Soft Computing

Soft computing differs from conventional or hard computing in that, unlike hard computing, it is
tolerant of imprecision, uncertainty, partial truth, and approximate reasoning. Conventional or
hard computing requires a precisely stated analytical model and a lot of mathematical and logical
computation.
Components of soft computing include:
Neural networks (NN),
Fuzzy logic (FL),
Evolutionary computation (EC), (which includes Evolutionary algorithms, Genetic algorithms,
Differential evolution),
Meta-heuristics and Swarm Intelligence, (which includes Ant colony optimization, Particle
swarm optimization, Firefly algorithm, Cuckoo search),
Support Vector Machines (SVM),
Probabilistic reasoning, (which includes Bayesian network), and Chaos theory.

NOTE:
SC is evolving rapidly

New techniques and applications are constantly being proposed

Generally speaking, soft computing techniques resemble biological processes more closely than
traditional techniques, which are largely based on formal logical systems, such as sentential logic
and predicate logic, or rely heavily on computer-aided numerical analysis (as in finite element
analysis). Soft computing techniques are intended to complement each other.
Unlike hard computing schemes, which strive for exactness and full truth, soft computing
techniques exploit the given tolerance of imprecision, partial truth, and uncertainty for a
particular problem. Another common contrast comes from the observation that inductive
reasoning plays a larger role in soft computing than in hard computing.

Hard Computing Soft computing

Precisely stated analytical model Tolerant to imprecision, uncertainty, partial
truth, approximation
Based on binary logic, crisp systems, Fuzzy logic, neural nets, probabilistic
numerical analysis, crisp software reasoning.
Programs are to be written Evolve their own programs
Two values logic Multi valued logic
Exact input data Ambiguous and noisy data
Strictly sequential Parallel computations
Precise answers Approximate answers

Rohan verma Page 1

 Many contemporary problems do not lend themselves to precise solutions such as Recognition
problems (handwriting, speech, objects, images), forecasting, combinatorial problems etc.
Soft Computing became a formal Computer Science area of study in the early 1990's.Earlier
computational approaches could model and precisely analyze only relatively simple systems.
More complex systems arising in biology, medicine, the humanities, management sciences, and
similar fields often remained intractable to conventional mathematical and analytical methods.
That said, it should be pointed out that simplicity and complexity of systems are relative, and
many conventional mathematical models have been both challenging and very productive. Soft
computing deals with imprecision, uncertainty, partial truth, and approximation to achieve
tractability, robustness and low solution cost.
Implications of Soft Computing
Soft computing employs NN, SVM, FL etc, in a complementary rather than a competitive way.
One example of a particularly effective combination is what has come to be known as "neuro-
fuzzy systems.” Such systems are becoming increasingly visible as consumer products ranging
from air conditioners and washing machines to photocopiers, camcorders and many industrial
applications.

Some Applications of Soft Computing

• Application of soft computing to handwriting recognition
• Application of soft computing to automotive systems and manufacturing
• Application of soft computing to image processing and data compression
• Application of soft computing to architecture
• Application of soft computing to decision-suppor systems
• Application of soft computing to power systems
• Neurofuzzy systems
• Fuzzy logic control

Overview of Techniques in Soft Computing

• Neural networks
• Fuzzy Logic
• Genetic Algorithms in Evolutionary Computation
• Support Vector Machines

Rohan verma Page 2

 Neural Networks:
Neural Networks, which are simplified models of the biological neuron system, is a massively
parallel distributed processing system made up of highly interconnected neural computing
elements that have the ability to learn and thereby acquire knowledge and making it available for
use. It resembles the brain in two respects: (1) Knowledge is acquired by the network through a
learning process and (2) Interconnection strengths known as synaptic weights are used to store
the knowledge

Fig. 1: A biological neuron connected to another neuron.

A neuron is composed of nucleus- a cell body known as soma. Attached to the soma are long
irregularly shaped filaments called dendrites. The dendrites behave as input channels, all inputs
from other neurons arrive through dendrites.
Another link to soma called Axon is electrically active and serves as an output channel. If the
cumulative inputs received by the soma raise internal electric potential of the cell known as
membrane potential, then the neuron fires by propagating the action potential down the axon to
excite or inhibit other neurons. The axon terminates in a specialized contact called synapse that
connects the axon with the dendrite links of another neuron. An artificial neuron model bears
direct analogy to the actual constituents of biological neuron. This model forms basis of
Artificial Neural Networks

Artificial Neural Networks (ANNs)

The ANN provides a general practical method for real-valued, discrete-valued, and vector-valued
functions from examples. The back propagation algorithm which is widely used in ANNs, uses
gradient descent to tune network parameters to best fit a training set of input-output pairs.

Rohan verma Page 3

 An artificial neural network (ANN) is a distributed computing scheme based on the structure of
the nervous system of humans. The architecture of a neural network is formed by connecting
multiple elementary processors, this being an adaptive system that has an algorithm to adjust
their weights (free parameters) to achieve the performance requirements of the problem based on
representative samples.
Examinations of the human's central nervous system inspired the concept of neural networks. In
an Artificial Neural Network, simple artificial nodes, known as "neurons", "neurodes",
processing elements" or "units", are connected together to form a network which mimics a
biological neural network.
There is no single formal definition of what an artificial neural network is. However, a class of
statistical models may commonly be called "Neural" if they possess the following characteristics:
1. consist of sets of adaptive weights, i.e. numerical parameters that are tuned by a learning
algorithm, and
2. are capable of approximating non-linear functions of their inputs.
The adaptive weights are conceptually connection strengths between neurons, which are
activated during training and prediction.
Neural networks are similar to biological neural networks in performing functions collectively
and in parallel by the units, rather than there being a clear delineation of subtasks to which
various units are assigned. The term "neural network" usually refers to models employed in
statistics, cognitive psychology and artificial intelligence. Neural network models which emulate
the central nervous system are part of theoretical neuroscience and computational neuroscience.
In modern software implementations of artificial neural networks, the approach inspired by
biology has been largely abandoned for a more practical approach based on statistics and signal
processing. In some of these systems, neural networks or parts of neural networks (like artificial
neurons) form components in larger systems that combine both adaptive and non-adaptive
elements. While the more general approach of such systems is more suitable for real-world
problem solving, it has little to do with the traditional artificial intelligence connectionist models.
What they do have in common, however, is the principle of non-linear, distributed, parallel and
local processing and adaptation. Historically, the use of neural networks models marked a
paradigm shift in the late eighties from high-level (symbolic) artificial intelligence, characterized
by expert systems with knowledge embodied in if-then rules, to low-level (sub-symbolic)
machine learning, characterized by knowledge embodied in the parameters of a dynamical
system.
Biological Motivation
 Human brain is a densely interconnected network of approximately 1011 neurons, each
connected to, on average, 104 others.
 Neuron activity is excited or inhibited through connections to other neurons.
 The fastest neuron switching times are known to be on the order of 10-3 sec.

