You are on page 1of 10

Neurocomputing 135 (2014) 69–78

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

An evolutionary neural system for incorporating expert knowledge


into the UA-FLP$
L. García-Hernández a,n, M. Pérez-Ortiz b, A. Araúzo-Azofra a,
L. Salas-Morera a, C. Hervás-Martínez b
a
Area of Project Engineering, University of Córdoba, Rabanales Campus, Leonardo Da Vinci Building, 14071 Córdoba, Spain
b
Department of Computer Science and Numerical Analysis, University of Córdoba, Spain

art ic l e i nf o a b s t r a c t

Article history: This paper presents a hybrid system for incorporating human expert knowledge into the unequal area
Received 23 October 2012 facility layout problem. A subset of facility designs is generated using a genetic algorithm and then
Received in revised form evaluated by a human expert. The hybrid system consists of assigning a mark, where the principal aim
16 January 2013
is to substitute the human expert0 s knowledge to avoid fatiguing or burdening him or her. The novel
Accepted 19 January 2013
Available online 4 January 2014
proposed approach was tested using a real case study of 365 facility layout designs for an ovine
slaughterhouse. The validation phase of the intelligent model presented was performed using a new
Keywords: subset of 181 facility layout designs evaluated by a different human expert. The results of the experiment,
Evolutionary Computation which validate the proposed approach, are presented and discussed in this study.
Artificial neural networks
& 2014 Elsevier B.V. All rights reserved.
Unequal area facility layout problem
Genetic algorithm
Heuristic search

1. Introduction on heuristic methods, such as tabu search [26,33], simulated


annealing [11], the ant system [20,39] and genetic algorithms
Plant layout designs are extremely important for production (GAs) [15]. The latter methods have been widely used because this
efficiency [21] to minimize material flow between facilities and type of problem does not require mathematical modeling instead
also to satisfy the logical relationships between them. It is only a fitness function is required to account for the limitation fact,
estimated that between 20% and 50% of production costs can be the success of these algorithms partially depends on the evalua-
attributed to material handling, and it is generally accepted that tion function employed [27]; the coding chosen for the set of
such costs can be reduced by between 10% and 30% through possible solutions (or individuals) and the selection and crossover
efficient design [37]. Different facility layout design problems operators used are also important. GAs can be applied successfully
involve different issues [12], which depend on the manufacturing to problems in which the search space is large or not well
plant characteristics (e.g., planning horizon, facility shape, number understood or the evaluation function is complex (noisy) and if
of floors, and aisle and elevator considerations, among others). the problem does not require a global optimum but rather only a
These problems may be modeled using linear integer program- sufficiently good solution to be determined, as is the case for
ming, mixed integer programming and graph theoretic methods facility layout design [28].
[18]. However, because the number of facilities limits the applica- In practice, layout designs are usually based on the available
tion of optimal methods, suboptimal methods have been devel- area to which the layout must be adapted; the prior data used are
oped to address more complex problems. Different techniques the plant surface area and shape, which is normally rectangular.
have been tested, such as the branch and bound method [34] This problem can be modeled as an unequal area facility layout
or the graph theory [19]. Recently, researchers have focused problem (UA-FLP) [5] in which the available rectangular spaces are
assigned to facilities with unequal areas to minimize a cost or
target function. Several authors have studied the UA-FLP using

This paper was invited to be included in the “Special Issue GAs. Tam [35] used a GA with a tree structure to represent a layout,
Neurocomputing-ISDA2011”.
n
and in the evaluation function, considered the material flow
Corresponding author. Tel.: þ 34 957 212 050; fax: þ34 957 218 550.
between each pair of areas. Similarly, Wu and Appelton [40]
E-mail addresses: ir1gahel@uco.es (L. García-Hernández),
i82perom@uco.es (M. Pérez-Ortiz), arauzo@uco.es (A. Araúzo-Azofra), proposed a GA for optimizing space layouts that consider material
lsalas@uco.es (L. Salas-Morera), chervas@uco.es (C. Hervás-Martínez). flows and isolated spaces between different work areas. Lee and

0925-2312/$ - see front matter & 2014 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.neucom.2013.01.068
70 L. García-Hernández et al. / Neurocomputing 135 (2014) 69–78

