Applied Soft Computing Journal 89 (2020) 106052

Contents lists available at ScienceDirect

Applied Soft Computing Journal

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

Configuration space evolutionary algorithm for multi-objective

unequal-area facility layout problems with flexible bays
∗
Jingfa Liu a,b , , Siyu Liu c , Zhaoxia Liu d , Bi Li a
a
School of Information Science and Technology, Guangdong University of Foreign Studies, Guangzhou 510006, China
b
Guangzhou Key Laboratory of Multilingual Intelligent Processing, Guangdong University of Foreign Studies, Guangzhou 510006, China
c
School of Computer & Software, Nanjing University of Information Science & Technology, Nanjing 210044, China
d
Network and Information Center, Guangdong University of Foreign Studies, Guangzhou 510006, China

article info a b s t r a c t

Article history: Facility layout problem (FLP) which deals with the layout of facilities within a given plant floor is an
Received 30 April 2019 NP-hard combinatorial optimization problem. This paper studies multi-objective unequal-area facility
Received in revised form 26 December 2019 layout problems (UA-FLPs) with the flexible bay structure (FBS), whose objectives refer to the material
Accepted 27 December 2019
handling cost, the closeness relationship, the distance requirement and the aspect ratio of facilities. In
Available online 13 January 2020
recent years, some successes have been achieved by multi-objective evolutionary algorithms (MOEAs)
Keywords: for solving various kinds of optimization problems with multiple conflicting objectives. However,
Multi-objective optimization traditional MOEAs face a great challenge in the convergence and diversity of solutions for UA-FLPs. In
Configuration space evolutionary algorithm this paper, a novel MOEA called the configuration space evolutionary (CSE) algorithm is developed
Flexible bay structure to solve the UA-FLPs with multiple objectives. We consider a mating pool called a configuration
Pareto-optimal solutions (solution) bank in the CSE, and use evolutionary operations (selection, novel crossover and mutation)
Facility layout problem to produce new configurations of the pool. By introducing a measure of the radius dspace of the
configuration bank, whose value is gradually reduced to narrow the search space, the convergence
of solutions in the CSE is controlled. A method of the nearest and farthest candidate solution based
on objective function normalization is combined with the fast non-dominated sorting to choose the
Pareto-optimal solutions, which is good for the algorithm to keep diversity of the obtained solutions.
The main contributions of this study lie in the use of a mechanism of evolution of population in the
algorithm based on a configuration bank, and the use of a selection strategy based on the nearest and
farthest candidate solution method, in order to improve the convergence and diversity of solutions.
Experiments are carried out on eight different representative instances and performance metrics from
the literature. Compared with the existing MOEAs, the CSE is able to find the better results and
show better performance. The numerical experiments confirm the effectiveness of the CSE for solving
multi-objective UA-FLPs.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction and reduce material handling costs at least 10% ∼ 30% [1]. Fa-
cility layout optimization design is one of the most important
With the growing intense competition in the international and complex problems in modern manufacturing. In general, the
market and the rapid development of social economy, an efficient facility layout problem (FLP) is the problem of determining of the
production system is very important to enhance the competitive- location of facilities within a given plant floor so as to satisfy some
ness of the enterprises. In order to reduce the production cost given objectives.
and enable quick changes to react to market demands, enterprises Facility location depends on its center coordinates which are
need to make full use of existing production workshop resources. continuous numerical variables. In theory, in the condition of
Research shows that about 20% ∼ 50% of the processing fees of known constraints, the FLP has the optimal solution. But con-
the enterprise’s production system are from the material han- tinuous center coordinates of facilities makes facility layout op-
dling. An excellent layout of facilities within the workshop can timization become one of the most complex combinatorial op-
enhance the efficiency of material handling between facilities, timization problems. In fact, the FLP belongs to the NP-hard
(non-deterministic polynomial-time hard) problem [2] which is
∗ Corresponding author at: School of Information Science and Technology, difficult to find the optimal solution in reality. Many scholars
Guangdong University of Foreign Studies, Guangzhou 510006, China. have widely researched it. General approaches to the FLP mainly
E-mail address: [email protected] (J. Liu). address the relative arrangement of equal-area facilities on a

https://doi.org/10.1016/j.asoc.2019.106052
1568-4946/© 2020 Elsevier B.V. All rights reserved.
2 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052

plant floor, which is called equal-area facility layout problem 2.1. Optimization approaches for UA-FLPs
(EA-FLP). The main shortcoming of this method is that equal-
area facility is a very poor assumption because few locations can To solve the UA-FLPs, various kinds of exact or approximate
accommodate some of the facilities. In fact, in reality, unequal- approaches have been proposed over the past several decades.
area facilities should be considered. This paper researches the Among these approaches, heuristics and meta-heuristics have
unequal-area facility layout problem (UA-FLP). At present, for attracted much attention due to their capability of strong global
the UA-FLP, most scholars just focus on reducing the material search. Genetic algorithms (GAs) are computational methods
handling cost between facilities within the workshop. In the based on the principles of natural selection and have been widely
actual process of facility layout, however, when the material han- applied to handle the UA-FLPs. Gómez et al. [7] gave a mix-
dling cost declines, it often can make other performances of the integer programming (MIP) model for optimizing the material
workshop equipment decrease [3]. In fact, the closeness relation- handling costs or the closeness relationship between facilities.
ship, the distance requirement among facilities and some other They applied GAs to solve layout problems in plant floors with
indicators directly affect the effectiveness of the workshop layout aisles. A set of 20 facilities was designed to test the effectiveness
and the investment cost. These factors should also be taken into of the GAs. Liu and Meller [8] put forward a GA-based heuris-
account when evaluating the best layout scheme. In this paper, tic which combined the sequence pair (SP) representation and
we consider simultaneously the material handling, the closeness heuristic strategy based the MIP model to solve the UA-FLP. For a
relationship, the distance requirement and the aspect ratio for all given SP, the corresponding layout was determined by the linear
facilities so as to establish a multi-objective optimization (MOO) programming relaxation of the MIP model. They showed the
model and obtain more scientific and reasonable layout scheme. effectiveness and efficiency of their algorithm by testing different
The weighted sum approach [4] which converts several objec- sized problems from both the literature and industrial applica-
tives to a single scalar objective by using weighted coefficients is tions. Gonçalves and Resende [9] presented a biased random-key
generally applied to solve the UA-FLPs with multiple objectives. GA for determining the order of placement and dimensions of
The poor operability of this method attributes to the fact that each facility in the UA-FLP whose objective is to minimize the
it is hard to standardize objective functions and determine the sum of the material handling costs between the centroids of the
weights beforehand. In order to overcome this defect of the facilities. In addition, a novel placement strategy was applied
classical weighted sum approach, Pareto optimization method [5] to position each facility and a linear programming model was
characterized by the evolution of the Pareto front which consists used to fine-tune the solutions. This combination improved the
of the fitness values of a set of individuals with most prof- results in most of the cases tested. Paes et al. [10] introduced
itable trade-offs between objectives can directly find the optimal two alternative heuristics for the UA-FLP: a basic GA, which relied
solutions in multi-objective optimization space and does not on a solution representation as a permutation of facilities and
need to convert multiple objectives into a single scalar objective, used a greedy heuristic for their insertion on the floor space,
so it can be used in more extensive applications. At present, and a GA with decomposition and reconstruction (DRGA). The
for multi-objective UA-FLPs, although all kinds of MOO meth- experimental results showed that the DRGA was more feasible
ods [6] based on Pareto optimization have been put forward, in addressing FLPs for larger-size instances than the basic GA.
they still have some shortcomings, such as slow convergence to Approaches based on traditional GAs usually exhibit a typical
the Pareto front and low efficient selection toward diversity of problem: premature convergence. Moreover, all the individuals in
solutions. Therefore, further improvement and development are a population normally need to be selected and evaluated in each
still awaited to enhance their effectiveness. This study proposes generation and they can be crossed and mutated, all of which
a novel MOEA called the configuration space evolutionary (CSE) requires additional CPU time when searching for good solutions.
algorithm which can effectively exploit and explore the solution Several GA variants have also been proposed to solve the
space for the UA-FLPs with four optimization objectives. The UA-FLPs. Ulutas and Kulturel-Konak [11] introduced an artifi-
main contributions of this study are as follows: (1) the use of cial immune system-based algorithm to solve the UA-FLP with
a mechanism of evolution of population in the algorithm based flexible by structure (FBS), where the proposed clonal selection
on a configuration bank, whose radius is gradually reduced to algorithm (CSA) had a new encoding and dummy facilities were
control the convergence of the CSE; (2) the use of a combination introduced to fill the empty space in the plant floor. In the
of the nearest and farthest candidate solution method based on CSA, the hypermutation and the receptor editing operators were
objective function normalization and fast non-dominated sorting distinctive from standard GAs where crossover and mutation
to choose the Pareto-optimal solutions, which is able to get a are the basic tools for creating new solutions. The algorithm
good spread in the Pareto front and keep the diversity of the showed consistent performance for the 25 problem cases stud-
solutions obtained. The numerical results show that the proposed ied. A new hybrid evolutionary algorithm (HEA) incorporating
CSE algorithm can find effectively the Pareto-optimal solutions of decision maker’s knowledge into the evolutionary algorithm was
the problem, and has better convergence and diversity than other proposed by García-Hernández, et al. [12] to solve the UA-FLP
approaches in the literature. with FBS. The novel hybrid system included a combination of
This paper is structured as follows. Section 2 reviews the an interactive GA with two different niching methods to pre-
related works in the literature. Section 3 describes the problem serve the diversity of solutions. This interactive algorithm might
statement for the multi-objective UA-FLP. To solve this problem, also execute slowly because they required the intervention of
the proposed CSE algorithm is introduced in Section 4 and perfor- a decision maker (generally, a human expert); furthermore, the
mances of the suggested algorithm and comparative experiments decision maker could be at risk of fatigue due to the amount
are presented in Section 5. Conclusions with a short summary are of information to be evaluated. Palomo-Romero, et al. [13] pre-
drawn in the last section. sented an island model genetic algorithm (IMGA) which used the
parallel evolution of a certain number of populations. The parallel
2. Related works approach helped to avoid premature convergence and excessive
execution time. An adaptive penalty function was used to direct
This section first introduces the optimization approaches for the search process to the feasible solution regions. In addition,
UA-FLPs and multi-objective UA-FLPs, and then handling methods this method improved search diversity, allowing it to explore a
of overlapping among facilities are introduced and analyzed. At larger search space and obtain better solutions. However, opti-
last, a survey of the related papers in the literature is given. mizing the number of islands (subpopulations) might be an issue
 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052 3