Rohan verma Page 4

 Genetic Algorithm:

Genetic algorithms (GAs) are search methods based on principles of natural selection and
genetics. GAs were First described by John Holland in the 1960s and further developed by
Holland and his students and colleagues at the University of Michigan in the 1960s and 1970s.
Holland's goal was to understand the phenomenon of \adaptation" as it occurs in nature and to
develop ways in which the mechanisms of natural adaptation might be imported into computer
systems. Holland's 1975 book Adaptation in Natural and Artificial Systems (Holland,
1975/1992) presented the GA as an abstraction of biological evolution and gave a theoretical
framework for adaptation under the GA.

The population size, which is usually a user-specified parameter, is one of the important
factors affecting the scalability and performance of genetic algorithms. For example, small
population sizes might lead to premature convergence and yield substandard solutions. On the
other hand, large population sizes lead to unnecessary expenditure of valuable computational
time. Once the problem is encoded in a chromosomal manner and a fitness measure for
discriminating good solutions from bad ones has been chosen, we can start to evolve solutions to
the search problem using the following steps:
1. Initialization: The initial population of candidate solutions is usually generated randomly
across the search space. However, domain-specific knowledge or other information can be easily
incorporated.
2. Evaluation: Once the population is initialized or an offspring population is created, the fitness
values of the candidate solutions are evaluated.
3. Selection: Selection allocates more copies of those solutions with higher fitness values and
thus imposes the survival-of-the-fittest mechanism on the candidate solutions. The main idea of
selection is to prefer better solutions to worse ones, and many selection procedures have been
proposed to accomplish this idea, including roulette-wheel selection, stochastic universal
selection, ranking selection and tournament selection, some of which are described in the next
section.
4. Recombination: Recombination combines parts of two or more parental solutions to create
new, possibly better solutions (i.e. offspring). There are many ways of accomplishing this (some
of which are discussed in the next section), and competent performance depends on a properly
designed recombination mechanism. The offspring under recombination will not be identical to
any particular parent and will instead combine parental traits in a novel manner
5. Mutation: While recombination operates on two or more parental chromosomes, mutation
locally but randomly modifies a solution. Again, there are many variations of mutation, but it
usually involves one or more changes being made to an individual‟s trait or traits. In other words,
mutation performs a random walk in the vicinity of a candidate solution.
6. Replacement. The offspring population created by selection, recombination, and mutation
replaces the original parental population. Many replacement techniques such as elitist
replacement, generation-wise replacement and steady-state replacement methods are used in
GAs.
7. Repeat steps 2–6 until a terminating condition is met. Goldberg has likened GAs to
mechanistic versions of certain modes of human innovation and has shown that these operators
when analyzed individually are ineffective, but when combined together they can work well.

Rohan verma Page 5

 Introduction to Taylor's theorem for multivariable functions

Remember one-variable calculus Taylor's theorem. Given a one variable function f(x), you can fit
it with a polynomial around x=a.

For example, the best linear approximation for f(x) is

f(x)≈f(a)+f′(a)(x−a).

This linear approximation fits f(x) (shown in green below) with a line (shown in blue) through
x=a that matches the slope of f at a.

We can add additional, higher-order terms, to approximate f(x) better near a. The best quadratic
approximation is

f(x)≈f(a)+f′(a)(x−a)+12f′′(a)(x−a)2

We could add third-order or even higher-order terms:

f(x)≈f(a)+f′(a)(x−a)+12f′′(a)(x−a)2+16f′′′(a)(x−a)3+⋯.

The important point is that this Taylor polynomial approximates f(x) well for x near a.

We want to generalize the Taylor polynomial to (scalar-valued) functions of multiple variables:

f(x)=f(x1,x2,…,xn).

We already know the best linear approximation to f. It involves the derivative,

f(x)≈f(a)+Df(a)(x−a).

where Df(a) is the matrix of partial derivatives. The linear approximation is the first-order Taylor
polynomial.

Rohan verma Page 6

 What about the second-order Taylor polynomial? To find a quadratic approximation, we need to
add quadratic terms to our linear approximation. For a function of one-variable f(x), the quadratic
term was

12f′′(a)(x−a)2.

For a function of multiple variables f(x), what is analogous to the second derivative?

Since f(x) is scalar, the first derivative is Df(x), a 1×n matrix, which we can view as an n-
dimensional vector-valued function of the n-dimensional vector x. For the second derivative of
f(x), we can take the matrix of partial derivatives of the function Df(x). We could write it as
DDf(x) for the moment. This second derivative matrix is an n×n matrix called the Hessian
matrix of f. We'll denote it by Hf(x),

Hf(x)=DDf(x).

When f is a function of multiple variables, the second derivative term in the Taylor series will
use the Hessian Hf(a). For the single-variable case, we could rewrite the quadratic expression as

12(x−a)f′′(a)(x−a).

The analog of this expression for the multivariable case is

12(x−a)THf(a)(x−a).

We can add the above expression to our first-order Taylor polynomial to obtain the second-order
Taylor polynomial for functions of multiple variables:

f(x)≈f(a)+Df(a)(x−a)+12(x−a)THf(a)(x−a).

The second-order Taylor polynomial is a better approximation of f(x) near x=a than is the linear
approximation (which is the same as the first-order Taylor polynomial). We'll be able to use it
for things such as finding a local minimum or local maximum of the function f(x).

Rohan verma Page 7

 INTRODUCTION TO THE CONCEPT OF F FUZZY LOGIC
In many cases in our day-to-day work we do not do exact measurements like washing a
cloth, driving a vehicle, brushing the teeth, adding salt or sugar to cooked food, etc. So also in
many cases of our day-to-day conversation we use statements like more or less tall, too fat, very
thin, slightly dwarf, etc which do not convey exact information, but conveniently tackle the
situations. In these situations it is not critical to be precise in data rather a rough answer solves
the problem. Fuzzy logic is all about this relative importance of precision. Hence fuzzy logics are
used in cases of imprecision and imprecise human reasoning and human perception. In cases of
high precision problems where exact mathematical models are already developed there the fuzzy
logics are not used. Still fuzzy logics are used to drill printed circuit boards under LASER to an
appreciate precision with low cost Hence for cost effectiveness we may prefer fuzzy than exact
mathematical model in many cases.
Applications of Fuzzy Logic:
Some already developed areas of applications of Fuzzy Logic are given below.
1) Fuzzy focusing and image stabilization.
2) Fuzzy air conditioner that controls temperature according to comfort index.
3) Fuzzy washing machine that washes clothes with help of smart sensors.
4) Fuzzy controlled sub-way systems.
5) Fuzzy controlled toasters, rice cookers, vacuum cleaners, etc.
Advantages of Fuzzy Logic
1) Fuzzy logic is conceptually easy to understand.
2) Fuzzy logic is flexible that can add more functionality on top of it without starting afresh.
3) Fuzzy logic is tolerant of imprecise data.
4) Fuzzy logic can model non-linear functions of arbitrary complexity.
5) Fuzzy logic can be added to conventional control system easily.
6) Fuzzy logic can be built on the top of the experience of experts.
7) Fuzzy logic is based on natural language.
FUZZY SET THEORY
Classical set: A classical or crisp set has a clear and unambiguous boundary that separates
the members of a classical set from members not belonging to the set. For example a classical set
of positive integers defined as:
Ac = { x | x < 5 } or Ac = {1, 2, 3, 4}
Here the whole numbers less than 5 belongs to the set Ac and the number 5 and more than 5 do
not belong to the set Ac. Hence number 5 is acting as a boundary.
These classical sets are suitable for various applications and have proven to be an
important tool for mathematics and computer science, but they do not reflect the nature of human
concepts and thoughts, which tends to be abstract and imprecise, as an example of a set of young
men.
Classical Set vs Fuzzy set with example of making a set of tall persons