Lee [23] also considered the material flow and free space between perspective of optimization. This is the case of [38], where a
activity areas. The design proposed by Lee et al. [22] included the near-optimum parallel algorithm is proposed for solving the
possibility of considering pre-existing walls or aisles. Gomez et al. quadratic assignment problem (QAP), by using two-dimensional
[16] used Muther0 s SLP method [29] and considered the material maximum neural networks and the objective of minimizing the
handling flow, closeness and distance ratios required in produc- material flow between facilities. In this case, the use of artificial
tion processes. Aiello et al. [2] proposed a method that combines neural networks provides a gradient descent methodology for
obtaining of a set of solutions using a GA and multi-criterion minimizing the material flow. However, as the authors argue,
selection using the Electre method. These authors considered the the optimization of multiple criteria (as the ones considered in this
costs of material flows between work centres, the shapes of the paper) could be difficult to achieve under this framework, due to
assigned work areas, and an adjacency function, which depends the complexity of both computing the cost function and satisfying
on the plant0 s requirements. Recently, they applied the algorithm the constraints. This is also the case of [7] where the optimization
using a slicing tree representation [3]. of ill-structured facility layouts is considered by means of a neuro
Most previous research has addressed this problem using quan- rule-based expert system, in order to deal with the noisy, incom-
titative performance criteria (e.g., material handling costs, closeness plete and uncertain information inherent to the facility layout
or distance relationships, adjacency requirements and aspect ratios), design. In this case, more factors are considered for the optimiza-
which are used in an optimization approach using some heuristic tion process, however, as said before, it can be considered as
methodology. These approaches, however, may not adequately a different philosophy to the one exposed in this paper, since in
represent all the relevant information that human experts involved this case the uncertainty is treated by the aid of the expert
in the design (e.g., engineers) would consider [6]. Consequently, the evaluation.
participation of a human expert (or experts) in the design process is The remainder of the paper is structured as follows. Section 2
essential to incorporate his or her experience and knowledge in the analyses and presents the proposed layout representation for
final layout design. Nevertheless, the evaluation of facility layouts assessing the UA-FLP, and formally presents the primary charac-
can be a complex and non-trivial process that can fatigue and teristics of the evolutionary artificial neural network used in
burden the human expert, especially when the number of facilities is the method. Section 3 presents the experiments performed and
high and there are complex relationships between them that must discusses the results obtained. Finally, Section 4 discusses with the
be satisfied. A promising approach to avoid this problem is to primary conclusions and future directions suggested by this study.
incorporate an intelligent system [1,10] that can learn from the
expert0 s knowledge and then replace the expert by the previously
trained predictive model, to evaluate a concrete layout design. To the 2. Methodology
best of our knowledge, no such approach for this type of problem
has been presented in the UA-FLP literature. In this section, the primary aspects of the proposal are explained.
Artificial neural networks (ANNs) are statistical-computational The proposed system requires the expert designer to evaluate
models [9] that use a highly interconnected network structure several facility layout designs. To do so, a three-stage system has
consisting of simple processing elements (nodes). These nodes are been developed (illustrated in Fig. 1).
capable of performing massively parallel computation for data
processing. Unlike other classification techniques, ANNs do not
impose any restriction on the type of relationship that governs the
dependence of the parameters on the running conditions. ANN
models were inspired by the physical structure and mechanisms
of the nervous system and the biological cognition and learning
process. Although ANNs are based on the functionality of the
nervous system, the cornerstone of an ANN is its structure.
In classification problems, ANNs, primarily multilayer feedfor-
ward networks, are frequently used. Typically, the network con-
sists of a set of sensory units (or source nodes), where each node is
associated with a variable or characteristic of the problem that
constitutes the input layer, one or more hidden layers of computa-
tion nodes, and an output layer of computation nodes, each of
which is associated with the evaluation of a given facility layout.
The input signal propagates through the network in the forward
direction on a layer-by-layer basis.
In this contribution, an intelligent system for addressing
the UA-FLP is proposed. This system allows to incorporate expert
knowledge in the final design (to evaluate a final solution using
the expert0 s experience). A system that includes a GA and an ANN
is employed. The GA is used to generate a number of solutions that
are obtained using an optimization process; then, these solutions
are evaluated by a human expert designer. To avoid overburdening
the human expert, an ANN system is applied in the second stage
for learning the expert knowledge and predicting the mark that he
or she would assign to each solution. This system enables one to
eliminate the human expert from the process after the ANN is
trained.
Although some previous works using neural networks for
solving a facility layout problem can be found in the literature,
they all assessed the problem from the most pure and strict Fig. 1. Graphical representation of the methodology.
L. García-Hernández et al. / Neurocomputing 135 (2014) 69–78 71

1. First, a GA that optimizes material flow, adjacency require- Table 2


ments and aspect ratios (concepts described in the next An individual codification example.
subsection) is used for obtaining the set of solutions to take
G H B D F C E A
under consideration.
2. Then, in order to capture the expert knowledge, the whole set 1 1 0 0 0 1 1 0
of facility layouts is evaluated according to the three factors 1 3 0 1 2 0 3 1
considered for the problem and the entire distribution. The
plant layouts are evaluated using a five-granularity Likert scale
(where the possible evaluations for a facility layout are 1:
strongly disagree, 2: disagree, 3: neutral, 4: agree, and 5: strongly
agree; and the expert answers the following question: Could
this plant layout be considered as a good solution for this
unequal-area facility layout problem?).
3. Finally, the data obtained are used for training an ANN, which
could act as a surrogate for the expert evaluation. The model
that performs best is used as a first validation for the metho-
dology to account for the evaluation of other human experts.
The primary objective of this experiment is to test the robust-
ness of a single human expert0 s opinion.

Fig. 2. A facility layout example.


The stages described previously are illustrated using Fig. 1.

2.1.1. Individual codification


2.1. Generating solutions To encode the possible GA solutions, the following representa-
tion structure has been proposed [32]. It uses a three-row matrix
The UA-FLP [5] considers a rectangular plant with fixed with n columns, as shown in Table 2.
dimensions (W  H) and a set of facilities, each with a required The first coding row indicates the order in which the facilities
area (Ai), where the sum of the facility areas must be less than or are arranged on the available surface (GHBDFCEA). The second row
equal to the plant area: contains binary values, where a group of equal values indicates a
group of facilities that are laid out together. In the example given in
n
∑ Ai o ¼ W  H ð1Þ Table 2, the first section (GH) is grouped because there are two
i 1 values at the beginning. Next are three zeros, which correspond to
the next three facilities (BDF). Then, (CE) and (A) are the other groups
The aim is to allocate the facilities in the plant to optimize a formed. The third row corresponds to the orientation: 0 indicates left-
given criterion (for example, material handling costs, adjacency vertical, 1 indicates right-vertical, 2 indicates bottom-horizontal, and
requirements, distance requests or desired aspect ratios) subject to 3 indicates top-horizontal. Thus, the matrix given in Table 2 repre-
the restriction that the facilities cannot overlap. sents a layout like the one shown in Fig. 2.
To solve the UA-FLP, a GA [32] has been designed according to If the available area is larger than necessary, the excess space
the following specifications: could remain in the centre or on any of the edges.

