Professional Documents
Culture Documents
net/publication/236268490
CITATIONS READS
11 3,642
3 authors:
Sanghamitra Bandyopadhyay
Indian Statistical Institute
434 PUBLICATIONS 14,249 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Advanced Machine Vision Paradigms for Medical Image Analysis View project
All content following this page was uploaded by Siddhartha Bhattacharyya on 16 May 2014.
Chapter 1
Soft Computing and
its Applications
Siddhartha Bhattacharyya
The University of Burdwan, India
Ujjwal Maulik
Jadavpur University, India
Sanghamitra Bandyopadhyay
Indian Statistical Institute, India
ABStRACt
Soft Computing is a relatively new computing paradigm bestowed with tools and techniques for handling
real world problems. The main components of this computing paradigm are neural networks, fuzzy logic
and evolutionary computation. Each and every component of the soft computing paradigm operates either
independently or in coalition with the other components for addressing problems related to modeling,
analysis and processing of data. An overview of the essentials and applications of the soft computing
paradigm is presented in this chapter with reference to the functionalities and operations of its con-
stituent components. Neural networks are made up of interconnected processing nodes/neurons, which
operate on numeric data. These networks posses the capabilities of adaptation and approximation. The
varied amount of uncertainty and ambiguity in real world data are handled in a linguistic framework
by means of fuzzy sets and fuzzy logic. Hence, this component is efficient in understanding vagueness
and imprecision in real world knowledge bases. Genetic algorithms, simulated annealing algorithm
and ant colony optimization algorithm are representative evolutionary computation techniques, which
are efficient in deducing an optimum solution to a problem, thanks to the inherent exhaustive search
methodologies adopted. Of late, rough sets have evolved to improve upon the performances of either
of these components by way of approximation techniques. These soft computing techniques have been
put to use in wide variety of problems ranging from scientific to industrial applications. Notable among
these applications include image processing, pattern recognition, Kansei information processing, data
mining, web intelligence etc.
DOI: 10.4018/978-1-61692-797-4.ch001
Copyright © 2011, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
Soft Computing and its Applications
2
Soft Computing and its Applications
Since the essence of neural network operation is network. These functions recognize specific
based on the behavior of human brain, these net- range of input signals and selectively tune the
works require a form of training or learning ability. neurons to respond to the input signals according
Once these are trained with the different aspects some learning algorithm. Most of these activa-
of the problem at hand, they can be used to solve tion functions take an input as an infinite range
similar problems given the immense generaliza- of activations (-∞, +∞) and squashes/transforms
tion capabilities embedded therein. Depending on them in the finite range [0, 1] or {-1, 1} [Le-
the type of learning procedure adopted, different ondes98]. Thus, these functions are able to map
neural network architectures have evolved from the input information into bipolar excitations.
time to time [Haykin99, Kumar2004]. Though these functions may vary from neuron
In the most general form, an artificial neural to neuron within the network, yet most network
network is a layered structure of neurons. It com- architectures are field-homogeneous i.e. all the
prises seven essential components [Kumar2004], neurons within a layer are characterized by the
viz., (i) neurons, (ii) activation state vector, (iii) same signal function. Some of the common neural
activation function, (iv) connection topology, network signal functions [Kumar2004, Haykin99]
(v) activity aggregation rule, (vi) learning rule include (i) binary threshold, (ii) bipolar threshold,
and (vii) environment. These components are (iii) linear, (iv) liner threshold, (v) sigmoid, (vi)
discussed in the following sections. hyperbolic tangent, (vii) Gaussian, (viii) stochastic
[Kumar2004] etc.
Neurons
Connection Topology
Neurons are the processing units of a neural net-
work. There are basically three types of neurons This refers to the interconnection topology of
viz. input, hidden and output. The input neurons the neural network architectures. These connec-
are designated to accept stimuli from the external tions may be either excitatory (+) or inhibitory
world. The output neurons generate the network (-) or absent (0). These connections, or synapses
outputs. The hidden neurons, which are shielded basically house the memory of the network. The
from the external world, are entrusted with the behavior of neural network architecture is decided
computation of intermediate functions necessary by its connection topology.
for the operation of the network.
Activity Aggregation Rule
Activation State Vector
This rule aggregates the activities of the neurons
Neural network models operate in a real n- at a given layer. It is usually computed as the
dimensional vector space Rn. The activation state inner product of the input vector and the neuron
vector, X = (x1, x2, …, xn)T ∈ Rn, is a vector of the fan-in interconnection strength (weight) vector.
activation levels xi of the individual neurons of An activation rule thereby, determines the new
the network. This state vector acts as the driving activation level of a neuron based on its current
force for a neural network. activation and external inputs.
The characteristic activation functions are used The neural network learning rules define an
to supplement the learning process of a neural architecture-dependent procedure to encode
3
Soft Computing and its Applications
pattern information into inter-neuron intercon- update their interconnection weights dur-
nections. This is a data driven process executed ing the process of self-organization.
by modifying the connection weights. Two types
of learning are in vogue; viz. supervised learning Environment
and unsupervised learning.
The operational environment of neural networks
• Supervised learning: Supervised learning can be either deterministic (noiseless) or stochastic
encodes a behavioristic pattern into a neu- (noisy). A neural network N is a weighted directed
ral network by attempting to approximate graph, where the nodes are connected as either
the function that best describes the data set
employed. For an input vector, Xk ∈ Rn (a • a feedforward architecture, in which
real n-dimensional vector space) related the network has no loops. Examples in-
to an output vector Dk ∈ Rp (in a real p- clude the perceptron, multilayer percep-
dimensional vector space), a supervised tron [Duda73], support vector machines
learning algorithm aims at deriving an un- [Cortes95] and radial basis function
known mapping function f: Rn→ Rp. The networks [Broomhead88], Kohonen’s
algorithm tries to reduce the error (Dk-Sk) Self Organizing Feature Map (SOFM)
in the system response, where Sk is the ac- [Kohonen95] etc., or
tual response of the system by employing • a feedback (recurrent) architecture, in
the desired output response of the system which loops occur in the network because
Dk (also referred to as the teaching input) of feedback connections. Examples in-
and the associate of Xk. Thus, input-output clude the Hopfield network [Hopfield84],
sample pairs are used to train/teach the net- BSB model [Hui92], Boltzmann machines
work through a simple form of error cor- [Ackley85], bidirectional associative
rection learning or gradient descent weight memories [Kosko88], adaptive resonance
adaptation. Hence, the system generates an theory [Carpenter95] etc.
output Dk in response to an input Xk. The
learning process achieves an association The following sections discuss the basic phi-
between Dk and Xk, when a stimulus X’k losophy of the supervised learning environment
close to Xk elicits a response S’k sufficiently with reference to the simple artificial neuron and
close to Dk. the multilayer perceptron.