Rohan verma Page 8

 Degree of Membership
Height of “tall men”
No Name
(cm)
Crisp Fuzzy

1 Boy 206 1 1

2 Martin 190 1 1

3 Dewanto 175 0 0.8

4 Joko 160 0 0.7

5 Kom 155 0 0.4

 What is fuzzy thinking

 Experts rely on common sense when they solve the problems
 How can we represent expert knowledge that uses vague and ambiguous terms in
a computer
 Fuzzy logic is not logic that is fuzzy but logic that is used to describe the
fuzziness. Fuzzy logic is the theory of fuzzy sets, set that calibrate the vagueness.
 Fuzzy logic is based on the idea that all things admit of degrees. Temperature,
height, speed, distance, beauty – all come on a sliding scale.
Jim is tall guy
It is really very hot today

Fuzzy Set: It is a set without crisp or sharp boundary. Here the transition from “belonging to a
set” to “not belonging to a set” is gradual. This is done by assigning a membership function to
the elements or members of the set. The members of a fuzzy set get a membership value within 0
and 1. Here it is obvious that the membership value 0 indicates that the element is not a member
of the fuzzy set.
Definition: If X is a collection of objects denoted generically by x, then fuzzy set Af in X is
defined as a set of ordered pairs:
Af = { ( x, µA(x) ) | x Є X }
where µA(x) is called the membership function for the fuzzy set Af.
The membership function (MF)maps each element of X to a membership grade or membership
value between 0 and 1.

Rohan verma Page 9

  Boolean logic
 Uses sharp distinctions. It forces us to draw a line between a members of
class and non members.
 Fuzzy logic
 Reflects how people think. It attempt to model our senses of words, our
decision making and our common sense -> more human and intelligent
systems

Several types of Fuzzy sets and their notations

(1) A Fuzzy set with discrete non-ordered Universe: Let X = {Kolkata, Delhi, Chennai,
Mumbai, Bangalore} be the set of cities one may choose to live in. The fuzzy set A =
“desirable city to live in” may be defined as (using comma notation):
A = { (Mumbai, 0.5), (Delhi, 0.6), (Chennai, 0.3), (Bangalore, 0.7), (Kolkata, 0.8)
}
(2) A Fuzzy set with discrete ordered Universe: Let „U‟ be a number of children from 1 to 5,
that is, U = {1,2, 3, 4, 5}. Now the fuzzy set B = “number of children a couple will wish
to have “ may be defined as (using summation notation):
B = ( 0.5/1 + 0.9/2 + 0.4/3 + 0.2/4 + 0.1/5 )

Rohan verma Page 10

 (3) A Fuzzy set with continuous Universe: Let X = R+ be the set of possible age of human
beings. Then the fuzzy set C = “a 50 year old man” may be defined as:
C = { ( x, µC (x) | x Є X }
where µC (x) = 1/ ( 1 + ( x – 50/10 )10 )

We observed that finding or designing an appropriate membership function has

importance on solution, but there is neither a unique solution nor a unique membership function,
to characterize a particular description. Once a fuzzy set is being defined by a membership
function, then the membership function are not fuzzy, rather they are precise mathematical
functions. Thus by having a fuzzy description with a membership function, we essentially
defuzzify the fuzzy description. In other words, fuzzy set theory does not fuzzify the world rather
they defuzzify the world by assigning a membership value to a fuzzy set elements.
Relation to probability theory: When we devise a fuzzy set, then there must be a unique
membership function associated with it, which expresses the abstract concept of the fuzzy
problem. Thus the membership functions are non-random in nature. Hence fuzzy should not be
compared with probability theory which mostly deals with random phenomena. Fuzziness
describes the ambiguity of an event where as randomness describes the uncertainty in the
occurrence of the event.
Some nomenclature used in Fuzzy sets:
Consider three fuzzy sets and their membership functions (MFs) corresponding to “young”,
“middle aged”, and “old” peoples which are plotted aside. Now we will define some
nomenclature used in fuzzy logic.
(1) Support: The support of a fuzzy set A, denoted as support(A) or supp(A), is the crisp set
of X whose elements all have non-zero membership values in A. Thus we can write
Support(A) = { x | µA(x) > 0 }

Rohan verma Page 11

 (2) Core: The core of a fuzzy set A is the set of all points x in X such that µA (x) = 1. Thus
we can write
Core(A) = {x | µA(x) = 1`}
(3) Normal fuzzy set and Sub normal fuzzy set: A fuzzy set A is called normal if there
exits an x Є X such that µA(x)= 1, that is, core(A) ≠ φ. Otherwise the set is called sub-
normal fuzzy set.
(4) Crossover Points: A crossover point of a fuzzy set A is a point x Є X at which µA(x) =
0.5. Thus we can write
Crossover(A) = { x | µA(x) = 0.5 }
(5) Fuzzy Singleton: A fuzzy set whose support is a single point in X with µA(x) = 1, is
called a fuzzy singleton.
(6) α- cut: The α- cut or α- level set of a fuzzy set A is a crisp set Aα or [A]α that contains all
the elements x in X such that µA(x) ≥ α. Thus we can write
Aα or [A]α = { x | µA(x) ≥ α}
(7) Strong α- cut: Strong α- cut or strong α- level set A‟α is defined as
Aα‟ = { x | µA(x) ≥ α}
Now we can write support(A) = A‟0 and core(A) = A1
(8) Convex fuzzy set: A convex fuzzy set A is described by a membership function whose
membership values are strictly monotonically increasing or whose membership values are
strictly monotonically decreasing or whose membership values are strictly monotonically
increasing and then strictly monotonically decreasing with increasing values of the
elements in the Universe. Thus for any elements x 1, x2 and x3 with relation x1 < x2 < x3
implies that
µA(x2) ≥ min { µA(x1), µA(x3) }
(9) Fuzzy Numbers: If a is a convex continuous point normal fuzzy set defined on the real
line (R), then A is often termed a fuzzy number.
(10) Quasi fuzzy number: A quasi fuzzy number A is a fuzzy set of real line with a
normal fuzzy convex and continuous membership function satisfying the limit condition
lim µA(x) =0 as x → ∞.
(11) Bandwidth: Band width of a normal and convex fuzzy set is defined as
the distance between two unique crossover points. That is
Width(A) = |x2 – x1| where µA(x1) = µA(x2) =0.5
(12) Symmetricity
(13) Open left
(14) Open Right
(15) Closed

