You are on page 1of 33

Accepted Manuscript

Integrated Slicing Tree Approach for Solving the Facility Layout


Problem with Input and Output Locations based on Contour Distance

Christian Friedrich, Armin Klausnitzer, Rainer Lasch

PII: S0377-2217(18)30001-8
DOI: 10.1016/j.ejor.2018.01.001
Reference: EOR 14902

To appear in: European Journal of Operational Research

Received date: 28 November 2016


Revised date: 23 December 2017
Accepted date: 2 January 2018

Please cite this article as: Christian Friedrich, Armin Klausnitzer, Rainer Lasch, Integrated Slicing Tree
Approach for Solving the Facility Layout Problem with Input and Output Locations based on Contour
Distance, European Journal of Operational Research (2018), doi: 10.1016/j.ejor.2018.01.001

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

• A concurrent method of designing the layout and material handling points is given.

• Distances along the perimeters of facilities are considered.

• Two improvement strategies for locating material handling points are proposed.

• Parallel tempering is introduced for the first time to the facility layout problem.

T
• The novel approach proved to be considerably better than comparable approaches.

IP
CR
US
AN
M
ED
PT
CE
AC

1
ACCEPTED MANUSCRIPT

Integrated Slicing Tree Approach for Solving the Facility Layout Problem
with Input and Output Locations based on Contour Distance

Christian Friedricha , Armin Klausnitzerb,∗, Rainer Laschb


a Chair of Management and Logistics, Technische Universität Darmstadt, Hochschulstr. 1, 64289 Darmstadt, Germany
b Faculty of Business and Economics, Technische Universität Dresden, Münchner Platz 1/3, 01062 Dresden, Germany

T
IP
Abstract

The designer of a plant layout faces mutually dependent subproblems that are traditionally solved in a

CR
sequential process. In this paper, a slicing tree based parallel tempering heuristic is used to solve the facility
layout problem, including the location of the material handling points of each facility. To determine the
input and output points on the perimeters of the departments for a given layout, two heuristic methods

US
are presented and compared to an integer linear programming approach. The shortest paths along the
perimeter of the departments are used to calculate the distance of the material flow paths from output to
AN
input points. The methods are embedded in a superordinate parallel tempering algorithm, which represents
a novel approach to solving the facility layout problem. To compare the efficiency of the superordinate
parallel tempering algorithm, two variants of the variable neighbourhood search are implemented as well.
M

The effectiveness of the approach in terms of the objective function value is shown by comparing the results
to those in the literature.
Keywords: Facilities planning and design, Integrated layout design, Unequal area facility layout problem,
ED

Slicing trees, Parallel tempering


PT

1. Introduction
CE

The facility layout problem (FLP), which describes the problem of finding the most efficient arrangement
of facilities in a given location, is an NP-hard optimisation problem, as shown by Sahni and Gonzalez
(1976). The FLP is concerned with determining a non-overlapping layout of n facilities that need to be
AC

placed within a limited or unlimited planar site while taking placement restrictions and the relationships
among the facilities into consideration.
During the past five decades, researchers have extensively addressed the FLP for various contexts and
problem specifications. In general, the FLP focuses on the efficient arrangement of machines or cells in a
manufacturing plant and is known to directly affect productivity, lead times, and manufacturing costs (Drira

∗ Correspondingauthor
Email address: armin.klausnitzer@tu-dresden.de (Armin Klausnitzer)

Preprint submitted to European Journal of Operational Research January 5, 2018


ACCEPTED MANUSCRIPT

et al., 2007). Closely related research has been conducted in other engineering contexts, including for the
efficient arrangement of temporary buildings on a construction site (Kumar and Cheng, 2015) and even for
the design of hospitals (Hahn and Krarup, 2001) and microprocessor layouts (Funke et al., 2016).
The primary tasks of layout planning are traditionally threefold and normally processed sequentially due
to the problem’s computational intractability. Once the shapes and locations of facilities are determined
within a block layout, the optimal input and output (I/O) locations, also called the pickup and delivery
points, are calculated. Ultimately, the design of material flow paths should define material transport routes

T
within the layout. Most often, the FLP is concerned with optimising the total transportation distance along

IP
the flow path or closely related costs of handling materials. Solving these subproblems sequentially can result
in far from optimal solutions (Hu et al., 2007). Once a block layout has been generated, each facility can

CR
be designed to accommodate the requirements of machines, equipment, and processes (Keller and Buscher,
2015).

US
In this paper, we present a novel, integrated approach for the efficient arrangement of facilities within a
floor area and their I/O locations by minimising the total transportation distance along the perimeters of
the facilities. We used two heuristic methods to calculate the positions of the I/O points: a constructive
AN
heuristic and a subordinate parallel tempering (PT) algorithm. To compare the effectiveness of the heuristics
with the optimal solution, we also applied an integer linear programming (ILP) formulation. These methods
were embedded within a superordinate PT algorithm to optimise layouts generated from a slicing tree based
M

on contour distances between the selected I/O points. To compare the efficiency of the superordinate PT,
we implemented two variants of variable neighbourhood search (VNS) as well.
ED

The paper is organised as follows: Section 2 provides a brief overview of the relevant literature for
layout planning, after which Section 3 explains the problem at hand and the related assumptions. Section 4
describes the applied coding and calculation scheme in detail, by first introducing a slicing tree approach
PT

and a construction method that afford efficient layout storage and construction and, later, by describing
a procedure to determine the I/O locations. Section 5 presents two heuristic methods to systematically
CE

explore the neighbourhoods: the PT approach, which is here applied to the FLP for the first time, and the
well-known VNS heuristic in two distinct variations. Section 6 reports a comparison of the performance of
our approach and of the solution methods, as well as the evaluation of two improvement strategies. Section 7
AC

concludes the paper with a summary of our findings and suggestions for future research.

2. Review of the Literature

In early facility layout approaches, the FLP was modelled as a quadratic assignment problem (QAP)
with equally sized discrete locations, only one of which each facility had to occupy. However, in the last two
decades, research on the FLP has shifted to continual representations of planar sites, in which the facilities

3
ACCEPTED MANUSCRIPT

can have different sizes. The FLP is commonly modelled as a mixed integer program (MIP), in which
restrictions prevent facilities from overlapping. Given the high complexity of the FLP, MIP models can be
solved only for small problem instances by using exact solution approaches. This limitation has generated
numerous studies on layout representations, encodings, and metaheuristics for the FLP (Anjos and Vieira,
2017).
Among recent research focused on applying novel solution approaches, some has involved integrating new
model assumptions to support the needs of practitioners. Although calculating the traveling distance metric

T
relates closely to considering I/O locations when optimising the placement of facilities, both processes have

IP
been treated differently. In general, the modelling of I/O points is simplified by locating the points at the
centroids of facilities, between which the material flow distance is calculated based on rectilinear distances

CR
(e.g. Gonçalves and Resende, 2015). Only a few approaches in the literature have deviated from this
assumption. For example, one considers I/O points at specific locations on the boundaries of the facilities,

US
meaning that either predefined I/O locations exist on the facilities’ boundaries or the assignment of I/O
points to potential positions is part of the optimisation problem. This extension of the FLP is appropriate
for calculating the transportation distance along paths and is more likely to meet practical requirements,
AN
because the location of I/O points on the boundaries guarantees their accessibility by material handling
devices.
Few authors have focused on the FLP with predefined I/O locations or potential positions. Wu and
M

Appleton (2002) took advantage of a slicing tree approach to minimise the contour-based distances, as
well as the number of flow paths, by assuming that the positions of I/O points within the facilities were
ED

already known. They applied an efficient genetic algorithm (GA), which coded the material paths within
a single string. Earlier works by Kim and Kim (1999) and Arapoglu et al. (2001), among others, planned
I/O points sequentially by assuming a given facility layout. Norman et al. (2001a) integrated information
PT

about material flow patterns (I- and U-shaped material flows) within the facilities by following a bottom-up
planning approach that allowed I/O points to be placed at distinct locations. They proposed an MIP for
CE

small problems and a GA with a subsequent greedy heuristic to improve the I/O points for larger problems
(Norman et al., 2001a).
The material flow path between two facilities is commonly based on the rectilinear distance between the
AC

centroids or I/O locations of the facilities (Drira et al., 2007). Less often, the distances are calculated based
on Euclidean distances (e.g. Ripon et al., 2013). In real layouts, if other facilities block the shortest path
between the facilities, then the rectilinear and Euclidean assumptions rarely hold up because the movement
of most types of material handling equipment such as forklifts or automated guided vehicles is restricted
to aisles and paths. Consequently, rectilinear or Euclidean-based layouts can be inferior to contour-based
layouts when applied to real situations (Figure 1). Such considerations initiated the adoption of the contour-
based distance measure for the FLP, which bars the material flows from crossing facilities and instead requires
4
ACCEPTED MANUSCRIPT

them to follow the perimeters of facilities.

Sawing
Milling

Warehouse

Drilling

T
Welding Assembly

IP
Shipping
Rectilinear distance
Euclidean distance

CR
Contour distance

Figure 1: Comparison of the Euclidean, rectilinear, and contour distance metrics.

