-position strategy can be applied in Problem DFLP-FZ
PT

as follows:

c ptx  0.5Lxt and c pty  0.5Lty ,


CE

(26)

where department p can be any department, and t is an arbitrarily selected period. Note that since the layouts in
AC

multiple periods are not independent of one another in the DFLP, symmetry-breaking constraints should be enforced

only in one period. Sherali et al. (2003) proposed symmetry-breaking constraints by fixing the relative position of two

specified departments, which is called pq-position strategy as follows:

(27) c ptx  cqtx and c pty  cqty ,

22
ACCEPTED MANUSCRIPT

where p and q are critical departments that can be selected based on the departmental flows and areas, and t can be any

selected period. For the flexible bay representation, Konak et al. (2006) proposed a tighter version of the p-position

strategy using the half width of the total layout area instead of the whole facility area as follows:

1
(28) c ptx 
2 Lxt
a
i
it and c pty  0.5Lty ,

T
where department p is selected as the department with the maximum amount of flows, and t can be any selected period.

IP
The above symmetry-breaking constraints improve the search efficiency of branch-and-bound procedures because

CR
they can eliminate up to three-fourths of the symmetric solution space. Our computational experiment also showed

that including symmetry-breaking constraints could reduce the computational time as much as 75% while solving FLP

instances to optimality in CPLEX.

US
In addition to the axis-based symmetries discussed above, diagonal symmetries also exist in the FLPs with a
AN
square-shaped facility. We introduce new symmetry-breaking constraints to eliminate the diagonal symmetry as

follows:
M

(29) c ptx  c pty and l ptx  l pty ,


ED

where p is an arbitrarily selected department, and t must be a period in which the facility has a square shape.

In addition to the symmetries due to the minor effect of the facility, another type of symmetry, referred to as parallel
PT

symmetry in this paper, may also exist when the area of the facility is much larger than the total area requirement of the

departments. A solution is called parallel symmetric if all departments can be moved in parallel toward any direction
CE

within the facility without changing the interdepartmental distances. Parallel symmetric solutions can be eliminated

by using the following constraints.


AC

(30) c ptx  c pty  0 and

(31) Let  cite  cejt  0.5lite  0.5l ejt i  j, e

where p is a selected department, and t is a selected period. Constraint (30) fixes the location of department p, e.g.,

the one with the largest area requirement, to the origin (0, 0) while Constraint (31) guarantees that the maximum

23
ACCEPTED MANUSCRIPT

widths/heights of the department do not exceed the width/length of the facility. Note that the variables for department

centers should be redefined to allow negative values in order to include parallel symmetry breaking constraints in the

model.

4. The problem evolution algorithm (PEA)

Problem DFLP-FZ is a very difficult problem to solve optimally with many binary and continuous decision variables,

T
which can generally be classified into three groups: (1) variables for zone partitions, i.e., wzt , wzt , wzt , wzt ,  zkt ,  zkt , (2)

IP
1x 2x 1y 2y x y

variables for department assignments, i.e., citx , city , litx , lity , dijtx , dijty , rizt , sijtx , sijty , and (3) dependent variables, i.e.,

CR
uitx , uity , vit , qit , o1ztx , ozt2 x , o1zty , ozt2 y , for calculating re-arrangement costs of zones and departments between two consecutive

US
periods. Among all these variables, the most important ones are sijtx and sijty which determine the relative positions

of department pairs in the facility. Our experiments showed that when we fixed variables sijtx and sijty to a set of
AN
feasible values, a MIP solver could find optimal values of the other variables in reasonable CPU times. This

observation makes it possible to develop a heuristic algorithm for optimizing sijtx and sijty and using a MIP solver to
M

determine the values of other variables for the fixed values of sijtx and sijty .
ED

The heuristic algorithm proposed in this paper is referred to as the Problem Evolution Algorithm (PEA). The

general principle of the PEA is to decompose the original problem (Pn), which can be large-sized and difficult to solve,
PT

into a series of sub-problems Pn-1, Pn-2, …, P2, and P1 such that problem Pi+1 is a slightly more difficult to solve than to

solve problem Pi. The PEA starts from P1 moves towards Pn by solving each sub-problem sequentially. First, the
CE

simplest problem P1 is solved with an optimal solution denoted as  1 . In the following steps, problem P2 is solved to

find a near-optimal solution  2 based on solution  1 , P3 is solved based on solution  2 , and so forth, until the original
AC

problem Pn is solved based on solution  n 1 of problem Pn-1. During this process, the incumbent problem evolves

gradually from P1 to Pn while the incumbent solution finally becomes the solution for Pn.

The proposed PEA is different from traditional evolutionary algorithms because its evolutionary strategy is

implemented on the problem itself, not on the problem's solutions as in the case of most traditional meta-heuristic

24
ACCEPTED MANUSCRIPT

algorithms such as GA, SA, PSO, etc. This evolutionary process of the PEA is an imitation of the learning process of

a human being. In the beginning, a person (i.e., any algorithm) may not be able to solve a difficult mathematical

problem (i.e., Pn). But, if he/she is trained on solving the easier version of the problem (i.e., P1) at the beginning, and

gradually toward solving more difficult problems (i.e., P2, P3, ...) based on his/her previous experiences on easier

versions of the problem, then he/she would eventually be able to solve the most difficult problem. Our experiments

T
show that the proposed PEA performs well for facility layout problems. The general framework of the PEA is given

IP
in Fig. 4.

CR
1) Determine a proper evolution strategy and define a serial of evolution problems {P1, P2, …, Pn}, where

P1 is the simplest problem while Pn is the original one.

1
2)

3)
Solve P1 to find the optimal solution

Let x←1 US
AN
4) Repeat the following steps until x=n

a) Let x←x +1

b) Construct an initial solution  x for Px based on  x 1


M

c) Implement a local search to improve x

5) Return  x as the final solution for problem Pn


ED

Fig. 4 A general framework of the PEA

The above framework presents the general steps for solving a complex problem using the problem evolution strategy.
PT

We use the PEA framework combined with a MIP solver, named as PEA-LP, as a general solution approach for FLPs.
CE

In Fig.5, we present the detailed steps of the PEA-LP algorithm for FLPs.
AC

25
ACCEPTED MANUSCRIPT

1) Determine an evolution index i , e.g., i   f ij or i  ai , for each department i  N


j

2) Let  be an ordered set of departments sorted according to the randomized i , i.e., i i , where  i is

a random number in [ 1   ,1   ] and 0    0.5

3) Define P={Px|2≤x≤n} as the set of sub-problems where Px contains only the first x departments in set  .

 x represent the solution of Px and sx  x indicate the relative positions of departments in  x .