Rohan verma Page 12

 Set-Theoretic Operations

A  B  μA  μB
Subset:

Complement: A  X  A  μ A (x)  1 μ A (x)

Union: C  A  B  μc (x)  max(μA (x), μB (x))  μ A (x)  μB (x)

Intersection: C  A  B  μc (x)  min(μA (x), μB (x))  μ A (x)  μB (x)

MF Formulation

Rohan verma Page 13

   xa cx 
Triangular MF: trimf (x; a, b, c)  , ,0
max  min   
  ba cb 
  xa d x 
Trapezoidal MF: trapmf (x; a,b, c, d )  ,1, ,0
max min   
 ba d c 
1  xc 2
  
2 σ 
Gaussian MF: gaussmf (x; a, b, c)  e

1
Generalized bell MF: gbellmf (x; a, b, c) 
x c
2b

1
b

Cylindrical Extension:

Rohan verma Page 14

 2D MFs:

Fuzzy Complement:

Rohan verma Page 15

 General requirements:
• Boundary: N(0)=1 and N(1) = 0
• Monotonicity: N(a) > N(b) if a < b
• Involution: N(N(a) = a
Two types of fuzzy complements:
1 a
Sugeno‟s complement: N (a) 
s
1 sa

Yager‟s complement: N w (a)  (1 a w )1/ w

Fuzzy Intersection: T-norm

Basic requirements:
• Boundary: T(0, 0) = 0, T(a, 1) = T(1, a) = a
• Monotonicity: T(a, b) < T(c, d) if a < c and b < d
• Commutativity: T(a, b) = T(b, a)
• Associativity: T(a, T(b, c)) = T(T(a, b), c)
Four examples (page 37):
• Minimum: Tm(a, b)
• Algebraic product: Ta(a, b)
• Bounded product: Tb(a, b)
• Drastic product: Td(a, b)

Rohan verma Page 16

 Fuzzy Union: T-conorm or S-norm

Rohan verma Page 17

 Basic requirements:
• Boundary: S(1, 1) = 1, S(a, 0) = S(0, a) = a
• Monotonicity: S(a, b) < S(c, d) if a < c and b < d
• Commutativity: S(a, b) = S(b, a)
• Associativity: S(a, S(b, c)) = S(S(a, b), c)
Four examples (page 38):
• Maximum: Sm(a, b)
• Algebraic sum: Sa(a, b)
• Bounded sum: Sb(a, b)
• Drastic sum: Sd(a, b)

Rohan verma Page 18

 Generalized DeMorgan’s Law:

T-norms and T-conorms are duals which support the generalization of DeMorgan‟s law:
• T(a, b) = N(S(N(a), N(b)))
• S(a, b) = N(T(N(a), N(b)))

Parameterized T-norm and S-norm:

Parameterized T-norms and dual T-conorms have been proposed by several researchers:
• Yager
• Schweizer and Sklar
• Dubois and Prade
• Hamacher
• Frank
• Sugeno
• Dombi

Rohan verma Page 19

 Fuzzy Rules and Fuzzy Reasoning

It includes the followings.

• Extension principle
• Fuzzy relations
• Fuzzy if-then rules
• Compositional rule of inference
• Fuzzy reasoning

Extension Principle:
A is a fuzzy set on X : A  μ A (x1 ) / x1  μ A (x2 ) / x2  L  μ A (xn ) / xn

The image of A under f(.) is a fuzzy set B: B  μA (x1 ) / y1  μA (x2 ) / y2   μA (xn ) / yn

where yi = f(xi), for i = 1 to n.

If f(.) is a many-to-one mapping, then μB ( y)  max μ A (x)

x  f 1 ( y )

Fuzzy Relations:
A fuzzy relation R is a 2D MF: R  {((x, y), μR (x, y)) | (x, y)  X Y }

Examples:
• x is close to y (x and y are numbers)
• x depends on y (x and y are events)
• x and y look alike (x and y are persons or objects)
• If x is large, then y is small (x is an observed instrument reading and y is a
corresponding control action which is inversely proportional to x.)
X is close to y is plotted as shown below.

(a) Fuzzy Relation F on X and Y

1
0.5
0
10

5
10
Y 5
0 0 X

Rohan verma Page 20

 Max-Min Composition:
The max-min composition of two fuzzy relations R1 (defined on X and Y) and R2 (defined on Y
and Z):
μ R ∘ R ( x, z)  [μ R ( x, y)  μ R ( y, z)]
1 2 y 1 2

• Associativity: R o (S o T )  (R o S ) o T

• Distributivity over union: R o (S UT )  (R o S ) U(R o T )

• Weak distributivity over intersection:

• Monotonicity: S  T  (R o S )  (R o T )

Max-Star Composition:
Max-product composition: μR oR (x, z)  [μR (x, y)μR ( y, z)]
1 2 y 1 2

In general, we have max * compositions: μR oR (x, z)  [μR (x, y)* μR ( y, z)]

1 2 y 1 2

where * is a T-norm operator.

Example of – Max * Compositions:

R1: x is relevant to y R2: y is relevant to z

z=a z=b
y=α y=β y=γ y=δ

x=1 0.1 0.3 0.5 0.7 y=α 0.9 0.1

x=2 0.4 0.2 0.8 0.9 0.2 0.3

y=β
x=3 0.6 0.8 0.3 0.2
y=γ 0.5 0.6

y=δ 0.7 0.2

How relevant is x=2 to z=a?

µR1ᵒR2(2,a) = 0.7 (by max-min composition)

µR1ᵒR2(2,a) = 0.63 (max-product composition)

Rohan verma Page 21

 Linguistic Variables:

A numerical variable takes numerical values: Age = 65

A linguistic variables takes linguistic values: Age is old
A linguistic value is a fuzzy set.
All linguistic values form a term set (set of terms):
T(age) = {young, not young, very young, middle aged, not middle aged, old, not old, very old,
more or less old, not very young and not very old, etc.}

Operations on Linguistic Values:

Concentration operation like very is defined as CON (A) = A2.
Dilation operation like more or less is defined as DIL (A) = A1/2.
 2 A2 , 0  μA (x)  0.5
Contrast intensification is defined as INT ( A) 

 2(A) , 0.5  μA (x)  1
2

Rohan verma Page 22

 Fuzzy If-Then Rules:
General format: If x is A then y is B
This is interpreted as a fuzzy set
Examples:
• If pressure is high, then volume is small.
• If the road is slippery, then driving is dangerous.
• If a tomato is red, then it is ripe.
• If the speed is high, then apply the brake a little.

Fuzzy If-Then Rules:

Two ways to interpret “If x is A then y is B”
A is coupled with B: (x is A)  (y is B)
A entails B: (x is not A)  (y is B)

Example:
if (profession is athlete) then (fitness is high)
Coupling: Athletes, and only athletes, have high fitness.
The “if” statement (antecedent) is a necessary and sufficient condition.
Entailing: Athletes have high fitness, and non-athletes may or may not have high fitness.
The “if” statement (antecedent) is a sufficient but not necessary condition.