US
Literature concerning the integrated planning of facility layouts and their I/O points remains scarce.
In such studies Banerjee et al. (1997) proposed a GA to create and evaluate graphs that illustrate the
AN
relative positions of facilities, with a subordinate mathematical programming routine that further improves
the layout. Later, Chittratanawat (1999) applied a tabu search heuristic (TS) to determine the location of
equally sized facilities and used a solver to locate the I/O points. Most recently, Kulturel-Konak (2017)
M

proposed an MIP for solving a zone-based dynamic FLP. The findings of their research suggest that a VNS
applying the energy function concept of simulated annealing (SA) can best search within a coding that
ED

represents a subset of decision variables of the MIP; in each iteration, the reduced problem is optimally
solved by using a solver. All three approaches employ a rectilinear distance metric. By some contrast, Kim
and Goetschalckx (2005) suggested using a sequence-pair coding scheme in combination with an SA and
PT

a linear program (LP) to arrange facilities. In this arrangement, I/O points are allowed to be located on
specific candidate positions. After the block layout and I/O points are determined by rectilinear distances,
CE

a heuristic approach can be applied to adjust the I/O positions in terms of a contour-based distance metric.
Norman et al. (2001b) suggested using a flexible bay model with the contour-based distance metric.
Unlike in other approaches, they simplified the solution approach by allowing an unrestricted number of I/O
AC

points for each facility. A GA was applied. Conversely, Lin and Lin (2015), who also used a flexible bay
structure, allowed only one I/O point per facility. They solved the problem by using a hybrid ant colony
algorithm (ACO), in which a constructive heuristic determines the locations of the I/O points.
Literature on the topic has identified four major aspects to support practical applications. Whereas using
the contour-based distance measure and the possibility of situating I/O points on the boundaries of facilities
can likely better model real-world requirements, designing the block layout and I/O locations requires less

5
ACCEPTED MANUSCRIPT

Table 1: Summary of related works.


Author Contour-based Block layout I/O location I/O points on facil- Solution approach
distances design design ities’ boundaries
Banerjee et al. (1997) – •** •*** • GA
Chittratanawat (1999) – •** •*** • TS
Kim and Kim (1999) • –** •*** • Branch and bound
Arapoglu et al. (2001) • –** •*** • SA, GA
Norman et al. (2001a) • –** •*** • GA + heuristic
Norman et al. (2001b) • •** •*** • GA
Wu and Appleton (2002) • •** –*** • GA

T
Kim and Goetschalckx (2005) • •** •*** • SA + LP
Ripon et al. (2013) – •** –*** – VNS
Gonçalves and Resende (2015) – •** –*** – GA

IP
Lin and Lin (2015) • •** •*** • ACO
Kulturel-Konak (2017) – •** • • VNS + SA + MIP
*Considers only equally sized facilities

CR
**Considers a flexible bay or a zone-based approach
***Includes an unrestricted number of I/O points

planning effort and can result in preferable layout solutions. Table 1 presents a summary of the relevant

US
literature. The fulfilment of aspects of an integrated approach based on contour distances is denoted by a •.
Our literature review revealed that studies on integrated approaches for planning facility layouts and I/O
AN
points are rare and that few apply integrated approaches that involve using contour distances. Although the
approach of Kim and Goetschalckx (2005) satisfies all of the identified requirements, the proposed layout
construction method shows two weaknesses. First, due to the use of an outer approximation method, areas of
M

facilities may deviate from the required areas. Second, each layout variant is evaluated twice, with rectilinear
and contour-based distances, which is usually considered to be the most time-consuming algorithmic step
(Sinha et al., 2017). In response, we propose an efficient layout construction procedure to overcome these
ED

disadvantages. Although metaheuristics have been extensively applied to solve the FLP, applying PT to the
FLP is novel and has not yet been studied in the literature.
PT

3. Problem Statement
CE

The problem at hand aims to optimise an FLP with unequal areas and variable I/O points at once by
using a contour-based distance metric. Our attempted solution is motivated by several design problems that
we have encountered in practice. In the metrological machinery manufacturer with small batch production
AC

that we studied, each product family is assigned to one self-managed manufacturing cell for assembly. Each
cell occupies a rectangular area and has one I/O location next to an aisle to which material is delivered
by either forklifts or tugger trains. Material flows within the cells are product dependent but begin and
end at the I/O locations. Several other industries have different enclosed areas as well, each of which
requires specific temperatures and clean room conditions (e.g. the medical equipment manufacturing and
the semiconductor industries). To minimise conditional fluctuations, often only one I/O location is supplied

6
ACCEPTED MANUSCRIPT

by forklifts or automated guided vehicles. Nevertheless, the problem specifications are also widely applicable
to other industrial problems, which are discussed below from a practical viewpoint:

1. All facility areas ai are rectangular. This assumption is reasonable because material flow paths usually
follow vertical and horizontal directions (Yang et al., 2005), meaning that irregularities are often used
as buffer zones and staging areas.
2. Each facility has a variable but restricted shape. Although a few approaches assume that shapes are

T
already known (e.g. Hu et al., 2007), in most situations the precise dimensions of facilities (e.g. cells,

IP
departments, and even machines) are unknown when planning the block layout (Scholz et al., 2009;
Keller and Buscher, 2015). Nevertheless, space and shape estimations can exist. For the latter, a

CR
maximum aspect ratio αimax can be applied that restricts the maximum permissible ratio between the
max{lis }
longest side max {lis } and shortest side min {lis }, i.e. αimax ≥ min{lis }
, in which s refers to the x and y
directions. Assuming a valid general aspect ratio for facilities of different areas, shape constraints can

US
be strengthened by a lower limit of the shortest side limin ≤ min {lis } of every facility i. This assumption
is reasonable because facilities can comprise machines or even cells that require a minimum length.
AN
3. The layout area is rectangular and described by its width wf and length hf . Although the layout area
is usually orthogonal due to the building’s construction, the area can have a different shape. To ensure
that the assumption accommodates real-world requirements, dummy facilities occupying forbidden
M

regions can be used (Scholz et al., 2010).


4. Each facility has one input point on the boundary that also functions as an output point. In practice,
the points for material deliveries or pickups often share locations and are close to main aisles in order
ED

to facilitate accessibility via discrete material handling systems. Standard unit loads often contain a
number of smaller ones (e.g. raw materials and semi-finished goods) that are usually moved separately
PT

within the facility according to procedural requirements. Therefore, the material flow within the
facility is often independent of the superordinate material handling system.
5. The shortest path along the object contours determines the bidirectional material flow path between two
CE

objects. This assumption is reasonable because floor-bound transports usually follow the main aisles
and cannot travel through walls, machinery, or fixed installations. Traditional algorithms measuring
AC

the rectilinear distances between the centroids of facilities cannot appropriately solve practical cases
with these requirements. When evaluating the resulting layouts using contour-based distances between
I/O locations on the boundaries of facilities, poor layout solutions can become apparent (Kim and
Goetschalckx, 2005).

Considering the planning details, the industry involved, and the objective pursued, the possibilities of
facility layout design are myriad. However, assuming that maximising profit is the overarching aim of a
company and that all possible layouts can generate the same revenue, the general objective is to minimise
7
ACCEPTED MANUSCRIPT

the total costs. Therefore, most FLP models aim to optimise the total transportation distances while
satisfying area and shape constraints (Scholz et al., 2009; Kim and Goetschalckx, 2005).
Total transportation distance (TTD) is defined as the sum of the material flows fij multiplied by the
corresponding transportation distances dij , as shown in the objective function (1).

n X
X n
Min TTD = dij fij (1)
i=1 j=1

T
j6=i

A common mathematical formulation of the problem using rectilinear distances between centroids was

IP
presented by Anjos and Vieira (2017). Instead of employing the common rectilinear or Euclidean distances

CR
between centroids, we consider contour-based distances between I/O points of each pair of connected facilities
i and j. An MIP was developed by Klausnitzer and Lasch (2017).

4. Coding and Calculating Layouts

4.1. Slicing Tree Representation


US
AN
The slicing tree approach, first applied to the FLP by Tam (1992), is a well-established method of
designing a floor plan of unequal areas. By dividing a rectangular area repeatedly in either vertical or
horizontal directions, each facility can be allocated to a resulting sub-area. This slicing operation is repeated
M

for each newly formed rectangle until every facility is allocated to a sub-area. Each potential floor plan can
be represented by a binary tree that shows the results of the recursive partitioning process. The internal
ED

nodes represent the slicing orientation. Each leaf is assigned a facility identifier that determines the position
for each facility. Thus, an FLP with n facilities is described by a tree structure of n leaves and n − 1 inner
nodes.
PT

The graphical representation of a slicing tree can be encoded in different ways. Similarly to Ripon et al.
(2013), we implemented a slicing structure encoding ε that consists of three parts. Figure 2 illustrates the
transformation from the coding expression to the slicing tree representation of n = 7 facilities. The first part
CE

represents the leaves of the tree, in which the facility identifier is assigned to one position. The second part
consists of n − 1 elements, in which each integer describes the position of a cut within the layout. Similar to
AC

the second part, the third part consists of n − 1 elements that represent the slicing orientation. Horizontal
(H) and vertical (V) cuts are denoted by 1 and 0, respectively. In the example in Figure 2, the first digit in
the slicing order is 4, which indicates an initial cut in the root node on the fourth position, located between
facilities 2 and 3. The orientation order applies a vertical cut, as indicated by the appearance of 0 as the
first digit.
Our approach assumes that the left branch below a node corresponds to the left layout partition or
upper layout partition in the case of a vertical or horizontal cut, respectively. Since the root node indicates

8
ACCEPTED MANUSCRIPT

a vertical cut, the available area is completely divided from the lower to the upper border. Facilities denoted
in the right branch are placed to the right of those of the left branch. The right branch indicates a horizontal
cut leading to the location of facility 2 above facilities 7 and 4, which are further separated by a horizontal
cut. The same procedure applies to the left branch below the root node.

1 5 6 3 2 7 4
V
2 7 4
1 5 6 3

T
Facility 1 5 6 3 2 7 4
H H
7 4
sequence 1 2 3 4 5 6 5 6 3
1 V 2 H

IP
Slicing order 4 1 5 3 6 2 5 6 H 3 7 4

Orientation 0 1 1 0 1 1 5 6

CR
(a) Layout encoding. (b) Slicing tree.