T
4) Let

IP
5) Let x←6

6) Use a MIP solver to solve Px to find the optimal solution x

CR
7) Repeat the following step a), b), c), and d) until x=n

a) Let x←x+1, add the xth department in  to Px-1 to form Px

b)

c)
US
Fix the relative positions of departments 1, 2, …, x-1 according to s x 1

Use a MIP solver to find an initial solution  x for Px


AN
d) Local search: Implement an iterated neighborhood search to improve x

8) Return  x as the final solution for problem Pn


M

Fig. 5 The PEA-LP for the facility layout problem

In Step 1, the evolution index i is defined for constructing a set of evolutionary sub-problems. The evolution
ED

index can be defined based on any metric indicating the importance of each department in the layout (e.g., area

requirements and material flows, etc.). In Step 2,  is an ordered set of the departments sorted in the descending
PT

order of i i where  i is a random number in [ 1   ,1   ] and 0    0.5 . The PEA-LP constructs sub-problems
CE

Px based on the order of departments given in set . Therefore, multiplying the evolution index with a random

number results in random sub-problems in each run of the algorithm while the departments that contribute to the total
AC

cost exceedingly are more likely to be considered in the earlier sub-problems.

In Step 4, the evolution process starts from P6 because an FLP instance with six departments can easily be solved to

optimality using any MIP solver. The local search algorithm used in Step 6d is an iterated neighborhood search (INS)

that optimizes the relative position relationships (RPR) of a department with respect to the other departments until no

further improvement can be made. In Fig. 6, we provide the detailed steps for the INS algorithm.

26
ACCEPTED MANUSCRIPT

INS for problem Px:

1) Repeat

2) Let Flag←0

3) For each department i in Px (in a random sequence)

4) Fix RPRs of all departments in Px

5) Unfix RPRs of department i to a number of neighboring departments in Px

T
6) Use a MIP solver to optimize the unfixed RPRs

IP
7) If the solution is improved Then let Flag←1

8) End For

CR
9) Until Flag =0

Fig. 6 The local search method INS

the MIP solver in Step 6.


US
Note that the total number of unfixed RPRs in Step 5 is the primary factor affecting the computational time used by

Therefore, this number should be small enough to ensure the computational efficiency of
AN
the local search.

To solve the DFLP-FZ, we can apply the PEA-LP sequentially to each period of the problem, and then apply the INS
M

algorithm on multiple periods as a whole. This process is described in Fig. 7.


ED

1) For each period t in T with random sequence

2) Use the PEA-LP algorithm in Fig.5 to determine the layout in period t

3) End For
PT

4) Repeat

5) Let Flag←0
CE

6) For each period t in T

7) Implement the INS procedure to improve the layout in period t


AC

8) If the solution is improved Then let Flag←1

9) End For

10) Until Flag=0

Fig. 7 The PEA-LP algorithm for FLPs with multiple periods

27
ACCEPTED MANUSCRIPT

5. A case study

Before studying the computational performance of the PEA-LP on FLP and DFLP instances from the literature, a

case study is presented to demonstrate the use of the zone-based DFLP approach. The case study is inspired from an

apparel manufacturer whose product demands are highly seasonal. All departments are grouped into three categories,

receiving/preparation units (departments 1, 8, and 10), finishing/packaging (departments 21, 24, and 25), and

T
manufacturing (all other departments). Another unique aspect of the problem is that the area requirements of some

IP
departments change significantly over the planning horizon. In addition, the facility has only limited empty space,

CR
making the problem quite challenging with 25 departments. Fig. 8 illustrates a layout found using a single zone.

This layout represents a solution that can be found by existing approaches to the unequal area DFLP on the continuous

some of them are located in the center of the facility. US


plane. In this layout, the receiving/preparation and finishing/packaging departments are not grouped together, and

Certainly, the presented layout is not a desirable one according


AN
to the requirement of the manufacturing system although it has a low cost. Fig. 9 illustrates a solution found by

grouping the departments into receiving/preparation, finishing/packaging, and manufacturing zones. Although this

layout has a higher cost, it satisfies the requirement of the manufacturing system. In addition, zone boundaries can be
M

used to locate aisles. In summary, the layout in Fig. 9 is a more practical dynamic block layout. As discussed in this
ED

case study, the zone-based DFLP approach defined in this paper allows facility designers to specify certain aspects of

the layout such as grouping of departments and assigning rough locations of zones in the facility.
PT
CE
AC

Fig. 8 A solution found by a single zone (Objective=65986.68).

28
ACCEPTED MANUSCRIPT

T
IP
Fig. 9 A solution found by three zones: receiving/preparation (departments 1, 8, and 10), finishing/packaging (departments 21, 24, and 25),

CR
and manufacturing (all other departments) (Objective= 68804.64).

6. Computational experiments

US
Several sets of well-known benchmark problems in the FLP literature were used to test the proposed methods and

algorithms in this paper. Although the PEA-LP is designed for the DFLP, it is also tested on the static FLP. The
AN
parameters and references of all test problems are summarized in Table 5 where the column Space ratio indicates the

total area requirement of departments to the available area of the plant floor. The PEA-LP was coded using
M

AMPL/CPLEX of version 12.6.1.0, and all experiments were carried out on a Mac computer with a 2.9GB memory and

Intel Core i7 CPU.


ED

Table 5. Summary of the parameters and data sources of the test problems.
Problem Dimensions Space Num. of Shape Num. of Num. of Data reference
ratio (%) Dept. constraint Periods Zones
F10 90x95 98.8 10 =3 1 1 (Montreuil et al. 2004)
PT

min
BA12 10x6 88.3 12 l =1 1 1 (Bazaraa 1975)
BA13 9x7 95.2 13 lmin=1 1 1
BA14 9x7 96.8 14 lmin=1 1 1
CE

LAC12x3 24x20 56.1 12 =2 3 1 Lacksonen (1997)


LAC12x5 24x20 60.4 12 =2 5 1
LAC20x3 30x20 80.8 20 =2 3 1
M11* 6x6.006 99.9 11 =4 1 1 (Meller 1992), (Liu and
AC

M15* 15x15.015 99.9 15 =5 1 1 Meller 2007)


M25* 15x5.005 99.9 25 =5 1 1
Tam20 40x35 85.7 20 =5 1 1 Tam (1992) and Gau &
Tam30 45x40 88.9 30 =5 1 1 Meller (1999)
SC30 15x12 90.6 30 =5 1 1 (Liu and Meller 2007)
SC35 16x15 80.0 35 =4 1 1

(1) Inner-approximation versus outer-approximation

29
ACCEPTED MANUSCRIPT

First, we carried out computational experiments on problem F10 to test the performance of the polyhedral

inner-approximation method proposed in this paper and compared it to the outer-approximation method of Sherali et al.