Rohan verma Page 23

 Fuzzy Reasoning:
Single rule with single antecedent Rule: if x is A then y is B
Premise: x is A‟, where A‟ is close to A
Conclusion: y is B‟
Use max of intersection between A and A‟ to get B‟

Single rule with multiple antecedents

Rule: if x is A and y is B then z is C
Premise: x is A‟ and y is B‟
Conclusion: z is C‟
Use min of (A  A‟) and (B  B‟) to get C‟

Multiple rules with multiple antecedents

Rule 1: if x is A1 and y is B1 then z is C1
Rule 2: if x is A2 and y is B2 then z is C2
Premise: x is A‟ and y is B‟
Conclusion: z is C‟
Use previous slide to get C1‟ and C2‟
Use max of C1‟ and C2‟ to get C‟ (next slide)

>> ruleview mam21 (Matlab Fuzzy Logic Toolbox)

Rohan verma Page 24

 Neural Network Representation
 An ANN is composed of processing elements called or neurons organized in different
ways to form the network‟s structure.
Processing Elements
An ANN consists of neurons. Each of the neurons receives inputs, processes inputs and delivers
a single output. A simplest ANN is a single neuron perceptron which is shown in Fig.2.

Fig. 2: A single neuron perceptron.

The input can be raw input data or the output of other perceptrons. The output can be the final
result (e.g. 1 means yes, 0 means no) or it can be inputs to other perceptrons.

The network:
 Each ANN is composed of a collection of perceptrons grouped in layers. A typical
structure is shown in Fig.3.
 Note the three layers of the shown ANN. They are input, intermediate (called the hidden
layer) and output layers.
 Several hidden layers can be placed between the input and output layers.

Rohan verma Page 25

 Fig. 3: A three layer ANN.

Perceptrons:
 A perceptron takes a vector of real-valued inputs, calculates a linear combination of these
inputs, then outputs
 a 1 if the result is greater than some threshold
 –1 otherwise.
 Given real-valued inputs x1 through xn, the output o(x1, …, xn) computed by the
perceptron is given as

o(x1, …, xn) = 1 if w0 + w1x1 + … + wnxn > 0

= -1 otherwise
where wi is a real-valued constant, or weight.
Notice the quantity (-w0) is a threshold that the weighted combination of inputs w1x1 + … + wnxn
must surpass in order for perceptron to output a 1.
 To simplify notation, we imagine an additional constant input x0 = 1, allowing us to write
the above inequality as
n
i=0 wixi >0
 Learning a perceptron involves choosing values for the weights w0, w1,…, wn.

Rohan verma Page 26

 Fig. 4: An n input percetron connected to a thresh-holding function for binary output.

Representation Power of Perceptrons:

 We can view the perceptron as representing a hyperplane decision surface in the n-
dimensional space of instances (i.e. points). The perceptron outputs a 1 for instances
lying on one side of the hyperplane and outputs a –1 for instances lying on the other
side, as in Figure 4. The equation for this decision hyperplane is

 Some sets of positive and negative examples cannot be separated by any hyperplane.
Those that can be separated are called linearly separated set of examples.

Fig.5: (a) Linearly separable classes (b) Not linearly separable classes

A single perceptron can be used to represent many boolean functions.

• AND function :
A decision hyper-plane described by w0 + w1 x1 + w2 x2 = 0 with w0 = -0.8, w1 =0.5 and w2 =
0.5, that is, the hyper-plane -0.8 + 0.5 x1 + 0.5 x2 = 0 can represent Boolean AND function.

Rohan verma Page 27

 x2
< Training exa m ples>

- +

- -
x1
-0.8 + 0.5 x1 + 0.5 x2 = 0

< T est R esults>

x1 x2 wixi o utp ut
0 0 -0.8 -1
0 1 -0.3 -1
1 0 -0.3 -1
1 1 0.2 1

OR function:
Similarly a Decision hyper-plane w0 + w1 x1 + w2 x2 = 0 with w0 = -0.3, w1 =0.5 and w2 = 0.5,
that is, the hyper-plane -0.3 + 0.5 x1 + 0.5 x2 = 0 can represent an OR function.

< Training exa m ples>

x1 x2 o utput
0 0 -1
0 1 1
1 0 1
1 1 1

< T est R esults>

x1 x2 wixi o utp ut
0 0 -0.3 -1
0 1 0.2 -1
1 0 0.2 -1
1 1 0.7 1
XOR function:
It‟s impossible to implement the XOR function by a single perception. So a two-layer network of
perceptrons can represent XOR function as per the equation

Rohan verma Page 28

 x2

+ -

- +
x1

Perceptron training rule:

Although we are interested in learning networks of many interconnected units, let us begin
by understanding how to learn the weights for a single perceptron. Here learning is to determine
a weight vector that causes the perceptron to produce the correct +1 or –1 for each of the given
training examples. Several algorithms are known to solve this learning problem. Here we
consider two: the perceptron rule and the delta rule. One way to learn an acceptable weight
vector is to begin with random weights, then iteratively apply the perceptron to each training
example, modifying the perceptron weights whenever it misclassifies an example. This process is
repeated, iterating through the training examples as many as times needed until the perceptron
classifies all training examples correctly. Weights are modified at each step according to the
perceptron training rule, which revises the weight wi associated with input xi according to the
rule.

wi  wi + wi
where wi = (t – o) xi
Here: t is target output value for the current training example „o‟ is perceptron output
 is small constant (e.g., 0.1) called learning rate
The role of the learning rate is to moderate the degree to which weights are changed at each
step. It is usually set to some small value (e.g. 0.1) and is sometimes made to decrease as the
number of weight-tuning iterations increases. We can prove that the algorithm will converge If
training data is linearly separable and η sufficiently small.
If the data is not linearly separable, convergence is not assured.
Gradient Descent and the Delta Rule:
Although the perceptron rule finds a successful weight vector when the training examples are
linearly separable, it can fail to converge if the examples are not linearly separately. A second
training rule, called the delta rule, is designed to overcome this difficulty. The key idea of delta

Rohan verma Page 29

 rule: to use gradient descent to search the space of possible weight vector to find the weights that
best fit the training examples. This rule is important because it provides the basis for the
backpropagration algorithm, which can learn networks with many interconnected units. The delta
training rule: considering the task of training an unthresholded perceptron, that is a linear unit,
for which the output o is given by:
o = w0 + w1x1 + ··· + wnxn (1)
Thus, a linear unit corresponds to the first stage of a perceptron, without the threshold.
In order to derive a weight learning rule for linear units, let specify a measure for the training
error of a weight vector, relative to the training examples. The Training Error can be computed
as the following squared error

where D is set of training examples, td is the target output for the training example d and Od is
the output of the linear unit for the training example d. Here we characterize E as a function of
weight vector because the linear unit output O depends on this weight vector.
Hypothesis Space: To understand the gradient descent algorithm, it is helpful to visualize the
entire space of possible weight vectors and their associated E values, as illustrated in Figure 5.