• Unsupervised learning: This paradigm
simply provides the system with an input Simple Artificial neuron
Xk, and allows it to self organize/self super-
vise its parameters to generate internal pro- The basic computational element of an artificial
totypes of the sample vectors. Such a para- neural network model is often referred to as a node
digm attempts to represent the entire data or unit. It receives inputs from some other units,
set by employing a smaller number of pro- or perhaps from an external source. The basic
totypical vectors. These prototypes are in a function of a single neuron is to add up its inputs,
state of continual updating as newer system and to produce an output if this sum is greater than
inputs enter into the system. This is often some value, known as the threshold value. The
driven by a complex competitive-cooper- basic neuron model, also known as “perceptron”
ative process where the individual neurons after Frank Rosenblatt [Haykin99, Rosenblatt58],
compete and cooperate with each other to is a binary classifier that maps real-valued vectored
4
Soft Computing and its Applications
5
Soft Computing and its Applications
E= ∑ (y i
− di )2 (4) During the training process of neural networks
i by the gradient descent mechanism, the gradient
G of the loss function (E) with respect to each
In words, it is the sum of the squared difference weight wijof the network is computed [Haykin99,
between the target value di and the perceptron’s Snyman2005]. This gradient indicates as to how
prediction yi (calculated from the input value xi) the small changes in the network weights affect
computed over all points i in the data set. For a the overall error E. Let, the loss function be rep-
linear model, the sum-squared error is a quadratic resented for each p training sample, as
function of the model parameters. The loss func-
tion E provides an objective measure of predictive
6
Soft Computing and its Applications
Table 1.
Begin
Initialize interconnection weights and threshold
Set wi(t=0), (0<= i <= n), to small random values Interconnection weights from input i at time t
Set w0 := k k is the bias in the output node
Set x0:=1 x is the input to the network
E= ∑E p
(5) The first factor of equation 8 can be obtained
p
by differentiating equation 6.
where, δE
= −(d o − yo ) (9)
δyo
1
Ep = ∑ (dop − yop )2
2 o
(6)
Using yo = ∑w oj
y j , the second factor be-
j
Hence,
Generalizing for all training samples and
decomposing the gradient into two factors using δE
chain rule, one gets = −(d o − yo ) yi (12)
δwoi
δE δE δyo
= (8) The gradient G for the entire data set can be
δwoi δyo δwoi
obtained by summing at each weight the contribu-
7
Soft Computing and its Applications
tion given by equation 12 over all the data points. classes is much more complex, the single layer
Then, a small proportion μ (called the learning rate) model fails abruptly.
of G is subtracted from the weights to achieve the
required gradient descent. Multilayer Perceptron
The Gradient Descent Algorithm In order to overcome the shortcomings of the single
layer perceptron model, the first and foremost
The loss function minimization procedure during way out is to resort to a multilayer model with the
the training of a neural network involves the com- threshold function slightly smoothed out to provide
putation of the gradient of the loss function with some information about the nonlinearly separable
time. The algorithm shown in Table 2 illustrates inputs [Haykin99]. This means that the network
the steps in determining the gradient for attain- will be able to adjust the weights as and when
ing the minimum of the loss function [Haykin99, required. A possible multilayer neural network
Snyman2005]. model comprising of three layers of nodes, viz.,
The algorithm terminates once the minimum the input layer node, the hidden layer node and the
of the error function, i.e., G=0 is reached. At this output layer node along with their characteristic
point the algorithm is said to have converged. activation functions and inter-connection weights,
An important consideration is the learning rate is shown in Figure 3. In the figure, an extra node
(μ), which determines by how much the weights with a nonlinear activation function has been
are changed at each step. If μ is too small, the inserted between input and output. Since such a
algorithm will take a long time to converge. Con- node is “hidden” inside the network, it is com-
versely, if μ is too large, the algorithm diverges monly referred to as a hidden unit. The hidden unit
leading to imprecise learning. also has a weight from the bias unit. In general,
However, the single layer perceptrons suffer all non-input neural network units have such bias
from several limitations. They learn a solution weights. For simplicity however, the bias unit and
if there is a possibility of finding it. They can weights are usually omitted from neural network
separate linearly separable classes easily enough, diagrams. The sole output layer node shown in
but in situations where the division between the the figure is characterized by a linear activation
function. Since the input layer node acts only as a
Table 2.
Begin
Initialize wij to small random values wij are the interconnection weights
Repeat until done
For each weight wij set Dwij:=0
For each data point (x, t)p
Set input units to x
Compute value of output units
For each weight wij set Dwij:= Dwij+(di - yi)yj
For each weight wij set wij:=wij+m Dwij m is the learning rate
End
8
Soft Computing and its Applications
9
Soft Computing and its Applications
cycles. Networks that respect this constraint are deh65] explains the varied nature of ambiguity
called feedforward networks and their connection and uncertainty that exist in the real world. This
pattern forms a directed acyclic graph or dag. is in sheer contradiction to the concept of crisp
For the purpose of training a multilayer feedfor- sets, where information is more often expressed
ward neural network by gradient descent, a training in quantifying propositions.
dataset consisting of pairs (x, d) is considered. The underlying concept behind the notion of
Here, vector x represents an input pattern to the fuzzy sets [Zadeh65] is that each and every observa-
network and vector d is the corresponding target tion, exists with a varied degree of containment in
output. The corresponding learning algorithm the universe of discourse. This degree of contain-
employing backprop is illustrated as shown in ment is referred to as the membership value of the
Table 3. observation. A fuzzy set is a mapping from an input
Neural networks are powerful and robust tools universe of discourse into the interval [0, 1] that
for the retrieval of incomplete data, finding pat- describes the membership of the input variable.
terns in datasets, and to mimic the human behavior This is referred to as fuzzification. The reverse
when it comes to the analysis and interpretation mechanism to revert to the crisp world is termed
of data. As such, they find wide use in processing, as defuzzification. Thus, it can be inferred that
retrieval and recognition of data patterns.
Whereas, crisp sets quantify quantities, fuzzy sets
qualify qualities.
