An Introduction To Soft Computing M A Tool For Building Intelligent Systems
An Introduction To Soft Computing M A Tool For Building Intelligent Systems
Intelligent Systems
In this paper terms associated with soft computing are defined and its
main components are introduced. It is argued, using a number of
practical applications, that the hybrid approach of soft computing can
provide a methodology for increasing machine intelligence.
I. Introduction
One of the primary issues in artificial intelligence (AI) has been the choice between
two fundamentally different (and often viewed as competing) approaches to building
intelligent systems - - traditional symbolic AI and numeric (sub-symbolic) artificial
neural networks (ANNs). This has been an issue engaging the AI community for three
decades, and there have been attempts to bridge the gap between these two paradigms
in order to take advantage of the relative merits of each [1].
In an attempt to model the human mind/brain it has been necessary to oversimplify
the structure (resulting in ANNs) and the function (resulting in precisely defined
symbolic- AI programmes) of the brain. Symbolic AI attempts to pre-program
intelligence into a deterministic algorithm. On the other hand, most ANNs are equipped
with relatively weak forms of learning (i.e. tuning a fixed set of parameters or weights).
It has been argued [2] that despite the seemingly different approaches that symbolic AI
and ANNs take to building intelligent systems, they both share common origins, and
are both based on the hypothesis that cognition can be modelled by computation. Tasks
performed by ANNs can be performed by symbolic AI and vice versa as both
192
There are many features that can be attributed to an intelligent system. Among them
one can mention robustness, adaptivity, autonomy and the ability to communicate,
including man/machine communication in multiple modalities. Dealing with real-
world uncertainty or robustness is one of the most important characteristics of an
intelligent system (Fig 1).
Uncertainty arises from many sources among which are nonlinear behaviour, time-
varying behaviour (e.g. degradation over time) and interaction with uncertain
environments. All of these features are present in human behaviour and therefore are
important in the context of machines that interact with, co-operate with or replace
humans in certain tasks. Humans do not seem to be as affected by uncertainty in
sensory data as present-day computing machines. One explanation is that humans do
193
intelligent
machine
not rely directly on raw data for decision making but on abstract, uncertain rules, e.g. in
the rule 'if it is cold, put on an extra jumper', the actual temperature is not important,
neither is the season nor the time of day. A definite advantage of using abstract rules is
that a large amount of irrelevant information can be filtered out and the decision-
making process is simplified. This is particularly important in the context of machines
that rely on search techniques. It can be argued that in many real applications the
relevant information belongs to a class that is not well defined, and its membership
changes from time to time. To use the above example 'cold' means one thing today and
its meaning may change next week, or next month. A fixed rule may be able to deal
with this particular task, but it lacks the degree of adaptability required to work in
changing environments. Humans seem to have the ability to change with their
environment. Adaptive behaviour can be captured in a machine by using symbolic
meta-level rules. For example, a rule can be def'med that adjusts other rules according
to a mean temperature based on the season, such that 'if it is winter then the mean
temperature is 10~ ', and 'in summer the mean temperature is 30~ ', and define cold
relative to these. This provides a fixed meta-level rule and an adaptive base-level rule.
This is a partial solution, but what if there is an exceptional circumstance such as a
particularly cold winter. This highlights one of the shortcomings of such an approach,
namely that fixed symbolic meta-level rules can be restrictive in some circumstances.
The question that arises is how are these rules derived? Humans develop general
rules from specific observations and then generalise from specific instances to new
situations. For example, it can be seen that touching a specific hot object will result in
pain and personal injury, so a general rule is developed - - 'if an object is hot, do not
touch it'. In this case it is assumed that we have some sensory information received by
one or several of our five senses as to the temperature of an object. On the other hand,
generally, humans do not develop rules for recognising friends' faces. Picking a
familiar face in a crowd is performed instantaneously. Humans recognise vast numbers
of patterns and exhibit many skills without having to develop rules for them or even
know the rules that would result in such behaviours. Studies of the human brain have
194
shown that the pre-attentive processing of stimuli is carried out in as few as 70 to 100
ms. We look, see, pay attention and then recognise without using rules. Then we
address the higher cognitive functions such as reasoning, decision making, planning
and control using rules. An intelligent machine therefore should be able to combine
signal-level (sub-symbolic) intelligence with symbolic, more abstract level intelligence
(rules). In this sense intelligence can become a property of a hybrid dynamical system.