Here the axes wo, w1 represents possible values for the two weights of a simple linear unit. The
wo,w1 plane represents the entire hypothesis space. The vertical axis indicates the error E relative
to some fixed set of training examples. The error surface shown in the figure summarizes the
desirability of every weight vector in the hypothesis space. For linear units, this error surface
must be parabolic with a single global minimum. And we desire a weight vector with this
minimum.

Rohan verma Page 30

 Derivation of the Gradient Descent Rule: This vector derivative is called the gradient of E
with respect to the vector <w0,…,wn>, written E .

] (3)

Notice E is itself a vector, whose components are the partial derivatives of E with respect to
each of the wi. When interpreted as a vector in weight space, the gradient specifies the direction
that produces the steepest increase in E. The negative of this vector therefore gives the direction
of steepest decrease. Since the gradient specifies the direction of steepest increase of E, the
training rule for gradient descent is

w w + w

where (4)

Here η is a positive constant called the learning rate, which determines the step size in the
gradient descent search. The negative sign is present because we want to move the weight vector
in the direction that decreases E. This training rule can also be written in its component form

wi wi + wi

where

(5)

which makes it clear that steepest descent is achieved by altering each component wi of weight
vector in proportion to E/wi. The vector of E/wi derivatives that form the gradient can be
obtained by differentiating E from Equation (2), as

Rohan verma Page 31

 (6)

where xid denotes the single input component xi for the training example d. We now have an
equation that gives E/wi in terms of the linear unit inputs xid, output od and the target value td
associated with the training example. Substituting Equation (6) into Equation (5) yields the
weight update rule for gradient descent.

(7)

The gradient descent algorithm for training linear units is as follows: Pick an initial random
weight vector. Apply the linear unit to all training examples, them compute Δwi for each weight
according to Equation (7). Update each weight wi by adding Δwi , them repeat the process. The
algorithm is given in Figure 6.

Because the error surface contains only a single global minimum, this algorithm will converge
to a weight vector with minimum error, regardless of whether the training examples are linearly
separable, given a sufficiently small η is used. If η is too large, the gradient descent search runs
the risk of overstepping the minimum in the error surface rather than settling into it. For this
reason, one common modification to the algorithm is to gradually reduce the value of η as the
number of gradient descent steps grows.

Stochastic Approximation to Gradient Descent:The key practical difficulties in applying

gradient descent are:Converging to a local minimum can sometimes be quite slow (i.e., it can
require many thousands of steps).If there are multiple local minima in the error surface, then
there is no guarantee that the procedure will find the global minimum.One common variation on
gradient descent intended to alleviate these difficulties is called incremental gradient descent (or
stochastic gradient descent). The key differences between standard gradient descent and
stochastic gradient descent are:

 In standard gradient descent, the error is summed over all examples before
upgrading weights, whereas in stochastic gradient descent weights are updated
upon examining each training example.

 The modified training rule is like the training example we update the weight
according to

wi = (t – o) xi (10)

Summing over multiple examples in standard gradient descent requires more computation per
weight update step. On the other hand, because it uses the true gradient, standard gradient
descent is often used with a larger step size per weight update than stochastic gradient descent

Rohan verma Page 32

 Stochastic gradient descent (i.e. incremental mode) can sometimes avoid falling into local
minima because it uses the various gradient of E rather than overall gradient of E to guide its
search. Both stochastic and standard gradient descent methods are commonly used in practice.

MULTILAYER NETWORKS AND THE BACKPROPOGATION ALGORITHM:

Single perceptrons can only express linear decision surfaces. In contrast, the kind of multilayer
networks learned by the backpropagation algorithm are capaple of expressing a rich variety of
nonlinear decision surfaces.This section discusses how to learn such multilayer networks using
a gradient descent algorithm similar to that discussed in the previous section.

Like the perceptron, the sigmoid unit first computes a linear combination of its inputs, then
applies a threshold to the result. In the case of sigmoid unit, however, the threshold output is a
continuous function of its input.The sigmoid function (x) is also called the logistic function.
Interesting property:

Output ranges between 0 and 1, increasing monotonically with its input.We can derive gradient
decent rules to train

Rohan verma Page 33

  One sigmoid unit

 Multilayer networks of sigmoid units  Backpropagation

The Backpropagation (BP)Algorithm:

The BP algorithm learns the weights for a multilayer network, given a network with a fixed set
of units and interconnections. It employs a gradient descent to attempt to minimize the squared
error between the network output values and the target values for these outputs.Because we are
considering networks with multiple output units rather than single units as before, we begin by
redefining E to sum the errors over all of the network output units

E(w) = ½   (tkd – okd)2 (13)

d D koutputs

where outputs is the set of output units in the network, and tkd and okd are the target and output
values associated with the kth output unit and training example d.

The BP algorithm is presented in Figure 8. The algorithm applies to layered feedforward

networks containing 2 layers of sigmoid units, with units at each layer connected to all units
from the preceding layer. This is an incremental gradient descent version of Backpropagation.
The notation is as follows:

xij denotes the input from node i to unit j, and wij denotes the corresponding weight.

n denotes the error term associated with unit n. It plays a role analogous to the quantity (t – o)
in our earlier discussion of the delta training rule

Initialize all weights to small random numbers. Until satisfied do

1. Input the training example to the network and compute the network output.
2. For each output unit k
3. For each hidden unit h

4. Update each network weight

In the BP algorithm, step1 propagates the input forward through the network. And the steps 2, 3
and 4 propagates the errors backward through the network.The main loop of BP repeatedly
iterates over the training examples. For each training example, it applies the ANN to the
example, calculates the error of the network output for this example, computes the gradient with
respect to the error on the example, then updates all weights in the network. This gradient

Rohan verma Page 34

 descent step is iterated until ANN performs acceptably well. A variety of termination conditions
can be used to halt the procedure.

One may choose to halt after a fixed number of iterations through the loop, or once the error on
the training examples falls below some threshold, or once the error on a separate validation set of
examples meets some criteria.

Adding Momentum

Because BP is a widely used algorithm, many variations have been developed. The most
common is to alter the weight-update rule in Step 4 in the algorithm by making the weight
update on the nth iteration depend partially on the update that occurred during the (n -1)th
iteration, as follows

+α

Here wi,j(n) is the weight update performed during the n-th iteration through the main loop of
the algorithm.

- n-th iteration update depend on (n-1)th iteration

- : constant between 0 and 1 is called the momentum.

Role of momentum term:

- keep the ball rolling through small local minima in the error surface.

- Gradually increase the step size of the search in regions where the gradient is unchanging,
thereby speeding convergence

Expressive Capabilities of ANNs

 Boolean functions: Every boolean function can be represented by network with two
layers of units where the number of hidden units required grows exponentially.

 Continuous functions: Every bounded continuous function can be approximated with

arbitrarily small error, by network with two layers of units.

 Arbitrary functions: Any function can be approximated to arbitrary accuracy by a

network with three layers of units .

Hidden layer representations

 Hidden layer representations

 This 8x3x8 network was trained to learn the identity function.