 It is based on n facilities with known surface areas Ai distrib-


uted in a rectangular area with dimensions H  W (height and 2.1.2. Factors to consider
width). The GA considers three aspects: material handling cost, com-
 The necessary closeness or distance ratios for each pair of pliance with closeness and distance ratios expressed in Table 1,
facilities are indicated in a relation matrix and expressed with and the rectangular shape assigned to each facility (aspect ratio).
the codes given in Table 1. These aspects are as follows:
 The density of material flows (for example in kg of material or
number of pieces) between each pair of facilities is also known. (a) Material handling cost: This aspect reflects the total material
This matrix is not required to be symmetric. handling cost for the plant layout. To calculate it, the following
equation is used:

In the definition contained in Table 1, the causes of closeness or M F ¼ ∑∑Dij  f ij  cij ; ð2Þ
distance requests may be, for example, the logical organization of i j

the production system, the absence or presence of noise or safety


reasons. Users may voluntarily or inadvertently assign the close- where Dij represents the distance between the centroids of the
ness requests expressed in Table 1 according to the flow of facilities measured using the Manhattan distance, fij represents
materials, which would be redundant, but this is not necessarily the material flow from facility i to facility j, and cij represents
problematic for generating layouts. the cost to move a unit of material from facility i to facility j.
(b) Closeness and distance ratios: This requirement corresponds
Table 1 to the closeness and separation requirements that can exist
Closeness or distance ratios. between a facility and others, e.g., between a noisy facility
and the offices. The following expression is used to evaluate
Symbol: closeness/distance ratios
compliance with the logical ratios between facilities:
A: essential closeness E: very important closeness I: important closeness
O: normal closeness U: indifferent closeness X: undesirable closeness Ad ¼ ∑∑Eij ; ð3Þ
i j
72 L. García-Hernández et al. / Neurocomputing 135 (2014) 69–78

where
8
>
> RELij  D2ij if RELij Z 0;
<
Eij ¼ jRELij j ð4Þ
>
> otherwise
: D2ij

The RELij values depend on the logical relationship that exists


between the pair of facilities ij. By default, these are assigned
the following values: A¼40, E¼ 12, I¼4, O¼1, U¼0 and
X¼  1. They may be modified by users according to the
characteristics of the problem.
(c) Aspect ratio: For each facility, an adequate aspect ratio is
desired, which is defined as
maxfhj ; wj g
ARj ¼ ð5Þ
minfhj ; wj g
where h and w are respectively the height and width of the
facility j.
To calculate the satisfaction of the aspect ratio, the adaptive
penalization proposed by Tate et al. [36] is used. The aspect
ratio is considered as an objective because some ratios are
better than others, but aspect ratio is also a restriction because
any solution becomes infeasible if the aspect ratio of any
facility surpasses its limit.

2.1.3. Genetic algorithm steps


The steps of the proposed GA are explained below:

1. An initial random population of N individuals is generated. Fig. 3. (a) An example of a satisfactory solution obtained by the GA. (b) An example
2. Tournament selection [13] is applied to select the individuals of a solution considered good by the human expert. (For interpretation of the
that will be involved in the evolutionary operations. references to colour in this figure caption, the reader is referred to the web version
3. Crossover and mutation operators are applied to the indivi- of this paper.)

duals with a probability given by the user.


4. Elitism [13] is applied to the new population. The process which satisfies all the closeness requirements. However, this
continues from step 2, until the specified number of iterations facility layout design was ranked by the GA as worse than the
is reached. previous one.
The graphical representation of facility layouts shown in Fig. 3
offers to the human expert all the information relevant to the
2.1.4. Crossover and mutation operators design, which would be difficult to represent in another way. Thus,
Crossover is performed by exchanging complete rows among a system for presenting this graphical information concerning the
the matrices that represent two selected individuals. A random facility layouts has been developed to help the expert easily
number between 0 and 2 is obtained. If the result is 0, the first row identify some measurable aspects of the plant layouts. Green lines
is exchanged; if it is 1, the second is exchanged; and if it is 2, the among facilities signify the amount of material flow between them
third row is exchanged. Similarly, for mutation, one of the three such that the thickness of the line expresses this quantity; thicker
rows is modified randomly. lines represent a higher flow. Blue lines indicate that the closeness
requirements are satisfied unless the facility background is coloured
2.2. Expert evaluation soft red. A red line between a pair of facilities indicates that the
distance requirement between them is not satisfied.
In this phase of the algorithm, the human expert applies his
or her experience and knowledge to the facility design. This is 2.3. Evolutionary artificial neural network algorithm
performed using an evaluation of the entire population of facility
layouts generated by the GA, as explained in the previous section. First, in this section, the terminology and notation that are used
As stated previously, the entire set of plant layouts is evaluated throughout this work are defined. The goal of classification is to
using a five-granularity Likert scale. The human expert assigns a assign an input vector x to one of Q discrete classes Cq, where
mark to each solution considering the material handling cost, the q A 1; …; Q . Thus, a formal framework for the classification problem
closeness and distance requirements and the aspect ratio. could be introduced by considering a training sample D ¼ fðxn ; yn Þ;
It is possible that one of the best solutions obtained by the GA n ¼ 1; …; Ng, where xn ¼ ðx1n ; …; xkn Þ is a random vector of mea-
is not assigned a ‘5’ value by the human expert. This is illustrated surements that have values in Ω  RK , and yn is the class level
by the facility layout design that is shown in Fig. 3(a). The design associated with the n-th pattern. The common technique for
has acceptable values for the three factors (material handling cost, representing the labels is used, i.e. a “1-of-Q” encoding vector is
closeness and distance requirements and aspect ratio), but, in this adopted in such a way that y ¼ ðy1 ; …; yQ Þ.
particular design, the human expert prefers solutions that better The evolutionary algorithm (EA) used to estimate the parameters
satisfy the closeness requirements. For that reason, he or she has and the structure of the different model neural networks (see Fig. 4),
assigned a mark of ‘3’ to this facility layout design. However, he or including the number of hidden nodes of each transfer and activa-
she assigned a mark of ‘5’ to the solution illustrated in Fig. 3(b), tion function, is also presented in this section. The objective is to
L. García-Hernández et al. / Neurocomputing 135 (2014) 69–78 73

Softmax

Bias 1 Bias 2

.
.
.