that should be considered. The procedure of the harmony search cases showed the effectiveness of the proposed algorithm. As
(HS) is analogous to an improvement operation for a GA. The main a representative of bio-inspired algorithm and meta-heuristic
difference between GA and HS is that the HS algorithm generates algorithm, PSO can also solve continuous optimization problems
a new solution after considering all of existing solutions in the and has good convergence speed. Asl and Wong [21] developed a
harmony memory, while the GA only considers the two parent modified PSO algorithm combining two local search methods to
solutions. In recent years, some researchers have introduced HS solve unequal area static and dynamic FLPs where the facilities
to solve the UA-FLPs. For example, Chang and Ku [14] developed have fixed shapes and areas throughout the time horizon. The
a heuristic method by combining the slicing tree representation objective of unequal area static FLP is to minimize the sum of the
and a quadratic constraint program (QCP) model with the HS to material handling costs while the objective of unequal area dy-
deal with the UA-FLP. The experimental results showed that this namic FLP is to minimize the sum of the material handling costs
approach was effective and could find optimal or near-optimal and rearrangement costs. The results showed that the proposed
solutions for most problems with fewer than 20 facilities. How- algorithm has created encouraging layouts in comparison with
ever, one limitation of this study is that the size of the instance other approaches.
is small, which should be expanded. Along the same direction,
Kang and Chae [15] proposed a novel HS method by incorporating 2.2. Optimization approaches of multi-objective UA-FLPs
some heuristic approaches to generate a quality solution of the
UA-FLP, where the structure of the slicing tree representation was Although the approaches mentioned above work relatively
modified and a re-adjustment operation was added to diversify well for the UA-FLPs, most researches focus only on the sin-
the possible range of solutions. Their approach used a penalty gle objective optimization of minimizing the material handling
scheme to produce the feasible solutions more effectively. Fur- costs or the sum of the material handling and rearrangement
thermore, they provided the results from experiments with a set costs which belongs to a quantitative factor. However, practical
of large-size problems that demonstrated the robustness of their UA-FLPs concern generally some competing objectives including
algorithm. In these proposals, the evolution of the population is both quantitative and qualitative ones. The closeness relationship,
guided by implementing subjective scores that evaluate a set of distance requirement, safety and so forth are important quali-
representative layouts, where the problem of becoming trapped tative factors in UA-FLPs. They have also significant impact on
in local optima is still easy to happen. the final location of facilities and should be taken into account
Meta-heuristics have global search ability and broad adapt- simultaneously. So, UA-FLPs generally belong to multi-objective
ability. Some meta-heuristics, such as ant colony optimization optimization problems (MOPs).
(ACO), simulated annealing (SA), particle swarm optimization In solving MOPs, an amount of MOO algorithms have been sug-
(PSO), and Tabu search (TS), seek to overcome the main disad- gested, which mainly include the following categories: EA [12],
vantages of sequential GAs. ACO is a bio-inspired optimization PSO [21,22], ACO [16,17,23], artificial immune systems (AIS) [11,
algorithm based on simulating the behavior of natural ants that 24], interval multi-objective quantum-inspired cultural algorithms
succeed in finding the shortest paths from their nest to food (IMOQCA) [25], brain storm optimization algorithms (BSO) [26]
sources. Some hybrid approaches based on ACO have also been and so on. In these algorithms, a category of the well-known
developed to handle the UA-FLPs. Komarudin and Wong [16] methods based on EA is the so-called multi-objective evolution-
put forward an ant system (AS) (one of the ACO variants) using ary algorithm (MOEA) [27] which is a kind of representative MOO
the slicing tree structure (STS) to represent the problems for method based on group and is suitable for solving constrained
solving UA-FLPs, without too restricting the solution space. An ant optimization problems with multiple objectives. Based on their
solution was given a penalty value that was proportional to the selection strategies, these MOEAs can be categorized into three
number of infeasible facilities that it contained. Guan and Lin [17] main classes [28]. The first category contains Pareto-dominance-
proposed a hybrid algorithm by combining the variable neighbor- based algorithms, e.g., non-dominated sorting genetic algorithm
hood search (VNS) and the ACO with a new pheromone rule and II (NSGA-II) [29], preference-inspired co-evolutionary algorithm
a reverse criterion to solve the single row FLP. In their approach, using goal vectors (PICEA-g) [30,31] and grid-based evolutionary
three neighborhood structures were utilized to enhance the ex- algorithm [32,33]. The second contains performance indicator-
ploitation ability of the algorithm. Scholz et al. [18] proposed based algorithms, e.g., approximated hyper volume-based evolu-
a TS algorithm based on slicing tree for solving the UA-FLP. tionary algorithm (HypE) [34]. The last contains decomposition-
They incorporated the possibility to specify various requirements based algorithms, e.g., multi-objective evolutionary algorithms
regarding shape and dimensions of each individual facility by based on decomposition (MOEA/D) [35], and NSGA-III [36]. These
using bounding curves. The experimental results showed that the algorithms decompose an MOP into a number of optimization
TS algorithm made great improvements compared with previ- problems and solve them using a population-based search in a
ous researches. SA is a probabilistic meta-heuristic to search for collaborative manner.
global optima in problems with complex search spaces. Kulturel- For multi-objective UA-FLPs, more and more researchers have
Konak and Konak [19] focused on solving four different UA-FLPs interested in Pareto optimization method. Liu and Liu [37] pro-
on the continuous plane for optimizing the sum of material posed an improved version of the multi-objective ant colony
handling costs and rearrangement costs, and proposed a large- optimization (MOACO) algorithm to solve the UA-FLP with two
scale hybrid simulated annealing (LS-HSA) algorithm based on objectives. In the modified MOACO algorithm, they put forward
SA and MIP. Recently, Hunagund et al. [20] presented a SA- a novel pheromone updating method, and combined the heuris-
based heuristic to solve unequal area dynamic FLPs with FBS. tic layout updating strategy, the Pareto optimization and the
The proposed SA heuristic gave new best solutions or the same niche technology to obtain Pareto-optimal solutions of the prob-
solutions as compared to other meta-heuristics for the tested lem. A combination of the local search based on the adaptive gra-
instances. Liu et al. [4] described a model where the sum of the dient method and the heuristic department deformation strategy
material handling and rearrangement costs was minimized. A hy- (DDS) was applied to deal with the non-overlapping constraint
brid algorithm which combined the Wang–Landau (WL) sampling between facilities. The experimental results on ten benchmark
algorithm and some heuristic strategies (vacant point strategy, instances showed that the proposed MOACO algorithm was an
pushing strategy, and pressuring strategy) was proposed to solve effective method for solving the UA-FLP with two objectives.
the UA-FLP. The computational results of the four groups of However, when the sum of the areas of all facilities is close
4 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052

or equal to the area of the plant floor, the efficiency is low 2.3. Handling methods of overlapping among facilities
for the MOACO to optimize the objectives and deal with the
interference between facilities. Zuo et al. [38] combined multi- For solving the UA-FLPs, another challenging issue except for
objective Tabu search with linear programming (MTS-LP) for an the optimization approach is how to handle the non-overlapping
extended double row layout problem where two objectives (ma- constraint among facilities. The traditional solution method usu-
terial movement and layout area) were considered. Experimental ally adopts line placement strategy [7,38,41,44] or penalty func-
results showed that MTS-LP improved upon NSGA-II and an exact tion method (PFM) [8–10,13,19,21,39]. However, their perfor-
approach for problems of no more than 30 machines. Saraswat mances are not always satisfactory because line placement meth-
et al. [39] presented a hybrid algorithm that combined the SP rep- ods limit the number of possible layouts generated and the PFMs
resentation, e-accurate model and multi-objective SA algorithm are hard to find appropriate penalty factors designed to direct
to solve the UA-FLP with three objectives: flow-distance, the the search toward the constrained optimal solutions. Flexible
number of required material handling devices and average work- bay structure (FBS) [11,12,20] is the expression of a continuous
facility layout. Based on the FBS, the layout area is divided into
in-process. However, this research neglected intra-departmental
several parallel horizontal or vertical bays (columns). Similarly,
queues when calculating the average WIP of a layout. Liu et al. [6]
an alternative decomposition scheme for facility layout design
studied a multi-objective UA-FLP where the objectives of the
is the slicing tree structure (STS) [14–16,18,40,42]. The STS is
problem aimed to minimize the material handling costs, maxi-
an encoding representation that organizes a layout into a tree
mize the total adjacency value and maximize the utilization ratio
structure where the relative locations of the facilities on the floor
of the plant floor (i.e. minimize the layout area). They put forward
are represented by a location matrix encoded in two chromo-
a heuristic configuration mutation operation and subsequent lo- somes. A slicing structure is obtained by cutting orthogonally
cal search to satisfy the non-overlapping constraint, and the an initial rectangle along horizontal or vertical direction (called
multi-objective particle swarm optimization (MOPSO) algorithm guillotine cut) and recursively executing to produce a series of
to obtain a set of Pareto-optimal solutions of the problem. The rectangles. In the STS, the plant floor is partitioned both in vertical
numerical results on three sets of different UA-FLPs from the and horizontal directions simultaneously, while in the FBS, the
literature with the size of the problem up to 62 facilities showed plant floor is partitioned either in vertical or horizontal direction,
that the proposed method was effective in solving the multi- but not in both directions. In general, the layout optimization
objective UA-FLP. Aiello et al. [40] proposed a Pareto based new approaches that apply the FBS or the STS to design layouts have
multi-objective genetic algorithm (MOGA) to solve UA-FLPs with good effectiveness for solving the UA-FLPs. The encoding vector
four objective functions (material handling costs, aspect ratio, of the STS is relatively long because it contains a variety of
closeness and distance requests). The suggested MOGA was based information about the consecutive operations of the slicing tree.
upon the STS. Results on an instance of 20 facilities confirmed The STS requires a (3n-2)-sized encoding vector [15] to represent
the effectiveness of MOGA in solving the multi-objective UA-FLP. a layout with n facilities, whereas the FBS requires only a (n+B)-
The main advantage of the proposed approach is its capability sized vector (where B is the number of bays in the layout).
to explore a wide space of solutions, preserving the practicabil- Although the STS can represent wider layouts some of which
ity of the design. But, a decision maker cannot choose optimal cannot be constructed in the FBS, the FBS is more suitable for
layouts from the same rank. Li et al. [41] applied NSGA-II to the layout of flexible facilities with fixed area. Furthermore, the
optimize a UA-FLP model referring to three objective functions blocky structure adopted by the FBS is more suitable for practical
(material handling costs, closeness and utilization ratio of plant application, and corridors can be set in the layout, which also
floor). The proposed algorithm adopted the strategy of the line- increases the universality of FBS application. Therefore, in this
by-line placement to avoid the overlapping of facilities. However, paper, we adopt the FBS method to eliminate overlapping in
the resulting layout was not compact enough. Ripon et al. [42] dealing with UA-FLP with constraints and obtain feasible layouts.
built a mathematic model for optimizing the material handling
2.4. A survey of the related references
costs and closeness rating, and proposed an adaptive VNS with
a modified 1-opt local search. Unlike conventional local search,
Table 1 gives a survey of the related papers in the literature
the proposed adaptive local search scheme could be used to
where the gaps and overlaps are identified with a spotlight on
automatically determine whether the local search should be used
the objective functions of different models, resolution approaches
in a GA loop or not. This can save the computation time wasted by
and methods of handling non-overlapping constraint among fa-
the traditional local search. García-Hernández et al. [43] proposed
cilities that have been applied to solve UA-FLPs. Several observa-
an evolutionary neural hybrid system by incorporating human ex- tions can be made regarding these research works. First, it can be
pert knowledge into the UA-FLP where four objectives (material noted that there are more researches focused on a single objective
handling costs, closeness and distance requirement, aspect ratio) in the mathematical models for the UA-FLPs while relatively less
were considered. A subset of facility designs was generated via on multiple objectives. Among multi-objective problems, most of
the GA and then evaluated by human expert’s knowledge learn- them aim to minimize the total material handling costs and maxi-
ing from an artificial neural network (ANN). The results of the mize the satisfaction of closeness relationship. Second, comparing
experiments on a real case of 365 facility layout designs validated meta-heuristics together, one finds that the procedures based on
the proposed approach. Vitayasak et al. [44] presented a modified GAs and their variants are most popular. Third, for multi-objective
backtracking search algorithm (BSA) which is a population-based UA-FLPs, it can be noted that Pareto-based approaches are rel-
iterative evolutionary algorithm for solving stochastic dynamic atively successful in exploring the Pareto-front and maintaining
FLPs with stochastic demand. The combination of material flow diversity. However, the current MOO algorithms still suffer from
and redesign costs were optimized. These proposals overcome a great challenge in the convergence to the Pareto optimal set and
the problems related to the traditional weighted sum approach, the maintenance of diversity of the Pareto-optimal solutions [45]
where it is difficult to determine the weight coefficients and when dealing with the UA-FLPs possessing more than three ob-
standardize objective functions. However, the convergence and jectives and the large-scale size of instances. Therefore, there is
diversity of solutions are hard to be guaranteed. Some effective a growing urgent need for developing a high-performance MOO
optimization techniques need to be explored further. algorithm for multi-objective UA-FLPs.
Table 1
Survey of papers related to UA-FLPs.
Reference Objective functions Resolution approaches Handling overlapping methods

J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052
Gómez et al. [7] Minimize MHC or Maximize CR GA Line placement
Liu and Meller [8] Minimize MHC GA + Sequence pair representation Penalty function
Gonçalves and Resende [9] Minimize MHC Biased random-key genetic algorithm Penalty function
Paes et al. [10] Minimize MHC Basic GA, GA with decomposition and reconstruction Penalty function
Ulutas and Kulturel-Konak [11] Minimize MHC Clonal selection algorithm FBS +Penalty function
García-Hernández et al. [12] Decision maker’s preferences Hybrid evolutionary algorithm FBS
Palomo-Romero et al. [13] Minimize MHC Island model genetic algorithm Penalty function
Chang and Ku [14] Minimize MHC Harmony search-based heuristic STS
Kang and Chae [15] Minimize MHC Harmony search-based heuristic STS+ Penalty function
Komarudin and Wong [16] Minimize MHC Ant system STS + Penalty function
Guan and Lin [17] Minimize MHC ACO + Variable neighborhood search \
Scholz et al. [18] Minimize MHC Tabu search STS
Kulturel-Konak and Konak [19] Minimize the sum of MHC and RC SA Penalty function
Hunagund et al. [20] Minimize the sum of MHC and RC SA FBS
Liu et al. [4] Minimize the sum of MHC and RC Heuristic Wang–Landau Heuristic approach
Asl and Wong [21] Minimize MHC or Minimize the sum of MHC and RC Modified PSO + local search methods Penalty function
Liu and Liu [37] Minimize MHC and Maximize CR MOACO LS+DDS
Zuo et al. [38] Minimize MHC and Minimize LA Multi-objective Tabu search + Linear programming Line placement
Saraswat et al. [39] Minimize MHC, Minimize NMHD, and Minimize WIP Multi-objective SA Penalty function
Liu et al. [6] Minimize MHC, Maximize CR, Minimize LA Multi-objective PSO LS
Aiello et al. [40] Minimize MHC, Maximize CR, Maximize DR, and Maximize AR Multi-objective GA STS
Li et al. [41] Minimize MHC and Maximize CR NSGA-II Line placement
Ripon et al. [42] Minimize MHC and Maximize CR Adaptive variable neighborhood search STS
García-Hernández et al. [43] Minimize MHC, Maximize CR, Maximize DR, and Maximize AR GA + artificial neural networks \
Vitayasak et al. [44] Minimize MHC and Minimize RAC Modified backtracking search algorithm Line placement
Liu et al. (this paper) Minimize MHC, Maximize CR, Maximize DR, and Maximize AR Configuration space evolutionary algorithm FBS
Notation conventions
MHC: the sum of material handling costs DR: the satisfaction of distance requirements RC: the sum of rearrangement costs WIP: the average work-in process
CR: the satisfaction of closeness relationship AR: the satisfaction of aspect ratio NMHD: the number of material handling devices LA: the layout area