Rohan verma Page 35

  8 training examples are used.

 After 5000 training iterations, the three hidden unit values encode the eight
distinct inputs using the encoding shown on the right.

Learning the 8x3x8 network

Most of the interesting weight changes occurred during the first 2500 iterations.

Figure 10.a The plot shows the sum of squared errors for each of the eight output units as the
number of iterations increases. The sum of square errors for each output decreases as the
procedure proceeds, more quickly for some output units and less quickly for others.

Rohan verma Page 36

 Fig-10.a

Figure 10.b Learning the 8  3  8 network. The plot shows the evolving hidden layer
representation for the input string “010000000”. The network passes through a number of
different encodings before converging to the final encoding

Fig 10.b

Rohan verma Page 37

 Generalization, Overfitting and Stopping Criterion

 Termination condition

 Until the error E falls below some predetermined threshold

 This is a poor strategy

 Overfitting problem

 Backpropagation is susceptible to overfitting the training examples at the cost of

decreasing generalization accuracy over other unseen examples.

 To see the danger of minimizing the error over the training data, consider how the
error E varies with the number of weight iteration.

Rohan verma Page 38

 The
generalization accuracy measured over the training examples first decreases, then increases,
even as the error over training examples continues to decrease. This occurs because the weights
are being tuned to fit idiosyncrasies of the training examples that are not representative of the
general distribution of examples.

Techniques to overcome overfitting problem

Weight decay: Decrease each weight by some small factor during each iteration. The motivation
for this approach is to keep weight values small.

Cross-validation: a set of validation data in addition to the training data. The algorithm monitors
the error w.r.t. this validation data while using the training set to drive the gradient descent
search.

How much weight-tuning iteration should the algorithm perform? It should use the number of
iterations that produces the lowest error over the validation set. Two copies of the weights are
kept: one copy for training and a separate copy of the best weights thus far measured by their
error over the validation set. Once the trained weights reach a higher error over the validation set
than the stored weights, training is terminated and the stored weights are returned.

NEURAL NETWORK APPLICATION DEVELOPMENT

The development process for an ANN application has eight steps.

Rohan verma Page 39

 Step 1: (Data collection) The data to be used for the training and testing of the network are
collected. Important considerations are that the particular problem is amenable to neural network
solution and that adequate data exist and can be obtained.

Step 2: (Training and testing data separation) Trainning data must be identified, and a plan must
be made for testing the performance of the network. The available data are divided into training
and testing data sets. For a moderately sized data set, 80% of the data are randomly selected for
training, 10% for testing, and 10% secondary testing.

Step 3: (Network architecture) A network architecture and a learning method are selected.
Important considerations are the exact number of perceptrons and the number of layers.

Step 4: (Parameter tuning and weight initialization) There are parameters for tuning the network
to the desired learning performance level. Part of this step is initialization of the network weights
and parameters, followed by modification of the parameters as training performance feedback is
received. Often, the initial values are important in determining the effectiveness and length of
training.

Step 5: (Data transformation) Transforms the application data into the type and format required
by the ANN.

Step 6: (Training) Training is conducted iteratively by presenting input and desired or known
output data to the ANN. The ANN computes the outputs and adjusts the weights until the
computed outputs are within an acceptable tolerance of the known outputs for the input cases.

Step 7: (Testing) Once the training has been completed, it is necessary to test the network. The
testing examines the performance of the network using the derived weights by measuring the
ability of the network to classify the testing data correctly. Black-box testing (comparing test
results to historical results) is the primary approach for verifying that inputs produce the
appropriate outputs.

Step 8: (Implementation) Now a stable set of weights are obtained.

Then the network can reproduce the desired output given inputs like those in the training set. The
network is ready to use as a stand-alone system or as part of another software system where new
input data will be presented to it and its output will be a recommended decision.

BENEFITS AND LIMITATIONS OF NEURAL NETWORKS

Benefits of ANNs

Rohan verma Page 40

 -Usefulness for pattern recognition, classification, generalization, abstraction and interpretation
of incomplete and noisy inputs. (e.g. handwriting recognition, image recognition, voice and
speech recognition, weather forecasting).

-Robustness. ANNs tend to be more robust than their conventional counterparts. They have the
ability to cope with incomplete or fuzzy data. ANNs can be very tolerant of faults if properly
implemented.

-Fast processing speed. Because they consist of a large number of massively interconnected
processing units, all operating in parallel on the same problem, ANNs can potentially operate at
considerable speed (when implemented on parallel processors).

- Flexibility and ease of maintenance. ANNs are very flexible in adapting their behavior to new
and changing environments. They are also easier to maintain, with some having the ability to
learn from experience to improve their own performance.

Limitations of ANNs

ANNs do not produce an explicit model even though new cases can be fed into it and new
results obtained.

 ANNs lack explanation capabilities. Justifications for results is difficult to obtain

because the connection weights usually do not have obvious interpretations.

 Providing some human characteristics to problem solving that are difficult to simulate
using the logical, analytical techniques of expert systems and standard software
technologies. (e.g. financial applications).

SOME ANN APPLICATIONS

 Tax form processing to identify tax fraud

 Enhancing auditing by finding irregularities
 Bankruptcy prediction
 Customer credit scoring

Rohan verma Page 41

  Loan approvals
 Credit card approval and fraud detection
 Financial prediction
 Energy forecasting
 Computer access security (intrusion detection and classification of attacks)
 Fraud detection in mobile telecommunication networks

Recent improvements
Computational devices have been created in CMOS, for both biophysical simulation and
neuromorphic computing. More recent efforts show promise for creating nano-devices for very
large scale principal components analyses and convolution. If successful, these efforts could
usher in a new era of neural computing that is a step beyond digital computing, because it
depends on learning rather than programming and because it is fundamentally analog rather than
digital even though the first instantiations may in fact be with CMOS digital devices.
Between 2009 and 2012, the recurrent neural networks and deep feedforward neural networks
developed in the research group of Jürgen Schmidhuber at the Swiss AI Lab IDSIA have won
eight international competitions in pattern recognition and machine learning. For example, multi-
dimensional long short term memory (LSTM) won three competitions in connected handwriting
recognition at the 2009 International Conference on Document Analysis and Recognition
(ICDAR), without any prior knowledge about the three different languages to be learned.
Variants of the back-propagation algorithm as well as unsupervised methods by Geoff Hinton
and colleagues at the University of Toront can be used to train deep, highly nonlinear neural
architectures similar to the 1980 Neocognitron by Kunihiko Fukushima, [17] and the "standard
architecture of vision",[18] inspired by the simple and complex cells identified by David H. Hubel
and Torsten Wiesel in the primary visual cortex.
Deep learning feedforward networks, such as convolutional neural networks, alternate
convolutional layers and max-pooling layers, topped by several pure classification layers. Fast
GPU-based implementations of this approach have won several pattern recognition contests,
including the IJCNN 2011 Traffic Sign Recognition Competition and the ISBI 2012
Segmentation of Neuronal Structures in Electron Microscopy Stacks challenge. Such neural

