MODULE-I (10 HOURS)

Introduction to Neuro, Fuzzy and Soft Computing, Fuzzy Sets : Basic Definition and Terminology,
Set-theoretic Operations, Member Function Formulation and Parameterization, Fuzzy Rules and
Fuzzy Reasoning, Extension Principle and Fuzzy Relations, Fuzzy If-Then Rules, Fuzzy Reasoning ,
Fuzzy Inference Systems, Mamdani Fuzzy Models, Sugeno Fuzzy Models, Tsukamoto Fuzzy Models,
Input Space Partitioning and Fuzzy Modeling.

LECTURE-1
INTRODUCTION:
What is intelligence?
Real intelligence is what determines the normal thought process of a human.

Artificial intelligence is a property of machines which gives it ability to mimic the human
thought process. The intelligent machines are developed based on the intelligence of a
subject, of a designer, of a person, of a human being. Now two questions: can we construct a
control system that hypothesizes its own control law? We encounter a plant and looking at
the plant behavior, sometimes, we have to switch from one control system to another control
system where the plant is operating. The plant is may be operating in a linear zone or non-
linear zone; probably an operator can take a very nice intelligent decision about it, but can a
machine do it? Can a machine actually hypothesize a control law, looking at the model? Can
we design a method that can estimate any signal embedded in a noise without assuming any
signal or noise behavior?

That is the first part; before we model a system, we need to observe. That is we collect certain
data from the system and How do we actually do this? At the lowest level, we have to sense
the environment, like if I want to do temperature control I must have temperature sensor.
This data is polluted or corrupted by noise. How do we separate the actual data from the
corrupted data? This is the second question. The first question is that can a control system be
able to hypothesize its own control law? These are very important questions that we should
think of actually. Similarly, also to represent knowledge in a world model, the way we
manipulate the objects in this world and the advanced is a very high level of intelligence that
we still do not understand; the capacity to perceive and understand.
What is AI ?
Artificial Intelligence is concerned with the design of intelligence in an artificial device.
The term was coined by McCarthy in 1956.
There are two ideas in the definition.
1. Intelligence
2. artificial device

What is intelligence?
–Is it that which characterize humans? Or is there an absolute standard of judgement?
–Accordingly there are two possibilities:
– A system with intelligence is expected to behave as intelligently as a human
– A system with intelligence is expected to behave in the best possible manner

– Secondly what type of behavior are we talking about?

 – Are we looking at the thought process or reasoning ability of the system?
– Or are we only interested in the final manifestations of the system in terms of its
actions?

Given this scenario different interpretations have been used by different researchers as defining
the scope and view of Artificial Intelligence.
1. One view is that artificial intelligence is about designing systems that are as intelligent as
humans. This view involves trying to understand human thought and an effort to build
machines that emulate the human thought process. This view is the cognitive science approach
to AI.

2. The second approach is best embodied by the concept of the Turing Test. Turing held that in
future computers can be programmed to acquire abilities rivaling human intelligence. As part
of his argument Turing put forward the idea of an 'imitation game', in which a human being
and a computer would be interrogated under conditions where the interrogator would not
know which was which, the communication being entirely by textual messages. Turing argued
that if the interrogator could not distinguish them by questioning, then it would be
unreasonable not to call the computer intelligent. Turing's 'imitation game' is now usually
called 'the Turing test' for intelligence.

3. Logic and laws of thought deals with studies of ideal or rational thought process and inference.
The emphasis in this case is on the inferencing mechanism, and its properties. That is how the
system arrives at a conclusion, or the reasoning behind its selection of actions is very
important in this point of view. The soundness and completeness of the inference mechanisms
are important here.

4. The fourth view of AI is that it is the study of rational agents. This view deals with building
machines that act rationally. The focus is on how the system acts and performs, and not so
much on the reasoning process. A rational agent is one that acts rationally, that is, is in the best
possible manner.

Typical AI problems
While studying the typical range of tasks that we might expect an “intelligent entity” to perform,
we need to consider both “common-place” tasks as well as expert tasks.
Examples of common-place tasks include
– Recognizing people, objects.
– Communicating (through natural language).
– Navigating around obstacles on the streets
These tasks are done matter of factly and routinely by people and some other animals.
Expert tasks include:
• Medical diagnosis.
• Mathematical problem solving
• Playing games like chess

These tasks cannot be done by all people, and can only be performed by skilled specialists.
Now, which of these tasks are easy and which ones are hard? Clearly tasks of the first type are
easy for humans to perform, and almost all are able to master them. However, when we look at
what computer systems have been able to achieve to date, we see that their achievements include
performing sophisticated tasks like medical diagnosis, performing symbolic integration, proving
theorems and playing chess.
On the other hand it has proved to be very hard to make computer systems perform many routine
tasks that all humans and a lot of animals can do. Examples of such tasks include navigating our
way without running into things, catching prey and avoiding predators. Humans and animals are
also capable of interpreting complex sensory information. We are able to recognize objects and
people from the visual image that we receive. We are also able to perform complex social
functions.
Intelligent behaviour
This discussion brings us back to the question of what constitutes intelligent behaviour. Some of
these tasks and applications are:
1. Perception involving image recognition and computer vision
2. Reasoning
3. Learning
4. Understanding language involving natural language processing, speech processing
5. Solving problems
6. Robotics

Practical applications of AI
AI components are embedded in numerous devices e.g. in copy machines for automatic
correction of operation for copy quality improvement. AI systems are in everyday use for
identifying credit card fraud, for advising doctors, for recognizing speech and in helping complex
planning tasks. Then there are intelligent tutoring systems that provide students with personalized
attention.
Thus AI has increased understanding of the nature of intelligence and found many applications. It
has helped in the understanding of human reasoning, and of the nature of intelligence. It has also
helped us understand the complexity of modeling human reasoning.
Approaches to AI
Strong AI aims to build machines that can truly reason and solve problems. These machines
should be self aware and their overall intellectual ability needs to be indistinguishable from that
of a human being. Excessive optimism in the 1950s and 1960s concerning strong AI has given
way to an appreciation of the extreme difficulty of the problem. Strong AI maintains that suitably
programmed machines are capable of cognitive mental states.
Weak AI: deals with the creation of some form of computer-based artificial intelligence that
cannot truly reason and solve problems, but can act as if it were intelligent. Weak AI holds that
suitably programmed machines can simulate human cognition.
Applied AI: aims to produce commercially viable "smart" systems such as, for example, a
security system that is able to recognise the faces of people who are permitted to enter a particular
building. Applied AI has already enjoyed considerable success.
Cognitive AI: computers are used to test theories about how the human mind works--for example,
theories about how we recognise faces and other objects, or about how we solve abstract
problems.
Limits of AI Today
Today‟s successful AI systems operate in well-defined domains and employ narrow, specialized
knowledge. Common sense knowledge is needed to function in complex, open-ended worlds.
Such a system also needs to understand unconstrained natural language. However these
capabilities are not yet fully present in today‟s intelligent systems.
What can AI systems do
Today‟s AI systems have been able to achieve limited success in some of these tasks.
• In Computer vision, the systems are capable of face recognition
• In Robotics, we have been able to make vehicles that are mostly autonomous.
• In Natural language processing, we have systems that are capable of simple machine
translation.
• Today‟s Expert systems can carry out medical diagnosis in a narrow domain
• Speech understanding systems are capable of recognizing several thousand words continuous
speech
• Planning and scheduling systems had been employed in scheduling experiments with
the Hubble Telescope.
• The Learning systems are capable of doing text categorization into about a 1000 topics
• In Games, AI systems can play at the Grand Master level in chess (world champion), checkers,
etc.
What can AI systems NOT do yet?
• Understand natural language robustly (e.g., read and understand articles in a newspaper)
• Surf the web
• Interpret an arbitrary visual scene
• Learn a natural language
• Construct plans in dynamic real-time domains
• Exhibit true autonomy and intelligence
Applications:
We will now look at a few famous AI system that has been developed over the years.
1. ALVINN:
Autonomous Land Vehicle In a Neural Network
In 1989, Dean Pomerleau at CMU created ALVINN. This is a system which learns to control
vehicles by watching a person drive. It contains a neural network whose input is a 30x32 unit
two dimensional camera image. The output layer is a representation of the direction the
vehicle should travel.
The system drove a car from the East Coast of USA to the west coast, a total of about 2850
miles. Out of this about 50 miles were driven by a human, and the rest solely by the system.
2. Deep Blue
In 1997, the Deep Blue chess program created by IBM, beat the current world chess
champion, Gary Kasparov.
3. Machine translation
A system capable of translations between people speaking different languages will be a
remarkable achievement of enormous economic and cultural benefit. Machine translation is
one of the important fields of endeavour in AI. While some translating systems have been
developed, there is a lot of scope for improvement in translation quality.
4. Autonomous agents
In space exploration, robotic space probes autonomously monitor their surroundings, make
decisions and act to achieve their goals.
NASA's Mars rovers successfully completed their primary three-month missions in April,
2004. The Spirit rover had been exploring a range of Martian hills that took two months to
reach. It is finding curiously eroded rocks that may be new pieces to the puzzle of the region's
past. Spirit's twin, Opportunity, had been examining exposed rock layers inside a crater.
5. Internet agents
The explosive growth of the internet has also led to growing interest in internet agents to
monitor users' tasks, seek needed information, and to learn which information is most useful
What is soft computing?
An approach to computing which parallels the remarkable ability of the human mind to
reason and learn in an environment of uncertainty and imprecision.
It is characterized by the use of inexact solutions to computationally hard tasks such as the
solution of nonparametric complex problems for which an exact solution can‟t be derived in
polynomial of time.
Why soft computing approach?
Mathematical model & analysis can be done for relatively simple systems. More complex
systems arising in biology, medicine and management systems remain intractable to
conventional mathematical and analytical methods. Soft computing deals with imprecision,
uncertainty, partial truth and approximation to achieve tractability, robustness and low
solution cost. It extends its application to various disciplines of Engg. and science. Typically
human can:
1. Take decisions
2. Inference from previous situations experienced
3. Expertise in an area
4. Adapt to changing environment
5. Learn to do better
6. Social behaviour of collective intelligence
Intelligent control strategies have emerged from the above mentioned characteristics of
human/ animals. The first two characteristics have given rise to Fuzzy logic;2nd , 3rd and 4th
have led to Neural Networks; 4th , 5th and 6th have been used in evolutionary algorithms.
Characteristics of Neuro-Fuzzy & Soft Computing:
1. Human Expertise
2. Biologically inspired computing models
3. New Optimization Techniques
4. Numerical Computation
5. New Application domains
6. Model-free learning
7. Intensive computation
8. Fault tolerance
9. Goal driven characteristics
10. Real world applications
Intelligent Control Strategies (Components of Soft Computing): The popular soft computing
components in designing intelligent control theory are:
1. Fuzzy Logic
2. Neural Networks
3. Evolutionary Algorithms
Fuzzy logic:
Most of the time, people are fascinated about fuzzy logic controller. At some point of time in
Japan, the scientists designed fuzzy logic controller even for household appliances like a
room heater or a washing machine. Its popularity is such that it has been applied to various
engineering products.
Fuzzy number or fuzzy variable:

We are discussing the concept of a fuzzy number. Let us take three statements: zero, almost
zero, near zero. Zero is exactly zero with truth value assigned 1. If it is almost 0, then I can
think that between minus 1 to 1, the values around 0 is 0, because this is almost 0. I am not
very precise, but that is the way I use my day to day language in interpreting the real world.
When I say near 0, maybe the bandwidth of the membership which represents actually the
truth value. You can see that it is more, bandwidth increases near 0. This is the concept of
fuzzy number. Without talking about membership now, but a notion is that I allow some
small bandwidth when I say almost 0. When I say near 0 my bandwidth still further increases.
In the case minus 2 to 2, when I encounter any data between minus 2 to 2, still I will consider
them to be near 0. As I go away from 0 towards minus 2, the confidence level how near they
are to 0 reduces; like if it is very near to 0, I am very certain. As I progressively go away
from 0, the level of confidence also goes down, but still there is a tolerance limit. So when
zero I am precise, I become imprecise when almost and I further become more imprecise in
the third case.

When we say fuzzy logic, that is the variables that we encounter in physical devices, fuzzy
numbers are used to describe these variables and using this methodology when a controller is
designed, it is a fuzzy logic controller.
Neural networks :
Neural networks are basically inspired by various way of observing the biological organism.
Most of the time, it is motivated from human way of learning. It is a learning theory. This is
an artificial network that learns from example and because it is distributed in nature, fault
tolerant, parallel processing of data and distributed structure.
The basic elements of artificial Neural Network are: input nodes, weights, activation function
and output node. Inputs are associated with synaptic weights. They are all summed and
passed through an activation function giving output y. In a way, output is summation of the
signal multiplied with synaptic weight over many input channels.

Fig. Basic elements of an artificial neuron

 Fig. Analogy of biological neuron and artificial neuron
Above fig. Shows a biological neuron on top. Through axon this neuron actuates the signal
and this signal is sent out through synapses to various neurons. Similarly shown a classical
artificial neuron(bottom).This is a computational unit. There are many inputs reaching this.
The input excites this neuron. Similarly, there are many inputs that excite this computational
unit and the output again excites many other units like here. Like that taking certain concepts
in actual neural network, we develop these artificial computing models having similar
structure.
There are various locations where various functions take place in the brain.

If we look at a computer and a brain, this is the central processing unit and a brain. Let us
compare the connection between our high speed computers that are available in the market
today and a brain. Approximately there are 10 to the power of 14 synapses in the human
brain, whereas typically you will have 10 to the power of 8 transistors inside a CPU. The
element size is almost comparable, both are 10 to the power minus 6 and energy use is almost
like 30 Watts and comparable actually; that is energy dissipated in a brain is almost same as
in a computer. But you see the processing speed. Processing speed is only 100 hertz; our
brain is very slow, whereas computers nowadays, are some Giga hertz.
When you compare this, you get an idea that although computer is very fast, it is very slow to
do intelligent tasks like pattern recognition, language understanding, etc. These are certain
activities which humans do much better, but with such a slow speed, 100 Hz. ….. contrast
between these two, one of the very big difference between these two is the structure; one is
brain, another is central processing unit is that the brain learns, we learn. Certain mapping
that is found in biological brain that we have studied in neuroscience is not there in a central
processing unit and we do not know whether self awareness takes place in the brain or
somewhere else, but we know that in a computer there is no self-awareness.

Neural networks are analogous to adaptive control concepts that we have in control theory
and one of the most important aspects of intelligent control is to learn the control parameters,
to learn the system model. Some of the learning methodologies we will be learning here is the
error-back propagation algorithm, real-time learning algorithm for recurrent network,
Kohonen‟s self organizing feature map & Hopfield network.
Features of Artificial Neural Network (ANN) models:
1. Parallel Distributed information processing
2. High degree of connectivity between basic units
3. Connections are modifiable based on experience
4. Learning is a continuous unsupervised process
5. Learns based on local information
6. Performance degrades with less units
All the methods discussed so far makes a strong assumption about the space around; that
is, when we use whether a neural network or fuzzy logic or …. and .… any method that may
have been adopted in intelligent control framework, they all make always very strong
assumptions and normally they cannot work in a generalized condition. The question is that
can they hypothesize a theory? When I design all these controllers, I always take the data; the
engineer takes the data. He always builds these models that are updated. They update their
own weights based on the feedback from the plant. But the structure of the controller, the
model by which we assume the physical plant, all these are done by the engineer and also the
structure of the intelligent controller is also decided by the engineer. We do not have a
machine that can hypothesize everything; the model it should select, the controller it should
select, looking at simply data. As it encounters a specific kind of data from a plant can it
come up with specific controller architecture and can it come up with specific type of system
model? That is the question we are asking now.
You will see that in the entire course we will be discussing various tools. They will only be
dealing with these two things; behaviour. These tools are actually developed by mimicking
the human behavior, but not the human way of working. An intelligent machine is one which
learns, thinks and behaves in line with the thought process. That we would like but we are
very far from it. At least, at the moment, we are very far from this target of achieving real
intelligence.

We perceive the environment in a very unique way, in a coherent manner. This is called unity
of perception and intelligence has also something to do with this unity of perception,
awareness and certain things are not very clear to us until now. So an intelligent machine is
one which learns, thinks & behaves in line with thought process.

Evolutionary algorithms:
These are mostly derivative free optimization algorithms that perform random search in a
systematic manner to optimize the solution to a hard problem. In this course Genetic
Algorithm being the first such algorithm developed in 1970‟s will be discussed in detail. The
other algorithms are swarm based that mimic behaviour of organisms, or any systematic
process.
LECTURE-2
Fuzzy Sets Basic Concepts
 Characteristic Function (Membership Function)
 Notation
 Semantics and Interpretations
 Related crisp sets
 Support, Bandwidth, Core, α-level cut
 Features, Properties, and More Definitions
 Convexity, Normality
 Cardinality, Measure of Fuzziness
 MF parametric formulation
 Fuzzy Set-theoretic Operations
 Intersection, Union, Complementation
 T-norms and T-conorms
 Numerical Examples
 Fuzzy Rules and Fuzzy Reasoning
 Extension Principle and Fuzzy Relations
 Fuzzy If-Then Rules
 Fuzzy Reasoning
 Fuzzy Inference Systems
 Mamdani Fuzzy Models
 Sugeno Fuzzy Models
 Tsukamoto Fuzzy Models
 Input Space Partitioning
 Fuzzy Modeling.

The father of fuzzy logic is Lotfi Zadeh who is still there, proposed in 1965. Fuzzy logic can
manipulate those kinds of data which are imprecise.

Basic definitions & terminology:

Fuzzy Number:
A fuzzy number is fuzzy subset of the universe of a numerical number that satisfies condition
of normality & convexity.It is the basic type of fuzzy set.

why fuzzy is used? Why we will be learning about fuzzy? The word fuzzy means that, in
general sense when we talk about the real world, our expression of the real world, the way we
quantify the real world, the way we describe the real world, are not very precise.

When I ask what your height is, nobody would say or nobody would expect you to know a
precise answer. If I ask a precise question, probably, you will give me your height as 5 feet 8
inches. But normally, when I see people, I would say this person is tall according to my own
estimate, my own belief and my own experience; or if I ask, what the temperature is today,
the normal answer people would give is, today it is very hot or hot or cool. Our expression
about the world around us is always not precise. Not to be precise is exactly what is fuzzy.