5
6 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052

In this paper, four different objectives of the UA-FLPs are

taken into account: (1) minimization of material handling costs;
(2) maximization of the satisfaction of weighted closeness rela-
tionship; (3) maximization of the distance requirements and (4)
maximization of the satisfaction of aspect ratio of facilities. The
above objectives are chosen because they are frequently cited in
the literature and involve quantitative (1) and qualitative (2), (3),
and (4) aspects. Along the line of thought of MOGAs (or MOEAs),
a novel MOEA, called the CSE algorithm, is proposed. A mecha-
nism of evolution of generations based on a mating pool named
configuration (solution) bank is proposed in the CSE algorithm,
whose convergence is controlled by a measure of the radius
of the configuration bank, whose value is gradually reduced to
narrow the search space. Also, a combination of the nearest and
farthest candidate solution method based on objective function
normalization and non-dominated sorting is used to choose the
Pareto-optimal solutions of the problem, which is able to get a Fig. 1. A schematic diagram of facility layout in a plant floor with five bays.
good spread in the Pareto front and keep the diversity of the
solutions obtained.
It is noteworthy that this paper distinguishes from our previ- define the UA-FLP configurations (in this problem, a configuration
ous publications [4,37], and [6] in objective functions, resolution represents a possible layout of all facilities and all configurations
approaches and handling overlapping constraint aspects although are dispersed in the solution space independently), it is prefer-
these four papers all study the UA-FLPs, as shown in Table 1. In able to construct a solution for the UA-FLP under the following
fact, four objectives (minimizing MHC, maximizing CR, maximiz- assumptions: (1) the shapes of facilities are either square or
ing DR, and maximizing AR) in this paper are taken into account, rectangular and their lengths and widths can be changed within a
while [4,37], and [6] studied a single objective (minimizing the certain range; (2) the area of the plant floor is larger than or equal
sum of MHC and RC), two objectives (minimizing MHC and max- to the sum of areas of all facilities; (3) facilities must be located
imizing CR) and three objectives (minimizing MHC, maximizing within a given plant floor without overlapping each other; (4) The
CR, minimizing LA), respectively. Furthermore, a so-called CSE al- route of handling material is horizontal or vertical, and material
gorithm is proposed in this paper, which combines a mechanism flow is from a facility’s center to another facility’s center; (5) too
of evolution of generations based on a mating pool, and a policy extreme shape of any facility cannot be accepted.
of updating the configuration bank based non-dominated sorting Considering the characteristics of the facility and the conve-
and the nearest and farthest candidate solution method, in order nience of description, we define a coordinates system (Fig. 1)
to improve the convergence and diversity of solutions. Whereas, which is used to describe the positions of n facilities and calculate
in [4], a hybrid algorithm which combined the Wang–Landau the center of the configuration. The bottom-left corner of the
sampling algorithm and some heuristic strategies (vacant point plant floor is set as the origin and the x-axis coincides with
strategy, pushing strategy, and pressuring strategy) was proposed the length edge of the plant floor and the y-axis coincides with
to solve the UA-FLP. In [37], an improved version of the MOACO the width edge of the plant floor. Let li and wi represent the
algorithm was proposed to solve the UA-FLPs. In the MOACO, actual length and the actual width of facility i, respectively. (xi ,
the Pareto optimization based on the local pheromone commu- yi ) denotes the coordinates of the center of facility i and X = (x1 ,
nication and the global search based on the niche technology y1 , l1 , w1 , x2 , y2 , l2 , w2 , . . . , xn , yn , ln , wn ) is called as a layout
were applied to obtain Pareto-optimal solutions of the problem. (i.e., a configuration or a solution) of all n facilities. Fig. 1 shows
In [6], a MOPSO algorithm was presented to solve the UA-FLPs. the scheme of a layout with fourteen facilities for a UA-FLP with
In the MOPSO, an objective space division method was applied FBS, where the number of bays is five.
to locate addresses of the new configurations of all the parti- This study refers to the optimization of the following four
cles in the objective space according to their values of objective objectives [40,43]: material handling costs, weighted closeness
functions and an external archive was used to store Pareto- relationship, distance requirement and aspect ratio requests of
optimal solutions of the problem. Finally, the FBS is adopted facilities. Their formal descriptions are presented as follows:
to satisfy the non-overlapping constraint in this study, while ∑∑
heuristic methods combining a pushing strategy and a pressur- minimize F1 (X ) = (fij cij )dmij (1)
ing strategy in [4] were used to deal with the non-overlapping i j

constraint. A combination of the local search based on the adap- ∑∑

tive gradient method and the heuristic department deformation maximize F2 (X ) = rij gij (2)
strategy in [37] was applied to deal with the non-overlapping i j
constraint between facilities. In [6], they put forward a heuristic
configuration mutation operation and subsequent local search to ∑∑
maximize F3 (X ) = sij deij (3)
satisfy the non-overlapping constraint.
i j

3. The problem statement Formula (1) denotes that the total material handling cost
between facilities should be small as far as possible. Formula (2)
Assume that there are n fixed-area facilities and a plant floor, denotes that the total weighted closeness relationship between
the length and the width of which are L and H, respectively. The facilities should be maximized. The third objective (3) is to max-
multi-objective UA-FLP with FBS is the problem of placing facil- imize the distance requirement between facilities. fij denotes the
ities in the given plant floor in the representation of FBS so that material flow between facilities i and j, and cij is the material
facilities do not overlap each other and are satisfied with some handling cost for moving one unit load per unit distance from
given objectives. Because of the complexity of the features that facility i to facility j. Closeness rating rij is based on the intimate
 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052 7

have shortcomings, such as slow convergence to the Pareto front

and low efficient selection toward diversity of solutions, espe-
cially when dealing with the UA-FLPs with three or more objec-
tives. In this section, a novel MOEA called configuration space
evolutionary (CSE) algorithm that can search the solution space
efficiently and find good spread of solutions is developed to solve
the UA-FLP with multiple objectives.

4.1. Framework of configuration space evolutionary algorithm

The configuration space evolutionary (CSE) algorithm (see Al-

Fig. 2. Satisfaction function of the aspect ratio.
gorithm 1) is a novel MOO method. At the beginning of the CSE,
we randomly generate k1 = 50 feasible configurations (called an
initial configuration bank Pinit in CSE) in the configuration space
degree between facilities, which describes the connection degree (see Section 4.2). On the basis of the center of each configuration
of the layout of facilities, and gij is the contact perimeter length in Pinit , we construct a circular region with the radius dspace =
between facilitiesi and j. Distance requirement factor sij is deter- dav g /2, where dav g equals the average value of Euclidean distances
mined by the environment relation between facilities according among centers of all configurations in the initial configuration
to a designer’s demands which can be the separation of some bank Pinit . The center o(x, y) of each configuration is computed
facilities from others. Reasons for such requirement are on the by:
consideration of environmental issues like vibration, pollution, ∑n ∑n
noise or aspects related to the security of workers or risks of xi yi
o(x, y) = ( i=1
, i=1
) (6)
fire or explosion, which requires maintaining a certain degree of n n
distance between some facilities. The distance dmij between fa- where n denotes the total number of facilities, and (xi , yi ) is the
cilities i and j is the Manhattan distance between their centers in coordinates of the center of the ith facility.
formula (1), while the Euclidean distance deij has been employed Select randomly k2 =10 configurations (called seed configura-
in formula (3). tion bank Pseed ) from the configuration bank Pnew (updated in each
For the UA-FLP with FBS, each facility needs to have reasonable iteration, and its initial value is Pinit ) to produce k3 = 200 new
aspect ratio. The aspect ratio is considered as an objective (such configurations (called evolution configuration bank Pev olv e ) by
as the case of Section 5.1) because some ratios are better than using evolutionary operations (selection, crossover and mutation)
others, but the aspect ratio is also a constraint (see the problems (see Section 4.3).
of Sections 5.1–5.3) because any solution becomes infeasible if Then, we use the fast non-dominated sorting method [29]
the aspect ratio of any facility surpasses its limit. The aspect ratio and the nearest and farthest candidate solution method (see
of facility i is defined as follows: Sections 4.4 and 4.5) to pick out k4 = 10 excellent configura-
max{li , wi } tions (called the test configuration bank Ptest ) from the evolution
γi = (4) configuration bank Pev olv e . Subsequently, we apply these 10 new
min{li , wi }
test configurations to update the configuration bank Pnew (see
For the case of 20 facilities with zero clearance [40] in Sec- Section 4.6). Once the configuration bank Pnew is updated 50
tion 5.1, if the aspect ratio of the facility is within a certain times, we let dspace = dspace /2. If there still exist configurations
interval it still is acceptable but its level of satisfaction will de- in Pnew , which are not used as the seed configurations, we select
crease as long as it deviates from the optimal shape. If the aspect randomly 10 (or the rest several) unselected configurations as
ratio exceeds the given interval, the shape of facility cannot the seed configurations. Repeat the above process until all con-
be accepted because resources for production cannot be placed figurations in the configuration bank Pnew are used as the seed
inside the facility anymore, and at that time, the value of the configurations. Thus, one round of iteration of the CSE is finished.
satisfaction function of the aspect ratio will drop to 0. A zero If the maximum number of iterations is not reached, we produce
value implies an unfeasible solution. Such a satisfaction function randomly 50 new feasible configurations (denoted by Pinit ) based
is given in Fig. 2 [40], where the horizontal coordinate γ denotes on the FBS, and reset the new initial configuration bank by select-
the aspect ratio of the facility, the vertical coordinate ars is its ing 50 configurations with better fitness values from Pnew ∪ Pinit
corresponding satisfaction function value, and γi,opt denotes the based on the fast non-dominated sorting method [29] and the
optimal aspect ratio of facility i. γil and γir represent the legal nearest and farthest candidate solution method. Repeat the whole
range of upper and lower bounds of aspect ratio for facility i, procedure until the maximum number of iterations is reached.
respectively. If the aspect ratio of each facility is in the interval Afterwards, we use the fast non-dominated sorting method [29]
between γil =1 and γir =2.5, the layout is feasible; otherwise it and the multiple attribute decision-making method [46] to select
does not satisfy the aspect ratio constraint of the problem. For a several excellent configurations as the Pareto-optimal solution set
feasible layout X , maximization of the satisfaction of the aspect of the problem. Obviously, in the CSE, the number of configura-
ratio for all facilities is considered a further objective: tions in Pnew remains unchanged, which is 50 always. Note that
∑n k1 , k2 , k3 and k4 in the CSE are some changeable positive integer
i=1 arsi
maximize F4 (X ) = (5) parameters.
n
where the satisfaction level of a layout is measured by the mean 4.2. Encoding of solution in flexible bay structure
of the aspect ratio satisfaction score ars of all facilities.
EAs are a kind of highly parallel, random and adaptive meta-
4. Configuration space evolutionary algorithm heuristic algorithms based on survival fitness in natural evolution.
EAs operate with populations, which are named chromosomes
MOEAs are widely used for solving optimization problems and encoded solutions of the problem. Each chromosome is made
with multiple conflicting objectives. However, traditional MOEAs of fundamental gens. It simulates the collective learning process
8 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052

of individual group, and each of these individuals represents a

point in solution space of the given problem. For each chro-
mosome, a fitness value determines its survival rate. Usually,
randomly generate the initial population and in order to obtain
new population, the evolution process performs by executing
encoding, selection, crossover, and mutation operators.
In the literature review on the UA-FLPs, some researchers
have been using an interesting approach called the flexible bay
structure (FBS) [12,20] to represent the scheme of layout. The FBS
is a continuous layout representation allowing the facilities to be
located only in parallel bays with varying widths. The width of
a bay depends on the division result of areas of facilities that it
contains, and all facilities in a bay have the same width (Fig. 3).
Bays are bounded by straight aisles on both sides, and facilities
are allowed to place only in one bay (i.e. no expansion over
multiple bays). The number of bays and facilities that are placed
in every bay are changeable. Therefore, the FLP is simplified as Fig. 3. Bay structure representation.
determining the number of bays and the facilities contained in
every bay. In the light of the architecture of the FBS, all facilities
in the plant floor are put from down to up, and from left to right. Optimal layout designs based on the FBS have some limits in
The layout based on the FBS can use simple numbers to describe layout form, compared with other layout forms such as the slicing
layout results. structure. However, the FBS has also some desirable features. The
 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052 9

Table 2
Crossover operator in the proposed CSE algorithm.
Parent 1 Parent 2 Random values Child 1 Child 2
1.25 1.42 0.26 1.38 1.29
2.24 2.45 0.12 2.42 2.27
2.62 3.25 0.12 2.05 2.84
3.55 3.83 0.53 3.68 3.70
4.12 4.45 0.22 4.38 4.19