SC enables the pre-processing of sensory information, reasoning using symbolic
rules, and learning directly from observations. Adaptive systems must be able to
develop rules for themselves and update the rules in view of new sensory data, i.e. learn
from their experience. Leaming (adaptability) is the second important feature of an
intelligent system (Fig 1). Learning can be viewed as change in system behaviour
based on experience. From a dynamical system's point of view, learning is the rate of
change of an analytic function describing the system's behaviour [5]. As the analytic
function is a mapping from the input to the output space, it can therefore be a collection
of rules or a mathematical function. Clustering techniques are one aspect of learning
addressed by SC.
Application of SC to real-world problems has been aimed at increasing machine
intelligence quotient (MIQ). MIQ is measured by the level of control that a system can
have over its own operation (autonomy) (Fig 1). For example a robot that can navigate
its way around obstacles has a certain MIQ, another that can navigate and cope with
unforeseen moving obstacles has a higher MIQ. Another measure for MIQ is the
degree to which the machine assists humans in a particular task, e.g. a washing
machine that chooses its own program has a certain degree of MIQ, another one that
can program itself and use cheap electricity has a higher MIQ.
There are two reasons for using fuzzy logic in real applications. Firstly, in certain
circumstances the definition of the problem is vague and uncertain. The information
available does not lend itself readily to precise mathematical reasoning as in rule-
based systems. A second class of applications are well defined but a precise solution is
not necessary; the tolerance for imprecision can be exploited to simplify the solution.
Most of the applications of fuzzy logic today fall into the second category.
Fuzzy logic can be viewed as a superset of Boolean logic in the sense that it can
handle the concept of partial truth [6]. This concept has been used to develop more
general extensions such as fuzzy calculus and fuzzy differential equations. Fuzzy logic
is based on the principle of fuzzy subsets. In classical set theory, based on Boolean
logic, membership of a subset Ucan be defined as a mapping from the elements of a set
Sto the elements of another set L with two members: 0 and 1. So an element is mapped
to 0 if it is not a member of S and to 1 if it is. In fuzzy sets, a similar mapping exists to
a set F with the difference that F contains all values between 0 and 1. This gives rise to
the concept of degree of truth. A mapping to 0.3 is less true than a mapping to 0.6. The
degree of membership of a set is defined by a membership function ~t. Boundaries of
195
fuzzy subsets are not sharp but fuzzy and overlapping. This implies that a particular
entity A could be a member of two subsets with different degrees of membership - -
~I(A) and ~2(A).
There are two important concepts that are central to the application of fuzzy logic:
9 a linguistic variable;
9 fuzzy if-then rules.
A linguistic variable is a variable that takes linguistic values such as height, age,
speed, quality, etc. Such variables can take linguistic values like tall, young, fast, good,
etc. A linguistic value is a label for a fuzzy set. Within fuzzy sets, degree of
membership is characterised by membership functions, e.g. a membership function
'tall' determines the degree of tallness of someone of a certain height as shown in Fig
2. For example, it depicts a person who is 180 cm as tall with a degree of membership
0.45, while someone who is 185 cm tall is 'tall' with a degree of membership 0.6.
tall
0.6
!
~t
0.45
180 185
height, cm
This introduces the idea of linguistic hedges or modifiers such as very, more or less
and almost. Such terms are used extensively in natural language in a purely subjective
way and therefore do not have a universal definition within different applications.
196
However, once they have been defined, consistency can be ensured within a particular
application. Some common definitions are:
~very(A) = O,(A)2
~ A x B ( X , D = P.A(X) ^ gB (}5
where ^ is the conjunction operator usually defined as min(AxB). For example, a fuzzy
rule can express a simple rule of thumb. If X is tall then X is a good basketball
player'.This rule can be visualised in terms of two membership functions defining 'tall'
and 'good_basketball_player' as shown in Fig 3. Any player of any height has a certain
degree of being a good basketball player between 0 and 1.
tall good_basketball_player
1.0 1.0
! I
0.0 0.C
height quality of play
A collection of fuzzy rules can be represented as a fuzzy graph. Fuzzy rules can be
written as i f X is A i then Y is Bi, i = 1...n. For example, consider the following simple
rules:
9 ifXis small then Yis large;
9 i f X is medium then Y is small;
9 i f X is large then Y is large.
This system can be represented by a fuzzy graph, f * as shown in Fig 4. A fuzzy
graph represents a coarse characterisation of functional dependency between X and Y.
In this context, interpolation of rules becomes an important issue, i.e. what value of Y
results if the inputXis not a perfect match with any of the antecedent variables defined.