Input layer Hidden layer


Output layer

Fig. 4. Artificial neural network algorithm used to train the proposed method. B represents a basis function. It should be noted that when the product unit basis is used, the
first bias is unnecessary.

design a neural network with optimal structure and weights for the cross-entropy error function lðθÞ or Q-class multinomial deviance is
classification problem treated. The algorithm is an example of an given by
evolutionary programming (EP) paradigm [4], in which the popula-
1 N Q ðlÞ
tion is subjected to the replication and mutation operators, but lðθÞ ¼  ∑ ∑ y log g l ðxn ; θl Þ; ð8Þ
Nn¼1l¼1 n
crossover is not considered, because this operation is usually
regarded as less effective for ANN evolution [14]. For this purpose,
where yðlÞ ðjÞ
n is the target value for class l in pattern n (yn ¼ 1 if x A C j
three types of hidden nodes are considered (sigmoidal Units, or SU;
product Unit, or PU; and radial basis functions, or RBF). The general and yðlÞ
n ¼ 0 otherwise) and N is the number of patterns. f l ðx n ; θ l Þ ¼

framework of the evolutionary programming (EP) algorithm is the βl0 þ ∑M


j ¼ 1 β j  Bj ðx; wÞ, θ ¼ ðθ 1 ; …; θ Q Þ are the corresponding para-
l

following: meters and Bj ðx; wÞ are the basis functions associated with the SU, PU
or RBF networks.
1. Generate a random population of size Np, where each indivi-
dual represents a model structure.
2.3.1. Parameters determining the structure of the neural networks
2. Repeat the following steps until the stopping criterion is
To define the topology of the neural networks generated in the
fulfilled.
evolution process, we consider three parameters: m, MI and ME
(a) Apply parametric mutation to the best 10% of individuals.
(which are the minimum number of hidden nodes in the model,
(b) Apply structural mutation to the remaining 90% of individuals.
the maximum number of hidden nodes in the initialization
(c) Calculate the fitness of each individual in the population.
process and the maximum number of hidden nodes in the entire
(d) Rank the individuals according to their fitness.
evolutionary process, respectively). To obtain an initial population
(e) The best 10% of individuals are replicated and the worst
formed by models simpler than the most complex model possible,
10% of individuals are replaced.
the parameters must fulfill m r M I rM E .
3. Select fittest individual of the final population as the solution.

To treat the classification problem outlined above, the outputs 2.3.2. Generation of an initial model structure
of the SU, PU or RBF models are interpreted probabilistically using The algorithm begins by generating a larger number of networks
the softmax activation function: than that is used during the evolutionary process. 10Np networks are
generated (where Np is the number of networks of the population
exp f l ðx; θl Þ
g l ðx; θl Þ ¼ ; l ¼ 1; …; Q ð6Þ used during the evolutionary process), and then, the best Np neural
∑Qj¼ 1 exp f l ðx; θl Þ networks are selected. For the network generation, the number of
nodes in the hidden layer is taken from a uniform distribution over the
where Q is the number of classes in the problem (5 in this case), interval ½m; M I . For the PU and SU hidden nodes, the number of
f l ðx; θl Þ is the output of output node l for pattern x and g l ðx; θl Þ is connections between each node of the hidden layer and the input
the estimated probability that a given pattern x belongs to class l. nodes is determined from a uniform distribution over the interval
The predicted class ^l is then obtained by the following decision ð0; kÞ, where k is the number of independent variables (three in this
rule ^l ¼ arg maxl g l ðx; θl Þ. problem). For the RBF hidden nodes, the number of connection is
The fitness measure is a strictly decreasing transformation of always k, because these connections represent the coordinates of
the entropy error lðθÞ, which is given by the centre of the neuron. The number of connections between each
hidden node and the output layer is selected from a uniform
1
AðgÞ ¼ ; ð7Þ distribution over the interval ð0; Q  1Þ.
1 þ lðθÞ The weights are initialized in manner that depends on the type
of hidden node generated. For the SU and PU hidden nodes, the
where g is the output of the neural network models, which is given weights are assigned using a uniform distribution defined over
by the multi-valued function gðx; θÞ ¼ ðg 1 ðx; θ1 Þ; …; g l ðx; θl ÞÞ. The two intervals, ½  I; I for connections between the input layer and
74 L. García-Hernández et al. / Neurocomputing 135 (2014) 69–78