(1) Selection. The elite strategy is used to reserve best indi-

viduals of the former population (all configurations in the seed
configuration bank Pseed ) with the elite rate kelite (kelite =0.2 in this
paper). The number of individuals selected by the elite strategy
Fig. 4. Layout generated from the chromosome representation.
is
Nelite = Nseed × kelite (7)
bay boundaries form the basis of an aisle structure that is in favor where Nseed denotes the amount of individuals in the Pseed .
of transferring from a block layout to the actual facility layout. (2) Crossover. Crossover is the process of taking two parent
Therefore, the difference between the material handling costs of a individuals (configurations) to produce two children. After the
theoretical block layout and its actual implementation is minimal selection process, the offspring are enriched with better individ-
for an FBS-based layout. In summary, the FBS representation is uals, but selection does not create new ones. Crossover operator
one of the most important layout representations that researchers is applied to the seed configuration bank Pseed with the hope
and practitioners continue studying and using. that it creates some better offspring. By crossover, new generated
The CSE algorithm uses the FBS to represent the UA-FLP. The configurations inherit good ingredients of the parents. In this
scheme of a layout is regarded as a chromosome, which is formed paper, we use a continuous uniform crossover.
by n genes. Each gene adopts real number encoding zi (i = 1,2,. . . , In the continuous uniform crossover we choose first two con-
n), which is a random number between 1 and B + 1 (B is the total figurations with crossover probability kcross from Pseed as parents,
number of the bays) whose integer part represents the number then a set of random values λi (i = 1, 2, . . . , n) with uniform
of bays and decimal part represents the order of facilities in each distribution between 0 and 1 are produced, in same length with
bay (smaller numbers have priority). In this paper, the numbering existent chromosome, and by using the following Eqs. (8) and (9),
of the bays is from left to right (from x = 0) and the order two children are created.
position of the facility is from bottom to top (from y = 0). Fig. 4
shows a layout plan for a FLP with seven facilities in the planning
′
z1i = λi z1i + (1 − λi )z2i , ∀ i = 1, 2, . . . , n (8)
horizon. In Fig. 4, the gene related to the sixth facility has 1.62
values. It means this facility will be arranged in the first bay and
after comparing with other facilities (the first and fourth facilities,
′
z2i = (1 − λi )z1i + λi z2i , ∀ i = 1, 2, . . . , n (9)
respectively, with 1.25 and 1.85 values) in this bay it has the where z1i and z2i denote the first and second parents’ genes,
second rank, in this layout, so it gains the second priority. ′
respectively; z1i ′
and z2i are the first and second genes of offspring,
At the beginning of the CSE, 50 random configurations based respectively. So we can generate a total of 200 new configurations
on the FBS are produced according to the following procedures. and gain an evolution configuration bank Pev olv e . The crossover
First, we produce randomly an integer B between 3 and 7, where operation for a problem with five facilities is illustrated in Table 2.
B is the total number of the bays. Then, n random real numbers zi (3) Mutation. After crossover, the genes are subjected to mu-
(i = 1,2,. . . , n) (1≤zi ≤B+1) are produced, which construct a chro- tation. Mutation is beneficial to increase the diversity of popula-
mosome. According to these random numbers zi (i = 1,2,. . . , n), tions, and overcome the defect of premature convergence because
facilities can be put orderly in the plant floor. Because the width adaptation values of offspring that crossover operation produces
of a bay is changeable, all n facilities will be placed within a given do not achieve optimum and not evolve any more.
plant floor without overlapping each other, resulting in a feasible We perform the fast non-dominated sorting [29] for the con-
configuration. But a configuration may become infeasible if the figurations obtained by selection and crossover, and the con-
aspect ratio (or the side length) of any facility surpasses its limit, figurations in the final level after non-dominated sorting will
and at that time, we will abandon this generated configuration be mutated. Mutation introduces new genetic structures in the
and produce randomly another chromosome again until a feasible population by randomly modifying some of its building blocks.
configuration is gained. Repeat the above steps until 50 feasible A gene is selected randomly by mutation probability kmutate and
configurations are produced. Obviously, the coding numbers of then is replaced with a random number between 1 and B+1,
genes in a chromosome generate randomly to increase the di- which is produced as following:
versity of solutions. In addition, considering the randomness of
zi + ((B + 1) − zi ) tanh(ku),
{
if tanh(ku) ≥ 0
the quantity of bays and the quantity of facilities placed in every zi′ = (10)
bay, the UA-FLP can be converted to optimize the number of bays, zi + (zi − 1) tanh(ku), else
the number of facilities contained in every bay, and the facilities’ where zi is the value of the selected randomly gene and zi′ is
placement order. the replaced value. u is a standard normal distribution random
number, and k is its coefficient.
4.3. Production of the configuration It is worth noting that whenever the crossover or mutation
operations produce infeasible configurations (the aspect ratio
In this section, we use the selection, crossover and mutation (or the side length) of any facility surpasses its limit), we will
in evolutionary algorithms to produce some new configurations abandon them and the crossover or mutation operations are
and enhance the diversity of solutions in the configuration space. repeated until new feasible configurations are gained. In the FBS
10 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052

Fig. 5. Five optimal solutions selected by different methods for 12 Pareto-optimal solutions.

representation, because all facilities are put in the plant floor from aforementioned operations until the number of solutions in set A
down to up, and from left to right, it is obvious that configurations reaches q.
are satisfied with the non-overlapping constraint among facilities. In the nearest and farthest candidate solution method, for two
solutions X s and X t , we need to define a distance calculation
4.4. Nearest and farthest candidate solution method method based on the objective functions. Due to huge differ-
ence between the different objective values, we should normalize
In MOPs, it is desired that an algorithm maintains a good these objective functions (see Section 4.5) before calculation. The
spread of solutions in the non-dominated solutions as well as distance between solutions X s and X t is computed by:
the convergence to the Pareto-optimal set. In NSGA-II [29], a 
 m
crowded degree comparison method is used to keep the spread ∑
distance(Xs , Xt ) = √ (F̄i (Xs ) − F̄i (Xt ))2 (11)
of solutions. Thereafter, the crowded degree comparison method
i=1
has been widely used for diversity maintenance of solutions in
many MOEAs. However, it has an obvious flaw that the solutions where F̄i (X ) is the value of objective function Fi (X ) after normal-
in the high-density space have lower chance to be selected so izing and m is the number of objective functions.
that the spread of solutions is not good enough. So, some scholars Fig. 5 shows an optimal solution set consisting of five Pareto-
have made improvements on the basis of the original crowded optimal solutions selected from twelve Pareto-optimal solutions,
degree method. Kukkonen and Deb [47] proposed an improved where two objectives F1 and F2 are considered. The solid circles
pruning of non-dominated solutions for bi-objective optimiza- in Fig. 5(a) is the optimal solution set obtained by the nearest
tion problems, which removed the solutions with the smallest and farthest candidate solution method, and the solid circles in
crowding distance value one by one. In [48], a fast and effective Fig. 5(b) is the optimal solution set obtained by the congestion
method for pruning of the non-dominated solutions based on the degree method in NSGA-II [29]. The numbers in the figure in-
crowded degree and the nearest neighborhood of solutions was dicate the order of selecting solutions. It is not hard to see that
proposed for many-objective problems. However, in some cases, the nearest and farthest candidate solution method can obtain a
these methods cannot still get a good spread of solutions. more uniform Pareto front than the traditional congestion degree
We replaces the crowded degree comparison with the nearest method in NSGA-II, and the Pareto-optimal solutions obtained is
and farthest candidate solution method in this paper, which can more diverse also.
solve above difficulty to a certain degree. Given a candidate solu-
tion set C, which contains p non-dominated individuals, we need 4.5. Normalization of the objective functions
to choose q (≤ p) Pareto-optimal solutions from C and obtain an
optimal solution set A. At the beginning of computation, let A be We use the method of normalization of the objective func-
an empty set. Firstly, calculate all the objective function values Fi tions in the NSGA-III [36,49]. First, calculate the least objective
(X j ) (i = 1,2,. . . , m) of each solution X j (j = 1,2,. . . , p) in set C and value Gmin
i of every objective function Fi (i = 1,2. . . , m) in the
find the minimal value of each objective, where m is the number candidate solution set C, and then construct the ideal point set
1 , G2 , . . . , Gm ). The objective value of each solution in
G = (Gmin min min
of objective functions. There are two cases:
(i) If q < m, we randomly select q solutions with the smallest C is then translated by subtracting objective value Fi (X ) by Gmin
i .
objective function values based on the preferences of different This translated objective is denoted by:
objectives into set A, and at the same time delete them from the Fi′ (X ) = Fi (X ) − Gmin (12)
i
candidate solution set C.
(ii) If q ≥ m, for each objective, we select a solution with where X is a layout of n facilities, that is, a configuration (or a
the smallest objective value and put it into A; at the same time solution). Therefore, the ideal point of translated C becomes a
delete it from C. The remaining u (which equal to q-m) population zero vector.
individuals are chosen from the updated candidate solution set C. Second, identify the extreme point Gi,max in each (ith) objective
For each solution X s in C (|C | = u) at this time, we calculate the axis by finding solution X that makes the following achievement
nearest distance (about normalized objective function) between scalarizing function (ASF) (formed with Fi′ (X ) and a weight vector
X s and all individuals in the optimal solution set A (see Eq. (11)). w close to the ith axis direction) minimum:
Choose solution X far with the farthest distance from C and add m Fi′ (X )
it into the optimal solution set A and delete it from set C. Repeat ASF (X , w ) = MAX (13)
i=1 wi
 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052 11

Fig. 7. Search procedure of the CSE in the configuration space.

Table 3
Parameter values of CSE.
Parameter Value
Initial configuration size (k1 ) 50
Seed configuration size (k2 ) 10
Evolution configuration size (k3 ) 200
Test configuration size (k4 ) 10
Elite rate (kelite ) 0.2
Fig. 6. Procedure of using extreme point to construct a hyper-plane for a Crossover probability (kcross ) 0.7
three-objective problem. Mutation probability (kmutate ) 0.2
Update frequency of dspace 0.5
Objective axis weight (wi ) 10−6

In the above function, for the ith translated objective direction

Fi (i = l), where l∈ {1, 2, . . . , m}, we set its corresponding weight
as 1, and for the rest objective direction Fi (i ̸ = l), we set wi = 0 Obviously, case (1) is more likely to happen when the circle
(usually we replace it with a small number 10−6 ). is large, and case (2) happens when the circle is small. Con-
Finally, according to these m extreme points we construct a sequently, a larger value of dspace produces more diversity of
m-dimensional hyper-plane. The intercept ai between this hyper- solutions, whereas a smaller value results in quicker conver-
plane and the ith objective axis can be computed. Note that if the gence toward sub-optimal configurations at the expense of get-
extreme value point cannot construct the hyper-plane, then let ai ting trapped in a basin. So, to search efficiently for the configu-
=Gi,max (see Fig. 6, where the objective is three-dimensional). The ration space, it is necessary to maintain the diversity of solutions
objective functions can then be normalized by the formula (14). in the early stages and then gradually shift the emphasis toward
sub-optimal configurations, which is realized by slowly reducing
Fi′ (X ) the value of dspace in the CSE (see Fig. 7).
F̄i (X ) = , i = 1, 2, . . . , m (14)
ai
5. Experimental results and analysis

4.6. Updating of the configuration bank To evaluate the performance of the CSE algorithm, three sets of
instances from the literature are performed. The first set contains
Updating the configuration bank is an essential step in the a case of 20 facilities with zero clearance. The second set contains
CSE. For each new test configuration T we calculate the Euclidean a case of 12 facilities with vertical and horizontal aisles and
distance between it and each configuration in the configuration the third set consists of six benchmark instances with different
bank Pnew . Choose the configuration with the nearest distance, number of facilities. The algorithm is compiled by Java language
named by Y . The distance between configurations T and Y is and all instances are solved by a PC with Core 2 Duo processor
denoted by D (T , Y ). (2.94 GHz) and 2.0 GB memory. The proposed CSE algorithm
(1) If D (T , Y ) ≤ dspace , the test configuration T is contained contains some parameters that have been tuned experimentally.
in the circle centered at the center of Y , which means these two The parameter values that provide good results are shown in
Table 3, and the maximum number of iterations for each instance
configurations are similar. If T dominates Y , we add T into the
is listed in the resultant tables of the corresponding problems.
configuration bank Pnew , and delete configuration Y from Pnew so
To avoid random effects, each instance is executed 10 times
as to keep the total number of the configurations in Pnew constant.
independently.
The center of circle moves from the center of Y to that of T . If
Y dominates T , we keep Y and abandon configuration T . If the 5.1. A case of 20 facilities with zero clearance
configurations T and Y do not dominate each other we randomly
select one from T and Y to update configuration bank. When T The first case is from [40,50]. Facility areas are reported in
is selected, the center of circle is moved to that of T . Table 4. Flows between facilities and unit material handling costs
(2) If D (T , Y ) > dspace , the test configuration T falls outside are the same as [50]. Closeness requirements and distance re-
of all existing circles. If T dominates a certain configuration Z quirements are shown in Table 5 and Table 6, respectively. The
in Pnew , we remove Z from Pnew and add T to the configuration satisfaction function of aspect ratio of each facility is shown in
bank Pnew . The center of circle moves from the center of Z to that Fig. 2.
of T . If T does not dominate any configuration in Pnew the new The feasible and Pareto-optimal solutions obtained by the CSE
configuration T is removed. algorithm, characterized by different non-dominated values of
12 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052

Table 4
Facility areas in the first case.
Facility 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Area 27 18 27 18 18 18 9 9 9 24 60 42 18 24 27 75 64 41 27 45

Table 5
Closeness requests between two facilities in the first case.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
6 4 4 4
9 8 8 8
11 2 2 2

Table 6
Distance requests between two facilities in the first case.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
1 8 8 8 8 8 8
2 2 2 2 2 2 2
3 4 4 4 4 4 4

Table 7
Best solutions obtained by approaches CSE and MOGA for the first case.
Algorithm Solution F1 (X ) (min) F2 (X ) (max) F3 (X ) (max) F4 (X ) (max) Iterations Time (s)
1 4914 213 3333 0.76
2 4979 198 3264 0.81
CSE 1500 10739
3 4879 162 2640 0.71
4 4698 162 2565 0.72
1∗ 5128 150 1227 0.71
MOGA 2500 –
2∗ 6211 91.5 2218 0.69