1
2

6
US 3
7

4
AN
(c) Slicing layout.

Figure 2: Example of the transformation of a layout encoding into a slicing tree and layout.
M

Although layouts constructed by a slicing tree are restricted because at least one cut is uninterrupted,
the approach is applicable for two reasons (Scholz et al., 2009). First, the optimal solutions generated by a
ED

solver using an MIP have often showed a slicing structure. Second, from the practitioners’ perspective, an
uninterrupted cut in the floor plan results in fewer nested aisles, which allows for smooth material transport.
PT

4.2. Layout Construction

A slicing tree illustrates how to partition a planning area but does not determine the coordinates of slices.
CE

To transform a slicing tree into a layout, a construction methodology is necessary. Scholz (2010) proposed
several methods for facilities with variable and fixed dimensions, and although the mentioned assumptions
AC

indicate that facilities possess no fixed dimensions, they often have predefined areas.
Initially, a worst case scenario for facility layout planning is considered, similar to the method used by
Pn
Ripon et al. (2013). It is assumed that the sum of facility areas i=1 ai is equal to the available area wf hf .
The coordinates of the left upper corners uli and right lower corners lri of every facility i are initialised by
(0, hf ) and (wf , 0), respectively.
Figure 3 illustrates the procedure for the first nodes and assumes a sequence of facility areas {750, 600,
800, 450, 450, 500, 450} and floor area dimensions {wf = 80, hf = 50}. The first node defines a vertical cut,

9
ACCEPTED MANUSCRIPT

1 5 6 3 2 7 4 1 5 6 3 2 7 4

V V V

H H H H H H

1 V 2 H 1 V 2 H 1 V 2 H

H 3 7 4 H 3 7 4 H 3 7 4

T
5 6 5 6 5 6

ul1 ul2

IP
uli ulj LC3
LC2 1 1
lr1 2
lr2
LC1 RC1 ulj

CR
ulj
1, 5, 6, 3 2, 7, 4 RC2 2, 7, 4
5, 6, 3 5, 6, 3 RC3
7, 4
lri lrj lrj
lrj

LC1 = {1, 5, 6, 3}
P
ai = 2500
uli : (0, 50)
lri : ( 2500
50
, 0)
RC1 = {2, 7, 4}
P
aj = 1500
ulj : ( 2500
50
lrj : (80, 0)
, 50)
LC2 = {1}
a1 = 750
ul1 : (0, 50)
lr1 : (50, 1750
50
)
US RC2 = {5, 6, 3}
P
aj = 1750
ulj : (0, 1750
50
lrj : (50, 0)
)
LC3 = {2}
a2 = 600
ul2 : (50, 50)
lr2 : (80, 900
30
)
RC3 = {7, 4}
P
aj = 900
ulj : (50, 900
30
lrj : (80, 0)
)
AN
∀i ∈ LC1 ∀j ∈ RC1 ∀j ∈ RC2 ∀j ∈ RC3

Figure 3: Example of the first 3 steps of the layout construction based on an encoded slicing tree.
M

which indicates that all facilities i on the left side of the cut belong to the left partition LC1 , whereas
those on the right side remain in the right partition RC1 . The location of each cut has to be determined
ED

in consideration of the area required by every partition. The x coordinates of the lower right corners
lri ∀i ∈ LC1 and upper left corners uli ∀i ∈ RC1 of the respective partitions need to be recalculated by
P
the proportion of the area belonging to the left partition and the total floor area length, i.e. i∈LC1 ai /hf .
PT

Accordingly, the first node of the left branch indicates a horizontal cut that divides LC1 into upper and
lower partitions LC2 and RC2 , respectively. Since the subsequent left branch is a leaf node, LC2 determines
the exact shape and position of the respective facility. Therefore, the y coordinates need to be similarly
CE

recalculated, and the described calculation procedure is recursively applied for every emerging partition.
The general procedure using the slicing tree encoding ε appears in Algorithm 4 in the Appendix.
AC

Since the layout construction procedure does not consider shape requirements (Assumption 2), violations
are considered while evaluating the layout (Section 4.4).

4.3. Input and Output Locations

Once the facilities have been arranged within the available area, I/O locations need to be defined.
Although they can be located at any position on the boundaries of the facilities, only points at which
horizontal and vertical boundaries of at least one facility meet may be considered (Kim and Kim, 1999).

10
ACCEPTED MANUSCRIPT

These locations may not be physically practical in later planning steps when designing each facility in detail;
however, placing the I/O locations at the nearest positions feasible will ensure near-optimal costs (Arapoglu
et al., 2001).
To determine I/O locations, the layout generated by the slicing tree needs to be transformed into an
undirected graph G = (V, E). The set of vertices V consists of intersection points connected by a set of
edges E corresponding to the borders of the facilities. Although every vertex of G can initially represent
a candidate for an I/O position, some candidates can nevertheless dominate others. To obtain a dominant

T
region, Kim and Kim (1999) suggested forming the smallest rectangle that borders at least one intersection

IP
of each department. Because departments cannot be divided, the corresponding vertices and edges not
located in the dominant region can be removed from graph G. Considering the layout constructed above,

CR
Figure 4(a) shows an example of the candidates, whereas Figure 4(b) shows the dominant region.

5
1
2
US 5
1
2
AN
7 7
3 3
6 4 6 4
M

(a) All candidates. (b) Dominant set of candidates.

Figure 4: Candidates for I/O positions.


ED

Once graph G with all candidates for I/O positions is constructed, the best set of I/O positions from
PT

graph G needs to be chosen. The problem can be modelled as a QAP, which is known to be an NP-hard
combinatorial problem (Arapoglu et al., 2001). An ILP formulation appears in the Appendix. However,
in its basic version, the approach applies a constructive heuristic for the I/O point selection, as shown in
CE

Algorithm 1.
The distance of the shortest paths is calculated by using the Floyd-Warshall algorithm. The heuristic
AC

procedure determines the I/O positions of facilities according to the descending order of the intensity of
their material flows. In every step, the pair of facilities with the highest material flow interaction is chosen.
Of the candidate positions of the corresponding facilities, those with the shortest distance are determined
to be I/O points, which do not change during the procedure. The algorithm stops when every facility is
assigned one I/O point.

11
ACCEPTED MANUSCRIPT

Algorithm 1 Constructive heuristic for input/output point selection


Let d be the matrix of minimum distances between all vertices calculated by the Floyd-Warshall algorithm and C be
the dominant set of possible I/O points.
1: procedure IO-Heuristic(d, f, C)
2: k←n
3: while k 6= 0 do . Stop when all objects have an I/O point.
4: i, j ← argmaxi,j∈[0,n) {fij } . Get (i, j) with the highest material flow between each other.
5: if ¬IOi ∧ ¬IOj then . Neither i nor j have an I/O point.
6: IOi , IOj ← argminp∈Ci {dpq } . Choose I/O points out of the candidates according to the shortest distance.
q∈Cj
7: k ← k − 2.
8: else if ¬IOi ∧ IOj then

T
9: IOi ← argminp∈Ci dp,IOj
10: k ← k − 1.

IP
11: else if IOi ∧ ¬IOj then
12: IOj ← argminq∈Cj dIOi ,q
13: k ← k − 1.
14: end if

CR
15: fij ← 0
16: end while
17: return I/O location of each facility i and dij between all connected I/O locations
18: end procedure

4.4. Calculation of the Objective Function Value US


Algorithm 4 provides the exact positions and shapes of facilities, whereas Algorithm 1 determines the
AN
I/O locations according to the shortest contour-based distances. To evaluate the quality of a layout coded
in ε, the objective function value (OFV) needs to be calculated. Since the proposed layout construction
method does not consider area and shape requirements, we convert the maximum aspect ratio and minimum
M

length requirements (Assumption 2) into a penalty term, which was first applied by Kulturel-Konak (2012).
The unconstrained objective function aims to minimise the sum of TTD and the penalty term, stated as
ED

follows:

n X
X n n
X
Min OFV = dij fij + pi + λ zi (2)
PT

i=1 j=1 i=1


j6=i
CE

where
AC

12
ACCEPTED MANUSCRIPT

pi = κ(max {0, lbxi − wi } + max {0, lbyi − hi }) (3)


n p o
ubxi = min wf , ai αimax (4)
n p o
ubyi = min hf , ai αimax (5)
 
x ai min
lbi = max ,l (6)
uby i

T
 i 
y ai min
lbi = max ,l (7)

IP
ubxi i


1 if pi > 0

CR
zi = (8)

0 if pi = 0

The penalty term (3) will increase the OFV (2) if any facility i violates the upper ubi and lower bounds

US
lbi of its dimensions. The difference of violation of each facility is multiplied by factor κ in order to
appropriately increase the OFV. Additionally, a static penalty λ is added to each invalid facility to allow
AN
penalisation regardless of the degree of deviation.
From a practical standpoint, the design generally aims to meet the estimated shape requirements. How-
ever, facility dimensions often need not comply exactly with the assumed shape restrictions (see Assump-
M

tion 2) if they result in considerably better solutions. In contrast to other penalty terms such as that of
Kang and Chae (2017), in which the extent of the violation is not assessed, the penalty term presented
here can allow small deviations from shape restrictions (e.g. 0.1%) by setting λ = 0 and κ to an acceptable
ED

level. Applying λ further permits shape deviations of some or of all facilities to be avoided if necessary.
Since appropriate values for both penalty parameters depend on the area and shape requirements as well
PT

as the proportion of penalties and the expected TTD, the evaluation of appropriate κ and λ levels require
problem-specific pre-tests. However, to compare the algorithms with other algorithms from the literature
and to exclude invalid solutions, the fixed additional penalty λ needs to be large enough to exclude invalid
CE

facilities.
For the sake of conciseness, the process of ascribing a layout coding ε to the corresponding OFV (Sec-
AC

tions 4.1 to 4.4) is hereafter denoted as OFV(ε).