the hidden layer and ½  O; O for connections between the hidden Table 3
layer and the output layer. For the RBF hidden nodes, the connec- Ovine slaughterhouse facilities.
tions between the input layer and the hidden layer are initialized
Id Facility Area Aspect ratio
using a clustering algorithm, such that the EP can start the
evolutionary process with well-positioned centres. The primary A Stables 570 4
idea is to cluster the input data in k groups, where k is the number B Slaughter 206 4
of hidden RBF neurons. Thus, each hidden RBF neuron can be C Entrails 150 4
D Leather and skin 55 4
positioned in the centroid of the corresponding cluster. Finally, the E Aeration chamber 114 4
radius of each hidden RBF neuron is calculated as the geometric F Refrigeration chamber 102 4
mean of the distances to the two closest centroids. The clustering G Entrails chamber 36 4
technique used (described in [8]) is a modification of the classic H Boiler room 26 4
I Compressor room 46 4
k-means algorithm, in which the initial centroids are calculated
J Shipping 109 4
using a specific initialization algorithm that avoids local minima K Offices 80 4
and thus increases the probability that the initial k cluster centres L Byproduct shipping 40 4
are not generated from a smaller number of clusters. Additionally,
the connections between the hidden layer and the output layer are
generated using a uniform distribution over the interval ½  O; O.
blue
60 60
2.3.3. Parametric mutations E F J
60
For the SU and PU hidden nodes, parametric mutation is
100
performed for each coefficient wji, βjl of the model using Gaussian
10 10
A B D
noise: 15 L
15
15
C G
wji ðt þ 1Þ ¼ wji ðtÞ þ ξ1 ðtÞ; βlj ðt þ1Þ ¼ βlj ðtÞ þ ξ2 ðtÞ ð9Þ
Fig. 5. Material flow among ovine slaughterhouse facilities.
where ξk ðtÞ A Nð0; αk ðtÞÞ, for each k¼1,2, represents a one-
dimensional normally distributed random variable with mean
0 and variance αk ðtÞ. If the hidden node mutated is an SU hidden
node, then βjl and wji are mutated with the same Gaussian noise
The parameters selected for generating the individuals using
ξ1 ðtÞ, thus βlj ðt þ 1Þ ¼ βlj ðtÞ þ ξ1 ðtÞ. Finally, the radius of the RBF the GA are the following: a large population of 1000 individuals
hidden nodes is mutated using a similar expression: r j ðt þ 1Þ ¼ was used to obtain enough solutions for training, probabilities of
r j ðtÞ þ ξ1 ðtÞ. 0.8 for crossover and 0.2 for mutation were used because these
values yielded good performance in previous studies [30], and a
2.3.4. Structural mutations low number of 100 generations was used because it only requires
Structural mutation modified the neural network structure and feasible solutions with diverse objective values.
allows for exploration of different regions in the search space For the sake of brevity and simplicity and because of the
while helping to maintain the diversity of the population. There difficulty and fatigue caused by evaluating multiple facility layout
are five different structural mutations: node deletion, connection solutions, the opinion of a single human expert in facility layout
deletion, node addition, connection addition and node fusion. design was considered for a dataset composed of 365 individuals,
Connection deletion and connection addition have no sense but additional opinions could be used. Instead, the best model
if the node mutated is an RBF, so we do not consider these obtained was tested using another dataset, which was composed
operations for this type of node. In the node fusion operation, of 181 facility layout designs for which a different expert0 s opinion
two randomly selected hidden nodes, a and b, are replaced by a was used to test the model robustness.
new node, c, which is a combination of the two. For further details
of the EA, see [17], [25] and [24].
3.2. Experimental design

3. Experimental results A simple linear rescaling of the input variables was performed
over the interval [0.1; 0.9], where X ni are the transformed variables.
In this section, the dataset constructed, the experiments The initial values for the algorithm parameters were the following:
performed and the results obtained by the artificial neural net- for fm; M I ; M E g, 2; 3; 4 were used for the SU nodes, 1; 2; 3 were used
work methodologies considered are described and analyzed. for the PU nodes, and 6; 7; 8 were used for the RBF nodes. For the
EP, the number of generations used for all neural network models
3.1. Dataset was 100. The connections between the input and hidden layers
were initialized in the ½  3; 3 interval for the SU and PU nodes and
The proposed approach was applied to a real-world layout ½  2; 2 for the RBF nodes, and the connections between the hidden
design problem [31]. The problem concerns to the plant layout and output layers were initialized in the interval ½  5; 5. The size
design for an industrial ovine slaughterhouse. The available space of the population was N ¼1000. For the structural mutation, the
for the design is a rectangular space with dimensions 30 m number of nodes that could be added or removed was in the
51:14 m, thus the total space for the design is 1534 m2 . The interval ½1; 2, and the number of connections that could be added
properties of the plant facilities are given in Table 3, the space or deleted in the hidden and output layers during structural
required for each facility and its aspect ratio limit are given. The mutations was in the interval ½1; 5.
material flow required starts with the flow from facility A to B, A stratified holdout technique was performed to divide the
where it is split into three lines, which connect B, E, F and J (60% of dataset, where 75% of the patterns were used for training the
the material), B, C, G and L (15%), and B, D and L (10%). Fig. 5 model, and the remaining 25% were used for testing it. Further-
illustrates the material flow in this problem. more, the algorithm was run on the dataset 30 times to achieve
L. García-Hernández et al. / Neurocomputing 135 (2014) 69–78 75