Table 8
Chromosomes of best solutions by approaches CSE and MOGA for the first case.
Algorithm Solution Chromosome
1 {6.4,2.7,6.0,3.2,2.1,3.9,3.8,3.4,5.4,5.2,4.9,4.6,1.3,3.0,4.2,6.2,1.8,1.5,2.5,5.7}
2 {6.4,2.6,6.0,3.4,2.0,3.9,2.9,3.6,5.3,5.1,4.8,4.5,1.2,3.1,4.1,6.2,1.7,1.5,2.3,5.7}
CSE
3 {2.8,1.5,1.7,1.2,5.0,3.3,3.1,2.0,3.6,2.6,6.3,6.0,3.9,3.8,5.6,4.7,4.5,5.3,6.1,2.3}
4 {2.9,1.5,1.7,1.2,4.5,3.3,3.1,2.0,3.6,2.6,6.3,6.1,3.9,3.8,5.3,5.9,4.9,4.6,5.5,2.3}
1∗ 0-0-0-0-0-15-8-7-6-4-1-20-11-19-2-16-9-5-14-17-10-12-3-18-13
MOGA
2∗ 0-0-0-0-0-1-11-20-15-8-7-6-19-9-4-2-17-16-12-5-3-14-10-18-13

the objective functions, are in the number of 28 on a population obtains better solutions than the MOGA in less iterations, demon-
of 50 individuals. Four of them are selected as optimal solutions strating the effectiveness of the CSE. Because of the differences
by the Electre III [46]. The objective values and the correspond- in the performance of the running computers and the program-
ing chromosomes of the four solutions obtained are reported in ming languages, the CPU time of other algorithms executed on
Table 7 and Table 8, respectively, and the corresponding layouts different computers or programming languages cannot be used
are shown in Fig. 8. These solutions are compared with those in this comparison. The running time by the MOGA has not been
obtained by a non-dominated ranking multi-objective genetic reported in [50]. But from Table 7, it is not hard to see that the
algorithm (MOGA) [50], which began from 50 randomly produced CSE can find the results in an acceptable time.
layouts based on the STS. Table 7 shows that all the solutions
reported dominate those obtained by the MOGA, verifying the 5.2. A case of 12 facilities with vertical and horizontal aisles
effectiveness of the proposed CSE algorithm. In Fig. 8, the facilities
having a closeness and distance requests (Tables 5 and 6) are
To verify the validity of the algorithm, we test another case
highlighted in dark gray and light gray areas, respectively. Solu-
from [51], whose objectives refer to minimization of material
tions 1∗ and 2∗ (Tables 7 and 8) are optimal solutions reported
handling costs F1 (X ), maximization of the weighted closeness
in [50].
From the results in Table 7, one can see that the optimal values relationship F2 (X ) and maximization of satisfaction of distance
of four objectives are 4698, 213, 3333 and 0.81, respectively. requirement F3 (X ). This model is different from the prior one. For
Compared with the best results by the MOGA in the litera- the convenience of handling materials, the vertical and horizontal
ture, material handling costs by the CSE reduces (5128-4698)/ corridors between facilities within the plant floor are set, with
5128×100%=8.39% and the closeness, distance and aspect ratio a certain width (Fig. 9). This is in accordance with the actual
increase (213–150)/150×100%=42%, (3333–2218)/2218×100%= situation.
50.27% and (0.81–0.71)/0.71×100%= 14.08%, respectively. About In this model, the longitudinal aisle width W is 25 m, and the
the number of generations of iterations, best solutions are reached lateral aisle width D between facilities within the same block is
by the MOGA in approximately 2500 iterations. The number of 15 m. Total number B of bays is set to 4. The area of the work-
iterations for the CSE to find the best solutions are about 1500 shop where there are a total number of 12 facilities is 420,000
interations, which are obviously less than that of the MOGA. square meters. The area of each facility is listed in Table 9. Flows,
Generally, achieving to better solutions can be resulted from more closeness relation and distance requirements for environment
iterations of the algorithm. Whereas the proposed CSE algorithm and security are shown in Table 10. Unit material handling cost
 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052 13

Fig. 8. Block layouts of the optimal solutions obtained by the CSE for the first case. Pictures (a), (b), (c) and (d) are four Pareto-optimal solutions in Table 7 and
Table 8, respectively.

where Ak is a penalty factor for material handling costs and R is a

big positive number. Assume F1k is the objective function value
about material flows for solution X k , then the objective about
material flows with a penalty term is:
′
F1k = F1k × Ak (16)
Because the optimization goal of material handling costs in
the model is the minimum, in the iterative process, if any facility
exceeds the boundary of the region, that is, a solution (layout)
does not satisfy the non-overlapping constraint, it will be obliged
to eliminate due to large objective value of material handling
costs.
The Pareto-optimal solutions obtained by the CSE are in the
number of 16 on a population of 50 individuals. Four of them are
selected as optimal solutions by the Electre III [46]. The optimal
objective values and the corresponding chromosomes by the CSE
Fig. 9. A layout with vertical and horizontal aisles. are given in Tables 11 and 12, respectively, in comparison with
those by the NSGA-II that started the search from 100 randomly
generated layouts based on the FBS. The corresponding layouts
is set to 1 and the aspect ratio range of the facility is between 0.3 obtained by the CSE are shown in Fig. 10.
and 5. From Table 11, it is not hard to see that the results by the CSE
To apply the CSE to solve the layout problem, we introduce can dominate all of the solutions by NSGA-II [51]. The optimal
material handling cost obtained by the CSE reduces (1.625–
the penalty function method in this experiment. The method is
1.569)/1.625×100%=3.446%, the best closeness and the best dis-
a generic constraint handling method, which adds the degree of
tance increase (22.7–14.8)/14.8×100%=53.378% and (25846−
constraint violation to punish a candidate solution.
25330)/25330×100%=2.037%, respectively, compared with the
Due to the order of the facility layout is along the direction
best results obtained by the NSGA-II. Especially, for the objec-
from bottom to up and from left to right, there will be no facilities tive of closeness, our results are much better than those in
beyond layout area in y direction. We need only to determine the literature. About the number of generations of iterations,
whether facilities of the last column in the x direction are beyond optimal solutions are reached by the NSGA-II in approximately
the boundary of the region. So we use the following penalty 500 iterations. The number of iterations for the CSE to find the
function to deal with constraint in the model [51]: optimal solution are about 400 iterations, which are obviously
{
1, if max (xi + 0.5li ) ≤ L less than that of the NSGA-II. In [51], they did not report the
i=1,2,...,n running time of the MOGA. But from Table 11, it is not hard to
Ak = (15)
R, else see that the CSE finds the results in an acceptable time.
14 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052

Table 9
Facility areas (m2 ) in the second case.
Facility 1 2 3 4 5 6 7 8 9 10 11 12
Area 10000 60000 65000 20000 50000 20000 35000 20000 35000 15000 15000 10000

Table 10
Flows, closeness relation and distance requirements in the second case.
Facility 1 2 3 4 5 6 7 8 9 10 11 12
Flows
2 1200 2200 500 4000
3 2000 5000 700 10000
4 1000 2500 3000
5 1500 4000 300 6000
6 200 2000
7 200 6000
8 500
Closeness relation
1 4
9 3 3 3
11 4 4 4 4
12 4 2
Distance requirements
1 6 6 2 2 4 4
10 2 2 2 2 2

Table 11
Best solutions obtained by approaches CSE and NSGA-II for the second case.
Algorithm Solution F1 (X ) (min) F2 (X ) (max) F3 (X ) (max) Iterations Time (s)
1 1.581 E+07 21.4 25740
2 1.569 E+07 20.6 25441
CSE 400 6781
3 1.613 E+07 21.6 25846
4 1.604 E+07 22.7 25535
1∗ 1.625 E+07 7.8 20137
NSGA-II 500 –
2∗ 2.117 E+07 14.8 17225
3∗ 1.752 E+07 8.6 25330

Table 12
Chromosomes of best solutions by approaches CSE and NSGA-II for the second case.
Algorithm Solution Chromosome
1 {4.79,1.67,2.97,1.38,3.31,3.62,2.31,1.98,2.60,4.00,4.30,4.60}
2 {4.77,1.68,2.64,1.38,3.04,3.30,2.08,3.71,2.43,4.03,4.29,4.57}
CSE
3 {4.79,1.95,2.89,1.62,3.42,3.63,2.26,1.35,2.57,4.25,3.95,4.53}
4 {4.78,1.78,2.02,1.37,3.73,3.37,2.62,3.02,2.39,4.27,3.97,4.55}
1∗ {1,5,3,9,7,12,6,8,4,10,11,2,2,3,4,3}
NSGA-II 2∗ {3,9,7,8,2,6,4,11,1,10,12,5,3,2,4,3}
3∗ {4,2,8,7,9,3,12,5,11,10,6,1,3,3,3,3}

5.3. A set of benchmark instances with different number of facilities six benchmark instances also need to satisfy specific constraints
that the maximum aspect ratio of each facility is 4 for O8, Ab20,
In order to investigate further the validity of the CSE, six SC35 and Du62, and the minimum side length of each facility is
benchmark instances O8 from [52], Ba14 from [53], Ab20 from 1 for Ba14, and the maximum aspect ratio of each facility is 5 for
[54], SC30 and SC35 from [18], and Du62 from [55] are tested SC30. The experimental results by the proposed CSE algorithm,
for optimizing objectives F1 (X ), F2 (X ) and F3 (X ). Dimensions of the NSGA-II and the MOGA in the literature are shown in Table 14
the workshop for this group of benchmark instances are shown for contrast, where the CSE and NSGA-II algorithms begin from
in Table 13. The material flows between facilities and the area the same search points based on 50 randomly produced layouts
of each facility for these instances could be found in [16]. Due using the FBS while the MOGA starts from 50 randomly produced
to very few benchmark problems for the UA-FLPs with multiple different initial layouts using the STS.
objectives can be found in the literature (especially for cases with Table 14 shows that the best values (Best), average values
closeness requests and distance requests), this paper generates (Avg) and standard deviations (SD) obtained by the CSE algorithm
test data for closeness relation and distance requirements as [37, for objective functions F1 (X ), F2 (X ) and F3 (X ) are better than re-
42], that is, some facilities are first chosen randomly and then sults of the NSGA-II and MOGA for all six instances. For example,
a set of integers in the interval [1,6] are produced as closeness the optimal material handling costs by the CSE for instance Du62
requests or distance requests of these facilities (details of pro- reduces (3423555.90–3267707.61)/3423555.90×100%= 4.6%, the
duced values are shown in Tables A.1–A.6). What is more, these best closeness and the best distance increase (1044.02–999.77)/
 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052 15

Fig. 10. Block layouts of the optimal solutions obtained by the CSE for the second case. Pictures (a), (b), (c) and (d) are four Pareto-optimal solutions in Table 11
and Table 12, respectively.

Table 13
Dimensions of the workshop for a set of benchmark instances with different number of facilities.
Instance Number of facilities Width of the workshop Length of the workshop
O8 8 13 11.31
Ba14 14 10 6
Ab20 20 3 2
SC30 30 12 15
SC35 35 15 16
Du62 62 137.18 100

Table 14
Comparison of results by CSE, NSGA-II and MOGA for a set of benchmark instances with different number of facilities.
Instance Algorithm F1 (X ) (min) F2 (X ) (max) F3 (X ) (max) Iterations
Best Avg SD Best Avg SD Best Avg SD (Scanning times)

O8 CSE 221.28 231.61 6.05 108.73 86.35 11.59 52.11 42.78 10.68 400 (2 400 784)
NSGA-II 225.50 235.87 6.74 105.48 84.19 11.86 51.99 40.87 11.46 750 (4 102 015)
MOGA 225.73 237.96 7.22 107.29 80.59 13.21 52.11 41.68 12.77 900 (5 793 380)

Ba14 CSE 5134.05 5300.88 99.06 78.09 68.07 6.54 110.30 95.13 7.09 600 (4 676 996)
NSGA-II 5442.18 5604.76 115.00 72.75 59.99 9.56 103.33 89.80 8.07 1100 (6 831 310)
MOGA 5587.63 5680.19 128.20 77.24 62.52 11.29 94.25 88.64 8.93 1800 (9 521 682)

Ab20 CSE 3770.04 4303.35 284.88 13.57 9.24 2.29 36.14 31.48 3.04 1500 (10 959 006)
NSGA-II 4332.90 4898.75 318.55 11.42 6.77 2.30 32.05 28.64 3.84 3000 (16 680 015)
MOGA 4412.32 4942.93 372.24 13.23 8.27 2.33 31.41 26.84 4.02 2500 (14 000 316)

SC30 CSE 5295.77 6193.97 489.17 128.66 98.90 13.80 468.87 398.90 35.78 1800 (13 048 194)
NSGA-II 6483.01 7999.44 929.44 107.61 81.70 14.72 434.77 371.13 39.81 4500 (31 363 401)
MOGA 6281.60 7343.72 832.43 110.32 80.14 15.79 446.25 368.78 39.92 5000 (35 261 297)

SC35 CSE 6964.75 7929.88 1120.93 100.02 94.81 3.01 327.75 306.62 8.63 2100 (18 115 973)
NSGA-II 8645.38 11039.86 1178.10 89.89 63.78 15.96 322.88 280.68 21.13 6000 (33 806 821)
MOGA 8827.41 9313.58 1145.88 91.42 68.35 16.32 318.37 273.35 19.35 8000 (41 241 394)

Du62 CSE 3267707.61 3441469.94 77796.23 1044.02 872.85 106.25 10907.6 9605.36 620.36 3000 (24 110 602)
NSGA-II 3423555.90 3533971.02 84201.77 999.77 781.82 120.00 10514.8 9304.43 716.73 8000 (57 477 519)
MOGA 3482942.48 3724182.37 100832.42 835.39 639.13 130.11 9990.53 9142.25 724.84 12000 (80 372 726)

999.77×100%=4.4%, and (10907.6–10514.8)/10514.8×100%=3.7%, obvious that the CSE can find optimal solutions with less scanning
respectively, compared with the best results obtained by the times of objective functions within less iterations than NSGA-II
NSGA-II. For two larger instances SC35 and Du62, two layouts and MOGA. In Table 14, the best results are shown in bold font.
with the smallest F1 (X ) as the optimal solutions for preference
which are found by the proposed CSE algorithm are shown in 5.4. Performance metrics
Fig. 11(a) and (b), respectively. In addition, the number of it-
erations and times of scanning objective functions to find the Convergence to the true Pareto front and uniformity (or di-
Pareto-optimal solutions in the CSE, NSGA-II and MOGA algo- versity) of the Pareto-optimal solutions are generally used as
rithms for each instance are also listed for comparison. It is distinct goals for testing the performance of MOO approaches in
16 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052

Fig. 11. Layouts with the smallest F1 (X ) as the optimal solutions for preference by the CSE algorithm for two larger instances. (a) Instance SC35 with F1 (X )=6964.75,
F 2 (X )=96.16, and F3 (X )=304.81. (b) Instance Du62 with F1 (X )=3267707.61, F 2 (X )=964.87, and F3 (X )=8082.639.