5. Solution Methods

Because every slicing structure encodes one facility layout, generating an efficient layout requires studying
different slicing trees. Starting with an initial solution, a set of neighbourhoods can be explored by applying
two solution approaches. First, we detail the concept of PT because it has never been applied to the FLP
before. Second, we briefly introduce the well-established VNS, which we have adapted to our problem.
13
ACCEPTED MANUSCRIPT

5.1. Neighbourhood

Local search methods are used to improve an initial solution by investigating the neighbourhood iter-
atively. To obtain an initial solution, we randomly generated the numerical sequence of the slicing tree
coding ε. A neighbourhood Nk (ε) in terms of a local search defines a subset of feasible solutions ε0 that can
be generated by modifying a solution ε by using the predefined moves k. We apply five moves k = 1, ..., 5,
which appear in Table 2. Moves 1 to 4 modify the facility sequences and the slicing order because their

T
coding needs to involve an integer between 1 and n. Conversely, move 5 alters the binary sequence of the
slicing orientation.

IP
Table 2: Comparison of neighbourhoods.

CR
Move Encoding part Example
Facility sequence,
Move 1 1 5 6 3 2 7 4
slicing order

Move 2
slicing order
US
Facility sequence,

Facility sequence,
1 5 6 3 2 7 4
AN
Move 3 1 5 6 3 2 7 4
slicing order

Facility sequence,
Move 4 1 5 6 3 2 7 4
slicing order
M

0
Move 5 Slicing orientation 1 0 0 1 1 0 1
ED

Move 1 allows the exchange of the position of two elements within the facility sequence and the slicing
order. Applying move 1 to the facility sequence causes minor changes in the layout, although the basic
PT

structure remains. However, an exchange of elements of the slicing order can prompt major changes,
particularly changes consisting of the first elements. In move 2, a neighbour is generated by cutting an
CE

element of the corresponding sequence and pasting it into a different position. Similarly, moves 3 and 4
change the position of two and three elements, respectively, of the facility sequence and slicing order. As
concluded in previous research, insertion neighbourhoods are particularly suitable for permutation encodings
AC

(Ripon et al., 2013). Move 5, which can be commonly observed in heuristic approaches to the FLP (e.g.
Scholz et al., 2009; Komarudin and Wong, 2010), aims to modify the slicing orientation of a cut by changing
one binary from a vertical cut (0) to a horizontal cut (1) or vice versa.

5.2. Parallel Tempering

Also called a replica exchange simulation, PT was originally developed by Swendsen and Wang (1986),
who presented an improved Monte Carlo method to efficiently simulate systems with quenched random
14
ACCEPTED MANUSCRIPT

interactions. Geyer (1991) expanded the concept to the familiar form that we use. Although PT was
initially applied in statistical physics, applications in chemistry, biology, material sciences, engineering, and
operations research are becoming increasingly common (Earl and Deem, 2005).
The concept of PT relates to the widely used SA, for both make use of the Metropolis method. Applied
to the FLP, the Metropolis method generates a sequence of layouts at a certain temperature T . At each
step, a trial move k to a neighbour ε0 is attempted, after which the trial is either accepted or rejected. The
type of move k is selected according to given probabilities P (k), which applies to a random chosen sub-string

T
of the supported encoding part of ε (Table 2). If the solution OFV(ε0 ) of the new layout is better than the

IP
previous one OFV(ε), i.e. ∆ OFV(ε) ≥ 0,then the new neighbour will be accepted immediately. However,
a neighbour with a worse solution could be accepted with a probability of PM etropolis :

CR
  
∆OFV(ε)
PM etropolis = min 1, exp − (9)
T

US
The primary function of accepting uphill moves with a certain probability is to avoid being trapped
in local optima. However, the probability of acceptance depends heavily upon the static temperature T ,
AN
which is a weakness of the original Metropolis method. Since T remains static during the search process,
its choice is crucial to the performance of the Metropolis method. A smaller temperature T allows only
small uphill moves, whereas a larger T permits larger ones. Thus, the algorithm can get trapped in a
M

local optimum or skip an optimal solution. In this context, the purpose of PT (Algorithm 2) is to process
m non-interacting Metropolis replicas simultaneously by using different temperatures Tr , Tr+1 , ..., Tm , in
which Tr < Tr+1 < ... < Tm . After a fixed number of Metropolis iterations imax , replicas of neighbouring
ED

temperatures Tr and Tr+1 , r = 1, 2, ..., m − 1 can be exchanged with a probability of Pswap :

1 1
PT

Pswap = min(1, exp(( − )(OFV(εr+1 ) − OFV(εr ))) (10)


Tr Tr+1

As illustrated in Figure 5, PT allows dynamic exchanges. The temperature parameter Tr is adjusted


CE

gradually and thus permits the investigation of some regions more intensively than others. Moves of exchange
permit a random walk within the temperature range:
AC

• If the replica with a higher temperature has a better solution, then the probability of exchange is
equal to 1. Thus, an exchange allows the replica with a worse solution to escape a local optimum. The
replica with the better solution drifts to a lower temperature, thereby allowing it to search within a
small solution space that could hold a global optimum.

• If the replica with a higher temperature has a worse solution, then the probability of exchange will
be less than 1, depending on the temperature difference and differences in OFVs. Thus, exchanges
of slightly worse solutions from higher temperatures with better solutions from lower temperatures
15
ACCEPTED MANUSCRIPT

Algorithm 2 Parallel Tempering


1: procedure ParallelTempering(m, Tmin , Tmax , imax )
2: for r from 1 to m do
3: εr ← random initial solution . Set a random initial solution for each replica.
−Tmin
4: Tr ← Tmin + (r − 1) Tmax m−1
. Set the equally distributed temperatures.
5: end for
6: t←0
7: while t < tmax do . Check, whether the maximum CPU time is met.
8: for r from 1 to m do . Start m parallel Metropolis searches.
9: i←1
10: while i < imax do
ε0r ← Shake(εr , k˜P (k))

T
11: . Apply move k to εr with a probability of P (k).
  
OFV(ε)−OFV(ε0 )
12: PM etropolis ← min 1, exp − T r

IP
13: if PM etropolis > rand(0, 1) then
14: εr ← ε0r
15: end if
16: i←i+1

CR
17: end while
18: end for
19: for r from 1 to m − 1 do . Start checking for possible swaps.
20: Pswap ← min(1, exp(( T1 − T 1 )(OFV(εr+1 ) − OFV(εr )))
r r+1

US
21: if Pswap > rand(0, 1) then
22: Swap Tr and Tr+1 . Swap the temperatures belonging to εr and εr+1 .
23: end if
24: end for
25: t ← CpuT ime()
AN
26: end while
27: return ε . Return the layout encoding ε which yields the best objective function value OFV(ε).
28: end procedure
M

are allowed. Consequently, the neighbourhoods of the slightly worse solution can be scanned more
thoroughly; however, the better solution, whose previous replica has searched within a small solution
space, is raised to a higher temperature in order to skip a local optimum.
ED

The proposed PT returns a layout coded in ε, which yields a minimised OFV. To achieve the best solutions
with low computational effort, choosing the number of replicas and the range of temperatures is important.
PT

Whereas the smallest temperature has to be small enough to not overlook good solutions, the highest
temperature has to be large enough to overcome local optima. Furthermore, a small number of replicas and
CE

thus large differences between their temperatures cause less swapping. However, a large number of replicas
requires large computational effort without returning a significant benefit. To thoroughly investigate a
promising neighbourhood, even the temperatures between the minimum and maximum temperature can be
AC

optimised.

5.3. Variable Neighbourhood Search

The VNS was proposed by Mladenovic̀ and Hansen (1997). The basic idea of the VNS is to systematically
investigate a set of predefined neighbourhoods through which a local search is processed to find the best
solutions. VNS considers the fact that various neighbourhoods in a local search can have different local

16
ACCEPTED MANUSCRIPT

Low temperature search of replica r Exchange of temperature High temperature search of replica
before exchange r + 1 before exchange

x1
x2

OFV

T
IP
High temperature search of replica r Low temperature search of replica
after exchange r + 1 after exchange

CR
Figure 5: Replica trajectory and temperature exchange in parallel tempering.

optima and that the global optimum is a local optimum within a neighbourhood. To date, several variants
of VNS have been developed (Hansen et al., 2010).
US
As shown in Algorithm 3, the VNS procedure begins with a randomly generated layout coding ε, from
AN
whose neighbourhood Nk=1 a random neighbour ε0 is generated. For this neighbour, either an extensive
best improvement (variable neighbourhood descent (VND)) or first improvement (reduced variable neigh-
bourhood search (RVNS)) is performed to modify the neighbour to search for a local improvement. The
M

best solution of either method is referred to as ε00 . Following the termination of the VND or RVNS method,
the OFVs of ε00 and the current best solution ε are compared; if a new best solution is found, then they are
updated. In the case of an improvement, i.e. OFV(ε) >OFV(ε00 ), the procedure repeats with the first neigh-
ED

bourhood structure k = 1. If an improvement of the current solution ε is impossible, the neighbourhood


structure changes from Nk to Nk+1 .
PT

Algorithm 3 Variable Neighbourhood Search


1: procedure VNS(kmax , tmax )
2: ε ← random initial solution
3: t←0
CE

4: while t < tmax do


5: k←1
6: while k ≤ kmax do
7: ε0 ← Shake (ε, k) . Apply move k to ε that generates a random neighbour ε0 .
8: ε00 ← VND or RVNS (ε0 , kmax ) . Analyse the neighbourhood of ε0 and store the best found neighbour as ε00 .
AC

9: if OFV(ε00 ) < OFV(ε) then . Check if ε00 is an improvement over ε.