(2003) and Castillo and Westerlund (2005). Problem DFLP-FZ was optimally solved using AMPL/CPLEX for

various values of  , i.e., the maximum allowed area deviation. For the outer-approximation method,  values were

from -3% to -0.01%. Similarly,  values were from 3% to 0.01% for the inner-approximation method. The

T
results are listed and compared in Table 6, where  indicates the deviation from the best solutions found with  =-0.01%

IP
for the outer-approximation method (i.e., 7649.99) and with  =0.01% for the inner-approximation method (i.e.,

CR
7653.29).

Table 6. Comparison of different linearization methods for F10 (=3)


Linearization method

Outer-approximation 
obj.
-3%

7397.33
-2%

6800.40
US
Optimal solutions (obj.) of F10 under different maximum area deviations (  )
-1%

7517.97
-0.5%

7557.9
-0.1%

7631.21
-0.05%

7646.9
-0.01%

7649.99
AN
 -3.30% -11.11% -1.73% -1.20% -0.25% -0.04% 0%
Inner-approximation  3% 2% 1% 0.50% 0.10% 0.05% 0.01%
obj. 8897.58 8079.6 7886.73 7732.27 7668.24 7659.92 7653.29
 16.26% 5.57% 3.05% 1.03% 0.20% 0.09% 0%
M

It can be observed in Table 6 that the linearization for department areas have an impact on the final solution. The

larger the parameter  is, the higher objective function value is in the case of inner-approximation.
ED

For the

inner-approximation method, the maximum deviation in the objective function can be as high as 16.26% (  =3%) and
PT

5.57% (  =2%) compared to that of  =0.01%. According to the Lemma defined in Section 3.2, the optimal solution

for F10 should be within the range of [7649.99, 7653.29]. Finally, we made more precise calculations by setting 
CE

=-0.001% and  =0.001% to reduce the range of the optimal solution to [7651.21, 7651.60].
AC

In Fig. 10, the CPU time used by the AMPL/CPLEX solver is plotted against different area deviations. It can be

observed that a longer CPU time is needed for a more precise approximation to area linearization. While the area

deviation is greater than 0.05% for the inner-approximation method or smaller than -0.05% for the outer approximation

method, there is not much difference in the CPU times. It means that fewer or more tangential and secant lines will

not make much difference on the CPU time used. An exceptional case is that when we set  =3%, the CPU time

30
ACCEPTED MANUSCRIPT

increased sharply to 3057 seconds because the secant lines forced departments to be much larger and resulted in a very

tight layout optimization problem within the limited facility space.

2000
3057

1500

T
CPU time (s)

IP
1000

CR
500

0
-3 -2 -1 US
-0.5 -0.1 -0.05 -0.01 -0.001 0.001 0.01 0.05
Maximum Area Deviation (%)
0.1 0.5 1 2 3
AN
Fig. 10. The computational time under different maximum area deviation

Next, we performed additional experimental computations to compare the outer-approximation and the
M

inner-approximation methods using =4 and =5. The results are shown in Table 7. It is noticeable that since the

space utilization ratio of F10 is 98.8%, the inner-approximation method under  =3% or 2% makes the problem with a
ED

larger aspect ratio very difficult to solve as CPLEX could not solve it to optimality even after several hours of CPU
PT

time.

Table 7. Comparison of the linearization methods for F10 (=4, =5)


CE

Outer-approximation Inner-approximation
 -3% -2% -1% -0.5% -0.1% -0.05% -0.01% 0.01% 0.05% 0.1% 0.5% 1% 2% 3%

=4 obj. 6544.89 6535.82 7168.25 7242.92 7404.26 7421.56 7422.63 7424.82 7427.65 7431.24 7454.62 7670.06 -- --
AC

 -11.83% -11.95% -3.43% -2.42% -0.25% -0.01% 0.00% 0.00% 0.04% 0.09% 0.40% 3.30% -- --
time 861 666 1376 2367 3747 1992 3964 7169 2846 2411 2356 6881 -- --
=5 obj 6620.26 5890.63 6735.16 6757.10 6888.72 6902.54 6913.90 6916.48 6922.46 6932.34 7131.37 7200.58 -- --
 -4.25% -14.80% -2.59% -2.27% -0.36% -0.16% 0.00% 0.00% 0.09% 0.23% 3.11% 4.11% -- --
time 2860 206 1804 1539 1424 1671 2068 4047 2363 1602 3162 4254 -- --

31
ACCEPTED MANUSCRIPT

(2) Removing symmetric solutions

In this section, we present the effectiveness of the new symmetry-breaking constraints introduced in the paper, i.e.

the diagonal and parallel symmetry-breaking constraints. We first modified problem F10 by expanding its facility

from 90x95 to 95x95 (denoted as problem F10*) so that the diagonal symmetry-breaking constraints are applicable in

problem F10*. Then, we solved problem F10* optimally with and without the parallel symmetry-breaking constraint

T
using AMPL/CPLEX. Next, we increased the facility size to 180x190 (denoted as problem F10**) so that the parallel

IP
symmetry-breaking constraint could be effective. Problem F10** was solved optimally with and without the parallel

CR
symmetry-breaking constraint. The results and CPU times of these experiments are listed in Table 8.

Table 8. Results of F10* and F10** with different symmetry-breaking constraints

Problem

F10*
Parameters

95x95,  =3 Solution
6713.06
Without
US
Diagonal symmetry-breaking constraint

CPU time (s)


151
Solution
6713.06
With
CPU time (s)
65
Time saved
57%
AN
Parallel symmetry-breaking constraint
Problem Parameters
Without With
Solution CPU time (s) Solution CPU time (s) Time saved
F10** 180x190,  =3
5745.06 58 5745.06 29 50%
M

A seen in Table 8, both symmetry-breaking constraints can reduce CPU times significantly. In the case of problem

F10*, the diagonal symmetry-breaking constraint reduced the CPU time by 57% while the parallel symmetry-breaking
ED

constraint reduced the CPU time by 50% for problem F10**. Note that the axis symmetry-breaking constraints were

also considered in these experiments.