Fuzzy logic is logic which is not very precise. Since we deal with our world with this
imprecise way, naturally, the computation that involves the logic of impreciseness is much
more powerful than the computation that is being carried through a precise manner, or rather
precision logic based computation is inferior; not always, but in many applications, they are
very inferior in terms of technological application in our day to day benefits, the normal way.

Fuzzy logic has become very popular; in particular, the Japanese sold the fuzzy logic
controller, fuzzy logic chips in all kinds of house hold appliances in early 90‟s. Whether it is
washing machine or the automated ticket machine, anything that you have, the usual house
hold appliances, the Japanese actually made use of the fuzzy logic and hence its popularity
grew.

Fig. Difference in Fuzzy and crisp boundary

As fuzzy means from precision to imprecision. Here, when I say 10, I have an arrow at 10,
pointing that I am exactly meaning 10 means 10.00000 very precise. When I say they are all
almost 10, I do not mean only 10, rather in the peripheral 10. I can tolerate a band from minus
9 to 9, whereas if I go towards 9 or 11, I am going away from 10, the notion of 10. That is
what is almost 10, that is around 10, but in a small bandwidth, I still allow certain bandwidth
for 10.

This concept to be imprecise is fuzzy or to deal with the day to day data that we collect or we
encounter and representing them in an imprecise manner like here almost 0, near 0, or hot,
cold, or tall; if I am referring to height, tall, short medium. This kind of terminology that we
normally talk or exchange among ourselves in our communication actually deals with
imprecise data rather than precise data. Naturally, since our communications are imprecise,
the computation resulting out of such communication language, the language which is
imprecise must be associated with some logic.
Fig. Sets: classical & fuzzy boundary

Set: A collection of objects having one or more common characteristics. For example, set of
natural number, set of real numbers, members, or elements. Objects belonging to a set is
represented as x belonging to A, where A is a set.
Universe of Discourse:
Defined as “a collection of objects all having the same characteristics”.
Notation: U or X, and elements in the universe of discourse are: u or x

Now, we will be talking about fuzzy sets. When I talked about classical set, we had classical
set of the numbers that we know, like we talked about the set of natural numbers, set of real
numbers. What is the difference between a fuzzy set and a classical set or a crisp set? The
difference is that the members, they belong to a set A or a specific set A or B or X or Y,
whatever it is, we define them; but the degree of belonging to the set is imprecise. If I say, a
universal set in natural numbers, all the natural numbers fall in this set. If I take a subset of
this natural number, like in earlier case, we put 1 to 11 in one set. When I ask, whether 12
belongs to set A, the answer is no; 13 belongs to set A? The answer is no; because, in my
natural number set, only 1 to 11 are placed. This is called classical set and their
belongingness here is one. They all belong to this set.

But in a fuzzy set, I can have all the numbers in this set, but with a membership grade
associated with it. When I say membership grade is 0 that means, they do not belong to the
set, whereas a membership grade between 0 to 1, says how much this particular object may
belong to the set.
The nomenclature/ Notation of a fuzzy set - how do we represent a fuzzy set there? One
way is that let the elements of X be x1, x2, up to xn; then the fuzzy set A is denoted by any of
the following nomenclature.
Mainly 2 types:
1. Numeric
2. Functional
Mostly, we
will be using either this or the first one, where you see the ordered pair x
1 µ A x1; x1 is member of A and x1 is associated with a fuzzy index and so forth, x2 and its
fuzzy index, xn and its fuzzy membership. The same thing, I can also write x1 upon µ A x1.
That means x1 is the member and this is the membership. The other way is here, in the third
pattern the membership is put first and in the bottom the member x1 with a membership, x2
with membership and xn with membership.
Every member x of a fuzzy set A is assigned a fuzzy index. This is the membership grade µA
x in the interval of 0 to 1, which is often called as the grade of membership of x in A. In a
classical set, this membership grade is either 0 or 1; it either belongs to set A or does not
belong. But in a fuzzy set this answer is not precise, answer is, it is possible. It is belonging to
set A with a fuzzy membership 0.9 and I say it belongs to A with a fuzzy membership 0.1;
that is, when I say 0.9, more likely it belongs to set A. When I say 0.1, less likely it belongs
to set A. Fuzzy sets are a set of ordered pairs given by A. The ordered pair is x, where x is a
member of the set. Along with that, what is its membership grade and how likely the subject
belongs to set A? That is the level we put, where x is a universal set and µx is the grade of
membership of the object x in A. As we said, this membership µ.
A x lies between 0 to 1; so, more towards 1, we say more likely it belongs to A. Like if I say
membership grade is 1, certainly it belongs to A.

For an example: a set of all tall people. Tall if I define, classically I would say above 6 is tall
and below 6 is not tall; that is, 5.9, 5 feet 9 inches is not tall and 6.1, 6 feet 1 inch is tall. That
looks very weird; it does not look nice to say that a person who is 6 feet 1 inch is tall and 5
feet 9 inches is not tall. This ambiguity that we have in terms of defining such a thing in
classical set, the difficulty that we face can be easily resolved in fuzzy set. In fuzzy set, we
can easily say both 6.1, 6 feet 1 inch as well as 5.9 inches as tall, but level this difference;
they are tall, but with a membership grade associated with this. This is what fuzzy set is.
Membership function - a membership function µ A x is characterized by µ A that maps all
the members in set x to a number between 0 to 1, where x is a real number describing an
object or its attribute, X is the universe of discourse and A is a subset of X.

Fig. Fuzzy Sets with Discrete Universes

Fuzzy set A = “sensible number of children”
X = {0, 1, 2, 3, 4, 5, 6} (discrete universe)
A = {(0, .1), (1, .3), (2, .7), (3, 1), (4, .6), (5, .2), (6, .1)}--(See discrete ordered pairs)(1st
expression)
or

Fig. Fuzzy Set with Cont. Universe

Fuzzy set B = “about 50 years old”

X = Set of positive real numbers (continuous)
B = {(x, µB(x)) | x in X}
µB(x)=f(x)

(2nd expression –with function that is subjective)

3rd expression of fuzzy set:

Linguistic variable and linguistic values:

Linguistic variable is a variable expressed in linguistic terms e.g. “Age” that assumes various
linguistic values like :middleaged, young, old. The linguistic variables are characterized by
membership functions.
Fig. A membership function showing support, bandwidth, core, crossover points
Support:
Support of a fuzzy set A is the set of all points x in X such that µA(x)>0.
Support (A)= {x| µA(x)>0}
Core:
The core of a fuzzy set A is the set of all poits x in X such that µA(x)=1
core (A)= {x| µA(x)=1}
Normality:
A fuzzy set A is normal if its core is nonempty. Always there is at least one x with µA(x)=1
then it is normal.
Crossover point:
A cross over point in fuzzy set A is the x with µA(x)=0.5
crossover (A)= {x| µA(x)=0.5}
Bandwidth:
For a normal & convex fuzzy set
Width(A)=|x2-x1|, where x2 & x1 are crossover points.
fuzzy singleton:
A fuzzy set whose support is a single point in X with µA(x)=1 is called a fuzzy singleton.
For the set given in figure we can find equivalence & write

Convexity:

Symmetry:
A fuzzy set is symmetric if its MF is symmetric about a certain point x=c such that,
µA(c+x)= µA(c-x) for all x in X
Comparison of the classical approach and fuzzy approach:
Let us say, consider a universal set T which stands for temperature. Temperature I can say
cold, normal and hot. Naturally, these are subsets of the universal set T; the cold temperature,
normal temperature and hot temperature they are all subsets of T.
The classical approach, probably, one way to define the classical set is cold. I define cold:
temperature T; temperature is a member of cold set which belongs to the universal set T such
that this temperature, the member temperature is between 5 degree and 15 degree centigrade.
Similarly, the member temperature belongs to normal, if it is between 15 degree centigrade
and 25 degree centigrade. Similarly, the member temperature belongs to hot set when the
temperature is between 25 degree centigrade and 35 degree centigrade. As I said earlier, one
should notice that 14.9 degree centigrade is cold according to this definition while 15.1
degree centigrade is normal implying the classical sets have rigid boundaries and because of
this rigidity, the expression of the world or the expression of data becomes very difficult. For
me, I feel or any one of us will feel very uneasy to say that 14.9 degrees centigrade is cold
and 15.1 degree centigrade is normal or for that matter, 24.9 degrees centigrade is normal and
25 degree or 25.1 degree centigrade is hot. That is a little weird or that is bizarre to have such
an approach to categorize things into various sets.
In a fuzzy set, it is very easy to represent them here. If the temperature is around 10 degree
centigrade, it is cold; temperature is around 20 degrees centigrade, it is normal and when
temperature is around 30 degree centigrade it is hot. In that sense, they do not have a rigid
boundary. If you say here, 25 degree centigrade, the 25 degree centigrade can be called
simultaneously hot as well as normal, with a fuzzy membership grade 0.5. 25 degrees
centigrade belongs to both normal as well as hot, but when I say 28 degree centigrade, this is
more likely a temperature in the category of hot, whereas the 22 degree centigrade is a
temperature that is more likely belonging to the set normal. This is a much nicer way to
represent a set. This is how the imprecise data can be categorized in a much nicer way using
fuzzy logic. This is the contrasting feature, why the fuzzy logic was introduced in the first
place.
Fuzzy sets have soft boundaries. I can say cold from almost 0 degree centigrade to 20 degree
centigrade. If 10 degree has a membership grade 1 and as I move away from 10 degree in
both directions, I lose the membership grade. The membership grade reduces from 1 to 0
here, and in this direction also from 1 to 0. The temperature, As I go, my membership grade
reduces; I enter into a different set simultaneously and that is normal. You can easily see, like
temperature 12, 13, 14, 15 all belong to both categories cold as well as normal, but each
member is associated with a membership grade; this is very important.
In a classical set, there are members in a set. Here, there are members in a set associated with
a fuzzy index or membership function.

LECTURE-3

Parameterization of Membership Function:

Once we talk about each member in a fuzzy set associated with membership function, you
must know how to characterize this membership function. The parameters are adjusted to fine
tune a fuzzy inference system to achieve desired I/O mapping. The membership functions
given here are one- dimensional. 2 dimensional MFs can be formed by cylindrical extension
from these basic MFs.

Where a<b<c & that are x coordinates of the corners of triangular MF

Where a<b<c<d & that are x coordinates of the corners of trapezoidal MF

Where c is the centre & a is adjusted to vary the width of MF, b controls slope at crossover
points.
Bell membership function is also termed as Cauchy MF.

Where c is the centre & Ϭ is the width of MF.

Left-Right MF:

Sigmoidal MF:
It can be open left or open right depending on sign of a.

Fig. Membership functions a. Triangle b. Trapezoidal c. Gaussian d. Bell, e. Left f. Right

LECTURE-4

Fuzzy set operations:

The main features of operation on fuzzy set are that unlike conventional sets, operations on
fuzzy sets are usually described with reference to membership function. When I say
operation, I do not do with the member itself, but I manipulate. When I say operation, I
manipulate the membership of the members in a set; members are not manipulated, rather the
membership function of the member is manipulated. This is very important; that is, x and µ(
x). In classical set what is manipulated is x.

If I say, x is 1 In classical set when I say x is 1 then, I would say 1 minus x is 0. In this, the
manipulation concerns with the member; whereas any kind of manipulation in fuzzy set does
not involve with x; rather it involves µx.

Containment or subset:

Three common operations: intersection which we say is the minimum function, union, which
we say is the maximum function and then fuzzy complementation

Standard fuzzy operations:

Intersection(Conjunction)or T-norm:
We can easily see that, the membership of A (green) intersection B(red) in fig. is all the
members that belongs to, that is common between A and B. Their membership will follow
these (blue) curves. There are two things we are doing. We have 2 sets. One is set A and the
other is set B. Classically, what we see is the common members between A and B. We are
not only seeing the common members, here we are also seeing, what is their membership
function.

Fig. Fuzzy set operations intersection & union

The membership function is computed minimum; that is, µA intersection B is minimum of
µA x and µ B x. That is the membership function. When there is a common member between
A and B, the membership function wherever is minimum that is retained and the other one is
thrown away. The member is retained; what is changing is the membership function.
Union(Disjunction) or T-co-norm or S-norm:
That is the meaning of these two curves that we have and then we are trying to find out what
the fuzzy union is. I have to find out In this the members are both belonging to A and B. But
their membership is maximum of both. if I have common members. I have set A and I have
set B; A union B is my union set. If x belongs to A and x belongs to B, then x also belongs to
A union B. But in fuzzy set, here this is µ A x and here it is µ x and in this case, this is
maximum of µA x and µB x; the membership function. That is the way it is assigned.

This candidate, when it comes to A union B take these two values of membership, find the
maximum which is 0.1 and assign here, which is 0.1. This is, µ union B is 0.1. This is the
meaning. This is a very important operation that we do. When we have two different fuzzy
sets, the operations are classical. The manipulation is among the membership functions;
otherwise, the notion of the classical fuzzy operation also remains intact, except that the
associated fuzzy membership gets changed.

Complement(Negation):

now it is fuzzy complementation. What is complement? This one, this particular triangular
function is my set R(red); fuzzy set R. The complement is like this; just inverse (blue). What
is 1 minus µA x; meaning 1 minus µA x.

Fig. Complement of fuzzy set

What is seen that the members remain intact in the set A, whereas the associated membership
functions got changed.

The other operations that we know for classical sets like De Morgan‟s law, the difference also
can be used for the sets like De Morgan‟s law.
Properties/ identities of fuzzy sets:

They are commutative. A union B is B union A; A intersection B is B intersection A. It is like

classical sets; fuzzy sets equally hold.

Associativity; A union B union C is A union B union C. Similarly, A union bracket B union

C is A intersection B intersection C is A intersection B combined with intersection C.
Distributivity: you can easily see that A union B intersection C is A union B intersection A
union C which is here. Similarly, here A intersection B union A intersection C. So, this is
distributivity.

Idempotency which is A union A is A and A intersection A is A.

Identity: A union null set is A, A intersection universal set is A, A intersection null set is null
and A union universal set is universal set X; here, X represents universal set.

The next step in establishing a complete system of fuzzy logic is to define the operations of
EMPTY, EQUAL, COMPLEMENT (NOT), CONTAINMENT, UNION (OR), and
INTERSECTION (AND). Before we can do this rigorously, we must state some formal
definitions:
Definition 1: Let X be some set of objects, with elements noted as x. Thus,
X = {x}.
Definition 2: A fuzzy set A in X is characterized by a membership function
mA(x) which maps each point in X onto the real interval [0.0, 1.0]. As
mA(x) approaches 1.0, the "grade of membership" of x in A increases.
Definition 3: A is EMPTY iff for all x, µA(x) = 0.0.
Definition 4: A = B iff for all x: µA(x) = µB(x) [or, µA = µB].
Definition 5: µA' = 1 - µA.
Definition 6: A is CONTAINED in B iff µA <= µB.
Definition 7: C = A UNION B, where: µC(x) = MAX(µA(x), µB(x)).
Definition 8: C = A INTERSECTION B where: µC(x) = MIN(µA(x), µB(x)).
Difference probability & fuzzy operations:
It is important to note the last two operations, UNION (OR) and INTERSECTION (AND),
which represent the clearest point of departure from a probabilistic theory for sets to fuzzy
sets. Operationally, the differences are as follows:
For independent events, the probabilistic operation for AND is multiplication, which (it can
be argued) is counterintuitive for fuzzy systems. For example, let us presume that x = Bob, S
is the fuzzy set of smart people, and T is the fuzzy set of tall people. Then, if µS(x) = 0.90
and µT(x) = 0.90, the probabilistic result would be:
µS(x) * µT(x) = 0.81
whereas the fuzzy result would be:
MIN(µS(x), µT(x)) = 0.90
The probabilistic calculation yields a result that is lower than either of the two initial values,
which when viewed as "the chance of knowing" makes good sense. However, in fuzzy terms
the two membership functions would read something like "Bob is very smart" and "Bob is
very tall." If we presume for the sake of argument that "very" is a stronger term than "quite,"
and that we would correlate "quite" with the value 0.81, then the semantic difference
becomes obvious. The probabilistic calculation would yield the statement If Bob is very
smart, and Bob is very tall, then Bob is a quite tall, smart person. The fuzzy calculation,
however, would yield If Bob is very smart, and Bob is very tall, then Bob is a very tall, smart
person.
Another problem arises as we incorporate more factors into our equations (such as the fuzzy
set of heavy people, etc.). We find that the ultimate result of a series of AND's approaches
0.0, even if all factors are initially high. Fuzzy theorists argue that this is wrong: that five
factors of the value 0.90 (let us say, "very") AND'ed together, should yield a value of 0.90
(again, "very"), not 0.59 (perhaps equivalent to "somewhat").
Similarly, the probabilistic version of A OR B is (A+B - A*B), which approaches 1.0 as
additional factors are considered. Fuzzy theorists argue that a sting of low membership grades
should not produce a high membership grade instead, the limit of the resulting membership
grade should be the strongest membership value in the collection.
The skeptical observer will note that the assignment of values to linguistic meanings (such as
0.90 to "very") and vice versa, is a most imprecise operation. Fuzzy systems, it should be
noted, lay no claim to establishing a formal procedure for assignments at this level; in fact,
the only argument for a particular assignment is its intuitive strength. What fuzzy logic does
propose is to establish a formal method of operating on these values, once the primitives have
been established.
Hedges :
Another important feature of fuzzy systems is the ability to define "hedges," or modifier of
fuzzy values. These operations are provided in an effort to maintain close ties to natural
language, and to allow for the generation of fuzzy statements through mathematical
calculations. As such, the initial definition of hedges and operations upon them will be quite a
subjective process and may vary from one project to another. Nonetheless, the system
ultimately derived operates with the same formality as classic logic. The simplest example is
in which one transforms the statement "Jane is old" to "Jane is very old." The hedge "very" is
usually defined as follows:
µ"very"A(x) = µA(x)^2
Thus, if mOLD(Jane) = 0.8, then mVERYOLD(Jane) = 0.64.
Other common hedges are "more or less" [typically SQRT(µA(x))], "somewhat," "rather,"
"sort of," and so on. Again, their definition is entirely subjective, but their operation is
consistent: they serve to transform membership/truth values in a systematic manner according
to standard mathematical functions.