This is carried out by considering each fuzzy rule and its degree of truth. Then
defuzzification of the outputs is performed using one of many available techniques.
The most widely used is the centre of gravity method. Interpolation is one of the most
important features of fuzzy systems which can be exploited in situations where
complex functional relationships are to be represented by a small number of fuzzy
rules. This has been demonstrated in a number of complex industrial problems where
the number of fuzzy rules have been typically between 10 and 20 [7].
small
x
large large x large
One of the central issues in fuzzy logic is how to induce rules from observations.
This is the problem of obtaining deep structure from surface structure [8]. It is
relatively easy to write down a set of fuzzy rules to describe a particular behaviour.
However, to calibrate these rules, i.e. to choose the type and characterisation of the
membership functions, is not a trivial problem. A number of techniques have been used
to solve this problem, such as dynamic and gradient programming (developed for
multi-stage optimisation), genetic algorithms, reinforcement learning, and trial and
error [8].
198
The structure of the brain has been the subject of intense research in the past several
decades. Many of the pioneers of AI drew upon biological inspiration for their work.
Analogies were established between artificial processing elements and real neurons,
between network connection and axions, and between connection strengths and
synapses. A key aspect of the brain that ANNs try to imitate is its parallelism. ANNs
achieve this by using densely interconnected simple processing units to store and
process information. Each aspect of the neuron is represented mathematically by real
numbers. The basic processing unit, or artificial neuron, is characterised by a set of
input connections, a set of output connections, an activation level, an output level, and
a bias value (Fig 5).
w
XY=f(~:wi
~ xi+b~i) ~y
Fig. 5. A simple artificial neuron.
The output level of a neuron is determined according to a function of the activation
level, which is a weighted sum of the signal from the input connections. ANNs have
many characteristics such as nonlinear mapping, self-organisation and learning.
Learning in ANNs is viewed as the problem of finding a set of connection weights so
that given a set of inputs the desired outputs are generated. ANNs effectively perform a
parallel version of curve fitting and their capabilities should be assessed as adaptive
function approximators [5]. When viewed in this manner they are powerful tools that
can be used in an intelligent system to give it the learning capability. Many learning
algorithms have been proposed, mostly network-architecture-specific. Supervised-
learning algorithms rely on a teacher module to provide a set of training data which
contains the input and the associated expected output. The learning algorithm then
minimises the difference between the network output and the expected output. A
possible application is a function- learning task. Unsupervised-learning algorithms take
only the input patterns as training data and try to organise the neurons which best
classify the data [9]. With reinforce-ment-leaming algorithms, instead of providing a
desired output for each input as in supervised learning, only a scalar reinforcement
signal is used, which may be available only occasionally. Typical application of
reinforcement learning algorithms is in automatic control applications like the pole
balancing problem [9].
A most common ANN architecture, called the feedforward net, arranges neurons
into layers, namely input, hidden and output layers. Connections are restricted to the
area between neurons in different layers. Many learning algorithms have been
developed to train such neural networks. Most applications that employ feedforward
net use the backpropagation algorithm for learning. In a multi-layer network, the input
199
5. Genetic Algorithms
Genetic algorithms (GAs) are search techniques which derive their inspiration from
biological natural selection and genetics. The starting point is a population of
individuals, each representing a possible solution to a problem. Each individual is
allocated a fitness measure according to the quality of the solution it produces. The
fittest individuals survive to the next generation while the individuals that produce
unsatisfactory solutions are eliminated. This represents survival of the fittest. The
transition from one generation to the next is by means of reproduction among the
survived individuals only. The reproduction results in new individuals as offspring who
share some features taken from each parent.
A basic GA processes a finite population of binary strings. There are three basic
operations - - selection, crossover and mutation. Selection chooses two individuals to
produce offspring. The primary objective of selection is to produce a partial ranking of
the population so that fitter individuals will have a higher chance to reproduce.
Crossover takes the two selected individuals and divides randomly their binary
representation into two sections, called heads and tails. The two tails are then swapped
to produce new individuals. For example 11111 and 00000 can produce two new
200
strings 11000 and 00111, or 11110 and 00001. Mutation is applied to each offspring
after crossover. It is an occasional alteration of a bit (gene) position. The quality of the
offspring is evaluated in the same way as the parents.
The GA's application domain is wide. It can be applied to any optimisation
problem and has produced impressive results in a number of applications [11]. General
Electric developed a CAD system that combined expert systems with genetic
algorithms. This system was used on a 100-variable portion of a gas-turbine design and
produced a 92% increase in efficiency over human designers.