PT

(3) Testing the PEA-LP algorithm


CE

Problem DFLP-FZ and the PEA-LP can be applied to solve the existing FLP test cases from the literature using only

one zone and a single period. This enables us to benchmark the performance of the PEA-LP to the existing
AC

approaches on a set of rigorously studied test problems from the literature. Therefore, we used the PEA-LP to solve

static FLPs, where the inner-approximation method was used to linearize the department area constraint with a

maximum permitted deviation  . In addition, we tested the PEA-LP on three DFLP test cases from Lacksonen

(1997), and a single zone was used for these problems as well. We set i  ai and used =0.1 to determine the

evolution sub-problems from P1 to Pn. Since the floor area is exactly equal to the total required area of the

32
ACCEPTED MANUSCRIPT

departments in problems M11*, M15*, and M25*, we increased the facility by 0.1% in the x-axis direction for these

three test problems so that feasible solutions can be found using the inner-approximation method. In the experiments,

each problem was solved for 10 random replications. The results are shown in Table 9 and compared to the best

solutions published in the literature. In Table 9, column Fs indicates the number of feasible solutions found in 10 runs

and column Dev.% indicates the percent deviation from the previous best-known solutions which can be found in

T
Kulturel-Konak and Konak (2013 and 2015) and Gonçalves and Resende (2015).

IP
Table 9. Overall performances of the PEA-LP algorithm
PEA-LP (10 runs)

CR
Problem % Fs Best Dev.% Average Dev.% CPU Time(s) Prev. Known Best
F10 0.1 10 7667.62 0.22 7667.62 0.22 6.8 7650.95
BA12 0.1 10 8021.00 0.00 8193.93 2.16 17.3 8020.97
BA13 0.1 10 4629.70 0.03 4644.48 0.35 17.8 4628.32
BA14
Lac12x3
Lac12x5
0.1
0.1
0.1
10
10
10
4629.70
6552.66
11074.97
0.02
-1.06
-2.96
US
4746.97
6633.00
11174.21
2.55
0.15
-2.09
20.1
135.7
195.2
4628.79
6622.82
11412.39
AN
Lac20x3 0.1 10 12339.09 1.57 12648.74 4.12 672.4 12148.60
M11* 0.01 8 1147.75 -1.40 1241.01 6.61 13.8 1164.07
M15* 0.01 8 24815.22 -2.75 28863.82 13.12 39 25515.84
M25* 0.01 9 1223.22 2.80 1365.16 14.73 425.7 1189.88
Tam20 0.1 10 8058.20 0.90 8307.41 4.02 169.6 7986.48
M

Tam30 0.1 10 18942.01 1.08 19338.01 3.19 614.3 18740.30


SC30 0.1 10 3321.79 0.09 3563.41 7.37 242.3 3318.76
SC35 0.1 10 3457.76 4.25 3518.90 6.09 414.4 3316.77
ED

Avg. 0.20 4.47 213.2


Note: Boldface font indicates the best-known value.

On the average, the best solutions found by the PEA-LP is only 0.20% higher than the previous best-known solutions
PT

in the literature. For problems Lac12x3, Lac12x5, M11*, and M15*, the PEA-LP found new best solutions. For

problems F10, BA12, BA13, and BA14, the PEA-LP found solutions with slightly larger objective function values than
CE

the best-known solutions. However, if we compare the layouts of new solutions with the previous ones, they have the

same relative positions of departments in all problems. These differences in objective function values are due to the
AC

different area approximation methods used. Therefore, we can claim that the PEA-LP algorithm also found the

best-known solutions for problems F10, BA12, BA13, and BA14. However, the PEA-LP algorithm did not obtain

solutions better than or equal to the previously-known best ones for problems Lac20x3, M25*, Tam20, Tam30, SC30,

and SC35. Most of these problems were solved in less than 10 minute CPU times. The computational efficiency of

33
ACCEPTED MANUSCRIPT

the PEA is also very good considering the fact that most of the best-known solutions were found by the methods

requiring CPU times longer than an hour.

We compare the performance of the PEA-LP with the existing solution approaches listed in Table 10. The

comparisons of the approaches based on the objective function values, percent improvements from the previously

reported solutions, and computing times used are shown in Table 11, 12, and 13, respectively. In Table 11, it can be

T
observed that the PEA-LP found solutions better than most of the benchmark approaches. Among the best-known

IP
solutions, the PEA-LP algorithm found 8 of them, BRKGA-LP found 6, and LS-HAS found 2. Table 12 shows the

CR
deviations of the solutions by different approaches from the previously best-known solutions. As seen Table 12, the

PEA-LP(b) has improved the best-known solutions of the four problems, and its average deviation, which is 0.2%, is

US
the lowest one among all benchmark approaches. Table 13 provides the computational times, if available, of the

benchmark approaches. It can be observed that the PEA-LP has very good computational efficiency and uses the least
AN
computing time among all approaches.

Table 10. Solution approaches in literature for comparison


Abbreviation Method description Source
M

CRAFT Steepest descent pairwise interchange Armour and Buffa (1963)


MIP-RB&B Mixed integer programming with revised branch and bound strategy Lacksonen (1997)
GA-ST Genetic algorithm with slicing-tree Gau and Meller (1999)
ED

ACO-ZB Ant colony optimization with zone-based layout Montreuil et al. (2002)
MIP-ɛ Mixed integer programming with maximally allowed area error Castillo and Westerlund (2005)
GA-MIP Genetic algorithm with mixed integer programming Liu and Meller (2007)
TS-ST TabuSearch with slicing-tree scholz, Petrick, and Domschke (2009)]
PT

AS-FBS Ant system with flexible bay structure Wong and Komarudin (2010)
ACO-ST Ant colony optimization with slicing tree Komarudin and Wong (2010)
ACO-LS Ant colony optimization with a local search Kulturel-Konak and Konak (2011b)
PSO-LS Particle swarm optimization with a local search Kulturel-Konak and Konak (2011a)
CE

GA-LP(b) Genetic algorithm with linear programming (best of 10 runs)


Kulturel-Konak and Konak (2013)
GA-LP(a) Genetic algorithm with linear programming (average of 10 runs)
LS-HAS(b) Large scale hybrid simulated annealing (best of 10 runs)
Kulturel-Konak and Konak (2015)
AC

LS-HAS(a) Large scale hybrid simulated annealing (average of 10 runs)


BRKGA-LP(b) BRKGA-LP (best of 10 runs) Gonçalves and Resende (2015)
PEA-LP(a) Problem evolution algorithm (average of 10 runs)
This study
PEA-LP(b) Problem evolution algorithm (best of 10 runs)

34
ACCEPTED MANUSCRIPT

Table 11. Comparison of the PEA-LP with the previous methods.