Cartesian Product & Co-product:

Let A & B be fuzzy sets in X & Y respectively, then Cartesian product of A & B is a fuzzy
set in the product space XxY with the membership function

Similarly, Cartesian co-product A+B is a fuzzy set

Both Product & Co-product are characterized by 2- dimensional MFs.

LECTURE-5

Fuzzy Extension Principle:

Consider a function y = f (x).
If we known x it is possible to determine y.
Is it possible to extend this mapping when the input, x, is a fuzzy value.
The extension principle developed by Zadeh (1975) and later by Yager (1986) establishes
how to extend the domain of a function on a fuzzy sets.
PRINCIPLE
Suppose that f is a function from X to Y and A is a fuzzy set on X defined as
A = μA(x1)/x1 + μA(x2)/x2 + . . . + μA(xn)/xn.
The extension principle states that the image of fuzzy set A under the mapping f (.) can be
expressed as a fuzzy set B defined as
B = f (A) = μA(x1)/y1 + μA(x2)/y2 + . . . + μA(xn)/yn
where yi = f (xi )
If f (.) is a many-to-one mapping, then, for instance, there may exist x1, x2 ∈ X, x1 6= x2,
such that f (x1) = f (x2) = y_, y_ ∈ Y . The membership degree at y = y∗ is the maximum of
the membership degrees at x1 and x2 more generally, we have μB(y_) = maxy=f (xi ) μA(x)

A point to point mapping from a set A to B through a function is possible. If it is many to one
for two x in A then the membership function value in set B is calculated for f(x) as max
value of MF.

Fuzzy Relation:
CRISP MAPPINGS:

Fig. Mapping a relation

Consider the Universe X = {−2,−1, 0, 1, 2}
Consider the set A = {0, 1}
Using the Zadeh notation A = { 0/ −2 + 0/−1 + 1/ 0 + 1/ 1 + 0/ 2}
Consider the mapping y = |4x| + 2
What is the resulting set B on the Universe Y = {2, 6, 10}
It is possible to achieve the results using a relation that express the
mapping y = |4x| + 2.
Lets X = {−2,−1, 0, 1, 2}.
Lets Y = {0, 1, 2, . . . , 9, 10}
The relation

B=A◦R
2011 13 / 62
Applying the Relation

Fuzzy Mappings:

Fig. Fuzzy arguments mapping

Consider two universes of discourse X and Y and a function y = f (x).
Suppose that elements in universe X form a fuzzy set A.
What is the image (defined as B) of A on Y under the mapping f ?
Similarly to the crisp definition, B is obtained as

@nce.ufrj.br (PPGI-UFRJ) Extension Principle September

2011 17 / 62
Fuzzy vector is a convenient shorthand for calculations that use matrix relations.
Fuzzy vector is a vector containing only the fuzzy membership values.
Consider the fuzzy set:

The fuzzy set B may be represented by the fuzzy vector b:

 EXTENSI
Now, we will be talking about fuzzy relation. If x and y are two universal sets, the fuzzy sets,
the fuzzy relation R x y is given. As this is all ordered pair, µR x y up on x y for all x y,
belonging to the Cartesian space x, you associate µ R x y with each ordered pair.
What is the difference between fuzzy and crisp relation? In fuzzy this is missing, where µ R x
y is a number in 0 and 1. µR x y is a number between 0 and 1. This is the difference between
crisp relation and fuzzy relation. In crisp relation, it was either 0 or 1. It is either completely
connected or not connected, but in case of fuzzy, connection is a degree; that is, it is from 0 to
1.

The example is, let x equal to 1 2 3. Then x has three members, y has two members 1 and 2.
If the membership function associated with each ordered pair is given by this e to the power
minus x minus y whole squared. I is seen that this is the kind of membership function that is
used to know, how close is the members of y are from members of x. Because, if I relate from
1 to 1 using this, then you can see 1 minus 1 is 0 that is 1 and 1 very close to each other;
whereas, 2 and 1 is little far and 3 1 one is further far. This is a kind of relationship we are
looking between these two sets.
Let us derive fuzzy relation. If this is the membership function, fuzzy relation is of course all
the ordered pairs. We have to find out 111 2 2 1 2 2 3 1 and 3 2. These are all the sets of
ordered pairs and associated membership functions. You just compute e to the power minus x
minus y whole square. Here, 1 1 1 minus 1 whole square, 1 2 1 minus 2 whole square, 2 1 2
minus 1 whole square, 2 two 2 minus 2 whole square, 3 1 3 minus 1 whole square, 3 2 3
minus 2 whole square and if you compute them, you find 1 0.4 3 0.4 3 1 0.1 6 0.4 3. This is
your membership function. This is one way to find relation.
Normally, I know, it is easier to express the relation in terms of a matrix instead of this
continuum fashion, where each ordered pair is associated with membership function. It is
easier to appreciate the relation by simply representing them in terms of matrix. How do we
do that? This is my x 1 2 3 y is 1 21 the membership function associated was 1 1 2
membership is 0.4 3 2 1 0.4 3 2 2 1 3 1 0.1 6 and 3 2 is 0.4 3 that you can easily verify here 1
3 0.4 3 0.1 6 and 1.
The membership function describes the closeness between set x and y. It is obvious that
higher value implies stronger relations. What is the stronger relation? It is between 1 and 1,
and they are very close to each other, and 2 and 2; they are very close to each other.
Closeness between 2 and
2, between 1 and 1 is actually 1 and 1. They are very close to each other; similarly, 2 and 2. If
I simply say numerical closeness, then 2 and 2 are the closest, and 1 and 1 are the closest.
That is how these are the closest. Higher value implies stronger relations.

This is a formal definition of fuzzy relation; it is a fuzzy set defined in the Cartesian product
of crisp sets; crisp sets x1 x2 until xn. A fuzzy relation R is defined as µR upon x1 to xn,
where x1 to xn belongs to the Cartesian product space of x1 until xn; whereas, this µR the
fuzzy membership associated is a number between 0 and 1.
LECTURE-6

Fig. Inferring Fuzzy relation

Max-min composition or Max-min product:

It is a sort of matrix multiplication but memberships are not multiplied.

We will now explain, max min composition operation using an example that makes things
much more clear. This is my matrix, relational matrix R1 relating x and y and R2 relating y
and z. I have to find out the relational matrix from x to z using fuzzy rule of composition. We
normally write R3 is R1 composition R2. Using max min composition, how do we compute
R3?
Fig. Example Max-min composition or Max-min product
I want to now build a relationship between R1 & R2. Membership associated with x1 is 0.1
and z1 is 0.9. Let me put it very precise, x1 x2 x3 z1 and z2; if you look at what we will be
doing here, This is my x1 row and this is my z1 column. What I do, x1 row and z1 column; I
put them parallel and find out what is minimum. Here, minimum is 0.1 and here minimum is
0.2. After that, I find out what is the maximum, which is 0.2. This is what maximum of
minimum 0.1. 0.9 is minimum 0.2. 0.7 is 0.2. This is how we found out. The easiest way if I
want to find out is this one; this x1 x2 and z1. x2 means this row which is 0.4 and 0.5 and x2
and z1. z1 is again 0.9 and 0.7. I will find out. Minimum here is 0.4, minimum here is 0.5 and
maximum here is 0.5. You get this 0.5.Similarly, we can compute all the elements in R3
using a max min composition operation. As usual, any max min composition can follow
certain properties, associative and distributive over union. That is P fuzzy composition Q
union R is P composition Q union P composition R.

Fig. Properties of max-min composition

Similarly, weekly distributed over union is P composition, Q intersection, R is a subset of P

composition. Q union P composition R monotonic Q is a subset of R implies that, P
composition Q is a subset of P composition R.

Max-product composition:

Now, again, the same example we have taken R

1, R2 and R3. Now, I want to find out from R1 and R2, what R3 using max product
composition is.
 Fig. Example Max-product composition

Let us say, this is x1 x2 x3 z1 z2 z1 z2 and this is x1 x2 x3 for x1. I take this row which is 0.1
0.2 and finding the relation the fuzzy membership associate x1 and z1. I take the column
from z1 which is 0.9 0.7 and I multiply them here 0.1 0.9 is point 0 9 0.2 0.7 is 0.1 4 and find
out what is the maximum. This is the maximum 0.1 4.
I take another example. Let us find out the relationship between x2 and z2; for x2 the row is
0.4 0.5 and z2 the column is 0.8 0.6. Corresponding to this, if I multiply I get 0.4 0.8 is 0.3 2
0.5 0.6 is 0.3. Maximum is 0.3 2. This is 0.4 3 0.3 2. This is where it is 0.1. The answer is
here, the R3 and if I go back, if I look, R3 here is different.

Fig. Projection of fuzzy relation

Projection of fuzzy relation:
A fuzzy relation R is usually defined in the Cartesian space x and x and y. Often a projection
of this relation on any of the sets x or y, may become useful for further information
processing.
The projection of R x y on x denoted by R 1 is given by µ R 1 x is maximum. So, y belongs
to y µR x y. The meaning is that if I have R, this is x1 and x2 and this is y1 and y2, and this is
0.1 0.4 and this is 0.5 0.6. If these are the membership functions associated with x1 y1 x2 y2
is 0.4 x2 y1 is 0.5 x2 y2 is 0.6.projection, which means for x projection, I find out what the
maximum is. Overall, y in this case maximum is 0.4 and for x2 the max maximum projection
is if I took it here, 0.6. Similarly, if I make projection of R, x, y over x, what is the
maximum? This is 0.5 and this is 0.6. This is called x projection and y projection of a relation
matrix R.

Fig. Example of projection

We repeat another example. We have x as 3 components 1 2 3, y has 2 components 1 and 2.
This is the previous example that we had 1 0.4 3 0.4 3 1 0.1 6 0.4 3. x projection would be 1
0.4 3 maximum 1 0.4 3 1 maximum 1 0.1 6 0.4 3 maximum 0.4 3. Above figure illustrates x
and y projection of fuzzy relation. For x projection, the maximum value in each row is
retained. What is the maximum value in each row? Here, x projection maximum value in
each row is retained, while the maximum value in each column is retained for y projection.

Fig. Definition of projection

This is our formal definition of a fuzzy relation, projection of a fuzzy relation R on to any of
its set in the Cartesian product space; that is in the Cartesian product space. This is our
Cartesian product space and for that, we can map this one to any of these i or j or k; whatever
it is, for any value, then is defined as a fuzzy relation Rp, where Rp is defined as maximum
over Xi until Xk, where this is our Xi Xj Xk and this is µRp.
First, we talked about fuzzy relation projection of fuzzy relation. Once we have projection of
fuzzy relation, we can extend the projection to again infer what should be the relation. This
kind of technique may be useful in coding the information, where we have a huge number of
information and we want to transfer such a kind of projection and from projection to
extension would be beneficial for coding operation.

The crisp relation and fuzzy relation:

the difference is that in crisp relation; the index is either 0 or 1 that is, either complete
relation or no relation. But in fuzzy the membership grade is either 0 or 1; Whereas, in fuzzy
the relation has a grade from 0 to 1. Fuzzy composition rule; max min composition max
product composition unlike in crisp relation, where both max min and max product gives you
the same answer; whereas in fuzzy composition, max min and max product will give two
different answers .

LECTURE-7

Fuzzy If-then rules:

If x is A then y is B

“x is A” is antecedent or premise which tells the fact

“y is B” is consequence or conclusion
The whole statement is the rule.

Eg. If tomato is red then it is ripe.

These if then rules are the base of fuzzy reasoning.

If then rules are of different types:

1. Single rule with single antecedent

2. Single rule with multiple antecedent
3. Multiple with multiple antecedent
Steps of Fuzzy reasoning:
Shown in fig. For 2 rules what will be the consequent MF after aggregation
1. Degree of compatibility
2. Firing strength
3. Qualified consequent MF
4. Aggregate all qualified consequent MFs to obtain an overall MF
 Fig. Fuzzy reasoning, deriving output

FUZZY MODELLING:
Fuzzy Inferencing
The process of fuzzy reasoning is incorporated into what is called a Fuzzy Inferencing
System. It is comprised of three steps that process the system inputs to the appropriate system
outputs. These steps are 1) Fuzzification, 2) Rule Evaluation, and 3) Defuzzification. The
system is illustrated in the following figure.
Each step of fuzzy inferencing is described in the following sections.
Fuzzification
Fuzzification is the first step in the fuzzy inferencing process. This involves a domain
transformation where crisp inputs are transformed into fuzzy inputs. Crisp inputs are exact
inputs measured by sensors and passed into the control system for processing, such as
temperature, pressure, rpm's, etc.. Each crisp input that is to be processed by the FIU has its
own group of membership functions or sets to which they are transformed. This group of
membership functions exists within a universe of discourse that holds all relevant values that
the crisp input can possess. The following shows the structure of membership functions
within a universe of discourse for a crisp input.
where:
degree of membership: degree to which a crisp value is compatible to a membership
function, value from 0 to 1, also known as truth value or fuzzy input.
membership function, MF: defines a fuzzy set by mapping crisp values from its domain to
the sets associated degree of membership.

Fig. Fuzzy inferencing system

crisp inputs: distinct or exact inputs to a certain system variable, usually measured
parameters external from the control system, e.g. 6 Volts.
label: descriptive name used to identify a membership function.
scope: or domain, the width of the membership function, the range of concepts, usually
numbers, over which a membership function is mapped.
universe of discourse: range of all possible values, or concepts, applicable to a system
variable.
When designing the number of membership functions for an input variable, labels must
initially be determined for the membership functions. The number of labels correspond to the
number of regions that the universe should be divided, such that each label describes a region
of behavior. A scope must be assigned to each membership function that numerically
identifies the range of input values that correspond to a label.
The shape of the membership function should be representative of the variable. However this
shape is also restricted by the computing resources available. Complicated shapes require
more complex descriptive equations or large lookup tables. The next figure shows examples
of possible shapes for membership functions.
When considering the number of membership functions to exist within the universe of
discourse, one must consider that:
i) too few membership functions for a given application will cause the response of the system
to be too slow and fail to provide sufficient output control in time to recover from a small
input change. This may also cause oscillation in the system.
ii) too many membership functions may cause rapid firing of different rule consequents for
small changes in input, resulting in large output changes, which may cause instability in the
system.
These membership functions should also be overlapped. No overlap reduces a system based
on Boolean logic. Every input point on the universe of discourse should belong to the scope
of at least one but no more than two membership functions. No two membership functions
should have the same point of maximum truth, (1). When two membership functions overlap,
the sum of truths or grades for any point within the overlap should be less than or equal to 1.
Overlap should not cross the point of maximal truth of either membership function.
The fuzzification process maps each crisp input on the universe of discourse, and its
intersection with each membership function is transposed onto the μ axis as illustrated in the
previous figure. These μ values are the degrees of truth for each crisp input and are associated
with each label as fuzzy inputs. These fuzzy inputs are then passed on to the next step, Rule
Evaluation.
Fuzzy If then Rules :
We briefly comment on so-called fuzzy IF-THEN rules introduced by Zadeh. They may be
understood as partial imprecise knowledge on some crisp function and have (in the simplest
case) the form IF x is A THEN y is B . They should not be immediately understood as
i i
implications; think of a table relating values of a (dependent) variable y to values of an
(independent variable) x:

A , B may be crisp (concrete numbers) or fuzzy (small, medium, …) It may be understood in

i i
two, in general non-equivalent ways: (1) as a listing of n possibilities, called Mamdani's
formula:

(where x is A and y is B or x is A and y is B or …). (2) as a conjunction of implications:

1 1 2 2
Rule Evaluation
Rule evaluation consists of a series of IF-Zadeh Operator-THEN rules. A decision structure
to determine the rules require familiarity with the system and its desired operation. This
knowledge often requires the assistance of interviewing operators and experts. For this thesis
this involved getting information on tremor from medical practitioners in the field of
rehabilitation medicine.
There is a strict syntax to these rules. This syntax is structured as:
IF antecedent 1 ZADEH OPERATOR antecedent 2 ............ THEN consequent 1 ZADEH
OPERATOR consequent 2..............
The antecedent consists of: input variable IS label, and is equal to its associated fuzzy input
or truth value μ(x).
The consequent consists of: output variable IS label, its value depends on the Zadeh Operator
which determines the type of inferencing used. There are three Zadeh Operators, AND, OR,
and NOT. The label of the consequent is associated with its output membership function. The
Zadeh Operator is limited to operating on two membership functions, as discussed in the
fuzzification process. Zadeh Operators are similar to Boolean Operators such that:
AND represents the intersection or minimum between the two sets, expressed as:

OR represents the union or maximum between the two sets, expressed as:

NOT represents the opposite of the set, expressed as:

The process for determining the result or rule strength of the rule may be done by taking the
minimum fuzzy input of antecedent 1 AND antecedent 2, min. inferencing. This minimum
result is equal to the consequent rule strength. If there are any consequents that are the same
then the maximum rule strength between similar consequents is taken, referred to as
maximum or max. inferencing, hence min./max. inferencing. This infers that the rule that is
most true is taken. These rule strength values are referred to as fuzzy outputs.
Defuzzification
Defuzzification involves the process of transposing the fuzzy outputs to crisp outputs. There
are a variety of methods to achieve this, however this discussion is limited to the process used
in this thesis design.
A method of averaging is utilized here, and is known as the Center of Gravity method or
COG, it is a method of calculating centroids of sets. The output membership functions to
which the fuzzy outputs are transposed are restricted to being singletons. This is so to limit
the degree of calculation intensity in the microcontroller. The fuzzy outputs are transposed to
their membership functions similarly as in fuzzification. With COG the singleton values of
outputs are calculated using a weighted average, illustrated in the next figure. The crisp
output is the result and is passed out of the fuzzy inferencing system for processing
elsewhere.