3. Fuzzy SEtS And Fuzzy logIC
Fuzzy logic is a collection of conditional
It may be pointed out that much of the informa- statements or fuzzy rules, which form the basis
tion available in the real world exhibit vagueness, of the linguistic reasoning framework, which
imprecision and uncertainty. In fact, the fuzzy embodies representation of shallow knowledge.
sets approach fits in with the linguistic modes The fundamental atomic terms in this linguistic
of reasoning that are natural to human beings. or natural language-reasoning framework are
The fuzzy set theory, introduced by Zadeh [Za- often modified with adjectives or “linguistic
10
Soft Computing and its Applications
Table 3.
Begin
dE dE
Define d j := the error signal for unit j; Dwij := − the (negative) gradient for weight wij.
d net j d wij
Let Ai={j:$ wij} represents the set of preceding nodes to unit i and Pj={i: $ wij} represents the set of succeeding nodes
to unit j of the network.
Dwij := d i y ' j
So,
Forward activation
Remark: The activity of the input units is determined by the network’s external input x. For all other units, the activity
is propagated forward as
yi := fi ( ∑ wij y j )
j∈ Ai
Remark: The activity of all the preceding nodes Ai to unit i must be known before calculating the activity of i.
Error backpropagation
Remark: The output error is propagated back for deriving the errors of the hidden units in terms of the succeeding
nodes.
d E d neti d y j
d j := − ∑ the first factor is the error of node i.
i∈Pj d neti d y j d net j
d neti d d yj d f j (net j )
The second factor,
d yj
:=
d yj
∑w
k∈ Ak
ik yk := wij ; The third factor,
d net j
:=
d net j
:= f '(net j )
If, the hidden units are characterized by the tanh activation function, then
∑
f '(neth ) := 1 − yh2 ; So, d j := f '(net j ) d i wij
i∈Pj
End
hedges”. These linguistic hedges have the effect which is similar to natural language expressions,
of modifying the membership function for a basic can be written as
atom. The general form of a fuzzy rule [Zadeh65],
11
Soft Computing and its Applications
IF premise (antecedent) THEN conclusion (con- that represents distinct α-cuts of a given fuzzy set
sequent) A, is called a level set of A, i.e.,
12
Soft Computing and its Applications
| A |= ∑µ A
( x) (25)
x∈U
13
Soft Computing and its Applications
14
Soft Computing and its Applications
these algorithms is to search the population space • Stochastic optimization, which is an um-
by the application of the embedded operators so as brella set of methods that includes simu-
to arrive at an optimal solution space. Generally, lated annealing and numerous other
two types of operators are used, viz. reproduc- approaches.
tion and evolution. The reproduction operator is
guided by a selection mechanism. The evolution The following sections discuss the operational
operator includes the crossover and mutation procedures of three popular heuristic search tech-
operators. The search technique is implemented niques, viz., genetic algorithms, simulated anneal-
through a series of iterations, whereby the differ- ing and ant colony optimization.
ent operators are applied in a loop on the initial
population. Each iteration is referred to as a gen- genetic Algorithms
eration. Each generation produces a new solution
space of parent individuals, which are selectively Genetic algorithms (GAs) [Goldberg89, Davis91,
chosen for participating in the next generation of Michal92, Bandyopadhyay2007a] are efficient,
the optimization procedure. The selection of the adaptive and robust multi-point search and op-
participating parents for the next generation is timization techniques guided by the principles
decided by a figure of merit, often referred to as of evolution and natural genetics. They provide
the objective function. This objective function is parallel near optimal and solutions of an objec-
entrusted with the evaluation of the fitness of the tive or fitness function in complex, large and
candidate solutions in a particular generation to multimodal landscapes. GAs are modeled on
qualify for the next generation of operations. Other the principles of natural genetic systems, where
notable and related search techniques include: the genetic information of each individual or
potential solution is encoded in structures called
• Quantum annealing [Apolloni89, chromosomes. They use some domain or problem
Das2005], which uses “quantum fluctua- dependent knowledge for directing the search in
tions” instead of thermal fluctuations to get more promising areas of the solution space; this
through high but thin barriers in the target is known as the fitness function. Each individual
function. or chromosome has an associated fitness function,
• Tabu search [Glover97], which normally which indicates its degree of goodness with respect
moves to neighboring states of lower en- to the solution it represents. Various biologically
ergy, but takes uphill moves when it finds inspired operators like selection, crossover and
itself stuck in a local minimum; and avoids mutation are applied on the chromosomes to yield
cycles by keeping a “taboo list” of solu- potentially better solutions.
tions already seen.
• Ant colony optimization (ACO) Basic Principles and Features
[Colorni91, Dorigo92], which uses many
ants (or agents) to traverse the solution A GA essentially comprises a set of individual
space and find locally productive areas. solutions or chromosomes (called the population)
• Harmony search, which mimics musicians and some biologically inspired operators that
in improvisation process where each mu- create a new (and potentially better) population
sician plays a note for finding a best har- from an old one. The different steps of a GA can
mony all together. be represented as follows.
15
Soft Computing and its Applications
16
Soft Computing and its Applications
For example, let the two parents and the crossover the population size, probabilities of performing
points be as shown below. crossover (usually kept in the range 0.6 to 0.8) and
mutation (usually kept below 0.1) and the termina-
1 0 0 1 1 | 0 1 0 tion criteria. Moreover, one must decide whether
0 0 1 0 1 | 1 0 0 to use the generational replacement strategy where
the entire population is replaced by the new popula-
After crossover the offspring will be the fol- tion, or the steady state replacement policy where
lowing: only the less fit individuals are replaced. Most of
such parameters in GAs are problem dependent,
1 0 0 1 1 1 0 0 and no guidelines for their choice exist in the lit-
0 0 1 0 1 0 1 0 erature. Therefore, several researchers have also
kept some of the GA parameters variable and/or
Some other common crossover techniques are adaptive [Baker85, Srinivas94].
two-point crossover, multiple point crossover, The cycle of selection, crossover and muta-
shuffle-exchange crossover and uniform crossover tion is repeated a number of times till one of the
[Davis91]. following occurs:
17
Soft Computing and its Applications
them to configurations with lower internal energy which implies that the system may migrate to the
than the initial one. new state even when it has a higher energy than
Similarly, in the SA algorithm, each step re- the current one. This prevents the process from
places the current solution by a random neighbor- becoming stuck in a local minimum. When T goes
ing solution. The probability of choosing such a to zero, the acceptance probability tends to zero if
neighboring solution depends on the difference g’ > g. P however, attains a positive value if g’ < g.
between the corresponding energy function Thus, the system favors transitions that go
values and a global parameter T referred to as to lower energy values, and avoid those that go
the temperature. This temperature is gradually higher for sufficiently small values of T. When
decreased during the cooling process. Thus, the T becomes 0, the procedure will ensure making
current solution changes almost randomly when the migration only if it goes to lower energy.