10: ε←ε 00 . Set ε00 as the new current solution.
11: k←1 . Reset k to 1.
12: else
13: k ←k+1
14: end if
15: end while
16: t ← CpuT ime()
17: end while
18: return ε . Return the layout encoding ε which yields the best objective function value OFV(ε).
19: end procedure

17
ACCEPTED MANUSCRIPT

To explore a neighbourhood with the VNS, we applied the two local search methods described below.
Further descriptions and pseudocodes of the two methods are available in Hansen et al. (2010).

Variable Neighbourhood Descent (VND). The VND (Algorithm 5 in the Appendix) completely
explores each neighbourhood to find the best neighbour and is thus also called the best improvement heuris-
tic. Similarly to the VNS, the VND sub-routine runs until a maximum neighbourhood Nkmax has been
reached. If no improvement can be found in the current neighbourhood structure, then the scanned neigh-

T
bourhood structure changes from Nk to Nk+1 . By contrast, whenever an improvement is made, the process

IP
will recommence with k = 1.

CR
Reduced Variable Neighbourhood Search (RVNS). The RVNS (Algorithm 6 in the Appendix)
also called the first improvement heuristic, does not explore an entire neighbourhood but compares only one
neighbour to the current solution. If an improvement is impossible, then a random neighbour of the next

US
neighbourhood is chosen for comparison until all neighbourhoods have been examined. Because it is less
time consuming, the RVNS method is particularly useful in investigating large instances.
AN
6. Computational Results

6.1. Data and Implementation


M

To evaluate the performance of the problem-solving methods and compare them to those of other ap-
proaches, we used instances tested in the literature, described in the following terms:
ED

• width wf and length hf of the floor plan

• area ai required by facility i


PT

• material flow fij between each pair of facilities ij


CE

• minimum length limin required by facility i

• maximum aspect ratio αimax required by facility i


AC

Although several instances met the requirements, few studies have considered our particular problem,
and because most research on the FLP concentrates on the rectilinear distance metric, few test instances
allow an appropriate comparison to our approach. Furthermore, because test instances from the literature
are often changed or applied to different circumstances, data about material flows are rounded to integer
values (Komarudin and Wong, 2010), material flow matrices are mirrored (Kim and Goetschalckx, 2005),
facility dimensions are changed (Wong and Komarudin, 2010), or empty spaces are added to the area of
facilities. Table 3 shows the specifications of the test instances that we used.
18
ACCEPTED MANUSCRIPT

Other than the number of facilities and their dimensional restrictions, the test instances differ in the
presence of empty spaces, especially when the available area of the floor plan exceeds the sum of the areas
Pm
of the facilities (i.e. wf hf > i ai ). Although MIP approaches (e.g. Castillo et al., 2005) can optimally
distribute empty spaces within the available floor space, empty spaces are not considered by construction
methods in slicing tree approaches and thus remain entirely in the last partition. To account for empty
spaces in our approach, the total empty space per test instance is divided into dummy facilities, each with
an area of 1 and freed from other dimensional restrictions.

T
IP
Table 3: Summary of the tested data sets.
Data Layout dimensions Empty Shape Max. CPU
n Source
sets wF × hF space requirements time [s]

CR
O7 7 8.54 × 13.00 - αmax = 4 3 Meller et al. (1999)
O8 8 11.31 × 13.00 - αmax = 4 6 Meller et al. (1999)
O9 9 12.00 × 13.00 - αmax = 4 13 Meller et al. (1999)
vC10 10 51.00 × 25.00 - lmin = 5 20 van Camp et al. (1991)
Ba12
Ba14
AB20
SC30
12
14
20
30
6.00
7.00
20.00
15.00
×
×
×
×
10.00
9.00
30.00
12.00
7
2
-
17
US
lmin = 1
lmin = 1
αmax = 4
αmax = 5
120
120
200
300
Bazaraa (1975)
Bazaraa (1975)
Armour and Buffa (1963)
Liu and Meller (2007)
AN

The algorithms presented in Section 5 were coded with Java 8 and made use of the Java libraries
JGraphT and JAMES (v1.2). JGraphT was applied to model layouts as a graph to calculate the shortest
M

paths between I/O points. JAMES, developed by De Beukelaer et al. (2015), is a framework for applying
metaheuristics to custom problems coded in Java and was used as a basis for the VNS and PT algorithms.
ED

The numerical tests were conducted on a computer with an Intel Core i5 CPU (2.3 GHz) and 8 GB of
physical memory on a Windows 10 operating system. To achieve statistically precise results, 30 runs were
performed for each approach.
PT

The penalty term can vitally affect the performance of the solution approaches because new neighbours
are accessible when allowing the violation of shape requirements during the neighbourhood search. Shape
CE

relaxations can afford superior OFVs but infeasible layout solutions as well. The algorithms are forced to
provide only layouts that hold the shape restrictions for comparison with the solutions of other approaches.
Hence, λ and κ are set to 1,000, which is appropriate to avoid shape violations considering the conditions
AC

of the test instances according to several conducted tests.

6.2. Solution Times

Various criteria exist to stop an algorithm in order to achieve good results while wasting as little com-
puting time as possible. Lin and Lin (2015) stopped their ACO after a maximum number of iterations. The
approach of Kim and Goetschalckx (2005) checks whether an appropriate improvement has been achieved

19
ACCEPTED MANUSCRIPT

after each iteration. Komarudin and Wong (2010) applied a combined criterion that terminates the algo-
rithm if either the maximum number of iterations has been processed or the maximum number of iterations
without improvement has been reached. The concepts of the proposed algorithms differ significantly. Since
we intended to compare different algorithms in which the computational costs per iteration were not equal,
stopping the algorithm after a maximum number of iterations was inappropriate. Therefore, the algorithms
stop after reaching a maximum CPU time adjusted by the size of the test instance (Table 3), which guar-
antees a fair comparison of our proposed algorithms.

T
IP
6.3. Parameter Choice

To perform effectively, a solution requires an appropriate set of parameters. Although manual iterated

CR
tuning can afford excellent performance, it is time intensive, often biased by personal experience, and severely
limited by the settings evaluated. In response to these obstacles, we used the automated tuning package
irace (v.2.3.1807), which employs the iterated racing procedure for metaheuristic parameter tuning. This

US
iterated procedure samples parameter configurations according to a particular distribution, evaluates them
by applying a t-test, and updates the sampling distribution for the refinement of sampling. Algorithmic
AN
details and proofs of efficiency appear in López-Ibáñez et al. (2016).
Due to the heterogeneity of the datasets and the different scales of OFV, we performed the tuning
procedure for each dataset separately with a computational budget of 500 to 5,000 evaluations for each
M

based on the computing time per dataset. Although higher budgets can afford better performance, an
increase in the budget revealed only small adjustments of parameters and weak benefits.
ED

6.3.1. Parallel Tempering


All parameters of PT were adjusted using the iterated racing procedure within the following ranges:
PT

• number of replicas m ∈ [2, 20] processed simultaneously

• maximum number of iterations imax ∈ {1, 2, 5, 10, 20, 40, 60, 80, 100} after which swaps can occur
CE

• minimal temperatures Tmin ∈ [1, 100] and maximal temperatures Tmax ∈ [150, 1000] to limit the
uniformly distributed temperature Tr of each replica r
AC

P5
• selection probability P (k) ∈ (0, 1), where k=1 P (k) = 1

Although each test instance requires a different set of parameters for a good performance of PT, the
results in Table 4 represent the findings of their general application. The number of replicas and the
maximum number of iterations needed to be no more than 16 and 10, respectively. Tmax and Tmin primarily
correlate with the scale of the OFV of the examined datasets. Datasets with greater intensities of material
flow, larger facility areas, and thus higher OFVs performed better at higher temperatures. It also became

20
ACCEPTED MANUSCRIPT

clear that move 1 should be used the most often, whereas move 3 seemed to be necessary only occasionally.
Nevertheless, the parameters of PT were relatively robust considering the tested instances.

Table 4: Overview of best parameter configurations for PT.


Data sets Budget m imax Tmin Tmax P (1) P (2) P (3) P (4) P (5)
O7 5000 2 2 3 230 0.4 0.08 0.08 0.24 0.2
O8 2500 2 5 8 444 0.44 0.18 0.09 0.20 0.09
O9 1500 3 1 5 350 0.22 0.35 0.12 0.09 0.22
vC10 1500 6 10 80 960 0.44 0.26 0.13 0.13 0.04

T
Ba12 750 16 1 30 310 0.29 0.26 0.17 0.08 0.2
Ba14 750 8 5 60 780 0.28 0.12 0.22 0.22 0.16
AB20 500 5 5 10 450 0.26 0.14 0.17 0.23 0.2

IP
SC30 500 5 5 25 750 0.22 0.09 0.25 0.25 0.19

CR
6.3.2. Variable Neighbourhood Search
Although the VNS does not require adjustable parameters, the order of moves and the corresponding

US
neighbourhood structures influence the quality of solutions and the computing time. Therefore, all possible
orders of moves were generated and tested 30 times per dataset in a random order using the VND sub-
routine. Results indicate that the move orders 5, 3, 2, 1, 4 and 5, 2, 3, 1, 4 resulted in fast convergence and
AN
¯
similar lowest mean values considering the test instances. For example, test instance O9 resulted in OFV
¯ = 58.60, respectively. On average, all move orders starting with move 5 resulted in lower
= 57.68 and OFV
mean values than move orders starting with another move, possibly because, by changing the orientation
M

of cuts, move 5 requires fewer evaluations than the other moves in the VND sub-routine. For all further
computational experiments, the move order 5, 3, 2, 1, 4 was used.
ED

6.4. Testing Results

As shown in Table 1, three integrated approaches based on contour distances are available to date. Due
PT

to their relaxed assumption of an unrestricted number of material handling points, the results of Norman
et al. (2001b) are not comparable. Table 5 shows the best results OFVmin of the three proposed algorithms
CE

compared to solutions reported by Kim and Goetschalckx (2005) and Lin and Lin (2015). The results of
Kim and Goetschalckx (2005) have been halved, because they used a symmetric material flow matrix that
doubled the results in comparison to the original problem.
AC

To evaluate the quality of the best solutions, we solved the MIP formulation presented in Klausnitzer
and Lasch (2017) using CPLEX v12.7.1 under reaching an optimal solution. We applied a time limit of
24 hours and set the maximum number of horizontal and vertical path segments to 6, which, according to
several pilot studies, was considered to be sufficiently high to ensure that the solution space for the tested
problems were not restricted by the number of path segments. Although the MIP uses a polyhedral outer
approximation of the area constraints, for which the number of discretisation points is set to 50 to ensure
area errors less than 0.07% (cf. Sherali et al., 2003), such use can also exert minor effects on the OFV. The
21
ACCEPTED MANUSCRIPT

optimal solutions OFVopt and their gap in relation to the best solutions obtained by the new approach were
achieved within 24 hours for test instances O7, O8, O9, and vC10 (Table 5). Figures of the best solutions
achieved in this study appear online in the Supplementary Material.