Methods F10 BA12 BA13 BA14 Lac12x3 Lac12x5 Lac20x3 M11* M15* M25* Tam20 Tam30 SC30 SC35
CRAFT 8756.00
MIP-RB&B 7094 12271 12903
GA-ST 8485.4 4804.1 29157.6 1591.3 9513.5 20658
ACO-ZB 8567
MIP-ɛ 8180 4919.47
GA-MIP 8702 4852 5004 1171 28526 1371 3707 3604

T
TS-ST 8264 4712.33

IP
AS-FBS 8252.67 4724.68 3868.55 4132.36
ACO-ST 8252.67 4724.68
ACO-LS 9020.75 8801.33 5088.05 4904.67 1268.55 27350.76 9003.82 19667.45 3794.82 4001.31

CR
PSO-LS 9020.75 8129 4780.91 8753.57 19462.41 3443.34 3700.75
GA-LP(a) 7878.17 8030.97 4745.73 4774.2 1173.92 26141.01 1271.36 8158.95 19133.61 3554.429 3493.01
GA-LP(b) 7651.28 8021 4628.32 4686.81 1164.07 25515.84 1189.88 8058.06 19009.9 3370.98 3385.48
LS-HAS(a)
LS-HAS(b)
BRKGA-LP(b) 7650.95 8020.97 4628.79
US
6728.29 11711.47 12325.80
6622.82 11412.39 12148.60
7986.48 18740.3
3,824.33 3,699.49
3318.76
3367.78
3469.4
3316.77
AN
PEA-LP(a) 7667.62 8193.93 4644.48 4746.97 6633 11174.21 12648.74 1241.01 28863.82 1365.16 8307.41 19338.01 3563.41 3518.9
PEA-LP(b) 7667.62 8021 4629.7 4629.7 6552.66 11074.97 12339.09 1147.75 24815.22 1223.22 8058.2 18942.01 3321.79 3457.76
Note: Boldface indicates the best-known value.
M
ED
PT
CE
AC

35
ACCEPTED MANUSCRIPT

Table 12. Deviations (%) from previous best-known objective values


Lac Lac Lac
Methods F10 BA12 BA13 BA14 M11* M15* M25* Tam20 Tam30 SC30 SC35 AVG
12x3 12x5 20x3
CRAFT 32.21 32.21
MIP-RB&B 7.11 7.52 6.21 6.95
GA-ST 5.79 3.79 14.27 33.74 19.12 10.23 14.49
ACO-ZB 11.97 11.97
MIP-ɛ 1.98 6.29 4.14
GA-MIP 8.49 4.83 8.11 0.60 11.80 15.22 11.70 8.66 8.68

T
TS-ST 3.03 1.80 2.42

IP
AS-FBS 2.89 2.07 16.57 24.59 11.53
ACO-ST 2.89 2.07 2.48

CR
ACO-LS 17.90 9.73 9.93 5.96 8.98 7.19 12.74 4.95 14.34 20.64 11.24
PSO-LS 17.90 1.35 3.30 9.60 3.85 3.75 11.58 7.33
GA-LP(a) 2.97 0.12 2.54 3.14 0.85 2.45 6.85 2.16 2.10 7.10 5.32 3.24

US
GA-LP(b) 0.00 0.00 0.00 1.25 0.00 0.00 0.00 0.90 1.44 1.57 2.07 0.66
LS-HAS(a) 1.59 2.62 1.46 15.23 11.54 6.49
LS-HAS(b) 0.00 0.00 0.00 0.00 4.61 0.92
BRKGA-LP(b) 0.00 0.00 0.00 0.00 0.00 1.48 0.00 0.21
AN
PEA-LP(a) 0.22 2.16 0.35 2.55 0.15 -2.09 4.12 6.61 13.12 14.73 4.02 3.19 7.37 6.10 4.47
PEA-LP(b) 0.22 0.00 0.03 0.02 -1.06 -2.96 1.57 -1.40 -2.75 2.80 0.90 1.08 0.09 4.25 0.20
M

Table 13. Computational times by different approaches


Methods F10 BA12 BA13 BA14 Lac12x3 Lac12x5 Lac20x3 M11* M15* M25* Tam20 Tam30 SC30 SC35
ED

MIP-RB&B 192 45196 69945


GA-ST 412 396 499 1397
ACO-ZB 26300
PT

GA-MIP 2988 2880 14652 57744


TS-ST 13.6 16.1
AS-FBS 164 292 19135 28671
CE

ACO-ST 4104 8568


ACO-LS 51 146 131 226 623 902 1185
PSO-LS 2 10 19 104 924 873 1842
AC

GA-LP(a) 300 600 750 2250 7350 2203.8 2953.9


LS-HAS(a) 2981.19 3951.35 6342.37 7799 9524
BRKGA-LP 49.15 113.81 160.46 686 3004.55 998.74 1556.03
PEA-LP(a) 6.8 17.3 17.8 20.1 135.7 195.2 672.4 13.8 39 425.7 169.6 614.3 242.3 414.4

To demonstrate the evolutionary process of the PEA-LP, we solved problems Tam30 and SC30 while recording the

changes in the objective function value of the incumbent solution during the entire search process. Fig. 12A and Fig.
36
ACCEPTED MANUSCRIPT

12B illustrate the shifts in the objective function values for problems Tam30 and SC30, respectively. These figures

reveal the detailed steps of evolving the incumbent problem from P1 toward Pn. Each jump in the objective function

value represents a move from Px to Px+1 because the PEA-LP adds a new department to the problem. The reductions

in the objective function values are due to exploring a better solution for Px+1 through the iterative neighborhood search

procedure.