Fuzzy Rule base and Approximate Reasoning an example:

What is fuzzy linguistic variable? Algebraic variables take numbers as values, while
linguistic variables take words or sentences as values.
For example, let x be a linguistic variable with a label „temperature‟. The universe of
discourse is temperature. In that universe, I am looking at a fuzzy variable x when I describe
the temperature. The fuzzy set temperature denoted as T can be written as T = very cold,
cold, normal, hot or very hot.
For each linguistic value, we get a specific membership function.
These are necessary because in the traditional sense, when we express worldly knowledge,
we express them in natural language. So here it is. From computational perspective, such
worldly knowledge can be expressed in terms of rule base systems.
Rule based systems:

Fig. Basics of rule based system

The above form is commonly referred to as the IF-THEN rule-based form. It typically
expresses an inference such that if we know a fact, we can infer or derive another fact. Given
a rule, I can derive another rule or given a rule, if I know a rule and the associated relation,
then given another rule, I can predict what should be the consequence.
 Fig.Fuzzy rules
This is a fuzzy rule base. Any worldly knowledge can be expressed in form in the form of a
rule base. Now, when I talk about fuzzy rule base, fuzzy information can be represented in
the form of a rule base, which consists of a set of rules in conventional antecedent and
consequent form such as if x is A, then y is B, where A and B represent fuzzy propositions
(sets). Suppose we introduce a new antecedent say A dash and we consider the following rule
if x is A dash, then y is B dash, from the information derived from rule 1, is it possible to
derive the consequent in rule 2, which is B dash?
The consequent B dash in rule 2 can be found from composition operation B dash equal to A
dash. This is called the compositional rule of inference, the compositional operator with R.

Fuzzy implication Relation:

A fuzzy implication relation is another category, which will call Zadeh implication. This is if
p implies q may imply either p and q are true or p is false. What we are saying is that just like
a local Mamdani rule, we say p and q are true imply either p and q are true or p is false. Thus,
p implies q means…. p and q are simultaneously true, which is Mamdani local rule or if p is
false, then p implies q has no meaning or p is false. This has taken an extra logic that is p and
q or not p.

Thus, the relational matrix can be computed as follows. If I look at this, what is p and q? p
and q means minimum of muA x and muB y. What is not p? 1 minus µA x. This entire thing
has to be maximum of minimum of these and this, which is this statement.µ, the relational
matrix elements are computed using this particular expression. Given a set of rules, we just
learnt various schemes by which we can construct a relational matrix between the antecedent
and the consequent. The next step would be to utilize this relational matrix for inference. This
method is commonly known as compositional rule of inference, that is, associated with each
rule we have a relational matrix. So, given a rule means given a relational matrix and given
another antecedent, we compute a consequent.
 Fig. Compositional rules
This is derived using fuzzy compositional rules. The following are the different rules for
fuzzy composition operation, that is, B equal to A composition R. R is the relational matrix
associated with a specific rule, A is a new antecedent that is known, R is known, B is the new
consequent for the new antecedent A. I have to find out what is B for this new A, given R.
That is computed by A composition R and we have already discussed in the relation class that
there are various methods and max-min is very popular.
First, we compute min and then max. Similarly, max-product: instead of min, we take the
product and compute what is the maximum value. Similarly, min-max: instead of max-min, it
is min-max. First, max and then min. Next, max-max and min-min. One can employ these
looking at the behavior of a specific data.

Fig. Example of Compositional rules

Now, we will take an example.

We are given a rule if x is A, then y is B, where A is this fuzzy set: 0.2 for 1, 0.5 for 2, and
0.7 for 3. This is a discrete fuzzy set. B is another fuzzy set that defines fuzzy membership
0.6 for 5, 0.8 for 7, and 0.4 for 9. The question is infer B dash for another rule if x is A dash,
then y is B dash, where A dash is known. A is known, B is known, and A dash is known.
What we have to find out is what B dash is. Infer B dash is the question that is being asked
Using Mamdani implication relation, first we will find out between A… the first rule, that is,
if x = A, then y is B. The relational matrix associated with this rule is…. For R, how do we
compute? A elements are 1, 2, and 3 and B elements are 5, 7, and 9. We have to find out now
for 0.2. Here, we compare with all the elements in point B and with each element, we found
what the minimum is. The minimum is always 0.2. Hence, the maximum of that is always
0.2. I have to find out the relational matrix between A and B.
The Mamdani principle means minimum, so between 1 and 5, 1 is associated with 0.2, and 5
is associated with 0.6, so the minimum is 0.2. Similarly, 1 is associated with 0.2, 7 is
associated with 0.8, so for 1 and 7, the minimum is 0.2. Similarly, 1 is associated with 0.2, 9
is associated with 0.4, so from 1 to 9, the minimum membership is 0.2. Similarly, you can see
that from 2 to all the elements 5, 7, 9, the minimum are 0.5, 0.5, and 0.4. Similarly, from 3 to
5, 7, and 9, we have 0.6, 0.7, and 0.4. These are the minimum fuzzy memberships between an
element in A to element in B. That is how we compute the relational matrix.
Once we compute the relational matrix, then we use max-min composition relation to find out
what is B dash, which is A dash (which is 0.5, 0.9, and 0.3) composition R and you can
compute. This is my R. I have to find out my matrix. This is 0.5, 0.9, and 0.3. So this
composition R is… you can easily see I take this row vector, put along the column matrix
and I see what is the minimum for each case. You can easily see 0.2 will be minimum here,
0.5 will be minimum here, 0.3 and maximum is 0.5.
The first element is 0.5. Again, I take this place in parallel with this column and then, I find
first minimum here is 0.2, here 0.5, here 0.3 and then maximum is again 0.5. Again, I take the
same row vector, put along this column vector and then, I find here the minimum is 0.2, here
minimum is 0.4, here minimum is 0.3 and the maximum is 0.4. This is the relation, this is the
answer. This is our B dash. Given A, this is my B dash using fuzzy compositional principle or
relation.

Fig. Comparison of compositional rules

There are other mechanisms also that we discussed. For the same example, if you use max-
min, you get B dash; for max-product, you get another B dash; for min-max, you get another.
Min-max and max are same for this example. Then, for max-max, you see that all the fuzzy
membership are the maximum values and for min-min, they are the minimum values here.
Approximate reasoning:
means given any logical system, we do not have, it is very difficult to make an exact result.
That is why from engineering perspective, we are more liberal. We do not want to be so
precise. As long as our system works, we are happy; if our control system works, we are
happy.

Fig. Approximate reasoning

Approximate reasoning. We have set up rules so we use a specific compositional rule of

inference and then we infer the knowledge or the consequence. Given a rule R (R is the
relational matrix associated with a specific rule) and given a condition A, the inferencing B is
done using compositional rule of inference B equal to A composition R. The fuzzy sets
associated with each rule base may be discrete or continuous, that is, A may be discrete or A
and B may be discrete or continuous.
A rule base may contain a single rule or multiple rules. If it is continuous, I cannot define
what the R relational matrix is. It is very difficult because it will have infinite values. R is not
defined. That is why for continuous, we apply compositional rule of inference but the method
to compute is different. A rule base may contain single rule or multiple rules. Various
inference mechanisms for a single rule are enumerated. Various mechanism means we talked
about min-max, max-min, max-max, min-min and so on. The inference mechanism for
multiple rules.
Single rule:
Now, we will take the examples one by one. Single rule with discrete fuzzy set. We talked
about a fuzzy set that may consist of a single rule or multiple rules. It can be discrete fuzzy
set or a continuous fuzzy set. We will try to understand how to make approximate reasoning
for such a rule base using the methods that we just enumerated. For each rule, we compute
what is the relational matrix if it is discrete fuzzy set and then we use compositional rule of
inference to compute the consequence given an antecedent. That is for discrete fuzzy set. We
have already talked about this but again, for your understanding, I am presenting another
example for single rule with discrete fuzzy set.
 Fig. Single rule

Rule 1: If temperature is hot, then the fan should run fast. If temperature is moderately hot,
then the fan should run moderately fast. In this example, we are given the temperature is in
degree Fahrenheit and the speed is expressed as 1000 rpm. The fuzzy set for hot H is for 70
degree Fahrenheit, 80 degree Fahrenheit, 90 degree Fahrenheit, and 100 degree Fahrenheit,
the membership values are 0.4, 0.6, 0.8, and 0.9. Similarly, for the fuzzy set F, for which the
fan should run fast, the fuzzy set is for 1000 rpm, the membership is 0.3, for 2000 rpm, the
membership is 0.5, for 3000 rpm, the membership 0.7, and for 4000 rpm, the membership is
0.9.
Given H dash, which is moderately hot, to be for 70… moderately hot means it is a little
more hot. So, same temperature obviously and their corresponding membership values will
reduce, because if I am describing moderately hot, they will have the same temperature but
the membership values will be less. You can easily see here that for 70, instead of 0.4, now it
is 0.2; for 80, instead of 0.6, it is 0.4; for 90, instead of 0.8, it is 0.6; for 100, instead of 0.9, it
is 0.8. This is moderately hot. Now, the question is find F dash.
I hope you are clear with this question. The question is very simple. We are given rule 1, we
have defined what is the fuzzy set hot and fuzzy set fast by these two statements and in the
second rule for moderately hot, we know the fuzzy set. We do not know what the fuzzy set is
corresponding to moderately hot, that is, moderately fast. We do not know moderately fast.
Find out F dash. If H, then F. If H dash, then F dash. Find out F dash. First, what do we do?
Corresponding to rule 1, we found out what is R. This is for rule 1. We knew that the
membership functions for H were 0.4, 0.6, 0.8, and 0.9, and for fast, the membership
functions where 0.3, 0.5, 0.7, and 0.9. If you look at this, these are my H values, the crisp
values: 70 degree Fahrenheit, 80 degree Fahrenheit, 90 degree Fahrenheit, and 100 degree
Fahrenheit. This is my speed: 1000 rpm, 2000 rpm, 3000 rpm, and 4000 rpm.
Between 70 and 1000 rpm, the entry would be minimum of these two (Refer Slide Time:
41:57), which is 0.3. Similarly, between 0.4 and 0.5, the minimum would be again 0.4 and
then between 0.4 and 0.7, it will be 0.4, and for 0.4 and 0.9, it is 0.4.
Similarly, we go to the next one, which is 0.6. For 0.6, 0.3 minimum 0.3, for 0.6 and 0.5, the
minimum is 0.5, for 0.6 and 0.7, minimum is 0.6, for 0.6 and 0.9, it is 0.6. Similarly, you can
fill all other cells here with their values: 0.3, 0.5, 0.7, 0.8, 0.3, 0.5, 0.7, and 0.9. This is my
relation matrix associated with rule 1: if H, then F. Now, what I have to do is I have to find
out F dash given H dash, using the fuzzy compositional rule of inference, which is
represented like this.

Fig. Relational matrix

F dash is H dash compositional rule of inference with R. This is max-min composition

operation. First, we take the min and then compute. H dash is given as 0.2, 0.4, 0.6, and 0.8.

Fig. Multiple rules

This is my H dash (moderately hot) and I have to do compositional inference between H dash
and R. Again, I am repeating so that you understand how to compute it. You put this row
vector along this column vector first . For each element, you find out what is the minimum.
You see that here it is 0.2, 0.3, 0.3, and 0.3 and the maximum of that is 0.3.
Similarly, you take again these values and put them here vertically. Here, the minimum is 0.2,
here 0.4, here 0.5, here 0.5, and maximum is 0.5. I am sure you will see here it is 0.7, but in
this case, you find that if you take this here, it is 0.2, here 0.4, here 0.6, here 0.8, and
maximum is 0.8. F dash is 0.3, 0.5, 0.7, and 0.8. That is how we infer or we do approximate
reasoning for a rule base. This is a very simple case.

Multiple rule:
There are two rules now. Rule 1 is if height is tall, then speed is high. Rule 2: if height is
medium, then speed is moderate. This is describing a rule for a person as to how fast he can
walk. Normally, those who are tall can walk very fast and those who are short, naturally their
speed will be less. This is one fuzzy rule that expresses the speed of a person while walking.
If height is tall, then speed is high and if height is medium, then speed is moderate. For this,
the fuzzy memberships are defined as tall, high, medium, and moderate.
Tall is 0.5, 0.8, and 1 for various feet like 5, 6, and 7. For speed is high, for 5 meter per
second, 7 meter per second, and 9 meter per second, the corresponding membership values
are 0.4, 0.7, and 0.9. For H2, which is medium height, the corresponding fuzzy
membership… you can easily see that when I say medium in this fuzzy set, 5 has 0.6, 6 has
0.7, and 7 has 0.6. The moderate speed is 0.6 for 5 meter per second, 0.8 for 7 meter per
second, and 0.7 for 9 meter per second. If this is the fuzzy set given, now the question is
given H dash, which is above average, and the corresponding fuzzy set is 0.5, 0.9, 0.8 for
three different heights, find S dash, the speed above normal. I hope the question is very clear
to you.

Fig. Relational matrix for 2 rules

We have two rules. If height is tall, then speed is high; tall is defined and high is defined. If
height is medium, then speed is moderate. I have already defined the fuzzy sets for both
medium as well as moderate. They are all discrete fuzzy sets. Now, you are presented with
new data and what is that new data? You are presented with a data called above average,
which is 0.5, 0.9, and 0.8 for three different heights for 5, 6, and 7. Then, find S dash equal to
above normal, that is, if height is above average, then the speed should be above normal.

This is the solution of this example. We have two rules. Naturally, we will have two
relational matrices: R1 for rule 1 and R2 for rule 2. I will not go in detail of how we compute.
You simply you go the antecedent and consequent, look at the membership function, find the
minimum for each entry. Here, these are the heights and these are the speeds; 5, 6, 7 feet is
the height and 5, 7, and 9 meter per second are the speeds of the individuals.
Now, you check the fuzzy sets and corresponding to each fuzzy set, find out what is the
minimum membership function. For 5, 5, you will find the membership function is 0.4,
minimum 0.5, 0.5, 0.4, 0.8, 0.8, 0.4, 0.8, 0.9. You can verify this. Similarly, R2 can be found
out. Taking the minimum membership entry between these two fuzzy sets, that is,
if I say this is H1 and S1 and this is H2 and S2. Look at these two fuzzy sets, find out what
the minimum entries are for each relation and then, how do we compute S dash above
normal? We have now two relational matrices. It is very simple. We do two composition
operations: H dash composition with R1 (this one) and again, H dash composition R2 and
then, we take the maximum of that, maximum of these two.

Fig. Multiple rule with continuous fuzzy sets

You can easily see that the maximum of H dash composition R1, H dash composition R2.
You can easily see that because H dash is common, this particular expression is the same as
H dash composition max of R1 and R2. This is R1 and R2. We look at all those entries
wherever it is the maximum: for 0.4 and 0.6, the maximum is 0.6; for 0.5 and 0.6, the
maximum is 0.6; for 0.5 and 0.6, the maximum is 0.6. You see the last element here 0.9 here
and 0.6, so this is 0.9. Like that, for all entries of R1 and R2, whatever the maximum values,
you put these values here (that is called maximum R1 and R2) and take a composition with H
dash. So H dash composition max of R1 and R2. H dash is already given as 0.5, 0.9, and 0.8.
If you do this composition, you get 0.6, 0.8, and 0.8. I hope this clears your concept of how
we compute or we do approximate reasoning in a rule base. Similarly, if there are multiple
rules, we have no problem and we can go ahead with the same principle.

The last section is the multiple rules with continuous fuzzy sets. We talked about discrete
fuzzy set, but if it is continuous fuzzy sets, how do we deal with that? Normally, a continuous
fuzzy system with two non-interactive inputs x1 and x2, which are antecedents, and a single
output y, the consequent, is described by a collection of r linguistic IF-THEN rules Where the
rule looks like this: If x1 is A1 k and x2 is A2 k, then y k is B k, where k is 1, 2 up to r. This
is the k th rule. Similarly, we can have rule 1, rule 2, rule 3, up to rule r. In this particular
rule, A1 k and A2 k are the fuzzy sets representing the k th antecedent pairs and B k are the
fuzzy sets representing the k th consequent. In the following presentation, what we will do
now is we will take a two-input system and two-rule system just to illustrate how we infer
from a rule base where the fuzzy sets are continuous. The inputs to the system are crisp
values and we use a max-min inference method.

Fig. Viewing multiple rules

We have two rules here represented graphically. You can see there are two variables x1 and
x2. There are two fuzzy variables and for each rule, we have a consequent y. The first rule
says that if x1 is A1 1 and x2 is A2 1, then y is B1.
Similarly, if x1 is A1 2, x2 is A2 2, then y is B2. Now, how do we infer? Given a crisp input,
a new input is given, crisp input in the domain of x1 and another crisp input in the domain of
x2. There can be a system whose two variables can be temperature as well as pressure. You
can easily think x1 to be the temperature and x2 to be the pressure. For example, for a
particular given system, you found out the temperature to be 50 degrees centigrade and
pressure to be some value. Given these two quantities, crisp quantities, how do we infer what
should be y?
The crisp input is given – temperature. Now, you find out corresponding membership values
here. Corresponding to this crisp input, we get the membership value in rule 1 as µA1 1 and
for the same crisp input, this rule 2 will provide you muA1 2. Now, in the second fuzzy
variable, given crisp input, rule 1 will compute µA2 1 and for the second one, the second
rule, the same crisp input would give this one, which is muA2 2. Once we find out these
membership values, what do we do? We graphically see which is minimum between µA1 1
and µA2 1. The minimum is µA2 1. We take that and we shade these areas in consequence.
Now, we take the second rule. We find between µA1 2 and µA2 2, the minimum is µA1 2.
We take that minimum and shade the area and consequent part of this rule 2. Now
graphically, we add these two taking the maximum. First, min and then max. You can easily
see that when I overlap this figure over this figure, I get this particular figure. You overlap
this second figure on the first figure or first figure on the second figure and take the resultant
shaded area. After taking this resultant shaded area…. Once you find this shaded area, the
next part is to see what is y given a crisp value. There are many methods, but we will focus in
this class or in this course on only one method, that is, center of gravity method(COG).
Obviously, if I take this figure and find out what is the center of gravity, it is this value y star.
The crisp output can be obtained using various methods. One of the most common method is
the center of gravity method. The resulting crisp output is denoted as y star in the figure. This
is y star. What we learnt in this is given a crisp input 1 and crisp input 2 and given two fuzzy
rules, how do we infer correspondingly a crisp output? Our data is crisp, but we are doing
fuzzy computation. Hence, rules are fuzzy. We take this data to the fuzzy rule base and then
fuzzify them through fuzzification process. Graphically, we find what is the net shaded area
using the max principle. We found out the shaded area for each rule in consequent taking the
min principle. Taking the max principle, we found out the resultant area and then, y star is the
center of gravity of these areas.

LECTURE-8

Fuzzy Control system or Fuzzy Inference System:

Categories:
1. Mamdani type and
2. Takagi–Sugeno type (T-S or TSK for short form T. Takagi, M. Sugeno, and K. T.
Kang).
Mamdani type fuzzy systems:
These employ fuzzy sets in the consequent part of the rules. This is a Mamdani type fuzzy
logic controller. What they do is that the consequent part itself takes the control action; the
incremental control action is described in the consequent part of each rule.