Table 5: Comparison of the best results.


Data Kim and Goet. Lin and Lin VNS+ VNS+ Improve- Gap to
PT OFVopt
sets (2005) (2015) VND RVNS ment [%] opt. [%]
O7 – 24.3 21.64 21.64 21.64 10.94 21.64 0.00

T
O8 – 52.2 52.09 52.09 52.09 0.21 51.71 0.73
O9 – 59.3 53.29 53.29 53.29 10.13 52.36 1.74

IP
vC10 3230.7 4411.97 3097.76 3097.76 3102.11 4.11 3080.52 0.56
Ba12 – 4990.84 3089.91 3586.58 3414.91 38.09 – –
Ba14 – 5259.38 2188.33 2291.72 2279.00 58.39 – –
AB20 1349.95 – 1185.99 1350.51 1230.14 12.15 – –

CR
SC30 – – 2018.60 4376.83 2018.60 – – –

Although the scope of this paper does not accommodate a direct comparison of the processing perfor-

US
mance of our proposed algorithms and the approaches from the literature due to differences in programming
language, development environment, and execution platforms used to program and run the algorithms, we
present the new best solutions for the tested datasets from the literature. In particular, the new application
AN
of PT was exceptional. Assuming that all approaches received sufficient time or a sufficient number of
iterations to achieve good solutions, the quality of solutions is crucial for practice, because the FLP man-
ages long-term design problems and all considered approaches could solve the test instances within several
M

seconds. Nevertheless, Kim and Goetschalckx (2005) achieved the results within 40 s and 727 s, whereas PT
required only 1 s and 66 s to generate the same mean solutions for vC10 and AB20, respectively (Figure 6).
ED

The construction procedure of our approach only applies one layout evaluation, which suggests the efficiency
of the approach (Section 2).
The differences in the results of our proposals compared to those of Lin and Lin (2015) could have arisen
PT

from their flexible bay approach, which is usually more restrictive than slicing tree approaches. A more
detailed examination of the layouts in the Supplementary Material reveals that the respective layouts could
CE

be modelled as flexible bay layouts in any case. Furthermore, Lin and Lin (2015) applied two methods to
address the empty spaces in Ba12. First, they allowed the facilities to exceed their required areas in order
to consider empty spaces; second, and similar to our approach, they divided the empty space into dummy
AC

objects. Table 5 includes their best solution resulting from the first method, which proved to be worse than
our solutions.
The quality of the solution of our proposed approach was fairly close to the optimal solutions. O7 was
optimally solved, and the solutions of O8, O9, and vC10 of our approach and the optimal solutions show a
difference of less than 1.8%.
It is impossible to make a general statement about the advantages of each of our algorithms that used

22
ACCEPTED MANUSCRIPT

3000
5500

KIM/GOET. (2005) KIM/GOET. (2005)


PT PT
VNS+RVNS VNS+RVNS
VNS+VND VNS+VND
5000

2500
4500
OFV

OFV

T
2000

IP
4000
3500

CR
5 10 15 20 1500 20 40 60 80 100 120

US
Max. CPU time [s] Max. CPU time [s]

(a) vC10. (b) AB20.

¯ from Kim and Goetschalckx (2005).


Figure 6: Comparison of the convergence times of PT, VND and RVNS and OFV
AN

only the best results OFVmin achieved, particularly when the results resembled each other. Table 6 shows
M

¯ and the standard deviation sOFV achieved with our proposed algorithms. PT showed
the mean values OFV
¯ for all datasets, and even considering the standard deviation, it achieved the most
a superior mean OFV
robust solutions in every simulation run.
ED

No distinct differentiation is possible between the effectiveness of VNS+VND and VNS+RVNS. The
results in Table 6 show that VNS+VND operates better with some medium-sized problems, whereas
PT

VNS+RVNS can manage very large and small problems more efficiently. This finding is reasonable since
VNS+VND explores a neighbourhood in its entirety and therefore needs more computing time. Clearly, the
time available to solve the largest test instances (e.g. SC30) is proportionally short, and VNS+RVNS is
CE

thus superior to VNS+VND because it explores neighbourhoods efficiently one neighbour at a time.

Table 6: Comparison of the solution performance.


AC

Data PT VNS + VND VNS + RVNS


sets OFVmin ¯
OFV sOFV OFVmin ¯
OFV sOFV OFVmin ¯
OFV sOFV
O7 21.64 21.64 0.05 21.64 22.59 2.75 21.64 22.08 1.64
O8 52.09 52.18 0.49 52.09 58.63 6.07 52.09 55.30 6.09
O9 53.29 54.06 1.66 53.29 57.68 3.85 53.29 56.73 3.71
vC10 3097.76 3182.32 77.22 3097.76 3751.69 471.55 3102.11 3741.98 496.42
Ba12 3089.91 3731.94 230.74 3586.58 4394.89 431.26 3414.91 4416.32 479.78
Ba14 2188.33 2395.24 94.50 2291.72 2690.22 166.28 2279.00 2752.37 265.11
AB20 1185.99 1421.05 119.92 1350.51 1565.78 153.44 1230.14 1627.28 201.45
SC30 2018.60 2534.74 288.21 4376.83 6283.63 1934.64 2018.60 2932.38 513.36

23
ACCEPTED MANUSCRIPT

6.5. Additional Improvement Strategies

To determine I/O points, we applied a simple construction heuristic, as shown in Section 4.3. By
extension, we sought to optimise the I/O locations and improve the OFV by using two different methods.

6.5.1. Sequential Improvement


First, we applied a step to improve the solutions of the approaches described in Section 5. After finishing
the respective algorithms, the best three solutions obtained from each heuristic were selected. Then, the

T
locations of I/O points were improved by applying another PT, while the locations and dimensions of

IP
facilities remained unchanged. We therefore applied Algorithm 2 to change the initial I/O locations (line 11:
IOij∈C ← Shake (IOij∈C , k IO ˜P (k IO )) ) according to predefined moves k IO with a selection probability

CR
of P (k IO ). The algorithm returned the I/O locations to yield a minimised OFV corresponding to each of
the three given layouts. Table 7 describes the parameter settings of the PT, which have been tuned with

US
irace for vC10. The termination time of PT was determined by various pilot studies to be 5, 6, 10, 20, 25,
25, 35, and 45 ms for test instances O7, O8, O9, vC10, Ba12, Ba14, Ab20, and SC30, respectively.
AN
Table 7: Parameters of PT to change the I/O locations.
Parameter Setting
Replicas m = 10
Iterations imax = 10
Temperatures Tmin = 50, Tmax = 320
M

Move type
Move kIO =1 Change I/O locations consider-
ing one facility
Move kIO =2 Change I/O locations consider-
ED

ing two facilities


Move kIO =3 Change I/O locations consider-
ing three facilities
Probability P (kIO ) of Moves 1-3 0.37, 0.45, 0.18
PT

To compare the results of the PT heuristic with the optimal solutions for the I/O selection problem,
the ILP formulation (Appendix) was solved using CPLEX. The results (Table 8) indicate a rather small
CE

potential for improvement as compared to the solutions achieved by the constructive heuristic (Table 5).
Only the solutions of O8, B14, AB20 and SC30 could be improved, and the gap between heuristic solutions
and the optimal solutions decreased.
AC

PT achieved the optimal solutions for the I/O location problem in all cases except SC30 for which
it achieved a slight difference of approximately 2%. However, the computing time to obtain the optimal
solutions for the I/O location problem drastically exceeded the time required by PT. For example, the
I/O point location problem was solved within 25 ms and 45 ms using PT for problems Ba12 and SC30,
respectively, whereas solving the ILP formulation required about 250 ms and 1047 ms. Such a disproportional
increase in computing time is reasonable because the problem is NP-hard.

24
ACCEPTED MANUSCRIPT

Table 8: Results of the sequential I/O improvement strategy.


Data Constructive Sequential improve. Improve- Gap to
OFVopt
sets heuristic PT ILP ment [%] opt. [%]
O7 21.64 21.64 21.64 0.00 21.64 0.00
O8 52.09 51.77 51.77 0.61 51.71 0.12
O9 53.29 53.29 53.29 0.00 52.36 0.07
vC10 3097.76 3097.76 3097.76 0.00 3080.52 0.56
Ba12 3089.91 3089.91 3089.91 0.00 – –
Ba14 2188.33 2175.99 2175.99 0.56 – –
AB20 1185.99 1182.82 1182.82 0.27 – –

T
SC30 2018.60 2007.51 1966.88 2.56 – –

IP
6.5.2. Simultaneous Improvement
The second method was designed to improve the I/O points within every iteration of the applied heuristic.