T
4000

IP
20000
3500
17500

CR
3000
15000
Objective function

Objective function 2500


12500

2000
10000

7500

5000 US 1500

1000
AN
2500 500

0 0
0 500 1000 1500 0 100 200 300
Time (s) Time (s)

(A) Tam30 (B) SC30


M

Fig. 12. The evolutionary processes of PEA

We also studied DFLP cases Lac12x3, Lac12x5, and Lac20x3 with multiple flexible zones. These modified
ED

problem instances are denoted as Lac12x3*, Lac12x5*, and Lac20x3*, respectively. Problems Lac12x3* and Lac12x5*

have three flexible zones in each period, and problem Lac20x3 has four flexible zones in each period. In these
PT

problems, no pre-specified relative position relations among the zones are defined, but departments 1, 2, and 3 are

required to be assigned to zones 1, 2, and 3, respectively. The best and average results found by the PEA-LP in 10
CE

random replications are given in Table 14. Compared to the best solutions found for problems Lac12x3, Lac12x5,

and Lac20x3 in Table 8, the solutions for problems Lac12x3*, Lac12x5*, and Lac20x3* have larger objective values as
AC

expected. Fig. 13 illustrates the layout of the best solution found for problem Lac12x3*. Although this solution has

a higher objective function value than the best-known solution (Kulturel-Konak and Konak, 2015) given in Table 8, it

has certain advantages. Foremost, the zone boundaries define possible locations to place aisles and material-handling

equipment. Therefore, the layout in Fig. 13 is more applicable in real-life cases. In addition, the DFLP-FZ is able to

account the cost of structural changes in the layout between two consecutive periods. In the best solution, for example,

37
ACCEPTED MANUSCRIPT

Zone No.1 is expanded from period 1 to period 2 to include department 4 without relocating departments 11, 6, 1, 12,

10, 5, and 4, and this will not result in any cost if the traditional relayout cost function is used. In practice, however,

such a change may require extending the horizontal aisle between zones 1 and 2 or making changes in the wall

structure of the layout. The DFLP-FZ can account for such types of costs.

In addition, the DFLP-FZ allows relocation a block of departments that should be grouped together because of

T
technical and practical reasons. The proposed modeling framework can handle rearrangement decisions for a group of

IP
departments, which has not previously considered in the literature.

CR
Table 14. Performances of the PEA-LP on DFLP-FZ
PEA-LP
Problem Best Avg. Avg. Time(s)
Lac12x3
Lac12x5
Lac20x3
*

*
6771.74
11154.24
13085.64
US 6903.55
11374.69
13588.92
200
298
500
AN
M
ED
PT
CE
AC

(A) Period 1 (B) Period 2

38
ACCEPTED MANUSCRIPT

T
IP
CR
(C) Period 3

7. Conclusions
US
Fig. 13. Best solution of Lac12x3* (objective=6771.74)
AN
In this study, we present a comprehensive MILP model for the dynamic facility layout problem with flexible zones

(DFLP-FZ) to model block layout designs of manufacturing facilities/departments in complex industrial environments.
M

With flexible zone partitioning of the facility, the DFLP-FZ model is applicable to many cases of industrial layout
ED

designs, such as traditional FLPs, zone-based FLP, dynamic FLP, multi-floor FLP, FLP with bays, FLP with a

non-rectangular floor, etc.


PT

The problem evolution algorithm (PEA) proposed in this paper is a promising approach for solving various types of

FLPs with good solutions and high computational efficiency. Computational experiments on the well-known
CE

benchmark cases of FLPs showed that the PEA-LP solved 9 problems out of 11 with better or same solutions than the

previous best-known solutions in relatively short CPU times. Besides, the PEA is also a new framework of
AC

meta-heuristics, which is applicable to other optimization problems where the problem can be decomposed into a set of

evolutionary sub-problems with an increasing order of computational complexity.

We also propose the polyhedral inner-approximation method based on secant lines for the linearization of the

non-linear area constraint of departments with a given maximum deviation. This inner-approximation method

guarantees that the actual area is greater than or equal to the required area, so it provides a safer option than the

39
ACCEPTED MANUSCRIPT

outer-approximation method. We provide some experimental data for the reader to integrate the inner-approximation

method in their applications quickly. Two new symmetry-breaking constraints, called diagonal symmetry-breaking

and parallel symmetry-breaking, are introduced to increase the computational efficiency of the exact algorithms in

cases where the facility has a square shape and/or a significant amount of empty space.

Furthermore, we consider re-layout of department blocks in the context of the DFLP for the first time. In our

T
approach, a group of departments can be pre-assigned to flexible zones, and then relayout of these flexible zones in

IP
addition to individual departments within the zones can be optimized throughout the planning horizon.

Promising future research lies in two aspects: (1) to develop better evolutionary techniques that can shorten the

CR
evolutionary process further or have higher computational efficiencies, and (2) to enhance the dynamic model with

further consideration such as demand uncertainties in particular industries. Furthermore, the presented model can be

US
expanded by considering input and output locations of departments and zones. This enhancement will make the

problem one step closer to an integrated facility design.


AN
Acknowledgments

This work is partially supported by the National Natural Science Foundation of China (Grant No.71271009 &
M

71501007) and the Aviation Science Foundation of China (Grant No.2014ZG51075).

References
ED

1. Armour G-C, Buffa E S. A heuristic algorithm and simulation approach to relative location of facilities. Management Science 1963;

9(1): 294–309.
PT

2. Arostegui Jr M A, Kadipasaoglu S N, Khumawala B M. An empirical comparison of Tabu Search, Simulated Annealing, and Genetic

Algorithms for facilities location problems. International Journal of Production Economics 2006; 103: 742-754.
CE

3. Balakrishnan J, Cheng C H. Dynamic layout algorithms: a state-of-the-art survey. Omega 1998; 26(4): 507-521.

4. Baykasoglu A, Dereli T, Sabuncu I. An ant colony algorithm for solving budget constrained and unconstrained dynamic facility layout

problems. Omega 2006; 34: 385-396.


AC

5. Balakrishnan J, Cheng C H. The dynamic plant layout problem: Incorporating rolling horizons and forecast uncertainty. Omega 2009;

37: 165- 177.

6. Balakrishnan J, Cheng C H, Conway D G, Lau C M. A hybrid genetic algorithm for the dynamic plant layout problem. International

Journal of Production Economics 2003; 86(2): 107-120.

7. Bazaraa M S. Computerized layout design: A branch and bound approach. AIIE Transactions 1975; 7: 432–438.

40
ACCEPTED MANUSCRIPT

8. Bazaraa M S, Elshafei A N. An exact branch and bound procedure for quadratic assignment problems. Naval Research Logistics

Quarterly 1979; 26: 109–121.

9. Bozer Y A, Meller R D. A re-examination of the distance based facility layout problem. IIE Trans. 1997; 29(7): 549–560.

10. Castillo I, Westerlund T. An ε-accurate model for optimal unequal area facility layout. Computers and Operations Research 2005; 32:

429–447.

11. Chae J, Peters B A. Layout design of multi-bay facilities with limited bay flexibility. Journal of Manufacturing Systems 2006; 25(1), 1–

T
11.

IP
12. Chen G Y H. A new data structure of solution representation in hybrid ant colony optimization for large dynamic facility layout

problems. International Journal of Production Economics 2013; 142: 362-371.