MOPs. Generally speaking, some statistical analysis tools should Table 15

The average value of each performance measure by the CSE over 10 independent
be applied to analyze the quality of MOO approaches. In this
runs for case 1 and case 2.
paper, four commonly used performance metrics including Pareto
Test case PR value GD value SP value OPS value
Ratio (PR) [56], Generational distance (GD) [57], Space (SP) [58]
Case 1 0.833 17.03 5.017 0.694
and Overall Pareto Spread (OPS) [59] are first used to measure
Case 2 0.698 112.59 9.163 0.581
the convergence and uniformity of the CSE algorithm. Then, the
non-parametric statistical analysis based on the Wilcoxon signed
ranks test [60] is adopted to compare the performance of the
proposed algorithm with those of the other algorithms. variance of each of the obtained solutions to its closest neighbor
PR metric [56] is used to compute the rate of production of is computed by Eq. (19).
non-dominated solutions at a given iteration of an optimization

 p
algorithm. PR is computed by the Eq. (17).
 1 ∑
SP = √ (D̄ − Di )2 (19)
|PPF (S)| p−1
i=1
PR = (17)
|S | (∑m ⏐
⏐ , i, j = 1, 2, . . . , p,
⏐)
where Di = minj k=1 Fk (Xi ) − Fk (Xj )
⏐
Here PPF is the set of the practical Pareto frontier that is close and p denotes the number of Pareto-optimal solutions by the
but different from the actual Pareto frontier of a problem. |PPF (S)| optimization algorithm and D̄ is the average of all Di . The closer
represents the number of solutions of the PPF in the solutions set the found solutions are to the uniform distribution, the smaller
S. S is the set of all the solutions computed by the multi-objective the corresponding value of SP is. When SP = 0, all non-dominated
optimization method at a given iteration, and |S | represents the solutions found are evenly distributed.
total number of solutions of S. The value of PR varies from 0 to 1 The OPS [59] is computed by Eq. (20). When the obtained
and the ideal value of a population is 1. solutions are distributed more widely, the value of OPS becomes
The GD proposed by Van et al. [57] is generally applied to larger.
evaluate the average distance between the Pareto-optimal solu- m
∏
tion set Q obtained by an algorithm and the actual Pareto-optimal OPS = OPSt (20)
solution set P ∗ . This distance is computed by Eq. (18). t =1
√∑
|Q |
i=1 d2i |maxX ∈P ∗ Ft (X ) − minX ∈P ∗ Ft (X )|
GD = (18) OPSt = (21)
|Q | |Ft (PB ) − Ft (PG )|
|P ∗ | Here PB and PG are the Utopia and Nadir design points [61], which
√∑
m
k =1 (Fk (Xi ) − Fk (Xj )), i = 1, 2, . . . , |Q | .
∗
Here di = minj =1 are composed of the best and worst values of the objectives in
∗
Fk (Xj ) means that the kth objective value of solution Xj in the the set of all the solutions, respectively. P ∗ is the Pareto-optimal
Pareto-optimal solution set P ∗ . m is the number of objective func- solution set found.
tions and di denotes the Euclidean distance between each solution Table 15 lists the average value of each performance measure
and the nearest one in the Pareto optimal set. The value of GD by the CSE over 10 independent runs for case 1 and case 2,
reflects the relative distance between the obtained solutions and and it is clear that the convergence, uniformity and diversity
the actual Pareto frontier. A smaller value of GD reveals a better distribution of the obtained solutions are satisfying. Table 16 lists
convergence to the Pareto-optimal solution set and when GD = 0 the statistical results for these two test cases, including the best,
all obtained solutions are located in the Pareto frontier. the worst, the mean values, and the standard deviation for each
The SP proposed by Schott [58] is another important metric objective function over 10 independent runs. For all objectives in
tool which is generally used to evaluate the uniformity of the these two cases, the standard deviations are small, indicating that
spread of the solutions obtained by an algorithm. The distance the objective values obtained by the CSE are not scattered.
 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052 17

Table 16
Statistics of results by the CSE for each objective function for case 1 and case 2.
Test case Objective Best value Worst value Average value Standard deviation
F1 (X ) 4698 5258 4868.14 190.220
F2 (X ) 213 142 178.06 12.764
Case 1
F3 (X ) 3333 2545 2754 175.872
F4 (X ) 0.81 0.65 0.69 0.021
F1 (X ) 1.569E+07 1.892 E+07 1.660E+07 0.112 E+07
Case 2 F2 (X ) 22.7 18.3 19.667 1.063
F3 (X ) 25846 22184 24189.31 394.201

Table 17 there exist equal ranks, those ranks are split evenly. Let h′ be
The average value of each performance measure by the CSE, the NSGA-II and the the number of test problems whose difference is not equal to 0,
MOGA over 10 independent runs for a set of benchmark instances with different
number of facilities.
and T = min(R+ , R− ). The null hypothesis will be rejected if the
Instance Algorithm PR value GD value SP value OPS value
test statistic T is not greater than the value of the distribution of
Wilcoxon for h′ degrees of freedom. The normal approximation
O8 CSE 0.822 2.55 1.216 0.795
NSGA-II 0.788 3.71 1.931 0.743 for the Wilcoxon T statistics is used for getting p-value. We use
MOGA 0.708 3.84 1.985 0.714 the R software packages to compute the p-value in this study. The
Ba14 CSE 0.732 5.42 1.802 0.676
level of significance α is assigned as 0.05, which indicates that
NSGA-II 0.690 8.63 3.874 0.669 if the p-value is smaller than α , there is a significant difference
MOGA 0.633 8.67 4.051 0.639 between the two algorithms.
Ab20 CSE 0.810 13.11 4.118 0.775 In order to analyze the non-parametric statistical test of the
NSGA-II 0.790 14.32 4.471 0.761 performances of the proposed CSE algorithm and other two algo-
MOGA 0.714 14.97 5.579 0.529 rithms in the literature, NSGA-II is run more ten times for each of
SC30 CSE 0.697 16.41 6.028 0.615 instances Case 2, O8, Ba14, Ab20, SC30, SC35, and Du62, MOGA for
NSGA-II 0.632 17.63 6.357 0.547 each of instances Case 1, O8, Ba14, Ab20, SC30, SC35, and Du62,
MOGA 0.552 16.88 6.085 0.551 and CSE for each of all eight instances. Statistical results by the
SC35 CSE 0.616 16.25 5.131 0.521 Wilcoxon signed ranks test for CSE, NSGA-II and MOGA are given
NSGA-II 0.570 19.02 5.859 0.505 in Table 18, where CSE versus NSGA-II considers the average
MOGA 0.418 24.65 7.820 0.500 results of 20 independent runs for each of four performance
Du62 CSE 0.608 77.03 6.183 0.571 measures on Case 2, O8, Ba14, Ab20, SC30, SC35, and Du62, while
NSGA-II 0.590 86.28 7.307 0.530 CSE versus MOGA on Case 1, O8, Ba14, Ab20, SC30, SC35, and
MOGA 0.474 101.58 9.576 0.570
Du62. When using the Wilcoxon’s test in our experiments, the
first step is to compute the R+ and R− related to the comparisons
between CSE and the other two algorithms. Once they have been
For a set of benchmark instances with different number of obtained, their associated p-values can be computed by the R
facilities, Table 17 lists the average value of the each performance software packages. It is obvious that for every comparison, there
measure by the proposed CSE algorithm, NSGA-II and MOGA in is the property R+ +R− =h′ *(h′ +1)/2. As Table 18 states, the CSE
the literature over 10 independent runs for comparison. From algorithm shows a significant improvement over the NSGA-II and
Table 17 where the best results are shown in bold font, it is ob- MOGA algorithms, with a level of significance α = 0.05.
vious that Pareto-optimal solutions found by the CSE have better To summarize the experimental results of three groups of
convergence, uniformity and diversity than the results found by instances from the literature for the UA-FLPs, the proposed CSE
the NSGA-II and the MOGA. algorithm successfully optimizes four objectives including ma-
The details of the final solutions obtained by the CSE in a terial handling costs (quantitative aspect), closeness requests,
running for each of two typical cases and a set of benchmark distance requests and aspect ratio satisfactions of facilities (qual-
instances with different number of facilities are illustrated in itative aspects). The all properties (convergence and diversity)
Figs. 12 and 13, respectively, where all circles denote the gained of the CSE algorithm overmatch those of other algorithms in
feasible solutions and the asterisk points indicate the Pareto- the literature. Furthermore, the obtained Pareto-optimal layout
optimal solutions obtained. From these figures, one can see that set provides a wide range of choices for the decision makers
the final solutions by the CSE are well spread. who could express their preferences based on the production
Besides the above performance measures, the non-parametric requirements or market circumstances.
Wilcoxon signed ranks test [60] is also used to analyze the per-
formances of the proposed algorithm and the other algorithms 5.5. Numerical experiments for the CSE under different parameters
statistically. It is a pairwise test, which aims to detect a significant
difference between the behaviors of two algorithms. From the In addition, the contrast experiments are designed for testing
statistical point of view, the test is safer since it does not assume and verifying the effects of the parameters in the CSE. The testing
normal distributions of samples. Also, the outliers (exceptionally parameters in the CSE include the initial configuration number k1 ,
good/bad performances of a few problems) have less effect on the the objective axis weight wi , and the update frequency of dspace .
Wilcoxon test than on the t-test. The Wilcoxon signed ranks test In order to test the effects of parameters on the results of the
is described briefly in the following. experiments, we select some representative values to calculate
Assume that there are h test problems, let di denote the the range of each parameter. The initial configuration number k1
difference between the performance scores of the two algorithms is set to 20, 50, 100 and 200, respectively, while wi is fixed to
on the ith test problem. Those ranks of di = 0 are neglected. 10− 6 and the update frequency of dspace is fixed to 1/2 (dspace =
Then, the rest of those differences are ranked according to their dspace /2). Similarly, the objective axis weight wi is set to 10−5 ,
absolute value in ascending order. The sum of positive ranks 10− 6 , 10− 7 and 10− 8 , respectively, while k1 is fixed to 50 and the
and negative ranks are denoted as R+ and R− , respectively. If update frequency of dspace is fixed to 1/2. The update frequency of
18 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052

Fig. 12. Details of final solutions obtained by the CSE in a running for each of two typical cases. (a) Case 1. (b) Case 2.