T is large, but the rate of change of states goes Thus, it is clear that the evolution of the state
down as T is reduced. depends crucially on the temperature T. Roughly
speaking, the evolution of a state ρ is sensitive to
Overview of SA coarser changes of energy variations when T is
large, and to finer changes of energy variations
The basic objective of SA is to minimize the system when T is small.
internal energy function F(ρ), where ρ corresponds
to each point of the search space. Thus, it aims to The Annealing Schedule
bring the system, from an arbitrary initial state, to Another essential feature of the SA method is that
a state with the minimum possible energy. the temperature (T) should be gradually reduced as
For this purpose, SA considers some user the simulation proceeds [Kirkpatrick83, Cerny85].
specified neighbors ρ’ of the current state ρ, and Initially, T is set to a high value (i.e. ∞), and then
probabilistically decides between migrating the it is decreased at each step according to some
system to either state ρ’ or to retain in state ρ. annealing schedule. The user generally specifies
The probabilities are chosen such that the system this schedule for the decrement of T. However, it
ultimately migrates to lower energy states. This must be ensured that the schedule should be such
step is repeated until the system reaches a state that it would end up with T = 0 towards the end
that is a good approximation to the required one, of the annealing process.
or until a prespecified limit to the approximation Thus, the system is expected to migrate initially
has been reached. The following sections highlight towards a broader region of the search space con-
the important aspects of simulated annealing. taining good solutions ignoring smaller features of
the energy function in the process. Subsequently,
Acceptance Probabilities it would drift towards the lower energy regions
This is the probability of migrating from the that become narrower and narrower. Finally, the
current state ρ to a candidate new state ρ’. This system migrates downhill according to the steepest
is decided by an acceptance probability function descent heuristic. However, the pure version of SA
P(g, g’, T), where g = F(ρ) and g’ = F(ρ’) and T does not keep track of the best solution obtained
is the temperature (mentioned earlier). The ac- in terms of the lower energy levels attained at
ceptance probability is usually chosen such that any point of time.
the probability of allowing a transition decreases
when the difference (g’− g) increases. This means SA Pseudocode
that smaller uphill migrations are more likely than The pseudocode shown in Table 4 implements
the larger ones. P must be nonzero when g’ > g, the simulated annealing heuristic, as described
18
Soft Computing and its Applications
above, starting from state ρ0 and continuing to a implementation of the algorithm. In other words,
maximum of kmax steps or until a state with energy the diameter of the search graph must be small to
gmax or less is found. facilitate faster transitions between states. Hence,
the choice of the search graph diameter is an es-
Selection of Operating sential criterion for successful operation of the
Parameters of SA simulated annealing algorithm.
Table 4.
19
Soft Computing and its Applications
that those candidate states ρ’ for which P[F(ρ), gbest, respectively and the annealing schedule is
F(ρ’), T] is large should be opted for first. restarted. The decision to restart could be based
on a fixed number of steps, or based on the current
Avoidance of Getting Stuck to Local Minima energy being too high from the best energy so far.
Another aspect of the selection of the candidate
generator function is to reduce the number of local Ant Colony optimization
minima, which may come up during the anneal-
ing process. Otherwise, the SA algorithm may be The ant colony optimization algorithm (ACO) is
trapped in these minima with a high probability another probabilistic computational search tech-
for a very long time. The probability of occurrence nique useful for finding the best possible paths
of such traps is proportional to the number of in search graphs. It is a member of the family of
states achieved in the local minimal state of SA. ant colony algorithms, which are referred to as
The time of trapping is exponential on the energy swarm intelligence methods. Initially proposed
difference between the local minimal state and its by Marco Dorigo in 1992 in his PhD thesis [Co-
surrounding states. These requirements, however, lorni91, Dorigo92], the first algorithm was aimed
can be met by resorting to slighter changes to the at searching for an optimal path in a graph based
candidate generator function. on the behaviors exhibited by ants while searching
for food out of their colony.
Cooling Schedule
The simulated annealing algorithm assumes that Overview of ACO
the cooling rate is always low enough such that
the probability distribution of the current state In their quest for food, ants generally start search-
remains near the thermodynamic equilibrium at ing randomly through all possible paths that lead
all times. But the time required for attaining the to food. Once food is found, they return to their
equilibrium state (referred to as the relaxation time) colony leaving behind pheromone trails. If the
after a change in temperature strongly depends on following ants make the path, they do not hover
the nature of the energy function, the current tem- randomly, instead follow the trail to find food.
perature (Tc) as well as on the candidate generator. This pheromone trail however, starts to evapo-
Hence, there is no basis for selecting an ideal rate with time, thereby reducing its attractive
cooling rate for the algorithm. It should be esti- strength. The more time it takes for an ant to travel
mated and adjusted empirically for a particular down the path and back again, the more time the
problem. pheromones have to evaporate. A short path, by
However, this problem has been taken care of comparison, gets marched over faster, and thus
by the thermodynamic simulated annealing algo- the pheromone density remains high as it is laid
rithm, which adjusts the temperature at each step on the path as fast as it can evaporate.
based on the energy difference between the two The phenomenon of pheromone evaporation
states according to the laws of thermodynamics prevents the convergence of the algorithm to a
instead of applying any cooling schedule. local optimum. In absence of any pheromone
evaporation the following ants would always be
Restarting of SA attracted to the paths traversed by the first/leading
Sometimes it is better to revert back to a solution ant thereby leading to a constrained solution space.