CR
13. Drira A, Pierreval H, Hajri-Gabouj S. Facility layout problems: A survey. Annual Reviews in Control 2007; 31: 255–267.

14. Dunker T, Radons G, Westkamper E. Combining evolutionary computation and dynamic programming for solving a dynamic facility

US
layout problem. European Journal of Operational Research 2005; 165 (1): 55–69.

15. Gau K-Y, Meller R D. An iterative facility layout algorithm. International Journal of Production Research 1999; 37: 3739–3758.

16. Gonçalves J F, Resende M G C. A biased random-key genetic algorithm for the unequal area facility layout problem. European Journal
AN
of Operational Research 2015; 246(1): 86–107.

17. Hitchings G G. Control, redundancy and change in layout systems. AIIE Transactions 1970; 2: 253-260.
M

18. Hosseini-Nasab H, Emami L. A hybrid particle swarm optimization for dynamic facility layout problem. International Journal of

Production Research 2013; 51(4): 4325-4335.


ED

19. Hungerländer P, Rendl F. A computational study and survey of methods for the single-row facility layout problem. Computational

Optimization and Applications 2013; 55(1): 1–20.

20. Komarudin, Wong K Y. Applying ant system for solving unequal area facility layout problems. European Journal of Operational
PT

Research 2010; 202: 730–746.

21. Konak A, Kulturel-Konak S, Norman B-A, Smith A-E. A new mixed integer programming formulation for facility layout design using
CE

flexible bays. Operations Research Letters 2006; 34: 660–672.

22. Koopmans T C, Beckmann M. Assignment problems and the location of economic activities. Econometrica 1957; 25: 53-76.
AC

23. Kulturel-Konak S. Approaches to uncertainties in the facility layout problems: perspectives at the beginning of the 21 st century. Journal

of Intelligent Manufacturing 2007; 18(2): 273-284.

24. Kulturel-Konak S, Konak A. A new relaxed flexible bay structure representation and particle swarm optimisation for the unequal area

facility layout problem. Engineering Optimisation 2011; 43: 1263-1287.

25. Kulturel-Konak S, Konak A. Unequal area flexible bay facility layout using ant colony optimisation. International Journal of

Production Research 2011; 49: 1877–1902.

41
ACCEPTED MANUSCRIPT

26. Kulturel-Konak S, Konak A. Linear programming based genetic algorithm for the unequal area facility layout problem. International

Journal of Production Research 2013; 51(14): 4302–4324.

27. Kulturel-Konak S. A linear programming embedded probabilistic tabu search for the unequal-area facility layout problem with flexible

bays. European Journal of Operational Research 2012; 223(3): 614–625.

28. Kulturel-Konak S, Konak A. A large scale hybrid simulated annealing algorithm for cyclic facility layout problem. Engineering

Optimization 2015; 47(7): 963–973.

T
29. Kusiak A, Heragu S. The facility layout problem. European Journal of Operational Research 1987; 29(3):229-253.

IP
30. Lacksonen T A, Emory Enscore Jr E. Quadratic assignment algorithms for the dynamic layout problem. International Journal of

Production Research 1993; 31(3): 503-517.

CR
31. Lacksonen T A. Static and dynamic layout with varying areas. Journal of the Operational Research Society 1994; 45: 59-69.

32. Lacksonen T A. Preprocessing for static and dynamic facility layout problems. International Journal of Production Research 1997;

35(4): 1095–1106.

US
33. Lawler E L. The quadratic assignment problem. Management Science 1963; 9(4): 586–599.

34. Lee Y H, Lee M H. A shape-based block layout approach to facility layout problems using hybrid genetic algorithm. Computers &
AN
Industrial Engineering 2002; 42: 237–248.

35. Lee K-Y, Roh M, Jeong H-S. An improved genetic algorithm for multi-floor facility layout problems having inner structure walls and
M

passages. Computers & Operations Research 2005; 32(4): 879–899.

36. Liu Q, Meller R D. A sequence-pair representation and MIP-model-based heuristic for the facility layout problem with rectangular
ED

departments. IIE Transactions 2007; 39: 377–394.

37. McKendall A R, Hakobyan A. Heuristics for the dynamic facility layout problem with unequal-area departments. European Journal of

Operational Research 2010; 201(1); 171–182.


PT

38. McKendall A R, Liu W-H. New Tabu search heuristics for the dynamic facility layout problem. International Journal of Production

Research 2012; 50(3): 867–878.


CE

39. Meller R D. Layout Algorithms for Single and Multiple Floor Facilities, University of Michigan, Doctoral Dissertation, 1992.

40. Meller R D. The multi-bay manufacturing facility layout problem. International Journal of Production Research 1997; 35(5): 1229–
AC

1237.

41. Meller R D, Narayanan V, Vance P H. Optimal facility layout design. Operations Research Letters 1999; 23: 117–127.

42. Meller R D, Bozer Y A. Alternative approaches to solve the multi-floor facility layout problem. Journal of Manufacturing Systems

1997; 16(3): 192–203.

43. Montreuil B, Ratliff H D. Utilizing cut trees as design skeletons for facility layout. IIE Transactions 1989; 21(2): 136-143.

42
ACCEPTED MANUSCRIPT

44. Montreuil B. A Modeling framework for integrating layout design and flow network design. Proceedings of Materials Handling

Research Colloq., Hebron, KY; 1990; 43-58.

45. Montreuil B. and Venkatadri U. Strategic interpolative design of dynamic manufacturing systems layouts. Management Science 1991;

37(6):682-694.

46. Montreuil B, Venkatadri U., Ratliff, H. D. Generating a layout design from a design skeleton 1993; IIE Transactions, 25(1): 3-15.

47. Montreuil B, Brotherton E, Marcotte S. Zone based layout optimization. In: Proceedings of 2002 industrial engineering research

T
conference, Orlando, USA; 2002. p.19-22.

IP
48. Montreuil B, Ouazzani N, Brotherton E, Nourelfath M. Antzone layout metaheuristic: coupling zone-based layout optimization, ant

colony system and domain knowledge. In: Progress in material handling research, Material Handling Institute of America, Charlotte,

CR
2004; p. 301-331.

49. Peters B A, Yang T. Integrated facility layout and material handling system design in semiconductor facilities. IEEE Transactions on

Semiconductor Manufacturing 1997; 10: 360-369.

US
50. Rosenblatt M J. The dynamics of plant layout. Management Science 1986; 32: 76–86.