Fig. Architecture of FLC

The actual data that the controller is receiving is crisp data or classical data that has a definite
value. That crisp data goes to the fuzzy logic controller and it has these four components that
you can see: fuzzifier, rule base, inference engine and defuzzifier.

Fuzzifier. In a fuzzy logic controller, the computation is through linguistic values, not
through exact computation. Naturally, the fuzzifier would fuzzify the crisp data. In case of
temperature, I can say it is hot, medium-hot, cold, medium-cold, very hot and normal. These
are the fuzzifier. That means given a crisp data or the value of temperature say 40 degrees,
then I have to now convert to various linguistic values and each linguistic value will be
associated with a specific membership function. That is fuzzifier.

Once the data has been fuzzified, then it goes to the rule base and using an inference
mechanism…. The inference is taking place in fuzzy term, not in classical term and after a
fuzzy inference takes place about the decision or about the control action, we place a
defuzzifier. What this defuzzifier does is it converts the fuzzy control action to a crisp control
action.

In general, what we can say is the principal design parameters of a fuzzy logic controller are
the following: fuzzification strategies and interpretation of a fuzzification operator. How do
we fuzzify a crisp data? In the database, the discretization or normalization of universe of
discourse is done, because we must know the range of data one will encounter in an actual
plant. Accordingly, the normalization must be done so that we are taking into account all
possible values of data that one may encounter in a physical plant.

Fuzzy partition of the input and output spaces:

If I know the dynamic range of an input to the controller and the input to the plant (input to
 Fig. Parameters to be designed in FLC

the plant is actually output to the controller)… if I know the dynamic range, then in that
dynamic range, I must learn how to do fuzzy partition of the input and output space and this
fuzzification suits such the process should be complete in the sense.… You see that I am
drawing a universe of discourse here. This is the real value for a specific variable x1. If I have
defined a fuzzy set like this and like this, you can easily see that this part of the data is not
associated with any fuzzy membership. This is µ and this is x1 and unfortunately, this part is
not associated with any membership.

This fuzzification process is not complete. That means the entire universe of discourse in a
specific domain, wherever there are control systems…. There are various kinds of control
systems: process control, robot control, and aircraft control. Every control system is
associated with some input data and some output data. All possible input data and all possible
output data should be associated with a specific linguistic value as well as a membership
function.

Rule base:
Once fuzzification is done, how do we create a rule base? As I said, typically, in the rule
base, the two variables that are most important are error and change in error and we also
showed why it is so. Rule base. Choice of process state input variables and control variables.
You know that if I am implementing a fuzzy state feedback controller, then, a fuzzy state
feedback controller u would be minus K x. So, x is the states of the system, whereas if I am
implementing a fuzzy PID controller, then it will be u old plus K delta u k. Here, this delta u
k is a function of error and change in
error, whereas, in a state feedback controller, this is a common signal r and so, the control
action is dependent on state x1, x2, and xn.
Source and derivation of fuzzy control rules.
How do I derive these rules? What is the basis? Types of fuzzy control rules. A type of fuzzy
control rule means whether it is a PID controller, fuzzy PID controller or it is a fuzzy state
feedback controller. Similarly, completeness of fuzzy control rules means given any crisp
data in the domain of input space as well as output space, do I have in my rule base a specific
rule associated with this data? If I do not have any rule for this data, then the FLC will fail.
That is meaning of completeness of fuzzy control rules.
Fuzzy inference mechanism:
We have already talked about what is fuzzy inference mechanism. Given multiple rules, how
do we infer the consequence part? Defuzzification strategies and the interpretation of
fuzzification operator. Once the fuzzy inference is done, from the fuzzy inference, how do I
get a crisp value or a crisp control action? This is called defuzzification.
This is how we fuzzify a crisp data to fuzzy data or we make them fuzzy, that is, the crisp
input for variable x1 and x2…. Actually, this is not x1 and x2 but e and delta e are converted to
fuzzy sets using triangular membership functions. It is not always triangular, it can be
anything, but normally in control literature, most of these membership functions are
triangular functions.

Fig. Defuzzification

Defuzzification. Once I know how to do the fuzzification, defuzzification is explained in the

following diagram. You see that among various defuzzification methods, the most popular is
center of gravity method. How do I do it? Crisp input is given at any situation, any k th
sampling instant and the fuzzy logic controller gets the command signal, gets the actual
output of the plant, computes the error, computes the change in error and then, those crisp
values are fed into the fuzzification layer. Then, you have the membership function. You pass
on those fuzzy data to the rule base and then, for a specific rule base… You see that in rule 1,
you see that if you compare the membership µ A1 1 and µ A2 1, µ A2 1 is the minimum and
correspondingly, you shade the zone of action. This is delta u. How much should be the
incremental control action? This is my shaded portion or shaded portion of my control action.
Now I take a second one, second rule and there again, I evaluate the fuzzy membership
function A1 2 and A2 2. You see that the membership function in A1 2 is less. Corresponding
to that, we shade the incremental control action. Now, you see that if I take the maximum of
these two shaded zones, I get this (Refer Slide Time: 38:03), maximum of this. After I get,
this is the fuzzy decision, this is the fuzzy incremental control action, but how do I convert
this fuzzy incremental control action to a crisp action? That is by the center of gravity
method. In the center of gravity method, I integrate µ delta u d delta u upon integration of µ
d delta u. If I integrate this function, I get somewhere here to be the center of gravity. delta µ
star is this value, which is graphically shown here. We discussed about Mamdani type fuzzy
logic controller.

LECTURE-9

Takagi–Sugeno fuzzy systems:

The number of rules required by the Mamdani model are reduced here. They employ
function of the input fuzzy linguistic variable as the consequent of the rules. What happens
here is
that a fuzzy dynamic model is expressed in the form of local rules. Local rule means if there
are two variables x1 = a and x2 = b, then the plant dynamics can be represented either as a
linear dynamical system or a nonlinear dynamical system, as a known dynamical system.
TSK Fuzzy Rule
• If x is A and y is B then z = f(x,y)
– Where A and B are fuzzy sets in the
antecedent, and
– Z = f(x,y) is a crisp function in the
consequence.
• Usually f(x,y) is a polynomial in the input variables
x and y, but it can be any function describe the
output of the model within the fuzzy region
specified by the antecedence of the rule.
First order TSK Fuzzy Model
• f(x,y) is a first order polynomial
Example: a two-input one-output TSK
IF x is Aj and y is Bk then zi= px+qy+r
The degree the input matches ith rule is typically computed using min
operator: wi = min(µAj(x), µBk(y))
• Each rule has a crisp output
• Overall output is obtained via weighted average (reduce computation time of
defuzzification required in a Mamdani model)
z = i wizi/ i wi
To further reduce computation, weighted sum may be used, I.e.
z = i wizi
Example #1: Single-input
• A single-input TSK fuzzy model can be expressed as
– If X is small then Y = 0.1 X +6.4.
– If X is medium then Y = -0.5X +4.
– If X is large then Y = X-2.
 Fig. Rules in first order TSK fuzzy model

Example #2 : Two-input
• A two-input TSK fuzzy model with 4 rules can be expressed as
– If X is small and Y is small then Z = -X +Y +1.
– If X is small and Y is large then Z = -Y +3.
– If X is large and Y is small then Z = -X+3.
– If X is large and Y is large then Z = X+Y+2.
Zero-order TSK Fuzzy Model
• When f is constant, we have a zero-order TSK fuzzy model (a special case of the
Mamdani fuzzy inference system which each rule‟s consequent is specified by a fuzzy
singleton or a pre defuzzified consequent)
• Minimum computation time Overall output via either weighted
average or weighted sum is always crisp
• Without the time-consuming defuzzification operation, the TSK (Sugeno) fuzzy model is by
far the most popular candidate for sample data-based fuzzy modeling.
A general Takagi–Sugeno model of N rules for any physical plant, a general T–S model of N
rules is given by Rulei. This is the i th rule. If x1 k is a specific fuzzy set M1 i and x2 k is
another specific fuzzy set M2 i and so on until xn k is another fuzzy set Mn i, then the system
dynamics locally is described as x k plus 1 is Ai x k plus Bi u k, where i equal to 1, 2 until N,
because there are N rules.
Advantages over Mamdani model:
1. Less computation
2. Less time consuming
3. Simple
4. Mostly used for sample data based fuzzy modelling
Tsukamoto Fuzzy Models:
• The consequent of each fuzzy if-then rule is represented by a fuzzy set with monotonical
MF
– As a result, the inferred output of each rule is defined as a crisp value induced by the
rules‟ firing strength.
• The overall output is taken as the weighted average of each rule‟s output.

Example: Single-input Tsukamoto fuzzy model

• A single-input Tsukamoto fuzzy model can be expresses as
– If X is small then Y is C1
– If X is medium then Y is C2
– If X is large then Y is C3
Example: Single-input Tsukamoto fuzzy model

Fig. Output of Tsukamoto fuzzy model

Input Space Partitioning:
The antecedent of a fuzzy rule defines a local fuzzy region, while the consequent describes
the behavior with in the region via various constituents. The consequent constituents are
different MF or equation or constant depending on the fuzzy model. But, antecedents of
fuzzy rules can be formed by partitioning the input space.
3 types-
Grid partition: Often chosen method. Applicable to small no. of input variables and MFs i.e.
curse of dimensionality .
Tree partition: Each region is specified uniquely along a corresponding decision tree.
Exponenential increase of no. of rules is reduced. More MFs are needed. Orthogonality
holds roughly. Used in CART algorithm.
Scatter partition:
Portioning is scattered. Orthogonality doesn‟t hold. The portioned regions are non
uniform.No. of rules is reduced, but overall mapping from consequent of each rule out put
is difficult to estimate.

Fig. (a) Grid partition (b) Tree partition (c) Scatter partition
If certain transformation of the input is done, more flexible boundaries and partition will be
obtained.

LECTURE-10

Defuzzification methods in detail:

References:
1. Chapter 1 to 4 J.S.R.Jang, C.T.Sun and E.Mizutani, “Neuro-Fuzzy and Soft
Computing”, PHI, 2004, Pearson Education 2004.
2. Chapter 6 & 7 S. Rajasekaran & GA Vijayalakshmi Pai “Neural Networks, Fuzzy
Logic, and Genetic Algorithms synthesis and application”, PHI
3. Timothy J.Ross, “Fuzzy Logic with Engineering Applications”, McGraw-Hill,
International Editions, Electrical Engineering Series, Singapore, 1997.
4. Internet sources.
MODULE-II (10 HOURS)
Neural networks: Single layer networks, Perceptrons: Adaline, Mutilayer Perceptrons
Supervised Learning, Back-propagation, LM Method, Radial Basis Function Networks,
Unsupervised Learning Neural Networks, Competitive Learning Networks, Kohonen Self-
Organizing Networks, Learning Vector Quantization, Hebbian Learning. Recurrent neural
networks,. Adaptive neuro-fuzzy information; systems (ANFIS), Hybrid Learning Algorithm,
Applications to control and pattern recognition.
LECTURE-1

NEURAL NETWORK INTRODUCTION:

What is a neuron? A neuron is the basic processing unit in a neural network sitting on our
brain. It consists of

1. Nucleus-

2. Axon- Output node

3. Dendrites-Input node

4. Synaptic junction

The dynamics of this synaptic junction is complex. We can see the signal inputs from the
action of a neuron and through synaptic junction an output is actuated which is carried over
through dendrites to another neuron. Here, these are the neurotransmitters. We learned from
our experience that these synaptic junctions are either reinforced or in the sense they behave
in such a way that the output of synaptic junction may excite a neuron or inhibit the neuron.
This reinforcement of the synaptic weight is a concept that has been taken to artificial neural
model.
The objective is to create artificial machine and this artificial neural networks are motivated
by certain features that are observed in human brain, like as we said earlier, parallel
distributed information processing.

Fig. ANN model

Artificial neural networks are among the most powerful learning models. They have the
versatility to approximate a wide range of complex functions representing multi-dimensional
input-output maps. Neural networks also have inherent adaptability, and can perform robustly
even in noisy environments.
An Artificial Neural Network (ANN) is an information processing paradigm that is inspired
by the way biological nervous systems, such as the brain, process information. The key
element of this paradigm is the novel structure of the information processing system. It is
composed of a large number of highly interconnected simple processing elements (neurons)
working in unison to solve specific problems. ANNs, like people, learn by example. An ANN
is configured for a specific application, such as pattern recognition or data classification,
through a learning process. Learning in biological systems involves adjustments to the
synaptic connections that exist between the neurons. This is true of ANNs as well. ANNs can
process information at a great speed owing to their highly massive parallelism.
A trained neural network can be thought of as an "expert" in the category of information it
has been given to analyse. This expert can then be used to provide projections given new
situations of interest and answer "what if" questions.
Advantages of ANN:
1. Adaptive learning: An ability to learn how to do tasks based on the data given for training
or initial experience.
2. Self-Organisation: An ANN can create its own organisation or representation of the
information it receives during learning time.
3. Real Time Operation: ANN computations may be carried out in parallel, and special
hardware devices are being designed and manufactured which take advantage of this
capability.
4. Fault Tolerance via Redundant Information Coding: Partial destruction of a network leads
to the corresponding degradation of performance. However, some network capabilities may
be retained even with major network damage.
Table- Difference between the brain and a digital Computer

Differences human brain & ANN:

1. Computer has such fast speed of GHz, a traditional computer, however, when it
comes to certain processing like pattern recognition and language understanding, the
brain is very fast.
2. Intelligence and self-awareness, are absent in an artificial machine.
 Fig. An artificial neuron
An Artificial Neuron:
Basic computational unit in an artificial neural network is neuron. Obviously, it has to be an
artificial neuron.

Fig. An artificial Neuron in linear NN

This artificial neuron has three basic elements:

1. Nodes,
2. Weights and
3. Activation function.
Between input nodes and output nodes, there are synaptic weights w1, w2, w3, w4, w5 and w6.
There can be as many weights and these weights are multiplied with the signal as they reach
the output unit, where the output is simply sum of the signal multiplied with the weights and
then this output goes to an activation function f.
 Fig. Basic processing unit- the neuron

In a simple neuron, if input signals be x1, x2, xn with weights w1, w2 and wn. The weighted sum
will activate this total output by an activation function f. That is your output. What you are
seeing is actually a nonlinear map from input vector x i to output y. A single neuron has single
output but multiple inputs. Inputs are multiple for a single neuron and the output is unique, y
and this output y and the input bear a nonlinear relationship, by f. Neural networks can be
built using this single neuron. We can use the single neuron and build neural networks.
Analogy to brain:
Artificial Neural Network (ANN) is a system which performs information processing. An
ANN resembles or it can be considered as a generalization of mathematical model of human
brain assuming that
1. Information processing occurs at many simple elements called neurons.
2. Signals are passed between neurons over connection links.
3. Each connection link has an associated weight, which in a typical neural net
multiplies the signal transmitted.

ANN is built with basic units called neurons which greatly resemble the neurons of human
brain. A neural net consists of a large number of simple processing elements called neurons.
Each neuron applies an activation function to its net input to determine its output signal.
Every neuron is connected to other neurons by means of directed communication links, each
with an associated weight. Each neuron has an internal state called its activation level, which
is a function of the inputs it has received. As and when the neuron receives the signal, it gets
added up and when the cumulative signal reaches the activation level the neuron sends an
output. Till then it keeps receiving the input. So activation level can be considered as a
threshold value for us to understand.
In general, a neural network is characterized by
1. Pattern of connections between the neurons called its architecture
2. Method of determining the weights on the connections called its training or learning
algorithm
3. Its internal state called its Activation function.
The arrangement of neurons into layers and the connection patterns within and between
layers is called the net architecture. A neural net in which the signals flow from the input
units to the output units in a forward direction is called feed forward nets.
Interconnected competitive net in which there are closed loop signal paths from a unit back to
it is called a recurrent network. In addition to architecture, the method of setting the values of
the weights called training is an important characteristic of neural nets. Based on the training
methodology used neural nets can be distinguished into supervised or unsupervised neural
nets. For a neural net with supervised training, the training is accomplished by presenting a
sequence of training vectors or patterns each with an associated target output vector. The
weights are then adjusted according to a learning algorithm. For neural nets with
unsupervised training, a sequence of input vectors is provided, but no target vectors are
specified. The net modifies the weights so that the most similar input vectors are assigned to
the same output unit. The neural net will produce a representative vector for each cluster
formed. Unsupervised learning is also used for other tasks, in addition to clustering.
LECTURE-2
Activation functions:
Architecture:

Fig. Architecture of multilayer Neural network

Artificial neural networks are represented by a set of nodes, often arranged in layers, and a set
of weighted directed links connecting them. The nodes are equivalent to neurons, while the
links denote synapses. The nodes are the information processing units and the links acts as
communicating media.
A neural network may have different layers of neurons like
1. input layer,
2. hidden layer,
3. output layer.
The input layer receives input data from the user and propagates a signal to the next layer
called the hidden layer. While doing so it multiplies the weight along with the input signal.
The hidden layer is a middle layer which lies between the input and the output layers. The
hidden layer with non linear activation function increases the ability of the neural network to
solve many problems than the case without the hidden layer. The output layer sends its
calculated output to the user from which decision can be made. Neural nets can also be
classified based on the above stated properties.

There are a wide variety of networks depending on the nature of information processing
carried out at individual nodes, the topology of the links, and the algorithm for adaptation of
link weights. Some of the popular among them include:
Perceptron: Definition: It‟s a step function based on a linear combination of real-valued
inputs. If the combination is above a threshold it outputs a 1, otherwise it outputs a –1. This
consists of a single neuron with multiple inputs and a single output. It has restricted
information processing capability. The information processing is done through a transfer
function which is either linear or non-linear.

Fig. A perceptron
A perceptron can learn only examples that are called “linearly separable”. These are
examples that can be perfectly separated by a hyperplane.
Perceptrons can learn many boolean functions: AND, OR, NAND, NOR, but not XOR
However, every boolean function can be represented with a perceptron network that has two
levels of depth or more.
The weights of a perceptron implementing the AND function is shown below.