Variations to the basic form of the GA described above include real number and
integer representation, different selection schemes that give various reproduction
advantage to fitter individuals, and crossover operators that divide a string into more
than two sections. Common to various forms of GA is their robustness in reaching an
optimal solution in the presence of minimal, if any, prior knowledge of the problem at
hand. It is also best used in situations which involve a large number of parameters. As a
result, the search conducted by a GA is very computationally intensive. Recent
research has produced an analytical theory of GAs based on stochastic differential
equations which may fiarther establish GAs as an efficient tool for optimisation and
simulation of distributed systems.
[14] as the first step of achieving biological, or human-level, intelligence (131), and it is
purely based on numerical computation using sensory signals. AI lies somewhere
between CI and BI and can be achieved by extending CI with symbolic representation
and manipulation of non-numeric data (see Table 2). Fuzzy models are particularly
suitable for a smooth transition between CI and AI because they can accommodate both
numeric and symbolic information in a common framework.
Complexity
Biological
Human knowledge BNN BI
+ sensory inputs
Artificial Symbolic
+ numeric + sensor ANN AI
data
Computational CNN CI
Numeric
Bezdek presents the case for combining symbolic and sub-symbolic techniques by
introducing a distinction between computational neural networks (CNNs) and artificial
neural networks (ANNs). He argues [14] that ANNs result from the combination of
CNNs, which are based purely on numerical processing of sensory data, and some
knowledge usually in the form of non-numeric rules.
The most widely researched hybrid system in the area of SC is neuro-fuzzy
systems. Here the learning capabilities of ANNs are exploited within the framework of
fuzzy logic. In some systems, ANNs can be used to generate and tune the membership
functions in a fuzzy system (Fig 6). A number of models have been suggested for such
hybrid systems, e.g. fuzzy ART [15], Fuzzy LVQ [16] and radial basis functions [17].
The process of obtaining and tuning the fuzzy rules is one that is particularly suitable
202
for ANNs, resulting in substantial reductions in cost and development time. Here,
gradient descend methods have been used to define the shape and position of
membership functions. This hybrid method has been used to design triangular,
Gaussian, sigmoidal and bell-shaped membership functions [ 18].
fuzzy system
ANN
A number of applications have been reported in which fuzzy systems and ANNs are
employed in series [19]. In such situations, either the sensor output is not suitable for
direct input to the fuzzy system in which case an ANN pre-processes the input to the
fuzzy system (Fig 7(a)), or the output of the fuzzy system is not suitable for direct
interface with the external devices and an ANN is used as a post-processor to perform a
mapping or conversion not easily achievable by other analytical techniques (see Fig
i ,uzztem
7(b)). For example, Toshiba's microwave-oven-cum-toaster estimates the temperature
and the number of pieces of bread using an ANN and decides the optimum toasting
time by using fuzzy reasoning, i.e. its model resembles closely Fig 7(a).
fuzzy system
--------ii-
output from an actual system. Symbolic knowledge is then acquired from the trained
ANN in fuzzy logic representation (Fig 9).
ANN
correcting
value
fuzzy system
"1
Fig. 8. Parallelhybrid ANN/fuzzysystems.
ANN
actual I
system
2t ,
exped
fuzzy system
crisp crisp
input output
Development tools are becoming available for integrating fuzzy and neural
networks which should pave the way for the exploitation of the available architectures
in information systems applications. Two such tools are NEFCON-I [21] and O'INCA
[19] by Intelligent Machines Inc.
Fuzzy-GA systems combine the optimisation capabilities of GAs with fuzzy logic
(Fig 11). Such systems can develop the best possible set of rules for use by a fuzzy
inference engine. They can also be used to optimise the choice of the membership
functions.
ANN/fuzzy
I heuristir
modification~
The applications of hybrid genetic and fuzzy systems are in adaptive process
control, pattern recognition, robot trajectory generation and face recognition. GAs can
be used to improve the performance of neural networks by changing their parameters,
topologies or both. Applications include structure organisation of fuzzy neural
networks, evolving ANNs and self-organising maps.
There are many other possible combinations of FL, ANN and GA. However, this
paper has only presented the most widely published ways of combining these
techniques within the context of SC. The references provide a more comprehensive list
for further reading.