that was significantly better rather than always This procedure adopted by real world ants is
migrate from the current state. This is called adapted in implementing the ant colony optimiza-
restarting. To do this ρ and g are set to ρbest and tion algorithm, which always leads to the short-
20
Soft Computing and its Applications
where, ψij(t+1) is the amount of pheromone on • The lower approximation of P with respect
that edge at time (t+1), ε is the rate of pheromone to R is the set of all objects which can be
evaporation and Δψij is the amount of pheromone certainly classified as belonging to P with
deposited. For the kth ant traveling on edge(i, j), respect to R,
• the upper approximation of P with respect
to R is the set of all objects which can be
21
Soft Computing and its Applications
R* ( x) = {R( x) : R( x) ⊆ P}
x∈U
(32)
R* ( x) = {R( x) : R( x) ∩ P ≠ ϕ}
x∈U
(33)
RN R ( x) = R* ( x) − R* ( x) (34)
22
Soft Computing and its Applications
23
Soft Computing and its Applications
Efforts have also been made by inducing proper analysis of uncertainty and vagueness in
multivalued logical reasoning to the existing color image information. Color image segmen-
neural networks for the purpose of classification tation techniques involving fuzzy set theory and
of multidimensional data. Multilevel thresholding fuzzy logic are also available in the literature
ability on input data has been induced in discrete- [Gillet2002, Chung2003, Ito95, Choi95].
time neural networks by means of synthesis of
multi-level threshold functions [Si91]. Other soft Soft Computing Applications to
computing techniques like genetic algorithms have kansei Information Processing
also been widely used in the domain of medical
image segmentation. An extensive review is avail- Another noteworthy application area where soft
able in [Maulik2009]. computing has made a mark is Kansei information
Of late, the processing of multichannel in- processing [Schutte2004]. Kansei engineering
formation has assumed great importance mainly invented in the 1970s by Professor Mitsuo Na-
due to the evolving fields of remote sensing, GIS, gamachi of Hiroshima International University, is
biomedical imaging and multispectral data man- a method for translating feelings and impressions
agement. Color image segmentation is a classical into product parameters. Kansei engineering es-
example of multichannel information processing. sentially deals with the processing and manipula-
The complexity of the problem of color image tion of nonverbal information such as voice pitch,
segmentation and object extraction is mainly due facial expressions, gestures as well as verbal
to the variety of the color intensity gamut. Due to information. Kansei encompass the total concept
the inherent parallelism and ability of approxima- of senses, consciousness, and feelings that relate
tion, adaptation and graceful degradation, neural to human behavior in social living.
networks are also suited for addressing the problem Since Kansei information has subjectivity,
of color image processing. Lee et al. [Lee96] em- human linguistic understanding plays a central
ployed a CNN multilayer neural network structure role in Kansei information processing. Fuzzy set
for processing of color images following the RGB theory [Zadeh65] has been effectively applied for
color model. In this approach, each primary color explaining the ambiguity and vagueness in Kan-
is assigned to a unique CNN layer allowing paral- sei information [Yan2008]. In [Hayashida2002],
lel processing of color component information. Hayashida and Takagi proposed an evolution-
Zhu [Zhu94] however, transferred the RGB color ary computation (EC) and interactive EC (IEC)
space to Munsell color space and used an adaptive system for visualizing individuals in a multi-
resonance theory (ART) neural network to clas- dimensional searching space thereby enabling the
sify objects into recognition categories. Roska envisioning of the landscape of an n-dimensional
et al. [Roska93] also applied a multilayer CNN searching space. The proposed system has been
structure for handling the enormous amount of found to be efficient than that obtained with the
data involved in the processing of color images. conventional genetic algorithm. Unehara and
Self-organizing neural network architectures Yamada [Unehara2008] introduced an Interactive
[Moreira96, Wu2000] have also been used for Genetic Algorithm (IGA) for the designing of
segmentation and classification of color images. products’ shapes through evaluation by Kansei.
The vagueness in image information arising out of Other notable applications of soft computing in
the admixtures of the color components has often Kansei information processing can be found in
been dealt with the soft computing paradigm. In the literature [Onisawa2005, Nagata94].
[Chen95], Chen et al. applied fuzzy set theory for
24
Soft Computing and its Applications
other Soft Computing Applications uncertainties in data without even requiring any
a priori information regarding the data content
Recently, soft computing tools are being widely and distribution. These techniques are applied
used in the domain of data mining and bioinfor- extensively in a wide variety of engineering and
matics. In data mining, genetic algorithms, neural scientific problems with astounding results.
networks and fuzzy logic are widely used in solving
many problems of optimization, feature selection,
classification and clustering [Bigus96, Cox2005, REFEREnCES
Bandyopadhyay2005]. In the domain of bioinfor-
matics, soft computing tools have been used for Abdallah, M. A., Samu, T. I., & Grisson, W. A.
sequence alignment, fragment assembly, gene (1995). Automatic target identification using
and promoter identification, phylogenetic tree neural networks. SPIE Proceedings on Intelligent
analysis, prediction of gene regulatory network, Robots and Computer Vision XIV, 2588, 556–565.
protein structure and function prediction, protein Ackley, D. H., Hinton, G. E., & Sejnowski, T.
classification, molecule design and docking, to J. (1985). A learning algorithm for Boltzmann
name just a few [Bandyopadhyay2007a, Bandyo- Machines. Cognitive Science, 9, 147–169. doi:.
padhyay2007b]. Web intelligence is another recent doi:10.1207/s15516709cog0901_7
area that has seen a spurt of successful application
of soft computing tools [Bandyopadhyay2007a]. Amari, S. (1988). Mathematical theory of self-
organization in neural nets. In Seelen, W. V.,
Shaw, G., & Leinhos, U. M. (Eds.), Organization
7. ConCluSIon of Neural Networks: Structures and Models. New
York: Academic Press.
A brief overview of the essence and applications
Antonucci, M., Tirozzi, B., & Yarunin, N. D.
of the soft computing paradigm is presented in
(1994). Numerical simulation of neural networks
this chapter. This computing framework is es-
with translation and rotation invariant pattern
sentially built around several intelligent tools and
recognition. International Journal of Modern
techniques. Notable among them are the neural net-
Physics B, 8(11-12), 1529–1541. doi:.doi:10.1142/
works, fuzzy sets, fuzzy logic, genetic algorithms
S0217979294000658
and rough sets. These tools form the backbone of
this widely used computing paradigm meant for Apolloni, B., Caravalho, C., & De Falco, D. (1989).
analysis and understanding of knowledge bases. Quantum stochastic optimization. Stochastic Pro-
Neural networks are used for the analysis of cesses and their Applications, 33, 233–244. doi:.
numeric data. These networks comprise intercon- doi:10.1016/0304-4149(89)90040-9
nected processing units or nodes in different to-
Baker, J. E. (1985). Adaptive selection methods
pologies depending on the problem at hand. Fuzzy
for genetic algorithms In J. J. Grefenstette (Ed.),
sets and fuzzy logic deal with the uncertainties
Proceedings of 1st International Conference on
inherent in data under consideration by represent-
Genetic Algorithms (pp. 101-111). Hillsdale, NJ:
ing the degree of vagueness in terms of linguistic
Lawrence Erlbaum Associates.
information. Genetic algorithm, simulated an-
nealing algorithm and ant colony optimization Bandyopadhyay, S., Maulik, U., Holder, L., &
algorithm are random search techniques meant Cook, D. (Eds.). (2005). Advanced Methods
for arriving at an optimum probable solution to for Knowledge Discovery from Complex Data.
a problem. Rough sets also handle the underlying London: Springer.