Table 18
Statistical results by the Wilcoxon signed ranks test for CSE, NSGA-II and MOGA, where CSE versus NSGA-II considers the average results of 20 independent runs on
Case 2, O8, Ba14, Ab20, SC30, SC35, and Du62 for each of four performance measures, while CSE versus MOGA on Case 1, O8, Ba14, Ab20, SC30, SC35, and Du62.
Metric name PR-metric GD-metric SP-metric OPS-metric
CSE vs. R+ R− p-value R+ R− p-value R+ R− p-value R+ R− p-value
NSGA-II 28 0 0.01563 28 0 0.01563 27 1 0.03125 26 2 0.04688
MOGA 28 0 0.01563 28 0 0.01563 26 2 0.04688 28 0 0.01563

Table 19
Computational results obtained by the CSE algorithm with different initial configuration number k1 when wi is fixed to 10−6 and the update frequency of dspace is
fixed to 1/2 (dspace = dspace /2).
Instance k1 F1 (X ) (min) F2 (X ) (max) F3 (X ) (max) Time (s)
Best Avg SD Best Avg SD Best Avg SD
O8 20 223.39 237.45 7.38 103.18 83.11 12.04 52.11 40.05 11.39 429
50 221.28 232.07 6.01 108.73 86.74 11.43 52.11 42.63 10.61 731
100 221.28 233.91 7.02 108.73 85.39 11.81 52.11 42.61 10.72 1255
200 221.28 233.74 6.67 108.73 84.88 12.64 52.11 41.36 11.12 1807

SC30 20 5337.90 6272.57 511.92 112.73 97.40 15.42 462.97 396.64 38.20 5480
50 5295.77 6199.24 488.58 128.66 99.34 13.72 468.87 396.51 35.34 8316
100 5295.77 6198.06 493.84 128.66 95.55 14.36 468.87 393.18 35.81 8675
200 5295.77 6184.48 507.29 128.66 97.74 15.10 468.87 394.15 36.62 10082

Du62 20 3267980.66 3442095.68 82037.16 997.04 854.00 109.25 10864.46 9569.17 632.69 12400
50 3267707.61 3441469.94 77796.23 1044.02 872.85 106.25 10907.60 9605.36 620.36 15280
100 3267822.71 3441486.76 77852.29 1037.33 869.95 112.77 10874.03 9535.53 632.98 18965
200 3267831.73 3441535.20 78069.96 1042.58 861.64 107.32 10891.23 9592.36 626.85 21114

Table 20
Computational results obtained by the CSE algorithm with different objective axis weight wi when k1 is fixed to 50 and the update frequency of dspace is fixed to
1/2 (dspace = dspace /2).
Instance wi F1 (X ) (min) F2 (X ) (max) F3 (X ) (max) Time (s)
Best Avg SD Best Avg SD Best Avg SD
O8 10−5 221.28 235.58 6.81 108.73 85.12 12.62 52.11 40.89 11.85 722
10−6 221.28 232.07 6.01 108.73 86.74 11.43 52.11 42.63 10.61 731
10−7 221.28 233.39 6.53 108.73 86.41 11.68 52.11 41.43 11.34 728
10−8 221.28 231.42 6.27 108.73 85.94 11.44 52.11 42.01 11.26 735

SC30 10−5 5323.65 6208.17 491.39 126.45 97.10 14.16 463.68 384.20 39.38 7503
10−6 5295.77 6199.24 488.58 128.66 99.34 13.72 468.87 396.51 35.34 8316
10−7 5301.90 6196.29 486.15 127.05 97.67 15.01 465.37 393.38 37.58 9575
10−8 5362.67 6213.06 503.61 121.58 96.14 15.21 462.43 387.78 41.04 10506

Du62 10−5 3267860.41 3441538.85 77946.69 1021.90 867.96 109.60 10886.16 9538.91 622.56 15138
10−6 3267707.61 3441469.94 77796.23 1044.02 872.85 106.25 10907.60 9605.36 620.36 15280
10−7 3267810.91 3441470.23 77801.60 1030.13 867.70 119.14 10896.63 9524.73 631.69 16486
10−8 3267831.73 3441511.49 78963.05 1028.75 870.30 125.96 10821.23 9584.11 631.98 17782
 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052 19

Fig. 13. Details of final solutions obtained by the CSE algorithm in a running for each of a set of benchmark instances with different number of facilities. (a) Instance
O8. (b) Instance Ba14. (c) Instance Ab20. (d) Instance SC30. (e) Instance SC35. (f) Instance Du62.

dspace is set to 2/3, 1/2, 1/3 and 1/5, respectively, while k1 is fixed parameters k1 , wi and the update frequency of dspace are listed in
to 50 and wi is fixed to 10− 6 . The maximum number of iterations Table 19, Table 20 and Table 21, respectively.
is set as 3000 to ensure that the algorithm reaches convergence Table 19 shows that for instances O8, SC30 and Du62, the
and the optimal solution of each objective can be obtained. The best values, the average values, the standard deviations for all
algorithm is run 10 times independently for each parameter of objectives are slightly different when the number k1 of initial
three representative instances O8, SC30 and Du62. The best val- configurations is set to 50, 100 and 200, respectively. But with
ues (Best), the average values (Avg), the standard deviations (SD) the increase of k1 , the average running time will also increase,
and the average of running time of the algorithm with different especially for instances SC30 and Du62. While k1 is set to 20,
20 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052

Table 21
Computational results obtained by the CSE algorithm with different the update frequency of dspace when k1 is fixed to 50 and wi is fixed to 10−6 .
Instance dspace F1 (X ) (min) F2 (X ) (max) F3 (X ) (max) Time (s)
Best Avg SD Best Avg SD Best Avg SD
O8 2/3 221.28 232.57 7.04 108.73 87.19 11.88 52.11 41.31 11.17 789
1/2 221.28 232.07 6.01 108.73 86.74 11.43 52.11 42.63 10.61 731
1/3 221.28 233.14 6.96 108.73 84.27 12.35 52.11 43.07 11.06 728
1/5 221.28 237.50 7.82 108.73 81.03 12.69 52.11 40.92 13.26 722

SC30 2/3 5295.77 6225.01 561.58 128.66 97.45 19.00 468.87 380.00 38.02 10778
1/2 5295.77 6199.24 488.58 128.66 99.34 13.72 468.87 396.51 35.34 8316
1/3 5313.36 6262.55 532.15 127.45 97.83 18.86 461.78 396.45 36.15 8037
1/5 5382.34 6214.33 698.05 114.88 93.06 14.91 452.36 397.94 46.89 7337

Du62 2/3 3267822.71 3441561.10 82820.41 1036.88 877.04 137.05 10783.42 9547.41 637.01 17308
1/2 3267707.61 3441469.94 77796.23 1044.02 872.85 106.25 10907.60 9605.36 620.36 15280
1/3 3267884.56 3441632.85 79731.53 1027.59 849.19 117.44 10880.58 9578.45 651.62 14709
1/5 3268761.69 3441668.66 93069.13 1033.91 834.34 162.30 10708.27 9594.75 702.05 14157

Table A.1 Table A.3

Closeness relation and distance requirements between two facilities for instance Closeness relation and distance requirements between two facilities for instance
O8. Ab20.
Facility 1 2 4 5 6 7 8 Facility 2 3 4 5 6 7 8 9 10 13 14 18 19
Closeness relation Closeness relation
1 5 6 2 2 5 3
2 4 3 1 3 3 4
5 3 4 3
Distance requirements 7 5 2
8 3
6 4 10 4
Distance requirements
Table A.2 5 2 1 3 2
Closeness relation and distance requirements between two facilities for instance 6 2 1
Ba14. 9 1 1
Facility 1 2 3 4 5 7 8 10 12 13 14 13 2

Closeness relation
1 3 5 4 4 Table A.4
2 3 3 2 Closeness relation and distance requirements between two facilities for instance
3 3 4 1 SC30.
5 2 Facility 1 2 3 4 7 8 9 10 11 12 13 14 15 17 18 25 29 30
Distance requirements Closeness relation
7 2 1 3 6
10 1 3 4 4 5 5
12 3 8 3
13 3 9 4
10 4
12 5
15 5
the results are worse than results of other values. For the evo- 25 2
lutionary algorithm, the production of the initial configurations 29 5

is important, because the subsequent crossover and mutation are Distance requirements
all based on the initial configurations for evolution and updation. 1 4 1 3 2 1 2
When k1 is set to 20, it is easy to result in the premature con- 2 1 2 3 1 1
7 1 1 3 2
vergence of the algorithm and the probability of obtaining the 11 2 1 3
optimal solution is lower because of the smaller number of initial 13 3 2
configurations. When k1 is 100 or 200, although the diversity of 14 1
initial configurations is increased the results do not turn better,
furthermore the running time is greatly increased. Therefore, on
the whole, when k1 is set to 50, the results of the algorithm are accuracy increases gradually, the running time of the algorithm
slightly better than those of other k1 values. is longer accordingly.
As can be seen from Table 20, the change of objective axis From Table 21, for the small instance O8, when the update fre-
weight wi has little influence on the results of instances O8, quency of dspace is 2/3, 1/2, 1/3 and 1/5, the best values, the
SC30 and Du62. But, on the whole, when wi is set to 10− 6 , the average values, the standard deviations of the three objectives
results of the algorithm are slightly better than those of the and the average running time are all similar. For two larger
algorithm with other wi values. For instance O8, the optimal instances SC30 and Du62, when the update frequency of dspace
solution of each objective can be obtained under the four values is 2/3, 1/2 and 1/3, the optimal values of the three objectives
of wi . For instances SC30 and Du62, one can see that although obtained are similar. However, when the update frequency of
the best values, the average values, and the standard deviations dspace is 2/3, the running time of the algorithm is relatively longer.
for all objectives of these two instances have little difference for When it becomes 1/5, the results of the algorithm for instances
different parameters of wi , as the decrease of wi , because the SC30 and Du62 are slightly worse than those by the algorithm
 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052 21

Table A.6
Closeness relation and distance requirements between two facilities for instance Du62.
Facility 4 5 13 16 18 19 20 21 22 29 31 34 35 37 38 50 53 56 57 59 60 61 62
Closeness relation
4 5 5 4 3 3 2 3 2 4 1
5 3 2 2 1 3 5 4 3 5
13 4 5 5 5 2 3 4
16 5 3 6 2 4 2 4 4 4 3 3
18 5 4 3 4 4 6 5 5
20 5 3 2 1 6 1 4 2 1
21 6 1 2 2 5 2 3 5
22 1 2 4 4 3 4 3
29 6 1 4 3 5 2
37 3 4 2 5
38 1 3 5
50 1 2 4
56 3 6
59 5
Distance requirements
19 4 3 5 6 2 4
31 5 4 4 5 3 1
34 6 4 4 5 5
35 4
53 5 4 4
57 6
60 5

Table A.5 But, in real-world situations, the layout with unequal-area fa-
Closeness relation and distance requirements between two facilities for instance cilities is more practical because of the geometric constraints
SC35.
of the equal-area ones. In addition, most of the existing ap-
Facility 1 3 4 7 9 13 15 18 23 24 25 26 27 28 29 32 33
proaches study either the quantitative single objective FLPs or
Closeness relation multi-objective FLPs which are converted into single objective
1 1 2 1 1 problem by using weighted coefficients. In comparison, this paper
4 4 5
involves the investigation of static multi-objective UA-FLPs based
15 5 4
24 2 on Pareto optimization method. Three sets of multi-objective
25 3 layout instances whose objectives refer to material handling costs
26 1 (quantitative aspect), closeness requests, distance requests and
27 2 satisfaction of aspect ratio of the facility (qualitative aspects) are
28 2
introduced.
29 5
For multi-objective UA-FLPs, the topic of MOO algorithms
Distance requirements
approached by GAs (such as MOGA, NSGA-II) is nowadays one of
3 2 3 1 2 1 the most promising and widely investigated research field. These
7 1 2 1 1
9 1 4
approaches are capable of finding a set of Pareto-optimal layouts
13 2 3 that optimizes the multiple objective functions simultaneously
throughout the entire evolutionary process, giving the decision
maker a restricted number of solutions among which one can
choose that he considers the best. However, the convergence
with other three update frequencies. When the update frequency and diversity of traditional MOO algorithms approached by GAs
of dspace is 2/3, the solution space shrinks slowly, and in turn largely depend upon the genetic operators. They have generally
the algorithm converges slowly so that the running time of the shortcomings, such as slow convergence to the Pareto front and
algorithm is longer. On the contrary, when the update frequency low efficient selection toward diversity of solutions, especially
of dspace is 1/5, the solution space shrinks quickly without full when dealing with the UA-FLPs with three or more objectives. In
update so that the algorithm is easy to premature converges, this paper, a novel MOO algorithm based on GA(EA), called con-
causing the running time becomes shorter and the results become figuration space evolutionary (CSE) algorithm which can search
worse accordingly. the solution space efficiently and find good spread of solutions, is
The experimental results of test instances illustrate the ef- developed to solve the static UA-FLPs with multiple objectives.
fectiveness of the settings of parameters in the CSE. In order In the CSE, we introduce a measure of the radius of the con-
to make the algorithm find the Pareto-optimal solutions effi- figuration bank, the value of which is gradually reduced until
ciently, we generally set the parameters in a certain range, for the algorithm attains the convergence. Considering the special
example, initial configuration number k1 ∈ [50, 200], the ob- characteristics of the problem, new encoding, selection, crossover,
jective axis weight wi ∈ [10−5 , 10−7 ], the update frequency of and mutation operators are used to produce new individuals in
dspace ∈ [1/3, 2/3]. In this paper, the parameters k1 , wi and the the course of evolution. Also, to get a good spread to Pareto
update frequency of dspace are set to 50, 10−6 and 1/2, respectively. front and maintain diversity of the obtained solutions, we adopt
a method of the nearest and farthest candidate solution based
6. Conclusions on objective function normalization and combine it with the fast
non-dominated sorting to choose the Pareto-optimal solutions.
Reasonable facilities layout of enterprises can effectively im- The experimental results on eight typical cases show that the
prove operational efficiency and reduce operating costs. Nowa- proposed CSE algorithm is an effective MOO algorithm for solving
days, most of classical methods focus on the equal-area FLPs. the UA-FLPs.
22 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052