CR
Since the ILP formulation for the optimal I/O location requires much longer computing times than PT and
showed a low benefit, the computational study on the simultaneous improvement was conducted using PT.

US
Arapoglu et al. (2001) also recommended applying a heuristic solution to this subproblem. Thus, after the
constructive heuristic used to locate I/O points was processed, a PT algorithm was chosen to improve them.
The PT algorithm for simultaneous improvement used the same termination times per I/O improvement
AN
and parameters as the PT algorithm described in Table 7. The maximum CPU times (Table 9) determined
by various pilot studies were considerably larger because an improvement-oriented procedure followed every
explored layout solution.
M

Table 9: Results of the integrated use of PT.


Data Max. CPU times PT Improve- Gap to
OFVopt
ED

sets Total [s] I/O improve. [ms] OFVmin ¯


OFV sOFV ment [%] opt. [%]
O7 9 5 21.64 23.93 3.60 0.00 21.64 0.00
O8 18 6 51.76 56.61 5.61 0.63 51.71 0.09
O9 39 10 52.40 56.85 3.29 1.67 52.36 0.07
vC10 60 20 3097.76 3675.78 389.74 0.00 3080.52 0.56
PT

Ba12 360 25 3477.5 3945.05 219.05 0.00 – –


Ba14 360 25 2240.07 2522.62 156.64 0.00 – –
AB20 600 35 1156.86 1509.45 146.39 2.46 – –
SC30 900 45 2545.19 3308.74 308.49 0.00 – –
CE

Table 9 shows the results of the integrated use of PT. Despite the increased computational effort, the
proposed simultaneous improvement strategy produced only three superior solutions compared to the al-
AC

gorithm without I/O point improvement. Only the OFVmin of O8, O9, and AB20 improved – and only
slightly – by using the integrated approach, although the mean values and standard deviations significantly
worsened.
Furthermore, only the best results of problems O9 and AB20 showed marginally better solutions than the
sequential improvement. Therefore, the simultaneous approach seems to offer only marginal benefits despite
the already tripled computing times. However, it is possible to extend the maximum allowable processing
times in order to assess the performance of the simultaneous approach.
25
ACCEPTED MANUSCRIPT

Just as the simultaneous improvement strategy afforded some slightly better solutions, their gap in
relation to the optimal solutions also decreased. The remaining gap might indicate further potential for
improvement. Nevertheless, even considering vC10, no significant difference emerged in the layouts between
the optimal solutions and the best solutions achieved by our approach. Hence, only the latter are shown in
the Supplementary Material. This finding could indicate that the better OFVs of the MIP resulted due to
the propensity of the polyhedral outer approximation of the area constraints to underestimate the areas of
facilities (Sherali et al., 2003).

T
For simultaneous improvement, the ILP formulation or other I/O selection procedures from the litera-

IP
ture can be applied. However, determining of the maximum CPU times is crucial for such improvement,
particularly if the solution time of the I/O setting procedure varies significantly. For example, an experi-

CR
ment of 1,500 generated layouts of SC30 revealed solution times for the ILP formulation between 1,047 ms
and 9,953 ms depending on the number of potential I/O points. By setting the maximum CPU times too

US
low when the required times for the I/O point selection are high, the number of iterations will decrease
significantly, and the solution quality will worsen. At the same time, determining the maximum termination
times according to the longest times required by the I/O selection procedure could result in impracticable
AN
computing times.

7. Conclusions
M

Extensive research on solving the FLP with unequal areas has been conducted in recent years. While
attention to real-life requirements has increased steadily in recent years, heuristic approaches have also
ED

improved, and numerous new, efficient heuristics have been proposed. As a contribution, we developed
three solution methods for the integrated FLP with I/O locations using contour distances that are more
PT

relevant to practical use. From a practical perspective, the proposed construction method can manage
shape restrictions – namely, if the assumed shape-related restrictions are hard constraints or could deviate
according to available information regarding shape.
CE

Other than applying the well-established VNS heuristic, we introduced the concept of PT to solving the
FLP for the first time. Although a direct comparison of computational effort required by our approaches
AC

and previous approaches is impossible, the layout solutions achieved by our approaches are considerably
better than those from the literature. Additionally, we proposed two strategies to improve the solutions.
First, to improve the I/O locations of the best-known solutions, we applied the PT and the ILP formulation,
which showed promising results and required little computing time. Second, we applied PT to improve the
I/O locations to every layout explored. Although slight improvements are possible, this procedure requires
far more computing time.
We considered different test instances from the literature. Additional studies with test instances from

26
ACCEPTED MANUSCRIPT

practice are needed to verify the research findings. However, since PT proved to be a very promising
heuristic, researchers could test it with related FLPs.
Although our approach proved to be highly efficient considering the tested instances, limitations should be
acknowledged, particularly regarding the slicing tree approach implemented. For one, an MIP can optimally
distribute empty spaces within the layout; however, in a slicing tree approach, the empty space would remain
in the upper right corner of the planning area, which might not be optimal. This problem can be mitigated
if the empty space is divided into smaller parts that can be considered as facilities without material flows, for

T
which it can be crucial to determine a reasonable number of facilities that represent the empty space and to

IP
keep the computational effort manageable. At the same time, slicing tree approaches require penalty terms
or modifications to the layout encoding in order to prevent infeasible layouts from being generated. Some

CR
practical cases require facilities with predefined shapes and locations that might significantly complicate the
search for good solutions. Other layout construction methods might be favourable and should be tested in
future research.

US
Although our approach solves several mutually dependent design problems at once, several other im-
portant design aspects should be considered. For example, there are practical cases where material flow
AN
paths are restricted to predefined structures (e.g. Kulturel-Konak, 2017). In the future, researchers should
consider path or aisle design restrictions, space for aisles, separated input and output locations, and various
kinds of geometric facility requirements for more practical use and to accelerate layout planning.
M
ED
PT
CE
AC

27
ACCEPTED MANUSCRIPT

Appendix

Algorithm 4 Layout construction using a slicing tree coding


1: procedure LayoutConst(ai, wf , hf , ε = {sequence {f s1 , ..., f sn }, order {so1 , ..., son−1 }, orientation {sr1 , ..., srn−1 }})
2: LC0 ← Facility sequence f si i ≤ n . Initialize the planning area containing all facilities.
3: RC0 ← ∅
4: for i from 1 to n do . Initialize the coordinates of the upper left and lower right corners of all facilities.
5: lri ← (wf , 0)
6: uli ← (0, hf )
7: end for

T
8: j←1
9: while j 6= n do . Stop when every partition contains one facility.

IP
10: k ←n−1
11: while k ≥ 0 do
12: if f ssoj ∈ LCk or RCk then . Select the last LCk or RCk containing f ssoj as partition R for slicing.
13: R ← LCk or RCk

CR
14: break
15: end if
16: k ←k−1
17: end while
18: LCj ← f si 0 ≤ i ≤ soj ∩ R . Assign facilities of the left part of facility sequence to new LCj .

19:
20:
21:
22:
RCj ← f si soj < i ≤ n ∩ R
if srj = 0 then
for all i  ∈ LCj do
lri ← xul i +
P
i∈LCj ai
ul yi
, yilr
 US
. Assign facilities of the right part of facility sequence to new RCj .

. Calculate
. The new width
P
i∈LCj ai
hR
. If srj is a vertical cut.
coordinates of partitions LCj and RCj .
gets added to the previous xul i .
AN
23: end for
24: for all i ∈
 RCj do
P 
i∈LCj ai
25: uli ← xul
i + , yiul
yiul
26: end for
M

27: else if srj = 1 then . If srj is a horizontal cut.


28: ∈ LCj do
for all i  P 
i∈LCj ai
29: lri ← xlr ul
i , yi − xlr
i
30: end for
ED

31: for all i ∈


 RCj do P 
i∈LCj ai
32: uli ← xul ul
i , yi − xlr
i
33: end for
34: end if
PT

35: j ←j+1
36: end while
37: return lri , uli of each facility i . Return the coordinates of each facility.
38: end procedure
CE
AC

28
ACCEPTED MANUSCRIPT

An integer linear programming model for the input-output location problem.


Notation:
fij Material flow between facility i and j.
dipjq Distance from candidate point p ∈ Ci of facility i to candidate point q ∈ Cj of facility j.
xip Binary variable that equals 1 if candidate point p ∈ Ci of facility i is chosen.
yipjq Binary variable that equals 1 if candidate point p ∈ Ci of facility i and candidate point q ∈ Cj
of facility j is chosen.
n X X
X n X

T
Min OFV = dipjq fij yipjq (11)
i=1 p∈Ci j=1 q∈Cj

IP
X
xip = 1 ∀i (12)
p∈Ci

CR
yipjq ≥ xip + xjq − 1 ∀i 6= j, p, q (13)

The objective function (11) minimises the sum of products of the material flow parameter, the distance

US
parameter, and the binary variable yipjq . The latter determines the I/O locations p and q from among
candidate points Ci and Cj of facility i and j. Constraint (12) ensures that each facility i uses only one I/O
AN
location. Constraint (13) forces yipjq to equal one only if candidate point p of facility i and candidate point
q of facility j is chosen (i.e. xip = xjq = 1).