Table 4 to ascertain the statistical significance of the observed differences


Results obtained for CCR and #Links for the three basis functions used and the test among the means. The comparisons were performed by considering a
set, where i; h; o are the number of input variables, the number of hidden nodes and
critical level for Snedecor0 s F values with 2 and 87 degrees of freedom.
the number of output nodes, respectively.
Because the significance level, α ¼ 0:05, is higher than the critical
Algorithms Network topology CCR (mean 7SD) #links (mean 7 SD) level, p¼0.00, we rejected the hypothesis that CCR and #link means
have identical values. Because this hypothesis was rejected, a Tukey
SU 3:2,3,4:4 87.107 2.35 25.577 3.30 test for ranking mean values was applied to obtain the corresponding
PU 3:1,2,3:4 82.077 2.78 18.20 7 2.58
RBF 3:6,7,8:4 81.88 73.33 45.337 4.40
homogeneous subsets. Table 5 presents the results obtained following
the above methodology, including the Tukey test performed and the
Best model obtained
performance ranking for the different basis functions. The results
SU 3:3:4 91.30 24
PU 3:3:4 90.22 18 indicate statistically significant differences between the SU neural
RBF 3:7:4 90.22 47 networks and the PU and RBF networks (the SU had a greater mean
CCR value). Conversely, regarding the number of links, there exist
some significant differences between the PU neural networks and the
SU and RBF networks (the PU networks had a lower mean value of
links). Because of these results, we decided to use sigmoidal units (SU)
Table 5 for the problem, because, although the mean number of links is
p-value of the Tukey tests for CCR and #links, subsets with α ¼ 0:05 and ranking of significantly higher compared with product units (25.27 versus 18.20),
the different basis functions based on this multiple comparison test. Note that the SU yields better performance as quantified by the CCR measure (87.10
binary relation r is not transitive. in contrast to the value of 82.07 obtained by PU).
Basis functions CCR #links
For the best model obtained using SU neural networks, Fig. 6
shows the four resultant discriminant functions. By using these
SU 87.10 – 25.57 – – discriminant functions and the softmax function, the correspond-
PU – 82.07 – 18.20 – ing estimated probabilities of each pattern belonging to each of
RBF – 81.88 – – 45.33
the 5 classes involved can be calculated.
p-value 1.00 0.96 1.00 1.00 1.00
3.4. Discussion
Mean ranking for CCR: μCCR ðRBFÞr μCCR ðPUÞ o μCCR ðSUÞ. Mean ranking for #links:
μ#links ðPUÞ o μ#links ðSUÞ o μ#links ðRBFÞ.
The confusion matrices presented in Table 6 demonstrate the
difficulty of capturing information from class 5 for the prediction;
more robust results because of the non-deterministic behaviour of this class was misclassified as the previous class, which could
artificial neural networks. be due to the unbalanced nature of the dataset. However, the
difference of facility layout solutions between classes 4 and 5 is
3.3. Results almost trivial. Therefore, patterns from both classes were treated
as “good” solutions for the proposed problem. The entire set of
Table 4 presents the results yielded for the network topologies patterns classified as class 4 could be analyzed by a human expert,
tested in all experiments using SU, PU or RBF hidden nodes (mean to choose an optimal solution and to consider other subjective
and standard deviation (mean7 SD) for 30 runs). Despite the fact factors, such as the distribution of the remaining space. Conver-
that ANNs have been proved to work well in an enormous quantity sely, the confusion matrices demonstrate that misclassified pat-
of problems, they also present some remarkable drawbacks. First of terns are always confused with a closed class, which indicates that
all, they usually tend to overfit, therefore not generalizing well on the model considers the order of categories.
unseen data. The most immediate solution to that issue would be The confusion matrix of the validation experiment is presented in
the use of cross-validation for choosing the networks parameters, Table 6. The presenting correct classification rate was 79.56%. From
however, that could be computational unaffordable when using this result, several conclusion can be drawn. For example, the fact
evolutionary algorithms. Other drawback to consider is the lack of that the opinion of a single human expert is very useful for obtaining
interpretability of this kind of techniques, which makes difficult to a robust evaluation methodology has been demonstrated by the
verify the plausibility of the model produced. Besides, the input validation using other human expert opinions. Furthermore, evolu-
variables must be numeric (indeed, nominal variables must be tionary artificial neural networks have demonstrated their ability to
transformed into binary ones), and have been previously scaled. In capture knowledge and compute from it a predictive model that
our case, the three criteria considered as input variables were generalizes to unseen data. Nonetheless, when comparing the
numeric, and they have been linearly rescaled before the training results with those achieved by the use of the same expert, one
process. Finally, the convergence to a solution is not guaranteed and may notice the differences in the prediction ability (91.30%), which
the training time can be high. Typically, the evolution stops after a are due to the subjectivity required for the evaluation and the fact
number of iterations, after a certain time, when the weights of the that different experts may yield different opinions about the same
network start to converge to fixed values, or when the error rate facility layout. Additionally, the study of the different expert opi-
starts to increase. In this case, the number of generations used for all nions could be a promising line of future work.
neural networks models has been prefixed to 100. However, in the As a final remark, other factors concerning the “kindness” of a
specific case treated in this paper the use of ANNs has been proved solution can be considered for the problem. However, the use of
to work perfectly, converging to a good solution, without overfitting the three factors chosen in this paper has been proved to perform
and generalizing well on unseen data. well when trying to capture the suitability of a facility layout for
First, the Kolmogorov–Smirnov (KS) test with a significance level the problem. Indeed, the use of more factors concerning a facility
α ¼ 0:05 was used to evaluate if the different performance metrics for could not ensure an improvement in the performance of the
all the methods followed a normal distribution. All methods obtained methodology, specially if those criteria are highly correlated.
a p-value greater than the critical level for the CCR and #links Regarding the Evolutionary Artificial Neural Networks, the use of
measures. Therefore, a normal distribution can be assumed for all more input variables could yield to an increase in the computa-
the cases. Then, we used the Analysis of Variance (ANOVA) technique tional time and the complexity of the model.
76 L. García-Hernández et al. / Neurocomputing 135 (2014) 69–78

Softmax
-1.10
Bias 1 0.05 Bias 2 4.37
Material flow 8.09
3.90 8.37
-7.67 0.76 -6.83
-4.83
4.60
-0.52
4.55
Adjacency
4.42
requirements 2.57
3.57
-5.34
-6.52
-5.19 -14.98
2.55
Aspect ratio
-6.38 9.81
-4.63

Input layer Hidden layer Output layer

Fig. 6. Graphical representation of the best model achieved where sðx; wÞ represents the sigmoidal unit. The input variables have been normalized in the interval ½0:1; 0:9,
thus, they are represented as xni .

Table 6
Confusion matrices and CCR values obtained for training, testing and validation using the SU basis function.

Train Predicted Test Predicted Valid. Predicted

Target 1 2 3 4 5 Target 1 2 3 4 5 Target 1 2 3 4 5

1 11 2 0 0 0 1 4 0 0 0 0 1 11 5 0 0 0
2 1 77 6 0 0 2 0 26 2 0 0 2 0 59 17 0 0
3 0 3 110 9 0 3 0 2 37 2 0 3 0 2 48 12 0
4 0 0 2 49 0 4 0 0 0 17 0 4 0 0 0 26 0
5 0 0 0 3 0 5 0 0 0 2 0 5 0 0 0 1 0

CCRTR ¼ 90:48% CCRTE ¼ 91:30% CCRV ¼ 79:56%

distance and aspect ratio factors. Thus, the ANN model might have
confused this solution as one of the fourth category instead of
classifying it as ‘5’.