205
Although SC is a relatively new field of research it has already been established as one
of the fastest growing areas of AI technology in terms of consumer and industrial
applications. Japanese companies lead the way both in research and development,
although many European and American companies are allocating large resources to
this technology and the products have already appeared in the market. The majority of
the applications are in the areas of expert systems, control, pattern recognition,
clustering and image processing. However, active research is being carried out in many
other areas such as decision support systems, user interfaces, speech recognition, face
recognition and natural language systems. This section gives a brief description of a
small number of recent applications of this technology. The emphasis here will be on
applications resulting from the fusion of SC techniques rather than applications of
individual techniques.
Hitachi has produced a washing machine that uses the parallel structure of Fig 8 [22].
The smart controller inside the washing machine determines the washing programme
automatically by measuring the amount and type of clothes placed inside the machine.
A fuzzy rule-based system then uses this information to control the water flow and the
programme parameters such as washing time, rinsing time and spinning time. A neural
network monitors the operation of the machine during a wash and uses the quality of
the water inside the drum to fine-tune the output parameters of the fuzzy system. In this
way, it acts as an adaptive correcting mechanism for the fixed fuzzy rules.
The electric fan of Sanyo is designed to face the user as the user moves inside a room
[23]. To solve this problem accurately a very sophisticated and expensive system is
required, which is not suitable for a relatively cheap consumer product. The main
problem is computing the distance of the user from the fan. Sanyo have designed a
fuzzy system to estimate the distance given readings from an infra-red sensor. A neural
network is then trained to use this information to compute the required turning angle of
the fan. This technique has not only produced a f'mancially viable solution but it has 2.5
times better accuracy compared to statistical regression methods.
The Matsushita Electric photocopier machine controls its operation with a set of fuzzy
rules [24]. All photocopier machines operate with a set of fixed parameters which can
be adjusted manually by an engineer or the user. However, the quality of the copies can
deteriorate with time or can be dependent upon the type of the original document. A
206
fuzzy rule-based system can use the information about the state of the machine, the
quality and type of the original document to make decisions regarding operating
parameters of the photocopier. Some of the fuzzy input parameters used by the rules
are temperature, humidity, toner density, image density and image background. The
output of the system controls the parameters such as exposure lamp, drum voltage and
toner density among others. Interestingly, the parameters of the fuzzy system were
designed automatically by neural networks, i.e. the position and width of the fuzzy
membership functions were tuned by a gradient method. It is important to note that the
neural network was used during the development stage and not during the operation.
The same approach has been used by Matsushita to design vacuum cleaners, rice
cookers and washing machines.
Hitachi have manufactured and run a rolling mill system since 1991 [25] whose aim is
to produce constant thickness metal rolls. The surface of the plate which is being driven
through 20 rolls is scanned. The scanned pattern is matched against standard template
pattems by a neural network. The standard templates are used as the antecedent of the
fuzzy rules, for example:
The level of matching identified by the neural network is the strength of each fuzzy
rule. The aggregated final output of the fuzzy system determines the output to the rolls.
This approach is called neural network-driven fuzzy reasoning.
of view. However, there are many active research areas within the soft computing
framework that are and will be producing new directions for exploiting this technology
in other challenging areas of applications.
This area of research can benefit from soft computing in many different ways because
of the inherent uncertainty and vagueness in natural language, image recognition, hand-
writing recognition, speech recognition and gesture understanding. The uncertainty is
either due to poor sensor technology and data, or lack of processing algorithms and
background information. There are already products available that use component
technologies of soft computing to perform many of the above mentioned tasks e.g.
neural networks have been used extensively for image and speech recognition, fuzzy
logic has been used in areas such as face recognition [28], hand-writing recognition and
speech recognition. However, as discussed in section 2, the merger of these technique
would improve the overall characteristics of the resulting system and therefore it is
anticipated that this will be an active area of research in the next few years. In the next
section we will briefly mention future trends in soft computing research.
The majority (70%) of the publications in this area are concerned with the fusion of
fuzzy systems and neural networks (FS-NN). About 25% of the publications are in the
area of combining neural networks and GAs (NN-GA), and the remainder are in the
area of merging fuzzy systems and GAs (GA-FS). The most promising areas in FS-NN
are in the automatic design of fuzzy systems using neural networks and in neural
networks whose structure is based on fuzzy rules (generally similar to that shown in Fig
10) which has produced results significantly superior compared to conventional neural
networks. Within NN-GA, GAs have been used for optimisation of synaptic weights in
ANNs and have produced better results when combined with back-propagation (BP)
compared to BP on its own. In the GA-FS area the performance of static GA has been
improved by incorporating a set of fuzzy rules to dynamically change the parameters of
the GA in order to improve its overall performance. On the other hand GAs have been
used to optimise the selection of best fuzzy rules as well as optimising the rules
themselves. In general, improvements are being made in the areas of soft computing
where individual components seem to have deficiencies. In summary the future
research directions are as follows:
208
9 a better understanding of the trade-offs between training time and size of neural
networks is necessary;
9 neural network implementations of fuzzy systems must be able to learn on-line in
order to respond to changes in their environment;
9 we must be able to extract the knowledge learnt by neural networks;
9 GAs must be able to handle qualitative fitness functions as well as quantitative
ones;
9 generic soft computing platforms are essential for further research in these areas.