25
Soft Computing and its Applications
Bandyopadhyay, S., Maulik, U., & Wang, J. T. Boskovitz, V., & Guterman, H. (2002). An adap-
L. (Eds.). (2007). Analysis of Biological Data: tive neuro-fuzzy system for automatic image
A Soft Computing Approach. Singapore: World segmentation and edge detection. IEEE Trans-
Scientific. actions on Fuzzy Systems, 10(2), 247–262. doi:.
doi:10.1109/91.995125
Bandyopadhyay, S., & Pal, S. K. (2007). Classi-
fication and Learning Using Genetic Algorithms: Broomhead, D. S., & Lowe, D. (1988). Multivari-
Application in Bioinformatics and Web Intel- ate functional interpolation and adaptive networks.
ligence. Germany: Springer. Complex Systems., 2, 321–355.
Baraldi, A., Binaghi, E., Blonda, P., Brivio, P. A., Carpenter, G. A. (1989). Neural network mod-
& Rampini, A. (2001). Comparison of the multi- els for pattern recognition and associative
layer perceptron with neuro-fuzzy techniques in memory. Neural Networks, 2(4), 243–258. doi:.
the estimation of cover class mixture in remotely doi:10.1016/0893-6080(89)90035-X
sensed data. IEEE Transactions on Geoscience
Carpenter, G. A., & Grossberg, S. (1991). Pattern
and Remote Sensing, 39(5), 994–1005. doi:.
recognition by self-organizing neural networks.
doi:10.1109/36.921417
Cambridge, MA: MIT Press.
Bezdek, J. C. (1992). On the relationship be-
Carpenter, G. A., & Ross, W. D. (1995). ART-
tween neural networks, pattern recognition and
EMAP: A neural network architecture for object
intelligence. International Journal of Approxi-
recognition by evidence accumulation. IEEE
mate Reasoning, 6, 85–107. doi:10.1016/0888-
Transactions on Neural Networks, 6(4), 805–818.
613X(92)90013-P
doi:.doi:10.1109/72.392245
Bhattacharyya, S., & Dutta, P. (2006). Designing
Cerny, V. (1985). A thermodynamical approach to
pruned neighborhood neural networks for object
the traveling salesman problem: an efficient simu-
extraction from noisy background. Journal of
lation algorithm. Journal of Optimization Theory
Foundations of Computing and Decision Sciences,
and Applications, 45, 41–51. doi:.doi:10.1007/
31(2), 105–134.
BF00940812
Bhattacharyya, S., Dutta, P., & Maulik, U. (2008).
Charalampidis, D., Kasparis, T., & Georgiopoulos,
Self Organizing Neural Network (SONN) based
M. (2001). Classification of noisy signals using
gray scale object extractor with a Multilevel Sig-
fuzzy ARTMAP neural networks. IEEE Transac-
moidal (MUSIG) activation function. Journal of
tions on Neural Networks, 12(5), 1023–1036. doi:.
Foundations of Computing and Decision Sciences,
doi:10.1109/72.950132
33(2), 131–165.
Chauvin, Y., & Rumelhart, D. E. (1995). Backpropa-
Bigus, J. P. (1996). Data Mining With Neural
gation: theory, architectures, and applications.
Networks: Solving Business Problems from Ap-
Hillsdale, NJ: L. Erlbaum Associates Inc.
plication Development to Decision Support. New
York: Mcgraw-Hill. Chen, C. T., Chen, K. S., & Lee, J. S. (2003). Fuzzy
neural classification of SAR images. IEEE Transac-
Bilbro, G. L., White, M., & Synder, W.(n.d.). Image
tions on Geoscience and Remote Sensing, 41(9),
segmentation with neurocomputers. In Eckmiller,
2089–2100. doi:.doi:10.1109/TGRS.2003.813494
R., & Malsburg, C. V. D. (Eds.), Neural comput-
ers. New York: Springer-Verlag.
26
Soft Computing and its Applications
Chen, Y., Hwang, H., & Chen, B. (1995). Color Davis, L. (Ed.). (1991). Handbook of Genetic
image analysis using fuzzy set theory. Proceedings Algorithms. New York: Van Nostrand Reinhold.
of International Conference on Image Processing.
Deneubourg, J.-L., Aron, S., Goss, S., & Pasteels,
(pp. 242-245).
J.-M. (1990). The self-organizing exploratory
Chiu, W. C., Hines, E. L., Forno, C., Hunt, R., & pattern of the Argentine ant. Journal of Insect
Oldfield, S. (1990). Artificial neural networks for Behavior, 3, 159. doi:.doi:10.1007/BF01417909
photogrammetric target processing. SPIE Pro-
Dorigo, M. (1992). Optimization, Learning and
ceedings on Close-Range Photogrammetry Meets
Natural Algorithms. PhD thesis. Politecnico di
Machine Vision, 1395(2), 794–801.
Milano, Italie.
Choi, Y., & Krishnapuran, R. (1995). Image en-
Dorigo, M., & Gambardella, L. M. (1997). Ant
hancement based on fuzzy logic. Proceedings of
Colony System: A Cooperative Learning Ap-
International Conference on Image Processing.
proach to the Traveling Salesman Problem. IEEE
(pp. 167-170).