51. Sahni S, Gonzalez T. P-complete approximation problems. Journal of Association for Computing Machinery 1976; 23: 555-565.
AN
52. Saraswat A, Venkatadri U, Castillo I. A framework for multi-objective facility layout design. Computers & Industrial Engineering 2015;

90: 167–176.
M

53. Scholz D, Petrick A, Domschke W. Stats: a slicing tree and Tabu search based heuristic for the unequal area facility layout problem.

European Journal of Operational Research 2009; 197: 166–178.


ED

54. Sherali H D, Fraticelli B M P, Meller R D. Enhanced model formulations for optimal facility layout. Operations Research 2003; 51:

629-644.

55. Solimanpur M, Vrat P, Shankar R. An ant algorithm for the single row layout problem in flexible manufacturing systems. Computers &
PT

Operations Research 2005; 32(3): 583–598.

56. Tam K Y. A simulated annealing algorithm for allocating space to manufacturing cells. International Journal of Production Research
CE

1992; 30: 63–87.

57. Ulutas B H, Kulturel-Konak S. An artificial immune system based algorithm to solve unequal area facility layout problem. Expert
AC

Systems with Applications 2012; 39: 5384–5395.

58. Ulutas B, Islier A. Dynamic facility layout problem in footwear industry. Journal of Manufacturing Systems 2015; 35: 55-61.

59. Urban T L. A heuristic for the dynamic facility layout problem. IIE Transactions 1993; 25(4): 57-63.

60. Wong K Y, Komarudin. Solving facility layout problems using flexible bay structure representation and ant system algorithm. Expert

Systems with Applications 2010; 37: 5523–5527.

43
ACCEPTED MANUSCRIPT

Appendix:

A1. New best solutions

T
3 4

IP
5

CR
4
6
2

3
y

2 10 9
1
5

US 7

8
AN
1

11
M

0
0 1 2 3 4 5 6
x
ED

M11* (1147.75)
PT
CE
AC

44
ACCEPTED MANUSCRIPT

15
3
7

12
13 11

10

8
9

T
9
y

IP
6
12

CR
3 2 1 15 14

0
0 3 6
US 9
5

12 15
AN
x
M15* (24815.22)

15
M
ED

12 11

16 26
27
14
7 8 9
PT

9
28
y

1
CE

5 12 15 18 4 3 29 30
6 10 13

25
AC

6 23
3 24

22
2
20 19

17 21
0
0 3 6 9 12 15
x
SC30 (3321.79)

45
ACCEPTED MANUSCRIPT

Period 1 Period 2
20 20

4
4

15 15

5 5
10 1 2 7 1 6 2 7
6 8 8

10 10
y

y
12

T
3 3
9 9
12

IP
5 5

11
11

CR
0 0
0 5 10 15 20 0 5 10 15 20
x x

US
Period 3
20
AN
10

15

12
1 5 7
M

3 6
10 8
y

9 4 2
ED

11
PT

0
0 5 10 15 20
x
CE

Lac12x3 (6552.66)
AC

46
ACCEPTED MANUSCRIPT

Period 1 Period 2
20 20

10
9 12 7
15 3 15

11

10 10
y

y
8 5 9 12
6 4 8 5 2 1 2
1

T
5 5

IP
4 11 3
6 10
7

CR
0 0
0 5 10 15 20 0 5 10 15 20
x x

Period 3 Period 4
20 20

15
7
US 15
7
AN
11 6 2 5 9 8 1
2 5
10 10
y

1
9
M

11 10
4 12
5 6 4 3 12 8 5
ED

3
10
0 0
0 5 10 15 20 0 5 10 15 20
x x
PT

Period 5
20

7 2
CE

15

4 9 5 6

8
AC

10
y

10
11 1
12

0
0 5 10 15 20
x

Lac12x5 (11074.97)

47
ACCEPTED MANUSCRIPT

A2. Detailed derivations of Eq. (2) and Eq. (3)

The area generated by the product of two coordinates of any point (x, y) on line y  kx  b, ( x  0, y  0) can be

expressed as A  xy  x(kx  b)  kx  bx , and the derivation of area A is A  2kx  b . Then, let A'=0, we can get
2

b b b2
x*  b / 2k and y  b / 2 .
*
So, the maximum area A* can be expressed as A*  x* y*   
2k 2 4k

T
The equation of the line crossing any two points, say (x0, y0) and (x1, y1), on the curve y=a/x can be written as

IP
a a a a a a b2 a( x0  x1 )2
y  kx  b   x  (  ) , where k   and b   . So we can derive A    ... 
*
.
4k 4 x0 x1

CR
x0 x1 x1 x0 x0 x1 x1 x0

( x0  x1 )2
Let  be the area deviation of A* from a, then we have   ( A  a ) / a  A / a  1   1 , and can get
* *

4 x0 x1

2 2 US
the equation x1  (2  4 ) x0 x1  x0  0 , and finally we can get the equation
AN
x1  (1  2  2    2 ) x0 (A1)

and
M

y1  a / x1 . (A2)
ED

Thus, we get the first cut line (x0, y0)-(x1, y1) on the curve y  a / x .

By using Equation (1A), we can get more cut point by the equation
PT

x p  (1  2  2    2 ) x p1 p  1 (A3)
CE

and

yp  a / xp . (A4)
AC

Let   1  2  2    2 , Equation (3A) can be simplified as x p   x p 1 , p  1 , and we can get

x p   p  x0 (A5)

log x p  log x0
and p  . (A6)
log 

48
ACCEPTED MANUSCRIPT

Let x0 be xmin, and to guarantee the furthest point (xmax, ymin) on the curve is covered, we must have

log x p 1  lg x1 log x max  log xmin log 


p max   1   1, (A7)
log  log  log 

where   1  2  2    2 and  is the maximum aspect ratio for a department. Thus, we get Equation (A7) to

estimate the maximum number of secant lines guaranteeing the maximum area deviation is under the given ratio  .

T
IP
By using Equation (A3), the following equations can be deduced.

CR
x p x p1   x 2p1   (  x p2 )2   3 x 2p2   3 (  x p3 )2   5 x 2p3    2 p1 x02

and

x p  x p1  (   1) x p1  (   1)  x p2  (   1)  2 x p3 


US  (   1)  p1 x0 .
AN
Thus, for the pth secant line, which is (xp-1, yp-1)- (xp, yp), the slop Kp and intercept Bp are calculated by

a a 
Kp     (A8)
x p 1 x p  2 p 1 2
x 0  2 p 1
M

a a x p 1  x p (   1)  p 1 x0 (   1) a (   1) max
Bp    a a  y (A9)
x p 1 x p x p 1 x p  x0
2 p 1 2
 p x0 p
ED
PT
CE
AC

49

You might also like