These are just some of the challenges faced by the soft computing research
community.
9. Conclusions
This paper has given a definition for soft computing and described its relevance to
intelligent systems. The principal aim of soft computing is to achieve robustness, low
solution cost and high machine IQ, through the exploitation of tolerance for
imprecision and uncertainty. The individual components of soft computing each
exhibit certain characteristics beneficial to the aim of increasing MIQ. Fuzzy logic
provides a model for approximate reasoning, as well as a representation for smooth
transition from a symbolic paradigm to a numeric one. Neural networks operate on
numeric data and provide low-level, fast-processing units that can adapt and learn. GAs
are used for optimisation to evolve better performance. A number of successful
applications, particularly in the area of consumer products, have shown that synergism
of these techniques can provide a route to building intelligent systems [29].
Acknowledgement
The authors wish to acknowledge the advice and encouragement provided by Professor
Lotfi Zadeh in the preparation of this paper.
References
1. Zadeh L A: 'The roles of fuzzy logic and soft computing in the conception, design and
deployment of intelligent systems', BT Technol J, 14, No 4, pp 32--36 (October 1996).
2. Uhr L and Honavar V: 'Introduction', in: 'Artificial Intelligence and Neural Networks - -
Steps Toward Principled Integration', Academic Press (1994).
209
3. Zadeh L A: 'Fuzzy Logic, Neural Networks and Soft Computing', Comm ofACM, 3._~7,No
3, pp 77--84 (March 1994).
6. Horstkotte E: http://www.quadralay.com/www/Fuzzy/Fuzzy.html
8. Zadeh L A: 'Soft Computing and Fuzzy Logic', IEEE Software, 1._[, No 6, pp 48--58
(1994).
9. Barto A, Sutton R and Anderson C: 'Neuro-like Adaptive Elements that can Solve Difficult
Control Problems', IEEE Tran on Systems, Man and Cybemetics, No 13 (1983).
11. Goldberg D E: 'Genetic and Evolutionary Algorithms Come of Age', Comm ofACM, 3._Z,
No 3, pp 113--119 (March 1994).
13. Fox J: 'Towards a reconciliation of fuzzy logic and standard logic', Int J of Man Machine
Studies, 1._55,pp 213--220 (1981).
16. Tsao E C K, Bezdek J C and Pal N R: 'Image Segmentation Using Fuzzy Clustering
Networks', in North American Fuzzy Information Processing, pp 98--107 (1992).
17. Wang L X: 'Training Fuzzy Logic Systems Using Nearest Neighbourhood Clustering',
Manuscript (1992).
18. Asakawa K and Takagi H: 'Neural Networks in Japan', Comm of ACM (March 1994).
20. Berenji H R: 'Fuzzy Systems That Can Learn', in: 'Computational Intelligence Imitating
Life', IEEE Press, pp 23--30 (1994).
22. Hitachi News Release: 'Neuro and fuzzy logic automatic washing machine and fuzzy logic
dryer', No. 91-024 (February 1991).
23. Sanyo News Release: 'Electric fan series in 1991', (March 1991).
24. Takagi H: 'Co-operative system of neural networks and fuzzy logic and its application to
consumer products', in: 'Industrial Applications of Fuzzy Control and Intelligent Systems',
IEEE Press (1994).
25. Nakajima M, Okada T, Hattori S and Morroka Y: 'Application of pattern recognition and
control techniques to shape control of the rolling mill', Hitachi Review, 75, No 2 (1993).
26. Takagi H: 'Survey of fuzzy logic applications in image processing equipment', in:
'Industrial Applications of Fuzzy Control and Intelligent Systems', IEEE Press (1995).
28. Baldwin J and Martin T: 'Basic concepts of fuzzy logic data browser with applications', in
Nwana H and Azarmi N (Eds): 'Software Agents and Soft Computing: Towards Enhancing
Machine Intelligence', Springer Verlag, Berlin (December 1996).