Transactions on Evolutionary Computation, 1(1),
Chua, L. O., & Yang, L. (1988). Cellular net- 53–66. doi:.doi:10.1109/4235.585892
work: Applications. IEEE Transactions on
Duda, R. O., & Hart, P. E. (1973). Pattern clas-
Circuits and Systems, 35(10), 1273–1290. doi:.
sification and scene analysis. New York: Wiley.
doi:10.1109/31.7601
Egmont-Petersen, M., & Arts, T. (1999). Rec-
Chua, L. O., & Yang, L. (1988). Cellular
ognition of radiopaque markers in X-ray images
network: Theory. IEEE Transactions on Cir-
using a neural network as nonlinear filter. Pat-
cuits and Systems, 35(10), 1257–1282. doi:.
tern Recognition Letters, 20(5), 521–533. doi:.
doi:10.1109/31.7600
doi:10.1016/S0167-8655(99)00024-0
Chung, F., & Fung, B. (2003). Fuzzy color quan-
Egmont-Petersen, M., de Ridder, D., & Handels, H.
tization and its application to scene change detec-
(2002). Image processing using neural networks - a
tion. Proceedings of MIR, 03, 157–162.
review. Pattern Recognition, 35(10), 2279–2301.
Colorni, A., Dorigo, M., & Maniezzo, V. (1991). doi:.doi:10.1016/S0031-3203(01)00178-9
Distributed Optimization by Ant Colonies, actes
Eshelman, L. J., & Schaffer, J. D. (1993). Real-
de la première conférence européenne sur la
coded genetic algorithms and interval schemata.
vie artificielle (pp. 134–142). France: Elsevier
In L. Whitley (Ed.), Foundations of Genetic
Publishing.
Algorithms (pp 187-202). 2, San Mateo, CA:
Cortes, C., & Vapnik, V. N. (1995). Support vector Morgan Kaufmann.
networks. Machine Learning, 20, 273–297. doi:.
Forrest, B. M. (1988). Neural network models. Par-
doi:10.1007/BF00994018
allel Computing, 8, 71–83. doi:.doi:10.1016/0167-
Cox, E. (2005). Fuzzy Modeling and Genetic Al- 8191(88)90110-X
gorithms for Data Mining and Exploration. San
Frege, G. (1893). Grundlagen der Arithmetik (Vol.
Francisco: Morgan Kaufmann.
2). Jena: Verlag von Herman Pohle.
Das, A., & Chakrabarti, B. K. (Eds.). (2005).
Quantum Annealing and Related Optimization
Methods. Lecture Note in Physics (Vol. 679).
Heidelberg: Springer. doi:10.1007/11526216
27
Soft Computing and its Applications
Fukushima, K. (1980). Neocognitron: A self- Ito, N., Shimazu, Y., Yokoyama, T., & Matushita,
organizing multilayer neural network model for Y. (1995). Fuzzy logic based non-parametric color
a mechanism of pattern recognition unaffected image segmentation with optional block processing
by shift in position. Biological Cybernetics, 36, (pp. 119–126). ACM.
193–202. doi:.doi:10.1007/BF00344251
Kamgar-Parsi, B. (1995). Automatic target extrac-
Gillet, A., Macaire, L., Lecocq, C. B., & Postaire, tion in infrared images (pp. 143–146). NRL Rev.
J. G. (2002). Color image segmentation by analy-
Kirkpatrick, S., Gelatt, C. D., & Vecchi, M. P.
sis of 3D histogram with fuzzy morphological
(1983). Optimization by Simulated Annealing.
filters. Studies in Fuzziness and Soft Computing,
Science. New Series 220., 4598, 671–680.
122, 153–177.
Kohonen, T. (1995). Self-organizing maps. Spring-
Glover, F., & Laguna, M. (1997). Tabu Search.
er Series in Information Sciences, 30.
Norwell, MA: Kluwer.
Kosko, B. (1988). Bidirectional associative memo-
Goldberg, D. E. (1989). Genetic Algorithms:
ries. IEEE Transactions on Systems, Man, and Cy-
Search, Optimization and Machine Learning.
bernetics, 18(1), 49–60. doi:.doi:10.1109/21.87054
New York: Addison-Wesley.
Kumar, S. (2004). Neural networks: A classroom
Goss, S., Aron, S., Deneubourg, J.-L., & Pasteels,
approach. New Delhi: Tata McGraw-Hill.
J.-M. (1989). The self-organized exploratory pat-
tern of the Argentine ant. Naturwissenschaften, Lee, C.-C., & de Gyvez, J. P. (1996). Color image
76, 579–581. doi:.doi:10.1007/BF00462870 processing in a cellular neural-network environ-
ment. IEEE Transactions on Neural Networks, 7(5),
Hayashida, N., & Takagi, H. (2002). Acceleration
1086–1098. doi:.doi:10.1109/72.536306
of EC convergence with landscape visualization
and human intervention. doi:10.1016/S1568- Leondes, C. T. (1998). Neural network techniques
4946(01)00023-0. and applications. In Image processing and pattern
recognition. New York: Academic Press.
Haykin, S. (1999). Neural networks: A compre-
hensive foundation (2nd ed.). Upper Saddle River, Lin, J.-S., Cheng, K.-S., & Mao, C.-W. (1996).
NJ: Prentice Hall. A fuzzy Hopfield neural network for medi-
cal image segmentation. IEEE Transactions
Hertz, J., Krogh, A., & Palmer, R. G. (1991).
on Nuclear Science, 43(4), 2389–2398. doi:.
Introduction to the theory of neural computation.
doi:10.1109/23.531787
Reading, MA: Addison-Wesley.
Lippmann, R. P. (1987). An introduction to com-
Hopfield, J. J. (1984). Neurons with graded response
puting with neural nets. IEEE ASSP Magazine,
have collective computational properties like those
3–22.
of two state neurons. In Proceedings of Nat (pp.
3088–3092). U. S: Acad. Sci. Lippmann, R. P. (1989). Pattern classification
using neural networks. IEEE Communications
Hui, S., & Zak, S. H. (1992). Dynamical analysis
Magazine, 27, 47–64. doi:.doi:10.1109/35.41401
of the Brain-State-in-a-Box (BSB) neural model.
IEEE Transactions on Neural Networks, 3, 86–94. Maulik, U. (2009). Medical Image Segmentation
doi:.doi:10.1109/72.105420 using Genetic Algorithms. IEEE Transactions
on Information Technology in Biomedicine, 13,
166–173. doi:.doi:10.1109/TITB.2008.2007301
28
Soft Computing and its Applications
Michalewicz, Z. (1992). Genetic Algorithms + Polkowski, L., & Skowron, A. (2001). Rough
Data Structures = Evolution Programs. New mereological calculi granules: a rough set ap-
York: Springer-Verlag. proach to computation. International Journal of
Computational Intelligence, 17, 472–479.