Algorithm 5 Variable Neighbourhood Descent


M

1: procedure VND(ε0 , kmax0 )


2: k0 ← 1
3: while k0 ≤ kmax
0 do
4: ε∗ ← arg minδ∈Nk0 (ε0 ) OFV(ε0 ) . Find the best neighbour ε∗ in neighbourhood Nk0 of ε0 .
ED

5: ∗ 0
if OFV(ε ) <OFV(ε ) then . Store solution ε∗ as the best current solution.
6: ε0 ← ε∗
7: k0 ← 1 . Reset k to restart the search with the current best solution.
8: else
9: k0 ← k0 + 1 . Increment k to change neighbourhood N 0 k .
PT

10: end if
11: end while
12: ε00 ← ε0
13: return ε00 . Return the coding ε00 which yields the best objective function value OFV(ε00 ) within the local search.
14: end procedure
CE

Algorithm 6 Reduced Variable Neighbourhood Search


1: procedure RVNS(ε0 , kmax0 )
AC

2: k0 ← 1
3: while k0 ≤ kmax
0 do
4: ε∗ ← Shake (ε0 , k0 ) . Apply move k0 to ε0 that generates a random neighbour ε∗ .
5: if OFV(ε∗ ) <OFV(ε0 ) then . Check if ε∗ is an improvement over ε0 .
6: ε00 ← ε∗ . Set ε∗ as the new current solution of the local search.
7: else
8: k0 ← k0 + 1
9: end if
10: end while
11: return ε00 . Return the coding ε00 which yields the best objective function value OFV(ε00 ) within the local search.
12: end procedure

29
ACCEPTED MANUSCRIPT

References

References

Anjos, M., Vieira, M., 2017. Mathematical optimization approaches for facility layout problems: The state-of-the-art and future
research directions. European Journal of Operational Research 261 (1), 1–16.
Arapoglu, R., Norman, B., Smith, A., Jun 2001. Locating input and output points in facilities design: A comparison of
constructive, evolutionary, and exact methods. IEEE Transactions on Evolutionary Computation 5 (3), 192–203.
Armour, G. C., Buffa, E. S., 1963. A heuristic algorithm and simulation approach to relative location of facilities. Management

T
Science 9 (2), 294–309.

IP
Banerjee, P., Zhou, Y., Montreuil, B., 1997. Genetically assisted optimization of cell layout and material flow path skeleton.
IIE Transactions 29 (4), 277–291.

CR
Bazaraa, M. S., 1975. Computerized layout design: A branch and bound approach. AIIE Transactions 7 (4), 432–438.
Castillo, I., Westerlund, J., Emet, S., Westerlund, T., 2005. Optimization of block layout design problems with unequal areas:
A comparison of MILP and MINLP optimization methods. Computers & Chemical Engineering 30 (1), 54–69.
Chittratanawat, S., 1999. An integrated approach for facility layout, P/D location and material handling system design.

US
International Journal of Production Research 37 (3), 683–706.
De Beukelaer, H., Davenport, G. F., De Meyer, G., Fack, V., 2015. JAMES: A modern object-oriented Java framework for
discrete optimization using local search metaheuristics. In: Doumpos, M., Grigoroudis, E. (Eds.), Conference proceedings:
AN
4th international symposium & 26th national conference on operational research. Hellenic Operational Research Society, pp.
134–138.
Drira, A., Pierreval, H., Hajri-Gabouj, S., 2007. Facility layout problems: A survey. Annual Reviews in Control 31 (2), 255–267.
Earl, D., Deem, M., 2005. Parallel tempering: Theory, applications, and new perspectives. Physical Chemistry Chemical Physics
M

7 (23), 3910–3916.
Funke, J., Hougardy, S., Schneider, J., 2016. An exact algorithm for wirelength optimal placements in VLSI design. INTE-
GRATION, the VLSI journal 52 (1), 355–366.
ED

Geyer, C. J., 1991. Markov chain Monte Carlo maximum likelihood. In: Keramidas, E. M. (Ed.), Computing science and
statistics : Proceedings of the 23rd Symposium on the Interface. Interface Foundation of North America, pp. 156–163.
Gonçalves, J., Resende, M., 2015. A biased random-key genetic algorithm for the unequal area facility layout problem. European
PT

Journal of Production Research 246 (1), 86–107.


Hahn, P., Krarup, J., 2001. A hospital facility layout problem finally solved. Journal of Intelligent Manufacturing 12 (5-6),
487–496.
CE

Hansen, P., Mladenovic̀, N., Prez, J., 2010. Variable neighborhood search. Annals of Operations Research 175 (1), 367–407.
Hu, G., Chen, Y., Zhou, Z., Fang, H., 2007. A genetic algorithm for the inter-cell layout and material handling system design.
International Journal of Advanced Manufacturing Technology 34 (11-12), 1153–1163.
Kang, S., Chae, J., 2017. Harmony search for the layout design of an unequal area facility. Expert Systems with Applications
AC

79, 269 – 281.


Keller, B., Buscher, U., 2015. Single row layout models. European Journal of Operational Research 245 (3), 629–644.
Kim, J.-G., Goetschalckx, M., 2005. An integrated approach for the concurrent determination of the block layout and the input
and output point locations based on the contour distance. International Journal of Production Research 43 (10), 2027–2047.
Kim, J.-G., Kim, Y.-D., 1999. A branch and bound algorithm for locating input and output points of departments on the block
layout. The Journal of the Operational Research Society 50 (5), 517–525.
Klausnitzer, A., Lasch, R., 2017. A new model formulation for the facility layout problem with material path design. No. 179/17
in Dresdner Beiträge zur Betriebswirtschaftslehre.

30
ACCEPTED MANUSCRIPT

Komarudin, Wong, K. Y., 2010. Applying ant system for solving unequal area facility layout problems. European Journal of
Operational Research 202 (3), 730–746.
Kulturel-Konak, S., 2012. A linear programming embedded probabilistic tabu search for the unequal-area facility layout problem
with flexible bays. European Journal of Operational Research 223 (3), 614–625.
Kulturel-Konak, S., 2017. The zone-based dynamic facility layout problem. INFOR: Information Systems on Operational
Research 0 (0), 1–31.
Kumar, S., Cheng, J., 2015. A BIM-based automated site layout planning framework for congested construction sites. Automa-
tion in Construction 59, 24–37.

T
Lin, Y.-Z., Lin, Y.-C., 2015. Applying an immune ant colony system algorithm to solve an integrated flexible bay facility layout
problem with input/output points design. Lecture Notes in Management Science 7, 56–62.

IP
Liu, Q., Meller, R. D., 2007. A sequence-pair representation and MIP-model-based heuristic for the facility layout problem
with rectangular departments. IIE Transactions 39 (4), 377–394.

CR
López-Ibáñez, M., Dubois-Lacoste, J., L. Pérez C., Birattari, M., 2016. The irace package: Iterated racing for automatic
algorithm configuration. Operations Research Perspectives 3, 43–58.
Meller, R. D., Narayanan, V., Vance, P. H., 1999. Optimal facility layout design. Operations Research Letters 23 (3-5), 117–127.

US
Mladenovic̀, N., Hansen, P., 1997. Variable neighborhood search. Computers & Operations Research 24 (11), 1097–1100.
Norman, B., Smith, A., Yildirim, E., Tharmmaphornphilas, W., 2001a. An evolutionary approach to incorporating intrade-
partmental flow into facilities design. Advances in Engineering Software 32 (6), 443–453.
Norman, B. A., Arapoglu, R. A., Smith, A. E., 2001b. Integrated facilities design using a contour distance metric. IIE Trans-
AN
actions 33 (4), 337–344.
Ripon, K. S. N., Glette, K., Khan, K. N., Hovin, M., Torresen, J., 2013. Adaptive variable neighborhood search for solving
multi-objective facility layout problems with unequal area facilities. Swarm and Evolutionary Computation 8, 1–12.
M

Sahni, S., Gonzalez, T., 1976. P-complete approximation problems. Journal of the ACM (JACM) 23 (3), 555–565.
Scholz, D., 2010. Innerbetriebliche Standortplanung: Das Konzept der Slicing Trees bei der Optimierung von Layoutstrukturen,
1st Edition. Gabler Research : Produktion und Logistik. Gabler, Wiesbaden.
ED

Scholz, D., Jaehn, F., Junker, A., 2010. Extensions to STaTS for practical applications of the facility layout problem. European
Journal of Operational Research 204 (3), 463–472.
Scholz, D., Petrick, A., Domschke, W., 2009. Stats: A slicing tree and tabu search based heuristic for the unequal area facility
layout problem. European Journal of Operational Research 197 (1), 166–178.
PT

Sherali, H., Fraticelli, B., Meller, R., 2003. Enhanced model formulation for the optimal facility layout. Operations Research
51 (4), 629–644.
Sinha, A., Malo, P., Deb, K., 2017. Evolutionary algorithm for bilevel optimization using approximations of the lower level
CE

optimal solution mapping. European Journal of Operational Research 275 (2), 395–411.
Swendsen, R. H., Wang, J.-S., Nov 1986. Replica Monte Carlo simulation of spin-glasses. Physical Review Letters 57 (21),
2607–2609.
AC

Tam, K. Y., 1992. Genetic algorithms, function optimization, and facility layout design. European Journal of Operational
Research 63 (2), 322–346.
van Camp, D. J., Carter, M. W., Vannelli, A., 1991. A nonlinear optimization approach for solving facility layout problems.
European Journal of Operational Research 57 (2), 174–189.
Wong, K. Y., Komarudin, 2010. Solving facility layout problems using flexible bay structure representation and ant system
algorithm. Expert Systems with Applications 37 (7), 5523–5527.
Wu, Y., Appleton, E., 2002. The optimisation of block layout and aisle structure by a genetic algorithm. Computers & Industrial
Engineering 41 (4), 371–387.

31
ACCEPTED MANUSCRIPT

Yang, T., Peters, B., Tu, M., 2005. Layout design for flexible manufacturing systems considering single-loop directional flow
patterns. European Journal of Operational Research 164 (2), 440 – 455.

T
IP
CR
US
AN
M
ED
PT
CE
AC

32

You might also like