Fig. AND operation on inputs by a single perceptron

Multi-layered Perceptron (MLP): It has a layered architecture consisting of input, hidden
and output layers. Each layer consists of a number of perceptrons. The output of each layer is
transmitted to the input of nodes in other layers through weighted links. Usually, this
transmission is done only to nodes of the next layer, leading to what are known as feed
forward networks. MLPs were proposed to extend the limited information processing
capabilities of simple perceptrons, and are highly versatile in terms of their approximation
ability. Training or weight adaptation is done in MLPs using supervised backpropagation
learning.
Adding a hidden layer:
The perceptron, which has no hidden layers, can classify only linearly separable patterns.
The MLP, with at least 1 hidden layer can classify any linearly non-separable classes also.
An MLP can approximate any continuous multivariate function to any degree of accuracy,
provided there are sufficiently many hidden neurons (Cybenko, 1988; Hornik et al, 1989). A
more precise formulation is given below.
A serious limitation disappears suddenly by adding a single hidden layer.
It can easily be shown that the XOR problem which was not solvable by a Perceptron can be
solved by a MLP with a single hidden layer containing two neurons.

Figure 6.2.1.1: MLP for solving Xor

Recurrent Neural Networks: RNN topology involves backward links from output to the
input and hidden layers. The notion of time is encoded in the RNN information processing
scheme. They are thus used in applications like speech processing where inputs are time
sequences data.
Fig. Multilayer feed back network (Recurrent Neural Network)

Self-Organizing Maps: SOMs or Kohonen networks have a grid topology, wit unequal grid
weights. The topology of the grid provides a low dimensional visualization of the data
distribution. These are thus used in applications which typically involve organization and
human browsing of a large volume of data. Learning is performed using a winner take all
strategy in a unsupervised mode. It is described in detail later.
Single layer Network:
A neural net with only input layer and output layer is called single layer neural network. A
neural network with input layer, one or more hidden layers and an output layer is called a
multilayer neural network. A single layer network has limited capabilities when compared to
the multilayer neural networks.

Fig. Single Layer feed foreward Neural Network

LECTURE-3

Steps in developing NN:

1.1. Network formation
Neural network consists of an input layer, an output layer and a hidden layer. While a neural
network is constructed, the number of neurons in each layer has to be fixed. The input layer
will have neurons whose number will be equal to the number of features extracted. The
number of neurons in the output layer will be equal to the number of pattern classes. The
number of neurons in the hidden layer is decided by trial and error basis. With a minimum
number of neurons in the hidden layer, the neural network will be constructed and the
convergence will be checked for. Then the error will be noted. The number of neurons for
which the error is minimum, can be taken and will be checked for reduced error criterion.
1.2. Data preprocessing and normalization
Data selection and pre processing can be a demanding and intricate task. Neural net is as
good as the input data used to train it. If important data inputs are missing, then the effect on
the neural network‟s performance can be significant. The most appropriate raw input data
must be preprocessed. Otherwise the neural network will not produce accurate results.
Transformation and normalization are two widely used preprocessing methods.
Transformation involves manipulating raw data inputs to create a single input to a net, while
normalization is a transformation performed on a single data input to distribute the data
evenly and scale it into an acceptable range for the network. Knowledge of the domain is
important in choosing preprocessing methods to highlight the features in the data, which can
increase the ability of the network to learn the association between inputs and outputs. Data
normalization is the final preprocessing step. In normalizing data, the goal is to ensure that
the statistical distribution of values should be scaled to match the range of the input neurons.
The simplest method of normalization can be done using the formula
X normalized = (X-μ) / σ where μ and σ are the mean and standard deviation of the input
data.
Perceptron Learning
Learning a perceptron means finding the right values for W. The hypothesis space of a
perceptron is the space of all weight vectors.
The perceptron learning algorithm can be stated as below.

1. Assign random values to the weight vector

2. Apply the weight update rule to every training example
3. Are all training examples correctly classified?
a. Yes. Quit
b. No. Go back to Step 2.
There are two popular weight update rules.
i) The perceptron rule, and
ii) Delta rule

The Perceptron Rule

For a new training example X = (x1, x2, …, xn), update each weight according to this rule:
wi = wi + Δwi
Where Δwi = η (t-o) xi
t: target output
o: output generated by the perceptron
η: constant called the learning rate (e.g., 0.1)
Comments about the perceptron training rule:
Example means training data.
• If the example is correctly classified the term (t-o) equals zero, and no update on the weight
is necessary.
• If the perceptron outputs –1 and the real answer is 1, the weight is increased.
• If the perceptron outputs a 1 and the real answer is -1, the weight is decreased.
• Provided the examples are linearly separable and a small value for η is used, the rule is
proved to classify all training examples correctly (i.e, is consistent with the training data).
The Delta Rule
What happens if the examples are not linearly separable?
To address this situation we try to approximate the real concept using the delta rule.
The key idea is to use a gradient descent search. We will try to minimize the following error:
E = ½ Σi (ti – oi) 2
where the sum goes over all training examples. Here oi is the inner product WX and not
sgn(WX) as with the perceptron rule. The idea is to find a minimum in the space of weights
and the error function E.
The delta rule is as follows:
For a new training example X = (x1, x2, …, xn), update each weight according to this rule:
wi = wi + Δwi
Where Δwi = -η E‟(W)/wi
η: learning rate (e.g., 0.1)
It is easy to see that
E‟(W)/ wi = Σi (ti – oi) (-xi)
So that gives us the following equation:
wi = η Σi (ti – oi) xi
There are two differences between the perceptron and the delta rule. The perceptron is based
on an output from a step function, whereas the delta rule uses the linear combination of inputs
directly. The perceptron is guaranteed to converge to a consistent hypothesis assuming the
data is linearly separable. The delta rules converges in the limit but it does not need the
condition of linearly separable data.
There are two main difficulties with the gradient descent method:
1. Convergence to a minimum may take a long time.

2. There is no guarantee we will find the global minimum.

These are handled by using momentum terms and random perturbations to the weight vectors.

LECTURE-4

ADALINE & MADALINE:

Tne Adaline networks(ADAptive LINear Element) an d Madaline (Multiple Adaline) were

developed by Widrow. The structures use neurons and step/ sigmoidal activation function.
ADALINE has one output neuron but MADALINE has many. The learning is different from
a perceptron. It is here by Widrow-Hoff or LMS (Least Mean Square error) rule. Analogical
input or out put can be found by this network as minimum error function is searched befor
applying activation function.
ADALINE:
The structure includes an adaptive linear combiner (ALC) to obtain linear response that can
be applied to other elements of bipolar commutation. IF O/P of ALC is +ve response of
ADALINE is +1, and if –ve result of ADALINE is -1. It is represented by:
(3.11)
Fig. MADALINE network

LECTURE-5

The Multi-layered Perceptron training:

Improvements over Perceptron:
1) Smooth nonlinearity - sigmoid
2) 1 or more hidden layers

Training the hidden layer:

Not obvious how to train the hidden layer parameters.
The error term is meaningful only to the weights connected to the output layer. How to adjust
hidden layer connections so as to reduce output error? – credit assignment problem.
Any connection can be adapted by taking a full partial derivative over the error function, but
then to update a single weight in the first stage we need information about distant
neurons/connections close to the output layer (locality rule is violated). In a large network
with many layers, this implies that information is exchanged over distant elements of the
network though they are not directly connected. Such an algorithm may be mathematically
valid, but is biologically unrealistic.

The Backpropagation Algorithm:

As in Perceptron, this training algorithm involves 2 passes:
The forward pass – outputs of various layers are computed
The backward pass – weight corrections are computed
Consider a simple 3-layer network with a single neuron in each layer.
Therefore,
w jkf   jf xk (6.2.2.1.16)
Similarly the update rule for the threshold term is,
 jf   jf (6.2.2.1.17)
Fig. Training using Back propagation algorithm
Traning:
Randomly initialize weights.
Train network using backprop eqns.
Stop training when error is sufficiently low and freeze the weights.
Testing
Start using the network.
Merits of MLP trained by BP:
a) A general solution to a large class of problems.
b) With sufficient number of hidden layer nodes, MLP can approximate arbitrary target
functions.
c) Backprop applies for arbitrary number of layers, partial connectivity (no loops).
d) Training is local both in time and space – parallel implementation made easy.
e) Hidden units act as “feature detectors.”
f) Good when no model is available

Problems with MLP trained by BP:

a) Blackbox approach
b) Limits of generalization not clear
c) Hard to incorporate prior knowledge of the model into the network
d) slow training
e) local minima

LECTURE-6
Architectures of MLP:
If there is no nonlinearity then an MLP can be reduced to a linear neuron.
1. Universal Approximator:

For the above theorem to be valid, the sigmoid function g(.) has to satisfy some conditions. It
must be: 1) non-constant, 2) bounded, 3) monotone-increasing and 4) continuous.
All the four transfer functions described in the section on Perceptrons satisfy conditions #1,2
and 3. But the hardlimiting nonlinearities are not continuous. Therefore, the logistic function
or the tanh function are suitable for use as sigmoids in MLPs.
2. In general more layers/nodes greater network complexity

Although 3 hidden layers with full connectivity are enough to learn any function often more
hidden layers and/or special architectures are used.

More hidden layers and/or hidden nodes:

3-layer network:
Arbitrary continuous function over a finite domain
4-layer network
Neurons in a 3-layer architecture tend to interact globally.
In a complex situation it is hard to improve the approximation at one point without worsening
it at another.
So in a 4-layer architecture:
st
1 hidden layer nodes are combined to construct locally sensitive neurons in the second
hidden layer.
Discontinuous functions:
learns discontinuous (inverse function of continuous function) functions also (Sontag, 1992)
For hard-limiting threshold functions:
st
1 hidden layer: semi-infinite regions separated by a hyper-plane
nd
2 hidden layer: convex regions
rd
3 hidden layer: non-convex regions also
Training MLP:
1. Initialization: is VERY important.
g‟(.) appears on the right side of all weight update rules.er sections 6.1.1, 6.1.2, 6.2.1). Note
that g‟(.) is high at the origin and falls on both sides. Therefore most learning happens when
the net input (h) to the neurons is close to 0. Hence it is desirable to make initial weights
small. A general rule for initialization of input weights for a given neuron is:

2. Batch mode and Sequential mode:

Epoch: presentation of all training patterns is called an epoch.

Batch mode:
Updating network weights once every epoch is called batch mode update.
- memory intensive
- greater chance of getting stuck in local minima

Sequential mode:
Updating the network weights after every presentation of a data point is sequential mode of
update.
- lesser memory requirement
- The random order of presentation of input patterns acts as a noise source lesser chance of
local minima

Rate of learning:
We have already seen the tradeoffs involved in choice of a learning rate.
Small learning rate η,approximate original continuous domain equations more closely but
slows down learning.
Large learning rate η ,oorer approximation of original equations. Error may not decrease
monotonically and may even oscillate. But learning is faster..
A good thumb rule for choosing eta 'η':
η = 1/m
Where „m‟ is the number of inputs to a neuron. This rule assumes that there are different η s
for different neurons.
3. Important tip relating learning rate and error surface:

Rough error surface,slow down, low η

Smooth (flat) error surface, speed up, high η
i) Momentum:

a) If |a| <1, the above time-series is convergent.

b) If the sign of the gradient remains the same over consecutive iterations the weighted sum
delta wji grows exponentially i.e., accelerate when the terrain is clear.
c) If the gradient changes sign in consecutive iterations, delta wji shrinks in magnitude i.e.,
slow down when the terrain is rough.
ii) Separate eta for each weight:

a) Separate η for each weight

b) Every eta varies with time
c) If delta(w) changes sign several time in the past few iters, decrease η
d) If delta(w) doesn‟t change sign in the past few iters, increase η

Stopping Criteria: when do we stop training?

a) Error < a minimum.
b) Rate of change in error averaged over an epoch < a minimum.
c) Magnitude of gradient ||g(w)|| < a minimum.
d) When performance over a test set has peaked.

Premature Saturation:
All the weight modification activity happens only when |h| is within certain limits.
g‟(h) ≈ 0, or delta(w) = 0, for large |h|.
NN gets stuck in a shallow local minimum.
Solutions:
1) - Keep a copy of weights
- Retract to pre-saturation state
- Perturb weights, decrease η and proceed
2) - Reduce sigmoid gain (lambda) initially
e) Increase lambda gradually as error is minimized

Network doesn‟t get stuck, but never settles either.

Testing/generalization:
Idea of overfitting or overtraining:
Using too many hidden nodes, may cause overtraining. The network might just learn noise
and generalize poorly.
Example of polynomial interpolation:
Consider a data set generated from a quadratic function with noise added. A linear fit is likely
th
to give a large error. Best fit is obtained with a quadratic function. Fit 10 degree might give
a low error but is likely to learn the variations due to noise also. Such a fit is likely to do
poorly on a test data set. This is called overfitting or poor generalization.
This happens because there are many ways of generalizing from a given training data set.
The above Venn diagram illustrates the possibility of generalizing in multiple ways from a
given training data set. U is the universe of all possible input-output patterns. F (the ellipse)
represents the set of I/O pairs that define the function to be learnt by the mlp. T (circle)
denotes the training data set which is a subset of F. X denotes the test data set. The dotted
rectangle denotes the actual function learnt by the NN, which is consistent with the training
set T, but is completely non-overlapping with the test set X, and very different from the
unknown function F.

Applications of MLP
Three applications of MLPs that simulate aspects of sensory, motor or cognitive functions are
described.
1. Nettalk
2. Past tense learning
3. Autonomous Land Vehicle in a Neural Network (ALVINN)
LECTURE-6

Multilayer Feed-Foreward Network:

Fig. Characteristics of Multilayer feed-foreward network

The algorithm that was derived using gradient descent for nonlinear neural networks with
nonlinear activation function is popularly known as back propagation learning algorithm,
although the learning algorithm still is derived using gradient descent rule.

Multilayer feed forward network has more hidden layers and again, when I say feed forward
network, the connections are all allowed only from any layer to its succeeding layer, but the
connections are not allowed from any layer to its preceding layer. The example is you see
here there are four layers. These are all inputs. First hidden layer, second hidden layer, third
hidden layer and this is output layer. When we say the number of layers, we do not count the
input layer as one of the layers. When I say two layered network, then I have only one hidden
layer and next layer becomes output layer.

Fig. Multilayer feed foreward network

This particular configuration means there are sub-units, sub-neurons here and this particular
configuration, if I connect you will see why I say feed forward network, because I am able to
connect any layer from its preceding layer. That means connections are allowed from the
preceding layer to any layer, but cannot allow the feedback connection. (Refer Slide Time:
30:54) This is called feedback connection; this is not allowed. This is allowed. From this
layer, I can connect to this layer. This is allowed, but I cannot allow from this layer to
connect to this layer. These are called feedback connections. They are not allowed and that is
why this is known as feed forward network.
Today, we will derive a two-layered feed forward neural network with sigmoid activation
function. We can very easily see that this is 1 layer; this is the only hidden layer and this is
the only output layer; output layer is always only one.
We have a certain convention that we will put while deriving a back propagation learning
algorithm for this. The same simple principle; given training data, we allow the input to pass
through the network, compute the error here, use the gradient descent rule and the back
propagated error are used to modify the weights here that is between output layer and hidden
layer and again another form of back propagated error here has to be used for modification of
the weights between input layer and hidden layer. This is again the convention that we will
use.
Fig. The Gradient descent rule

After choosing the weights of the network randomly, the backpropagation algorithm is used
to compute the necessary corrections. The algorithm can be decomposed in the following four
steps:
i) Feed-forward computation
ii) Backpropagation to the output layer
iii) Backpropagation to the hidden layer
iv) Weight updates
The algorithm is stopped when the value of the error function has become sufficiently small.

In the case of p > 1 input-output patterns, an extended network is used to compute the error
function for each of them separately. The weight corrections The Backpropagation Algorithm
are computed for each pattern and so we get, for example, for weight w(1)ij the corrections

The necessary update in the gradient direction is then

We speak of batch or off-line updates when the weight corrections are made in this way.
Often, however, the weight updates are made sequentially after each pattern presentation (this
is called on-line training). In this case the corrections do not exactly follow the negative
gradient direction, but if the training patterns are selected randomly the search direction
oscillates around the exact gradient direction and, on average, the algorithm implements a
form of descent in the error function. The rationale for using on-line training is that adding
some noise to the gradient direction can help to avoid falling into shallow local minima of the
error function. Also, when the training set consists of thousands of training patterns, it is very
expensive to compute the exact gradient direction since each epoch (one round of
presentation of all patterns to the network) consists of many feed-forward passes and on-line
training becomes more efficient.