Moreira, J., & Costa, L. D. F. (1996). Neural-based
color image segmentation and classification us- Rosenblatt, F. (1958). The Perceptron: A Probabi-
ing self-organizing maps (pp. 47–54). Anais do listic Model for Information Storage and Organi-
IX SIBGRAPI. zation in the Brain. Cornell Aeronautical Labora-
tory. Psychological Review, 65(6), 386–408. doi:.
Nagata, T., Kakihara, K., Ohkawa, T., & Tobita,
doi:10.1037/h0042519
N. (1994). Concept Space Generation Oriented
Design Using Kansei by Individual Subjectivity. Roska, T., Zarandy, A., & Chua, L. O. (1993).
Journal of IEEJ, 116(4). Color image processing using multilayer CNN
structure. In Didiev, H. (Ed.), Circuit theory and
Onisawa, T., & Unehara, M. (2005). Application
design. New York: Elsevier.
of Interactive Genetic Algorithm toward Human
Centered System. Journal of SICE, 44(1), 50–57. Ross, T. J., & Ross, T. (1995). Fuzzy logic with
engineering applications. New York: McGraw
Pao, Y. H. (1989). Adaptive pattern recognition
Hill College Div.
and neural networks. New York: Addison-Wesley.
Roth, M. W. (1990). Survey of neural network
Pawlak, Z. (1982). Rough sets. International
technology for automatic target recognition. IEEE
Journal of Computer and Information Sciences,
Transactions on Neural Networks, 1(1), 28–43.
11, 341–356. doi:.doi:10.1007/BF01001956
doi:.doi:10.1109/72.80203
Pawlak, Z., & Skowron, A. (1994). Rough mem-
Rumelhart, D. E., Hinton, G. E., & Williams,
bership function. In Yeager, R. E., Fedrizzi, M., &
R. J. (1986). Learning representations by back-
Kacprzyk, J. (Eds.), Advances in the Dempster-
propagating errors. Nature, 323, 533–536. doi:.
Schafer Theory of Evidence (pp. 251–271). New
doi:10.1038/323533a0
York: Wiley.
Schutte, S., Eklund, J., Axelsson, J. R. C., &
Perlovsky, L. I., Schoendor, W. H., & Burdick,
Nagamachi, M. (2004). Concepts, methods and
B. J. (1997). Model-based neural network for
tools in Kansei Engineering. Theoretical Issues
target detection in SAR images. IEEE Transac-
in Ergonomics Science, 5(3), 214–232. doi:.
tions on Image Processing, 6(1), 203–216. doi:.
doi:10.1080/1463922021000049980
doi:10.1109/83.552107
Scott, P. D., Young, S. S., & Nasrabadi, N. M. (1997).
Pham, D. T., & Bayro-Corrochano, E. J. (1998).
Object recognition using multilayer Hopfield neural
Neural computing for noise filtering, edge detec-
network. IEEE Transactions on Image Processing,
tion and signature extraction. Journal of Systems
6(3), 357–372. doi:.doi:10.1109/83.557336
Engineering, 2(2), 666–670.
Si, J., & Michel,A. N. (1991). Analysis and synthesis
Polkowski, L. (2002). Rough Sets, Mathemati-
of discrete-time neural networks with multi-level
cal Foundations. Advances in Soft Computing,
threshold functions. Proceedings of IEEE Interna-
Physica – Verlag, A Springer-Verlag Company.
tional Symposium on Circuits.
29
Soft Computing and its Applications
Snyman, J. A. (2005). Practical Mathematical Op- Unehara, M., & Yamada, K. (2008). Interactive
timization: An Introduction to Basic Optimization Conceptual Design Support System Using Hu-
Theory and Classical and New Gradient-Based man Evaluation with Kansei. Proceesings of 2nd
Algorithms. New York: Springer Publishing. International Conference on Kansei Engineering
and Affective Systems, (pp. 175-180).
Srinivas, M., & Patnaik, L. M. (1994). Adap-
tive probabilities of crossover and mutation in Wu, Y., Liu, Q., & Huang, T. S. (2000). An adap-
genetic algorithm. IEEE Transactions on Sys- tive self-organizing color segmentation algorithm
tems, Man, and Cybernetics, 24, 656–667. doi:. with application to robust real-time human hand
doi:10.1109/21.286385 localization. Proceedings of Asian Conference on
Computer Vision.
Stutzle, T., & Hoos, H. H. (2000). MAX MIN Ant
System. Future Generation Computer Systems, 16, Yan, H.-B., Huynh, V.-N., Murai, T., & Nakamori,
889–914. doi:10.1016/S0167-739X(00)00043-1 Y. (2008). Kansei evaluation based on prioritized
multi-attribute fuzzy target-oriented decision
Tang, H. W., Srinivasan, V., & Ong, S. H. (1996).
analysis. International Journal of Information
Invariant object recognition using a neural template
Sciences, 178(21), 4080–4093.
classifier. Image and Vision Computing, 14(7),
473–483. doi:.doi:10.1016/0262-8856(95)01065-3 Zadeh, L. A. (1965). Fuzzy sets. Information and
Control, 8, 338–353. doi:.doi:10.1016/S0019-
Tatem, A. J., Lewis, H. G., Atkinson, P. M., &
9958(65)90241-X
Nixon, M. S. (2001). Super-resolution target
identification from remotely sensed images using Zhang, G. P. (2000). Neural networks for clas-
a Hopfield neural network. IEEE Transactions on sification: a survey. IEEE Transactions on
Geoscience and Remote Sensing, 39(4), 781–796. Systems, Man and Cybernetics. Part C, Ap-
doi:.doi:10.1109/36.917895 plications and Reviews, 30(4), 451–462. doi:.
doi:10.1109/5326.897072
Tsao, E. C. K., Lin, W. C., & Chen, C.-T. (1993).
Constraint satisfaction neural networks for image Zhu, Z. (1994). Color pattern recognition in an
recognition. Pattern Recognition, 26(4), 553–567. image system with chromatic distortion. Optical
doi:.doi:10.1016/0031-3203(93)90110-I Engineering (Redondo Beach, Calif.), 33(9),
3047–3051. doi:.doi:10.1117/12.177509
Tzeng, Y. C., & Chen, K. S. (1998). A fuzzy neu-
ral network to SAR image classification. IEEE
Transactions on Geoscience and Remote Sensing,
36(1), 301–307. doi:.doi:10.1109/36.655339
30