4. Conclusions

In this work, an intelligent system for addressing the UA-FLP is


proposed. This system allows to incorporate expert knowledge in the
final design (to evaluate a final solution using the expert's experience).
A system that includes a GA and an ANN is employed. The GA is used
Fig. 7. Best facility layout design found by the model used. (For interpretation of
to generate a number of solutions that are obtained using an
the references to colour in this figure caption, the reader is referred to the web
version of this paper.) optimization process; then, these solutions are evaluated by a human
expert designer. To avoid overburdening the human expert, an ANN
system is applied in the second stage for learning the expert knowl-
To properly analyze the results, the facility layout with the highest edge and predicting the mark that he or she would assign to each
probability of belonging to the fourth class for the validation set is solution.
presented in Fig. 7. More specifically, the solution corresponds to the Evolutionary neural networks based on sigmoidal units success
pattern associated with class 5. This solution can be considered as modelling human expert evaluation of UA-FLP solutions in the test
satisfactory for this problem because all closeness and distance case carried out; as the proposed methodology achieved good
requirements are satisfied (no red lines between facilities or red results: 91.30% correct classification in the testing set —where the
facility backgrounds are shown), and the aspect ratio is acceptable. same expert evaluates the layouts— and 79.56% correct classification
Furthermore, facilities that have higher material flows are almost in the validation step —where a different human expert evaluates
always connected (except for the A-Stables). However, because of the them. These results also indicate that the opinion of a single human
high flows reported for all solutions of this problem, the human expert yields a fairly robust classification. Furthermore, the results
expert assigns more importance to the satisfaction of the closeness, allow replacing the human expert with the neural network (reducing
L. García-Hernández et al. / Neurocomputing 135 (2014) 69–78 77