From the executing process of the CSE, it is not hard to see [12] L. García-Hernández, J.M. Palomo-Romero, L. Salas-Morera, A. Arauzo-
that the proposed CSE algorithm is easy to be extended to inves- Azofra, H. Pierreval, A novel hybrid evolutionary approach for capturing
decision maker knowledge into the unequal area facility layout problem,
tigate other multi-objective or even many-objective optimization
Expert Syst. Appl. 42 (10) (2015) 4697–4708.
problems in various real-world scenarios. For further work, this [13] J.M. Palomo-Romero, L. Salas-Morera, L. García-Hernández, An island model
study can be expanded in many ways. It can be extended to genetic algorithm for unequal area facility layout problems, Expert Syst.
add the input/output points for facilities. It is also suggested Appl. 68 (2017) 151–162.
that other MOO algorithms can be utilized to solve the same [14] M.S. Chang, T.C. Ku, A slicing tree representation and QCP-model-based
heuristic algorithm for the unequal-area block facility layout problem,
problems. Furthermore, in today’s market, with increasing global
Math. Probl. Eng. 2013 (4) (2013) 1–19.
competition and short life cycle of production, material flows [15] S. Kang, J. Chae, Harmony search for the layout design of an unequal area
among facilities change during the planning horizon. Hence, fu- facility, Expert Syst. Appl. 79 (2017) 268–281.
ture research should be directed toward dynamic FLPs rather than [16] K.Y. Komarudin, Applying ant system for solving unequal area facility
static FLPs. Along this research line, the idea of the proposed layout problems, European J. Oper. Res. 202 (3) (2010) 730–746.
[17] J. Guan, G. Lin, Hybridizing variable neighborhood search with ant colony
algorithm will be applied to solve the multi-objective or even
optimization for solving the single row facility layout problem, European
many-objective unequal-area dynamic FLPs with variable shapes J. Oper. Res. 248 (3) (2016) 899–909.
and areas of facilities throughout the time horizon. [18] D. Scholz, A. Petrick, W. Domschke, Stats: A slicing tree and tabu search
based heuristic for the unequal area facility layout problem, European J.
Declaration of competing interest Oper. Res. 197 (1) (2009) 166–178.
[19] S. Kulturel-Konak, A. Konak, A large-scale hybrid simulated annealing
algorithm for cyclic facility layout problems, Eng. Optim. 47 (7) (2015)
No author associated with this paper has disclosed any po- 963–978.
tential or pertinent conflicts which may be perceived to have [20] I.B. Hunagund, V.M. Pillai, U.N. Kempaiah, A simulated annealing algorithm
impending conflict with this work. For full disclosure statements for unequal area dynamic facility layout problems with flexible bay
refer to https://doi.org/10.1016/j.asoc.2019.106052. structure, Int. J. Ind. Eng. Comput. 9 (3) (2018) 307–330.
[21] A.D. Asl, K.Y. Wong, Solving unequal-area static and dynamic facility layout
problems using modified particle swarm optimization, J. Intell. Manuf. 28
Acknowledgments (2015) 1317–1336.
[22] H.L. Wei, N.A.M. Isa, Bidirectional teaching and peer-learning particle
This work is supported by the Natural Science Foundation of swarm optimization, Inform. Sci. 280 (4) (2014) 111–134.
Jiangsu Province (Grant No. BK20181409), the Major Program [23] Y. Luo, Y.P. Waden, The improved ant colony optimization algorithm for
of the National Social Science Foundation of China (Grant No. MLP considering the advantage from relationship, Math. Probl. Eng. 2017
(2017) 1–11.
16ZDA047), the Special Foundation of Guangzhou Key Labo- [24] Z.H. Hu, A multi-objective immune algorithm based on a multiple-affinity
ratory of Multilingual Intelligent Processing, China (Grant No. model, European J. Oper. Res. 202 (1) (2010) 60–72.
201905010008), and the Program of Basic and Applied Basic [25] Y.N. Guo, P. Zhang, J. Cheng, C. Wang, D.W. Gong, Interval multi-objective
Research of Guangzhou, China. quantum-inspired cultural algorithms, Neural Comput. Appl. 30 (3) (2018)
709–722.
[26] J. Cheng, J.J. Chen, Y.N. Guo, S. Cheng, L.K. Yang, P. Zhang, Adaptive
Appendix
CCR-ELM with variable-length brain storm optimization algorithm for
class-imbalance learning, Nat. Comput. (2019) http://dx.doi.org/10.1007/
See Tables A.1–A.6. s11047-019-09735-9.
[27] C.M. Fonseca, P.J. Fleming, An overview of evolutionary algorithms in
References multi-objective optimization, Evol. Comput. 3 (1) (2014) 1–16.
[28] R. Wang, P.J. Fleming, R.C. Purshouse, General framework for localised
multi-objective evolutionary algorithms, Inform. Sci. 258 (3) (2014) 29–53.
[1] B. Marcello, Z. Simone, Z. Lucio, Layout design in dynamic environ-
[29] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multi-
ments: Strategies and quantitative indices, Int. J. Prod. Res. 41 (5) (2003)
objective genetic algorithm: NSGA-II, IEEE Trans. Evol. Comput. 6 (2) (2002)
995–1016.
[2] H. Pourvaziri, B. Naderi, A hybrid multi-population genetic algorithm for 182–197.
the dynamic facility layout problem, Appl. Soft Comput. 24 (24) (2014) [30] R. Wang, R.C. Purshouse, P.J. Fleming, Preference-inspired co-evolutionary
457–469. algorithms for many-objective optimization, IEEE Trans. Evol. Comput. 17
[3] J.A. Tompkins, J.A. White, Y.A. Bozer, E.H. Frazelle, J.M.A. Tanchoco, J. (4) (2013) 474–494.
Trevino, Facilities Planning, Wiley, 1996. [31] R. Wang, M.M. Mansor, R.C. Purshouse, P.J. Fleming, An analysis of
[4] J.F. Liu, D.W. Wang, K. He, Y. Xue, Combining Wang-Landau sampling parameter sensitivities of preference-inspired co-evolutionary algorithms,
algorithm and heuristics for solving the unequal-area dynamic facility Internat. J. Systems Sci. 46 (13) (2015) 2407–2420.
layout problem, European J. Oper. Res. 262 (1) (2017) 1052–1063. [32] S. Yang, M. Li, X. Liu, J. Zheng, A grid-based evolutionary algorithm for
[5] K.S.N. Ripon, K. Glette, O. Mirmotahari, M. Hovin, J. Torresen, Pareto many-objective optimization, IEEE Trans. Evol. Comput. 17 (5) (2013)
optimal based evolutionary approach for solving multi-objective facility 721–736.
layout problem, in: Proceedings of the 16th International Conference [33] M. Li, S. Yang, X. Liu, Shift-based density estimation for pareto-based
on Neural Information Processing, ICONIP, Berlin, Heidelberg, 2009, pp. algorithms in many-objective optimization, IEEE Trans. Evol. Comput. 18
159–168. (3) (2014) 348–365.
[6] J.F. Liu, H.Y. Zhang, K. He, S.Y. Jiang, Multi-objective particle swarm opti- [34] J. Bader, E. Zitzler, Hype: An algorithm for fast hypervolume-based
mization algorithm based on objective space division for the unequal-area many-objective optimization, Evol. Comput. 19 (1) (2014) 45–76.
facility layout problem, Expert Syst. Appl. 102 (2018) 179–192. [35] T.C. Wang, R.T. Liaw, C.K. Ting, MOEA/D using covariance matrix adaptation
[7] A. Gómez, Q.I. Fernández, D.D.L.F. Garcıa, P.J. Garcıa, Using genetic algo- evolution strategy for complex multi-objective optimization problems, in:
rithms to resolve layout problems in facilities where there are aisles, Int. Proceedings of 2016 IEEE Congress on Evolutionary Computation, CEC,
J. Prod. Econ. 84 (3) (2003) 271–282. IEEE, Vancouver, BC, Canada, 2016;, pp. 983–990.
[8] Q. Liu, R.D. Meller, A sequence-pair representation and MIP-model-based [36] K. Deb, H. Jain, An evolutionary many-objective optimization algorithm us-
heuristic for the facility layout problem with rectangular departments, IIE ing reference-point-based non-dominated sorting approach, part I: Solving
Trans. 39 (4) (2007) 377–394. problems with box constraints, IEEE Trans. Evol. Comput. 18 (4) (2014)
[9] J.F. Gonçalves, M.G.C. Resende, A biased random-key genetic algorithm for 577–601.
the unequal area facility layout problem, European J. Oper. Res. 246 (1) [37] J.F. Liu, J. Liu, Applying multi-objective ant colony optimization algorithm
(2015) 86–107. for solving the unequal area facility layout problems, Appl. Soft Comput.
[10] F.G. Paes, A.A. Pessoa, T. Vidal, A hybrid genetic algorithm with decom- 74 (2019) 167–189.
position phases for the unequal area facility layout problem, European J. [38] X.Q. Zuo, C.S. Murray, A.E. Smith, Solving an extended double row layout
Oper. Res. 256 (3) (2017) 742–756. problem using multiobjective tabu search and linear programming, IEEE
[11] B.H. Ulutas, S. Kulturel-Konak, An artificial immune system based algo- Trans. Autom. Sci. Eng. 11 (2014) 1122–1132.
rithm to solve unequal area facility layout problem, Expert Syst. Appl. 39 [39] A. Saraswat, U. Venkatadri, I. Castillo, A framework for multi-objective
(5) (2012) 5384–5395. facility layout design, Comput. Ind. Eng. 90 (2015) 167–176.
 J. Liu, S. Liu, Z. Liu et al. / Applied Soft Computing Journal 89 (2020) 106052 23

[40] G. Aiello, G.L. Scalia, M. Enea, A multi objective genetic algorithm for the [49] H. Jain, K. Deb, An evolutionary many-objective optimization algorithm
facility layout problem based upon slicing structure encoding, Expert Syst. using reference-point based non-dominated sorting approach, Part II:
Appl. 39 (12) (2012) 10352–10358. Handling constraints and extending to an adaptive approach, IEEE Trans.
[41] A.P. Li, Z.Y. Yan, N. Xie, J.Z. Huang, A research of multi-objective facility Evol. Comput. 18 (4) (2014) 602–622.
layout based on NSGA-? Mach. Des. Research 28 (3) (2012) 90–95. [50] G. Aiello, G.L. Scalia, M. Enea, A non dominated ranking multi objective
[42] K.S.N. Ripon, K. Glette, K.N. Khan, M. Hovin, J. Torresen, Adaptive variable genetic algorithm and electre method for unequal area facility layout
neighborhood search for solving multi-objective facility layout problems problems, Expert Syst. Appl. 40 (12) (2013) 4812–4819.
with unequal area facilities, Swarm Evol. Comput. 8 (1) (2013) 1–12. [51] Z. Chang, J. Lu, Analysis of multi-objective facility layout problem using
[43] L. García-Hernández, M. Pérez-Ortiz, A. Araúzo-Azofra, L. Salas-Morera, C. flexible bay structure, Oper. Res. Manage. Sci. 24 (2) (2015) 128–134.
Hervás-Martínez, An evolutionary neural system for incorporating expert [52] R.D. Meller, V. Narayanan, P.H. Vance, Optimal facility layout design, Oper.
knowledge into the UA-FLP, Neurocomputing 135 (2014) 69–78. Res. Lett. 23 (3–5) (1998) 117–127.
[44] S. Vitayasak, P. Pongcharoen, C. Hicks, A tool for solving stochastic dynamic [53] D.J. Van Camp, A Nonlinear Optimization Approach for Solving Facility
facility layout problems with stochastic demand using either a genetic Layout Problem (thesis), University of Toronto, Canada, 1989.
algorithm or modified backtracking search algorithm, Int. J. Prod. Econ. [54] G.C. Armour, E.S. Buffa, A heuristic algorithm and simulation approach to
190 (2017) 146–157. relative allocation of facilities, Manage. Sci. 9 (2) (1963) 294–309.
[45] T. Irohara, H. Yamashita, Y. Ishizuka, Facility layout problem with buffer [55] T. Dunker, G. Radons, E. Westkämper, A coevolutionary algorithm for a
space allocation for throughput and material handling cost, J. Japan Ind. facility layout problem, Int. J. Prod. Res. 41 (15) (2003) 3479–3500.
Manage. Assoc. 58 (2) (2017) 87–96. [56] Y. Collette, P. Siarry, Three new metrics to measure the convergence of
[46] C. Giannoulis, A. Ishizaka, A web-based decision support system with metaheuristics towards the Pareto frontier and the aesthetic of a set
electre III for a personalised ranking of British universities, Decis. Support of solutions in bi-objective optimization, Comput. Oper. Res. 32 (2005)
Syst. 48 (3) (2010) 488–497. 773–792.
[47] S. Kukkonen, K. Deb, Improved pruning of non-dominated solutions based [57] D.A. Van, V. Gary, B. Lamont, Multi-objective evolutionary algorithm
on crowding distance for bi-objective optimization problems, in: Pro- research: A history and analysis, Evolutionary Computation 8 (2) (1998)
ceedings of 2006 IEEE Congress on Evolutionary Computation, 2006, pp. 125–147.
1179–1186. [58] J.R. Schott, Fault tolerant design using single and multi-criteria genetic
[48] S. Kukkonen, K. Deb, A fast and effective method for pruning of algorithm optimization, Cell. Immunol. 37 (1) (1995) 1–13.
non-dominated solutions in many-objective problems, in: Proceedings [59] K. Zheng, R.-J. Yang, H.Y. Xu, J. Hu, A new distribution metric for comparing
of International Conference on Parallel Problem Solving from Nature, pareto optimal solutions, Struct. Multidiscip. Optim. 55 (2017) 53–62.
Springer-Verlag, 2006, pp. 553–562. [60] D.J. Sheskin, Handbook of Parametric and Nonparametric Statistical
Procedures, third ed., Chapman & Hall/CRC, 2003, pp. 609–632.
[61] M. Makowski, Multi-objective decision support including sensitivity
analysis, Encycl. Life Support Syst. (2004) 1–24.