Back Propagation Neural Network

Backpropagation is a training method used for a multi layer neural network. It is also called
the generalized delta rule. It is a gradient descent method which minimizes the total squared
error of the output computed by the net. Any neural network is expected to respond correctly
to the input patterns that are used for training which is termed as memorization and it should
respond reasonably to input that is similar to but not the same as the samples used for training
which is called generalization. The training of a neural network by back propagation takes
place in three stages 1. Feedforward of the input pattern 2. Calculation and Back propagation
of the associated error 3. Adjustments of the weights After the neural network is trained, the
neural network has to compute the feedforward phase only. Even if the training is slow, the
trained net can produce its output immediately.
Architecture
A multi layer neural network with one layer of hidden unitss is shown in the figure. The
output units and the hidden units can have biases. These bias terms are like weights on
connections from units whose output is always 1. During feedforward the signals flow in the
forward direction i.e. from input unit to hidden unit and finally to the output unit. During
back propagation phase of learning, the signals flow in the reverse direction.
Algorithm
The training involves three stages 1. Feedforward of the input training pattern 2. Back
propagation of the associated error 3. Adjustments of the weights. During feedforward, each
input unit (Xi) receives an input signal and sends this signal to each of the hidden units Z1,
Z2, …Zn. Each hidden unit computes its activation and sends its signal to each output unit.
Each output unit computes its activation to compute the output or the response of the neural
net for the given input pattern.
During training, each output unit compares its computed activation yk, with its target value tk
to determine the associated error for the particular pattern. Based on this error the factor ∂k
for all m values are computed. This computed ∂k is used to propagate the error at the output
unit Yk back to all units in the hidden layer. At a later stage it is also used for updation of
weights between the output and the hidden layer. In the same way ∂j for all p values are
computed for each hidden unit Zj. The values of ∂j are not sent back to the input units but are
used to update the weights between the hidden layer and the input layer. Once all the ∂ factrs
are known, the weights for all layers are changed simultaneously. The adjustment to all
weights wjk is based on the factor ∂k and the activation zj of the hidden unit Zj. The change
in weight to the connection between the input layer and the hidden layer is based on ∂j and
the activation xi of the input unit.
Activation Function
An activation function for a back propagation net should have important characteristics. It
should be continuous, Differentiable and monotonically non- decreasing. For computational
efficiency, it is better if the derivative is easy to calculate. For the commonly used activation
function, the derivative can be expressed in terms of the value of the function itself. The
function is expected to saturate asymptotically. The commonly used activation function is the
binary sigmoidal function.
Training Algorithm
The activation function used for a back propagation neural network can be either a bipolar
sigmoid or a binary sigmoid. The form of data plays an important role in choosing the type of
the activation function. Because of the relationship between the value of the function and its
derivative, additional evaluations of exponential functions are not required to be computed.
Algorithm
Step 0: Initialize weights
Step 1: While stopping condition is false, do steps 2 to 9
Step 2: For each training pair, do steps 3 - 8 Feed forward
Step 3: Input unit receives input signal and propagates it to all units in the hidden layer
Step 4: Each hidden unit sums its weighted input signals
Step 5: Each output unit sums its weighted input signals and applied its activation function to
compute its output signal.
Backpropagation Step 6: Each output unit receives a target pattern corresponding to the input
training pattern, computes its error information term δk = ( tk – yk) f‟ (y_ink) Calculates its
bias correction term ΔWok = αδk And sends δk to units in the layer below
Step 7: Each hidden unit sums its delta inputs Multiplies by the derivative of its activation
function to calculate its error information term Calculates its weight correction term Δvij =
αδjxi And calculates its bias correction term Δvoj = αδj Update weights and biases
Step 8: Each output unit updates its bias and weights Wjk(new) = wjk(old) + Δ wjk Each
hidden unit updates its bias and weights Vij (new) = vij (old) + Δvij
Step9:Test stopping condition

LECTURE-7

Radial Basis Function Networks:

Fig. RBF network

• These are 3-layer networks that can approximate any continuous function through a basis
function expansion.
• The basis functions here (which are data dependent as earlier) exhibit some radial
symmetry.
• These networks have the so called perfect interpolation property.
The function represented by an RBF network with p hidden nodes can be written as

X is the input to the network.

• wj is weight from jth hidden node to the output.
• Á (||X −  j ||) is the output of the jth hidden node
and  j is the parameter vector associated with jth
hidden node, j = 1, ・ ・ ・ , p.
A very popular model is the Gaussian RBF network.

• Here the output is written as

• The  j is called the center of the jth hidden or RBF node and  is called the width.
• We can have different  for different hidden nodes.
We next consider learning the parameters of a RBF network from training samples.
• Let {(Xi, di), i = 1, ・ ・ ・ ,N} be the training set.
• Suppose we are using the Gaussian RBF.
• Then we need to learn the centers (  j ) and widths (  ) of the hidden nodes and the weights
into the output node (wj).

Like earlier, we can find parameters to minimize empirical risk under squared error loss
function.
• Same as minimizing sum of squares of errors. Let

J is a function of  , wj ,  j , j = 1, ・ ・ ・ , p.
We can find the weights/parameters of the network to minimize J.
• To minimize J, we can use the standard iterative algorithm of gradient descent.
• This needs computation of gradient which can be done directly from the expression for J.
• For this network structure there are no special methods to evaluate all the needed partial
derivatives. Such a gradient descent algorithm is certainly one method of learning an RBF
network from given training data.
• This is a general-purpose method for learning an RBF network.
• Like in the earlier case, we have to fix p, the number of hidden nodes.
• Such procedure would have the usual problems of converging to a local minimum of the
error function.
• There are also other methods of learning an RBF network.
• If we have the basis functions, Áj , then it is exactly same as a linear model and we can use
standard linear least squares method to learn wj .
• To fix Áj , we need to essentially fix  j (and may be  ).
• So, if we can somehow fix centers and widths of the RBF nodes, then we can learn the wj
very easily.
As we have discussed earlier, these RBF networks use „local‟ representations.
• What this means is that  j should be „representative‟ points of the feature space and they
should „cover‟ the feature space.
• Essentially, the proof that these networks can represent any continuous function is based on
having such centers for RBF nodes.
• We can use such ideas to formulate methods for fixing centers of RBF nodes.
One simple method of choosing centers,  j , is to randomly choose p of the training
examples.
• We know that with N hidden nodes and centers same as training examples, we get perfect
interpolation.
• Hence we can take some of the training examples as centers.
• There can be some variations on this theme.
• However, such a method does not, in general, ensure that we have representative points in
the feature space as centers.
When we have p hidden nodes, we need p „centers‟.
• Hence we are looking for p number of „representative‟points in the feature space.
• The only information we have are the N training examples.
• Hence the problem is:
given N points, Xi, i = 1, ・ ・ ・ ,N in <m, find p „representative‟ points in <m.
• This is the „clustering problem‟ This is a problem of forming the data into p clusters.
• We can take the „cluster centers‟ to be the representative points.
• The kind of clusters we get depends on how we want to formalize the notion of the p points
being representative of the N data points.
• We now look at one notion of clustering that is popular.
Let 1 , ・ ・ ・ ,  p represent the p cluster centers.
• Now we need an objective function that specifies how representative these are of the data
Xi, i = 1, ・ ・ ・ ,N.
Now we can define a cost function as

• The J is a function of  j , j = 1, ・ ・ ・ , p. (Note that Sj are also functions of the  j ‟s).

• For a given set of centers, {  j }, J gives us the total error in approximating each of the
training data by its nearest cluster center.
• Hence we want to choose centers to minimize J We now discuss a simple algorithm to find
centers to minimize J.
• This is known as K-means clustering algorithm.
(Originally proposed by Lloyd in the context of vector quantization).
• We are given N data points, Xi, i = 1, ・ ・ ・ ,N.
We want to find p cluster centers  j , j = 1, ・ ・ ・ , p, to minimize J.
• We first rewrite J in a different form to motivate our algorithm.
We think of the problem as finding the centers  1 , ・ ・ ・ , p and assigning Xi to these
clusters.
• Let j , n = 1, ・ ・ ・ ,N, j = 1, ・ ・ ・ , p be indicators of the cluster assignment.
• That is, if we assign Xn to cluster j, then we would have 1 = 1 and p = 0,
• Now we can rewrite J as

We now have to find a way of minimizing J wrt all nj and j .

Note that for a given n, nj is 1 for exactly one j (and it is zero otherwise).
• Thus the μj would be the mean of all data vectors assigned to the jth cluster.
• This is the reason for the name K-means clustering.
• What we derived are optimum values for nj keeping
μj fixed and optimum values for μj keeping nj fixed.
• Hence, in an algorithm we do this repeatedly.
• This is like the EM algorithm.
LECTURE-8
Experiences or learning:
Learning algorithms use experiences in the form of perceptions or perception action pairs to
improve their performance. The nature of experiences available varies with applications.
Some common situations are described below.
Supervised learning: In supervised learning a teacher or oracle is available which provides
the desired action corresponding to a perception. A set of perception action pair provides
what is called a training set. Examples include an automated vehicle where a set of vision
inputs and the corresponding steering actions are available to the learner.

Fig. Supervised learning

Unsupervised learning: In unsupervised learning no teacher is available. The learner only
discovers persistent patterns in the data consisting of a collection of perceptions. This is also
called exploratory learning. Finding out malicious network attacks from a sequence of
anomalous data packets is an example of unsupervised learning.
Active learning: Here not only a teacher is available, the learner has the freedom to ask the
teacher for suitable perception-action example pairs which will help the learner to improve its
performance. Consider a news recommender system which tries to learn an users preferences
and categorize news articles as interesting or uninteresting to the user. The system may
present a particular article (of which it is not sure) to the user and ask whether it is interesting
or not.
Reinforcement learning: In reinforcement learning a teacher is available, but the teacher
instead of directly providing the desired action corresponding to a perception, return reward
and punishment to the learner for its action corresponding to a perception. Examples include
a robot in a unknown terrain where its get a punishment when its hits an obstacle and reward
when it moves smoothly.
In order to design a learning system the designer has to make the following choices based on
the application.

LECTURE-9

Unsupervised Learning in Neural Networks:

Unsupervised learning mechanisms differ from supervised learning in that there is no

"teacher" to instruct the network.
 Competitive Learning:
Competitive learning is a form of unsupervised learning which performs clustering over the
input data. In a competitive learning network with n-output neurons, each output neuron is
associated with a cluster. When a data point from a cluster is presented to the network, only
the neuron corresponding to that cluster responds, while all other neurons remain silent. The
single neuron that responds is often called a “winner” and therefore a competitive learning
network of the kind just described is also known as a “winner-take-all” network.
It is easiest to introduce CL mechanism as a slight variation of Hebb‟s rule.

Kohonen Self-organizing Map:

It is also known as Kohonen feature map or topology-preserving map or Kohonen Self-
organizing .
Information is often represented spatially in the two-dimensional neuronal sheets in the brain,
in both the cortex and subcortical structures. We have learnt about the somtosensory, motor
and visual maps in the corresponding sensory cortices in the brain. A map, in its ordinary
sense, denotes a two-dimensional representation of a real-world domain, such that nearby
points in the domain are mapped onto nearby points in the map.
Due to this “adjacency-preserving” property, these maps are also called topographic maps.
Self-organizing maps (SOM) are models of the topographic maps of the brain, first proposed
by Teovo Kohonen.
The SOM model can be presented as an extension of the competitive learning model
described in the previous section. It is constructed by adding a biologically-relevant feature
that is not originally present in the competitive learning network.
A key property of the SOM is that nearby or similar inputs activate nearby neurons in the
map. The competitive learning network does not have this property.
Consider a hypothetical competitive learning network with 3 output neurons. The input space
is two-dimensional. The weight vectors w1, w2, w3 lie on a line as shown in Fig., with w1 in
between w2 and w3. Note that such an arrangement is possible since there is no relation
between the spatial position of the weight vectors and their indices.

Fig. weight vectors and their indices when not related

The essence of the modification proposed in the SOM model, is a mechanism that ensures
that the weight vectors remain spatially ordered, while they also move towards the data points
that activate them maximally.
Unlike a competitive learning network, which consists of a single row of output neurons, a
SOM consists of a m-dimensional grid of neurons. Usually two-dimensional SOMs are
studied since SOMs were originally inspired by the two-dimensional maps in the brain. The
topology of the grid is usually rectangular, though sometimes hexagonal topologies (Fig.) are
also considered.
Figure: Rectangular and hexagonal trajectories of Kohonen‟s network

As in the case of competitive learning, the weight vector of the winner is moved towards the
input, x. But addition, neurons close to the winner in the SOM are also moved towards the
input, x, but with a lesser learning rate. Neurons that are nearby in the SOM are defined by a
neighborhood N .

Fig. For the neuron in white (center) the neurons in red represent the neighborhood if we
consider the neighborhood radius to be 1

Neighborhood size is large in the early stages, and is decreased gradually as training
progresses.
Learning Vector Quantization(LVQ):
Vector quantization is noting but clustering, where Given a set of vectors {x}, find a set of
representative vectors {wm; 1 ≤m ≤M} such that each x is quantized into a particular wm.
{wm} locate at the mean (centroid) of the density distribution of each cluster. LVQ is an
unsupervised pattern classifier where the actual class membership information is not used.

Fig. Clusters of data

Applications of LVQ:

Speech Recognition
• Robot Arm control
• Industrial process control
• automated synthesis of digital systems
• channel equalization for telecommunication
• image compression
• radar classification of sea-ice
• optimization problems
• sentence understanding
• classification of insect courtship songs

LECTURE-10

Linear neuron model: (Hebbian Learning)

Hebb described a simple learning method of synaptic weight change. In Hebbian learning,
when 2 cells have strong responses, and fire simultaneously, their connection strength or
weight increases. The weight increase is proportional to the frequency at which they fire
together.

Fig. A simple network topology for Hebbian Learning, where W ij resides between two
neurons

Where η is the learning rate, f (.) is the neuron function, x is the input to the jth neuron.
Since the weights are adjusted according to the correlation formula is a type of correlational
learning rule.
A sequence of learning patterns indexed by p is presented to the network. Initial weights are
taken zero. So updated weight after entire data set is:

Frequent input patterns have more impact on weights, giving largest output at end.
The objective function is maximized to maximize output.

This rule causes unconstrained growth of weights. Hebbian rule was modified by Oja by
normalization.
Modified Hebbian Learning:

For small learning rate expanding in Taylor‟s series weight update rule becomes

Here, a weight decay proportional to the squared output is added to maintain weight vector
unit length automatically.

LECTURE-11

ANFIS: Adaptive Neuro-Fuzzy Inference Systems:

ANFIS are a class of adaptive networks that are functionally equivalent to fuzzy inference
systems.
• ANFIS represent Sugeno & Tsukamoto fuzzymodels.
• ANFIS uses a hybrid learning algorithm

Fig. Architecture of ANFIS

Ol,i is the output of the ith node of the layer l.

• Every node i in this layer is an adaptive node with a node function
O1,i = μAi(x) for i = 1, 2, or O1,i = μBi−2(x) for i = 3, 4
• x (or y) is the input node i and Ai (or Bi−2) is a linguistic label associated with this node
• Therefore O1,i is the membership grade of a fuzzy set (A1,A2,B1,B2).
Typical membership function is Gaussian.
Every node in this layer is a fixed node labelled Prod.
• The output is the product of all the incoming signals.
• O2,i = wi = μAi(x) ・ μBi(y), i = 1, 2
• Each node represents the fire strength of the rule
• Any other T-norm operator that perform the AND operator can be used
Every node in this layer is a fixed node labelled Norm.
• The ith node calculates the ratio of the ith rulet‟s firing strenght to the sum of all rulet‟s
firing strengths.
• O3,i = wi = wi w1+w2 , i = 1, 2
• Outputs are called normalized firing strengths.
Every node i in this layer is an adaptive node with a node function:
O4,1 = wifi = wi(px + qiy + ri)
• wi is the normalized firing strenght from layer 3.
• {pi, qi, ri} is the parameter set of this node.
• These are referred to as consequent parameters.
The single node in this layer is a fixed node labeled sum, which computes the overall output
as the summation of all incoming signals:
• overall output = O5,1 =Pi wifi = Pi wifi Pi wi
Hybrid Learning Algorithm:
The ANFIS can be trained by a hybrid learning algorithm presented by Jang in the chapter 8
of the book.
• In the forward pass the algorithm uses least-squares method to identify the consequent
parameters on the layer 4.
• In the backward pass the errors are propagated backward and the premise parameters are
updated by gradient descent.

Fig. Two passes in the hybrid learning algorithm for ANFIS.

Suppose that an adptive network has L layers and the kth layer has #(k) nodes.
• We can denote the node in the ith position of the kth layer by (k, i).
• The node function is denoted by Oki .
• Since the node output depends on its incoming signals and its parameter set (a, b, c), we
have

• Notice that Oki is used as both node output and node function. Assume that a training data
set has P entries.
• The error measure for the pth entry can be defined as the sum of the squared error

Tm,p is the mth component of the pth target.

• OLm,p is the mth component the actual output vector.
• The overall error is
In order to implement the gradient descent in E we calculate the error rate E
O for the pth training data for each node output O.
• The error rate for the output note at (L, i) is

For the internal node at (k, i), the error rate can be derived by the chain rule:

where 1 ≤ k ≤ L − 1
• The error rate of an internal node is a linear combination of the error rates of the nodes in
the next layer.

Consider _ one of the parameters.

• Therefore

where S is the set of nodes

• The derivative of the overall error with respect to _ is

The update formula for  is

If the parameters are to be updated after each input-output pair (on-line training) then the
update formula is:

With the batch learning (off-line learning) the update formula is based on the derivative of
the overall error with respect to α:

Problems of the gradient descent are:

The method is slow.
• It is likely to be trapped in local minima.
Hybrid Learning Rule:
Combines:
• the gradient rule;
• the least squares estimate.
Considere that the adptive network has only one output.
• output = F(I, S)
• I is the vector of input variables.
• S is the set of parameters.
• F is the function implemented by the ANFIS.
• If there exists a function H such that the composite function H ◦ F is linear in some elements
of S then these elements can be identified by LSM.
More formally, if the parameter set S can be decomposed into two sets S = S1 ⊕ S2 (⊕
direct sum), such that H ◦ F is linear in the elements of S2
• then applying H to output = F(I, S) we have H(output) = H ◦ F(I, S) (7) which is linear in the
elements of S2.
• Given values of elements of S1, it is possible to plug P training data in equation 7.
• As a result we obtain a matrix equation A_ = y where _ is the unknown vector whose
elements are parameters in S2.
• This is the standard linear least-square problem.
Combining LSE and gradient descent:
- forward pass
In batch mode, each epoch is composed of a forward pass and a backward pass.
• In the forward pass an input vector is presented and the output is calculated creating a row
in the matrices A and y.
• The process is repeated for all training data and the parameters S2 are identified by BLS or
RLS.
• After S2 is identified the error for each pair is computed.
Combining LSE and gradient descent:
- backward pass
The derivative of the error measure with respect to each node output propagate from the
output toward the input.
• The derivatives are:

The parameters in S2 are updated by the gradient method

Applications of ANFIS:
1. Printed Character recognition
2. Inverse Kinematics
3. Nonlinear System identification
4. Channel Equalization
5. Feed back control system
6. Adaptive noise cancellation

References:

5. Chapter 8,9,11,12,19 J.S.R.Jang, C.T.Sun and E.Mizutani, “Neuro-Fuzzy and Soft

Computing”, PHI, 2004, Pearson Education 2004.
6. Chapter 2, 3 & 7 S. Rajasekaran & GA Vijayalakshmi Pai “Neural Networks, Fuzzy
Logic, and Genetic Algorithms synthesis and application”, PHI
7. Chapter 2& 3 Stamatios V. Kartalopoulos “Understanding Neural Networks and Fuzzy
Logic Basic concepts & Applications”, IEEE Press, PHI, New Delhi, 2004.
8. Internet sources.
MODULE-III (10 HOURS)
Derivative-free Optimization Genetic algorithms: Basic concepts, encoding, fitness
function, reproduction. Differences of GA and traditional optimization methods. Basic
genetic programming concepts Applications.
LECTURE-1
Introduction:
Most real world optimization problems involve complexities like discrete, continuous or
mixed variables, multiple conflicting objectives, non-linearity, discontinuity and non-convex
region. The search space (design space) may be so large that global optimum cannot be found
in a reasonable time. The existing linear or nonlinear methods may not be efficient or
computationally inexpensive for solving such problems. Various stochastic search methods
like simulated annealing, evolutionary algorithms (EA) or hill climbing can be used in such
situations. EAs have the advantage of being applicable to any combination of complexities
(multi-objective, non-linearity etc) and also can be combined with any existing local search
or other methods. Various techniques which make use of EA approach are Genetic
Algorithms (GA), evolutionary programming, evolution strategy, learning classifier system
etc.