the cost and the problems associated with fatigue), even with the [26] A.R. McKendall Jr., J.R. Jaramillo, A tabu search heuristic for the dynamic space
best performing network that is quite simple. allocation problem, Comput. Oper. Res. 33 (3) (2006) 768–789.
[27] Z. Michalewicz, D. Dasgupta, R.G.L. Riche, M. Schoenauer, Evolutionary
As previously mentioned, the evaluation of further experts could algorithms for constrained engineering problems, Evolut. Comput. 4 (1996)
be dealt with in future work, and the best model could be validated 1–32.
by including it in a genetic algorithm for studding and analyzing the [28] M. Mitchell, An Introduction to Genetic Algorithms, MIT Press, Cambridge,
MA, USA, 1998.
best facility layouts obtained. Additionally, it would be a promising [29] R. Muther, Systematic Layout Planning, Cahners Books, Boston, MA, 1973.
future research line the study of different classification models to [30] M. Pérez-Ortiz, L. García-Hernández, L. Salas-Morera, A. Arauzo-Azofra,
improve the result percentage. Furthermore, it would be reasonable C. Hervás-Martínez, An ordinal regression approach for the Unequal Area
Facility Layout Problem, in: Proceedings on the International Conference on
a wider generalization of the results with a set of UA-FLP cases. Soft Computing Models in Industrial and Environmental Applications
(SOCO'12), 2012.
Acknowledgments [31] L. Salas-Morera, Computer-aided plant layout, Inf. Tecnologica 7 (4) (1996)
39–46.
[32] L. Salas-Morera, L. Garcia-Hernandez, A. Arauzo-Azofra, An evolutionary
This work was partially supported by the Spanish Inter- algorithm for the unequal area facility layout problem, in: 11th International
Conference on Intelligent Systems Design and Applications (ISDA), 2011,
Ministerial Commission of Science and Technology under Project
pp. 414–419.
TIN2011-22-794, the European Regional Development Fund, and [33] D. Scholz, A. Petrick, W. Domschke, Stats: a slicing tree and tabu search based
the “Junta de Andalucía”, Spain, under Project P08-TIC-3745. heuristic for the unequal area facility layout problem, Eur. J. Oper. Res. 197 (1)
(2009) 166–178.
[34] M. Solimanpur, A. Jafari, Optimal solution for the two-dimensional facility
layout problem using a branch-and-bound algorithm, Comput. Ind. Eng. 55 (3)
References (2008) 606–619.
[35] K.Y. Tam, Genetic algorithms, function optimization, and facility layout design,
[1] A. Abraham, E. Corchado, J.M. Corchado, Hybrid learning machines, Neuro- Eur. J. Oper. Res. 63 (2) (1992) 322–346.
computing 72 (13–15) (2009) 2729–2730. [36] D.M. Tate, A.E. Smith, S.M. Iie, S.M. Iie, Unequal area facility layout using
[2] G. Aiello, M. Enea, G. Galante, A multi-objective approach to facility layout genetic search, IIE Trans. 27 (1994) 465–472.
problem by genetic search algorithm and electre method, Robotics Comput. [37] J. Tompkins, J. White, Y. Bozer, J. Tanchoco, Facilities Planning, 4rd ed., Wiley,
Integr. Manuf. 22 (2006) 447–455. New York, 2010.
[3] G. Aiello, G.L. Scalia, M. Enea, A multi objective genetic algorithm for the [38] K. Tsuchiya, S. Bharitkar, Y. Takefuji, A neural network approach to facility
facility layout problem based upon slicing structure encoding, Expert Syst. layout problems, Eur. J. Oper. Res. 89 (3) (1996) 556–563.
Appl. 39 (2012) 10352–10358. [39] B.H. Ulutas, S. Kulturel-Konak, An artificial immune system based algorithm to
[4] P.J. Angeline, P.J. Angeline, G.M. Saunders, G.M. Saunders, J.B. Pollack, solve unequal area facility layout problem, Expert Syst. Appl. 39 (5) (2012)
J.B. Pollack, An evolutionary algorithm that constructs recurrent neural net- 5384–5395.
works, IEEE Trans. Neural Netw. 5 (1994) 54–65. [40] Y. Wu, E. Appleton, The optimisation of block layout and aisle structure by a
[5] G.C. Armour, E.S. Buffa, A heuristic algorithm and simulation approach to genetic algorithm, Comput. Ind. Eng. 41 (4) (2002) 371–387.
relative location of facilities, Manag. Sci. 9 (1963) 294–309.
[6] M. Babbar-Sebens, B.S. Minsker, Interactive genetic algorithm with mixed
initiative interaction for multi-criteria ground water monitoring design, Appl.
Soft Comput. 12 (1) (2012) 182–195. Laura García-Hernández received M.Sc. in Computer
[7] Y.-K. Chung, A neuro-based expert system for facility layout construction, Science in 2007 and Ph.D. in the field of Engineering in
J. Intell. Manuf. 10 (1999) 359–385. 2011 from the University of Córdoba (Spain). She is an
[8] S. Cohen, N. Intrator, Global Optimization of RBF Networks, 2000. Associate Professor of University of Córdoba, in the Depart-
[9] E. Corchado, B. Baruque, Wevos-visom: an ensemble summarization algorithm ment of Rural Engineering, Area of Project Engineering.
for enhanced data visualization, Neurocomputing 75 (1) (2012) 171–184. Her primary areas of research are engineering design,
[10] E. Corchado, M. Graña, M. Wozniak, Editorial: new trends and applications on development of intelligent user adaptive systems for
hybrid artificial intelligence systems, Neurocomputing 75 (1) (2012) 61–63. automated facility layout design, interactive evolutionary
[11] R. Şahin, K. Ertoğral, O. Türkbey, A simulated annealing heuristic for the computation, project management, risk prevention in
dynamic layout problem with budget constraint, Comput. Ind. Eng. 59 (2) automatic systems and educational technology.
(2010) 308–313.
[12] A. Drira, H. Pierreval, S. Hajri-Gabouj, Facility layout problems: a survey, Annu.
Rev. Control 31 (2) (2007) 255–267.
[13] A. Eiben, J. Smith, Introduction to Evolutionary Computing, Springer-Verlag,
Berlin Heidelberg, 2003.
[14] D.B. Fogel, Evolutionary Computation: Toward a New Philosophy of Machine
Intelligence, EEE Press, Piscataway, NJ, USA, 1995. María Pérez-Ortiz was born in Córdoba, Spain, in 1990.
[15] D.E. Goldberg, Genetic algorithms in Search, Optimization, and Machine She received his B.S. degree in Computer Science from
Learning, Adisson-Wesley Professional, New York, 1989. the University of Córdoba, Spain, in 2011. She is
[16] A. Gomez, Q.I. Fernandez, D. De la Fuente Garcia, P.J. Garcia, Using genetic currently working toward her Ph.D. degree in the
algorithms to resolve layout problems in facilities where there are aisles, Int. Department of Computer Science and Numerical Ana-
J. Prod. Econ. 84 (3) (2003) 271–282. lysis (University of Cordoba, Spain), in the area of
[17] P.A. Gutiérrez, C. Hervás-Martínez, M. Carbonero-Ruz, J.C. Fernandez-Caballero, computer science and artificial intelligence. Her current
Combined projection and kernel basis functions for classification in evolutionary interests include a wide range of topics concerning
neural networks, Neurocomputing 72 (13–15) (2009) 2731–2742. the machine learning and the pattern recognition
[18] S.S. Heragu, A. Kusiak, Efficient models for the facility layout problem, Eur. paradigms.
J. Oper. Res. 53 (1) (1991) 1–13.
[19] J.Y. Kim, Y.D. Kim, Graph theoretic heuristics for unequal-sized facility layout
problems, Omega 23 (4) (1995) 391–401.
[20] Komarudin, K.Y. Wong, Applying ant system for solving unequal area facility
layout problems, Eur. J. Oper. Res. 202 (3) (2010) 730–746.
[21] P. Kouvelis, A.A. Kurawarwala, G.J. Gutierrez, Algorithms for robust single and
multiple period layout planning for manufacturing systems, Eur. J. Oper. Res.
63 (2) (1992) 287–303. Antonio Araúzo-Azofra received M.Sc. in Computer
[22] K.-Y. Lee, M.-I. Roh, H.-S. Jeong, An improved genetic algorithm for multi-floor Science in 2001 and Ph.D. in the field of Artificial
facility layout problems having inner structure walls and passages, Comput. Intelligence in 2006 from the University of Granada
Oper. Res. 32 (4) (2005) 879–899. (Spain). He is currently an Assistant Professor of the
[23] Y.H. Lee, M.H. Lee, A shape-based block layout approach to facility layout Project Engineering department at the University of
problems using hybrid genetic algorithm, Comput. Ind. Eng. 42 (2-4) (2002) Cordoba (Spain). His areas of research are feature
237–248. selection applied to classification problems for intelli-
[24] A.C. Martínez-Estudillo, F.J. Martínez-Estudillo, Evolutionary product unit gent data mining in information systems and the
based neural networks for regression, Neural Netw. 19 (4) (2006) 477–486. development of intelligent user adaptive systems for
[25] F.J. Martínez-Estudillo, C. Hervás-Martínez, P.A. Gutiérrez, A.C. Martínez- automated facility layout design.
Estudillo, Evolutionary product-unit neural networks classifiers, Neurocom-
puting 72 (1-3) (2008) 548–561.
78 L. García-Hernández et al. / Neurocomputing 135 (2014) 69–78

Lorenzo Salas-Morera received a M.Sc. in Agricultural César Hervás-Martínez was born in Cuenca, Spain. He
Engineering in 1989 and a Ph.D. in Agricultural Engi- received the B.S. degree in Statistics and Operating
neering in 1993 from the University of Córdoba (Spain). Research from the Universidad Complutense, Madrid,
He is a Professor at the University of Córdoba in the Spain, in 1978 and the Ph.D. degree in Mathematics
Department of Rural Engineering, Area of Project Engi- from the University of Seville, Seville, Spain, in 1986. He
neering. He is currently the Director of Academic is a Professor with the University of Córdoba in the
Organization of Córdoba University, and previously, he Department of Computing and Numerical Analysis in
was the Director of the Polytechnical Superior School of the area of computer science and artificial intelligence
Córdoba University from 2002 to 2010. His primary and an Associate Professor in the Department of
areas of research are engineering design, development Quantitative Methods in the School of Economics. His
of intelligent user adaptive systems for automated current research interests include neural networks,
facility layout design, project management and educa- evolutionary computation, ordinal classification and
tional technology. the modelling of natural systems.

You might also like