The principle of Darwinian evolution theory i.e., survival of the fittest is evaluated by a
fitness function derived from objective function. Every individual in a population searches to
be the best according to a fitness function in its own way(randomly).
Basic Concepts:
Optimization means to make the objective function max or min. That means in evolutionary
computing where the individuals/ elements represent possible solutions, an element exists
such that the fitness of that element is the maximum or minimum among all others‟ fitness
depending on it is maximization or minimization problem.
Optimization can be classified as:
1. Deterministic-Uses derivative or gradient to reach final solution
2. Stochastic- Derivative free optimization, a type of random search, suitable for non-
linearity, discontinuity escape from local optima and non-convex region
Components of Genetic Algorithm:
The individuals are genes which encode a trait or a parameter. The design space is to be
converted to genetic space. It is parallel processing by a population used when single point
approach of traditional methods cannot find a possible solution with in the required time
frame.
Important common aspects of evolutionary/swarm optimization algorithms: It is an
iterative process where best solution is searched by a population in search space evaluating a
fitness function.
1. Search space-Space for all feasible solutions is called search space.
2. Solution- It is the point with maximum or minimum value of fitness function.
3. Fitness function- A function derived from objective function
4. Population size- A number of points in a search space used in parallel for computing is
called population, generally ranging from 30 to 200.
5. Constraints- Lower and upper bounds
6. Stopping criteria- it can be no. of iterations, or minimum value of error in fitness or
minimum improvement from previous iteration
LECTURE-2

Fig. Basic cycle of EA

Basic flow chart of EA:
The initial population is usually generated randomly in all EAs. The termination condition
may be a desired fitness function, maximum number of generations etc. In selection,
individuals with better fitness functions from generation „i' are taken to generate individuals
of „i+1‟th generation. New population (offspring) is created by applying recombination and
mutation to the selected individuals (parents). Recombination creates one or two new
individuals by swaping (crossing over) the genome of a parent with another. Recombined
individual is then mutated by changing a single element (genome) to create a new individual.
Finally, the new population is evaluated and the process is repeated. Each step is described in
more detail below with special reference to GA. GA was proposed by Davis E. Goldberg.
LECTURE-3
Special features of GA:
Encoding-
Objects forming possible solution sets to the original problem is called phenotype and the
encoding (representation) of the individuals in the EA is called genotype. In GA each
possible solution is coded in to genetic space. The coding may be binary coding, real coding,
hexadecimal coding, value coding and tree coding.

Binary coding:
If each design variable is given a string of length „l‟, and there are n such variables, then the
design vector will have a total string length of „nl‟. For example, let there are 3 design
variables and the string length be 4 (not necessarily fixed for all problems, depends on
accuracy in representing variable)for each variable. The variables are x1=4,x2=7 & x3=1.
Then the chromosome length is 12, where 4 bit in binary representing x1=0100 , x2=0111,
x3=0001 are genes. So each string/ chromosome represents a different solution.
 An individual consists a genotype and a fitness function. Fitness represents the quality of the
solution (normally called fitness function). It forms the basis for selecting the individuals and
thereby facilitates improvements.

Fig. Flow chart of GA

Decoding:
If xiL & xiU correspond to 0000 to 0111 ; ni the bit length of coding decoded value xi will be
LECTURE-4

PARENT SELECTION:
After fitness function evaluation, individuals are distinguished based on their quality.
According to Darwin's evolution theory the best ones should survive and create new offspring
for the next generation. There are many methods to select the best chromosomes.
1. Roulette wheel selection
2. Boltzmann selection
3. Tournament selection
4. Rank selection
5. Steady state selection
The first one is briefly described.
Roulette Wheel Selection: Parents are selected according to their fitness i.e., each individual
is selected with a probability proportional to its fitness value. In other words, depending on
the percentage contribution to the total population fitness, string is selected for mating to
form the next generation. This way, weak solutions are eliminated and strong solutions
survive to form the next generation. For example, consider a population containing four
strings shown in the Table 1. Each string is formed by concatenating four substrings which
represents variables a,b,c and d. Length of each string is taken as four bits. The first column
represents the possible solution in binary form. The second column gives the fitness values of
the decoded strings. The third column gives the percentage contribution of each string to the
total fitness of the population. Then by "Roulette Wheel" method, the probability of
candidate 1 being selected as a parent of the next generation is 28.09%. Similarly, the
probability that the candidates 2, 3, 4 will be chosen for the next generation are 19.59, 12.89
and 39.43 respectively. These probabilities are represented on a pie chart, and then four
numbers are randomly generated between 1 and 100. Then, the likeliness that the numbers
generated would fall in the region of candidate 2 might be once, whereas for candidate 4 it
might be twice and candidate 1 more than once and for candidate 3 it may not fall at all.
Thus, the strings are chosen to form the parents of the next generation.

The one with higher probability pi calculated is judged as per

where Fi the fitness and n is the number of chromosome.

Choice of Methods of selecting chromosomes to be parents depends on
1. Population diversity- good ones exploited but new areas are explored but slow
2. Selective pressure-better individuals are preferred but chance of getting local max/min
but fast
Generation:
Population of design vector obtained after one computation on all individuals.
Generation gap:
Proportion of individuals in the population that are replaced in each generation. It should be
increased to improve results.
Reproduction:
Generally after choosing a method of selection population with above average fitness
function in one generation are taken as parents and inserted in next population in
multiple(clones) so that good trait is transferred to child.

LECTURE-5

CROSSOVER:
It is a recombination operator. Selection alone cannot introduce any new individuals into the
population, i.e., it cannot find new points in the search space. These are generated by
genetically-inspired operators, of which the most well known are crossover and mutation.
Types-
1. One-point
2. Two-point
3. Uniform
4. Arithmetic
5. Heuristic
6. Matrix
In one point crossover, selected pair of strings is cut at some random position and their
segments are swapped to form new pair of strings.

Fig. One point crossover

In two-point scheme, there will be two break points in the strings that are randomly chosen.
At the break-point, the segments of the two strings are swapped so that new set of strings are
formed. For example, let us consider two 8-bit strings given by '10011101' and '10101011'.
Then according to one-point crossover, if a random crossover point is chosen after 3 bits from
left and segments are cut as shown below: 100 | 11101 101 | 01011 and the segments are
swapped to form 10001011 10111101 According to two-point crossover, if two crossover
points are selected as 100 | 11 | 101 101 | 01 | 011 Then after swapping both the extreme
segments, the resulting strings formed are 10001101 10111011 Crossover is not usually
applied to all pairs of individuals selected for mating. A random choice is made, where the
probability of crossover Pc being applied is typically between 0.6 and 0.9.
MUTATION:
Mutation is applied to each child individually after crossover. This creates diversity. It
randomly alters each gene with a small probability Pm(generally not greater than 0.01).
Types-
1. Flip-bit
2. Boundary
3. Uniform
4. Non- Uniform
5. Gaussian
It injects a new genetic character into the chromosome by changing at random a bit in a string
depending on the probability of mutation. Example: 10111011 is mutated as 10111111 It is
seen in the above example that the sixth bit '0' is changed to '1'. Thus, in mutation process,
bits are changed from '1' to '0' or '0' to '1' at the randomly chosen position of randomly
selected strings. The mutation may be for one individual or a group.

Fig. Mutation bits shown in red

LECTURE-6

ADVANTAGES AND DISADVANTAGES OF EA:

EA can be efficiently used for highly complex problems with multi-objectivity, non-linearity
etc. It provides not only a single best solution, but the 2nd best, 3rd best and so on as
required. It gives quick approximate solutions. EA methods can very well incorporate with
other local search algorithms. There are some drawbacks also in using EA techniques. An
optimal solution cannot be ensured on using EA methods, which are usually known as
heuristic search methods. Convergence of EA techniques are problem oriented. Sensitivity
analysis should be carried out to find out the range in which the model is efficient. Also, the
implementation of these techniques requires good programming skill.
Differences and similarities between GA and other traditional methods:
Differences:
1. GA uses coding which discretizes search space even though function is continuous
2. A discrete function can be handled with no extra cost
3. Works with a population of points instead of single, so multiple optimal solutions can
be captured at a time, reducing no. of run of algorithm
Similarities:
1. Search direction in traditional algorithms is used to find new point, where as 2 points
are used to define search direction a in crossover in GA
2. Search direction is not fixed for all points, as mutation works in GA
New Variants of GA:
According to the encoding, selection, crossover or mutation methods and adaptive changing
of the probabilities with convergence many variants like continuous GA, binary GA, RCGA,
NSGA etc have been developed. In addition the new ones are
1. Messy GA
2. Parallel GA
3. Multiobjective GA
LECTURE-7
Issues for GA Parameter settings:

Choosing basic implementation issues:

1. Representation
2. Population size, mutation rate, ...
3. Selection, deletion policies
4. Crossover, mutation operators
5. Termination Criteria
6. Performance, scalability
7. Solution is only as good as the evaluation function (often hardest part)

Benefits of Genetic Agorithms

1. Concept is easy to understand

2. Modular, separate from application
3. Supports multi-objective optimization
4. Good for “noisy”environments
5. Always gives answer; answer gets better with time
6. Inherently parallel; easily distributed
7. Multiple ways to speed up and improve a GA-based application as knowledge about
problem domain is gained
8. Easy to exploit previous or alternate solutions
9. Flexible building blocks for hybrid applications
10. Substantial history and range of use

Shortcomings of GA:

1. Minimal deception problem- Some objective functions may be very difficult to optimize
by GA. Representing the solution accuracy depends on coding.
2. GA drift(Bias)- Loss of population diversity may seek suboptimal solution with a smaller
population size
3. Real time& online issues- It does not guarantee response time, which is vital in real time
issues. Works offline satisfactorily.
4. Computationally expensive and time consuming
5. Issues in representation of problem
6. Proper writing of fittness function
7. Proper values of size of population, crossover and mutation rate
8. Premature Convergence
9. No one mathematically perfect solution since problems of biological adaptation don't have
this issue

LECTURE-8 to 10

Applications of Genetic Algorithms:

1. Optimization and design

 numerical optimization, circuit design, airplane design, factory scheduling, drug design,
network optimization
2. Automatic programming
evolving computer programs (e.g., for image processing), evolving cellular automata
3. Machine learning and adaptive control
robot navigation, evolution of rules for solving “expert”problems, evolution of neural
networks, adaptive computer security, adaptive user interfaces

4. Complex data analysis and time-series prediction

prediction of chaotic systems, financial-market prediction, protein-structure prediction
5. Scientific models of complex systems
economics, immunology, ecology, population genetics, evolution, cancer

References:
1. Chapter 7 J.S.R.Jang, C.T.Sun and E.Mizutani, “Neuro-Fuzzy and Soft
Computing”, PHI, 2004, Pearson Education 2004.
2. Chapter 8 & 9 S. Rajasekaran & GA Vijayalakshmi Pai “Neural Networks, Fuzzy
Logic, and Genetic Algorithms synthesis and application”, PHI
3. Internet sources
MODULE-IV (10 HOURS)
Evolutionary Computing, Simulated Annealing, Random Search, Downhill Simplex Search, Swarm
optimization
LECTURE-1

Drawback of traditional techniques

Computing tasks have to be
 well-defined
 fairly predictable
 computable in reasonable time with serial computers

Analogy-based algorithms
For any natural phenomenon you can think of, there will be at least one AI research
group that will have a combinatorial optimization algorithm “based” on “analogies”
and “similarities” with the phenomenon. Here’s the beginning of the list…
• Metal cooling annealing
• Evolution / Co-evolution / Sexual Reproduction
• Thermodynamics
• Societal Markets
• Management Hierarchies
• Ant/Insect Colonies
• Immune System
• Animal Behavior Conditioning
• Neuron / Brain Models

• Hill-climbing (okay, that’s a stretch…)

• Particle Physics

LECTURE-2
Advantages of Simulated Annealing
1. Simulated annealing is sometimes empirically much better at avoiding local
minima than hill-climbing. It is a successful, frequently-used, algorithm. Basic hill
climbing algorithm is so prone to getting caught in local optimums.

This is because a hill climber algorithm will simply accept neighbour solutions that
are better than the current solution. When the hill climber can't find any better
neighbours, it stops.

2. Not much opportunity to say anything formal about it (though there is a proof
that with an infinitely slow cooling rate, you’ll find the global optimum).
.
LECTURE-2

Random search algorithms

Random search algorithms are useful for many ill-structured global optimization
problems with continuous and/or discrete variables. Typically random search
algorithms sacrifice a guarantee of optimality for finding a good solution quickly with
convergence results in probability. Random search algorithms include simulated
annealing, tabu search, genetic algorithms, evolutionary programming, particle
swarm optimization, ant colony optimization, cross-entropy, stochastic
approximation, multistart and clustering algorithms, to name a few. They may be
categorized as global (exploration) versus local (exploitation) search, or instance-
based versus model-based. However, one feature these methods share is the use of
probability in determining their iterative procedures. This article provides an overview
of these random search algorithms, with a probabilistic view that ties them together.
A random search algorithm refers to an algorithm that uses some kind of
randomness or probability (typically in the form of a pseudo-random number
generator) in the definition of the method, and in the literature, may be called a
Monte Carlo method or a stochastic algorithm. The term metaheuristic is also
commonly associated with random search algorithms.
Generic Random Search Algorithm
Step 0. Initialize algorithm parameters  0, initial points X0  S and iteration index k
= 0.
Step 1. Generate a collection of candidate points Vk+1  S according to a specific
generator
and associated sampling distribution.
Step 2. Update Xk+1 based on the candidate points Vk+1, previous iterates and
algorithmic
parameters. Also update algorithm parameters  k+1.
Step 3. If a stopping criterion is met, stop. Otherwise increment k and return to Step
1.
This generic random search algorithm depends on two basic procedures, the
generator in
Step 1 that produces candidate points, and the update procedure in Step 2.

LECTURE-3
LECTURE-4

Swarm Optimization
Swarm intelligence (SI) is the collective behavior of decentralized, self-
organized systems, natural or artificial. SI systems consist typically of a population of
simple agents or boids interacting locally with one another and with their
environment. The inspiration often comes from nature, especially biological systems.
The agents follow very simple rules, and although there is no centralized control
structure dictating how individual agents should behave, local, and to a certain
degree random, interactions between such agents lead to the emergence of
"intelligent" global behavior, unknown to the individual agents. Examples in natural
systems of SI include ant colonies, bird flocking, animal herding, bacterial growth,
fish schooling and Microbial intelligence. It is infact a multi-agent system that has
self-organized behaviour that shows some intelligent behaviour.
Two principles in swarm intelligence
self-organization is based on:
activity amplification by positive feedback
activity balancing by negative feedback
amplification of random fluctuations
multiple interactions
stigmergy - stimulation by work - is based on:
work as behavioural response to the environmental state
an environment that serves as a work state memory
work that does not depend on specific agents

Roots in models of social insects behavior:

  Foraging behavior
 Division of labor and task allocation
 Cemetery organization
 Nest building
 Collaborative transport

Properties of collective intelligence systems:

 Distributed computation
 Direct and indirect interactions
 Agents equipped with simple computational capabilities
 Robustness
 Adaptiveness

Multiple interactions among agents

Simple agents (e.g., rule based)
Sistems composed of many agents
Positive feedback
Amplification of random fluctuations and structure formation
Reinforcement of most common behavior patterns
Negative feedback
Saturation
Competition
Resource exhaustion

Particle Swarm Optmization

 Population initialized by assigning random positions and velocities; potential
solutions are then flown through hyperspace.
 Each particle keeps track of its “best” (highest fitness) position in hyperspace.
 This is called “pbest” for an individual particle
 It is called “gbest” for the best in the population
 It is called “lbest” for the best in a defined neighborhood
 At each time step, each particle stochastically accelerates toward its pbest
and gbest (or lbest).

LECTURE-5

Ant Algorithm:
Algorithms inspired by the behavior of real ants
Examples:
 Foraging
 Corpse clustering
 Division of labor
While walking ants deposit a substance called pheromone on the ground They
choose with higher probability paths that are marked by stronger pheromone
concentrations
Cooperative interaction which leads to the emergence of short(est) paths
LECTURE-6

Bees Algorithm:
 The queen moves randomly over the combs eggs are more likely to be layed
in the neighbourhood of brood
 honey and pollen are deposited randomly in empty cells
 four times more honey is brought to the hive than pollen
 removal ratios for honey: 0.95; pollen: 0.6
 removal of honey and pollen is proportional to the number of surrounding cells
containing brood

The above are few examples of SI, there are numerous others.
LECTURE-7-10
Swarm Optimization Applications:
1. Combinatorial optimization
2. Mixed integer-continuous optimization
3. Networks: AntNet
4. Data clustering and exploratory data analysis
5. Coordinated motion
6. Self-assembling

No free Lunch Theorem:

A number of “no free lunch” (NFL) theorems are establish that for any algorithm, any
elevated performance over one class of problems is offset by performance over
another class. These theorems result in a geometric interpretation of what it means
for an algorithm to be well suited to an optimization problem. Applications of the NFL
theorems to information-theoretic aspects of optimization and benchmark measures
of performance are also presented.

References:
1. Yamille del Valle, Ganesh Kumar Venayagamoorthy, Salman Mohagheghi,
and Ronald G. Harley, “Particle Swarm Optimization: Basic Concepts,
Variants and Applications in Power Systems, IEEE Transactions On
Evolutionary Computation, Vol. 12, No. 2, April 2008, pp. 171-195.
2. Krause, J., Ruxton, G. D., & Krause, S. (2010). Swarm intelligence in animals
and humans. Trends in ecology & evolution, 25(1), 28-34.
3. Internet sources.
4. Chapter 7 J.S.R.Jang, C.T.Sun and E.Mizutani, “Neuro-Fuzzy and Soft
Computing”, PHI, 2004, Pearson Education 2004.