Rohan verma Page 42

 networks also were the first artificial pattern recognizers to achieve human-competitive or even
superhuman performance on benchmarks such as traffic sign recognition (IJCNN 2012), or the
MNIST handwritten digits problem of Yann LeCun and colleagues at NYU.
Successes in pattern recognition contests since 2009
Between 2009 and 2012, the recurrent neural networks and deep feedforward neural networks
developed in the research group of Jürgen Schmidhuber at the Swiss AI Lab IDSIA have won
eight international competitions in pattern recognition and machine learning. For example, the
bi-directional and multi-dimensional long short term memory (LSTM) of Alex Graves et al. won
three competitions in connected handwriting recognition at the 2009 International Conference on
Document Analysis and Recognition (ICDAR), without any prior knowledge about the three
different languages to be learned. Fast GPU-based implementations of this approach by Dan
Ciresan and colleagues at IDSIA have won several pattern recognition contests, including the
IJCNN 2011 Traffic Sign Recognition Competition, the ISBI 2012 Segmentation of Neuronal
Structures in Electron Microscopy Stacks challenge, and others. Their neural networks also were
the first artificial pattern recognizers to achieve human-competitive or even superhuman
performance on important benchmarks such as traffic sign recognition (IJCNN 2012), or the
MNIST handwritten digits problem of Yann LeCun at NYU. Deep, highly nonlinear neural
architectures similar to the 1980 neocognitron by Kunihiko Fukushima and the "standard
architecture of vision" can also be pre-trained by unsupervised methods of Geoff Hinton's lab at
University of Toronto. A team from this lab won a 2012 contest sponsored by Merck to design
software to help find molecules that might lead to new drugs.
Models
Neural network models in artificial intelligence are usually referred to as artificial neural
networks (ANNs); these are essentially simple mathematical models defining a function
or a distribution over or both and , but sometimes models are also intimately
associated with a particular learning algorithm or learning rule. A common use of the phrase
ANN model really means the definition of a class of such functions (where members of the class
are obtained by varying parameters, connection weights, or specifics of the architecture such as
the number of neurons or their connectivity).
Network function
The word network in the term 'artificial neural network' refers to the inter–connections between
the neurons in the different layers of each system. An example system has three layers. The first
layer has input neurons which send data via synapses to the second layer of neurons, and then via
more synapses to the third layer of output neurons. More complex systems will have more layers
of neurons with some having increased layers of input neurons and output neurons. The synapses
store parameters called "weights" that manipulate the data in the calculations.
An ANN is typically defined by three types of parameters:
1. The interconnection pattern between the different layers of neurons
2. The learning process for updating the weights of the interconnections
3. The activation function that converts a neuron's weighted input to its output activation.

Rohan verma Page 43

 Mathematically, a neuron's network function is defined as a composition of other functions
, which can further be defined as a composition of other functions. This can be conveniently
represented as a network structure, with arrows depicting the dependencies between variables. A
widely used type of composition is the nonlinear weighted sum, where ,
where (commonly referred to as the activation function) is some predefined function, such as
the hyperbolic tangent. It will be convenient for the following to refer to a collection of functions
as simply a vector .

ANN dependency graph

This figure depicts such a decomposition of , with dependencies between variables indicated by
arrows. These can be interpreted in two ways.
The first view is the functional view: the input is transformed into a 3-dimensional vector ,
which is then transformed into a 2-dimensional vector , which is finally transformed into . This
view is most commonly encountered in the context of optimization.
The second view is the probabilistic view: the random variable depends upon the random
variable , which depends upon , which depends upon the random variable .
This view is most commonly encountered in the context of graphical models.
The two views are largely equivalent. In either case, for this particular network architecture, the
components of individual layers are independent of each other (e.g., the components of are
independent of each other given their input ). This naturally enables a degree of parallelism in
the implementation.

Two separate depictions of the recurrent ANN dependency graph

Networks such as the previous one are commonly called feedforward, because their graph is a
directed acyclic graph. Networks with cycles are commonly called recurrent. Such networks are
commonly depicted in the manner shown at the top of the figure, where is shown as being
dependent upon itself. However, an implied temporal dependence is not shown.
Learning
What has attracted the most interest in neural networks is the possibility of learning. Given a
specific task to solve, and a class of functions , learning means using a set of observations to
find which solves the task in some optimal sense.

Rohan verma Page 44

 This entails defining a cost function such that, for the optimal solution ,
– i.e., no solution has a cost less than the cost of the optimal solution (see
Mathematical optimization).
The cost function is an important concept in learning, as it is a measure of how far away a
particular solution is from an optimal solution to the problem to be solved. Learning algorithms
search through the solution space to find a function that has the smallest possible cost.
For applications where the solution is dependent on some data, the cost must necessarily be a
function of the observations, otherwise we would not be modelling anything related to the data. It
is frequently defined as a statistic to which only approximations can be made. As a simple
example, consider the problem of finding the model , which minimizes , for
data pairs drawn from some distribution . In practical situations we would only have
samples from and thus, for the above example, we would only minimize
. Thus, the cost is minimized over a sample of the data rather than the
entire data set.
When some form of online machine learning must be used, where the cost is partially
minimized as each new example is seen. While online machine learning is often used when is
fixed, it is most useful in the case where the distribution changes slowly over time. In neural
network methods, some form of online machine learning is frequently used for finite datasets.
See also: Mathematical optimization, Estimation theory and Machine learning
Choosing a cost function
While it is possible to define some arbitrary ad hoc cost function, frequently a particular cost will
be used, either because it has desirable properties (such as convexity) or because it arises
naturally from a particular formulation of the problem (e.g., in a probabilistic formulation the
posterior probability of the model can be used as an inverse cost). Ultimately, the cost function
will depend on the desired task. An overview of the three main categories of learning tasks is
provided below:
Learning paradigms
There are three major learning paradigms, each corresponding to a particular abstract learning
task. These are supervised learning, unsupervised learning and reinforcement learning.
Supervised learning
In supervised learning, we are given a set of example pairs and the aim is to find a
function in the allowed class of functions that matches the examples. In other words, we
wish to infer the mapping implied by the data; the cost function is related to the mismatch
between our mapping and the data and it implicitly contains prior knowledge about the problem
domain.
A commonly used cost is the mean-squared error, which tries to minimize the average squared
error between the network's output, f(x), and the target value y over all the example pairs. When
one tries to minimize this cost using gradient descent for the class of neural networks called
multilayer perceptrons, one obtains the common and well-known backpropagation algorithm for
training neural networks.

Rohan verma Page 45

 Tasks that fall within the paradigm of supervised learning are pattern recognition (also known as
classification) and regression (also known as function approximation). The supervised learning
paradigm is also applicable to sequential data (e.g., for speech and gesture recognition). This can
be thought of as learning with a "teacher," in the form of a function that provides continuous
feedback on the quality of solutions obtained thus far.
Unsupervised learning
In unsupervised learning, some data is given and the cost function to be minimized, that can be
any function of the data and the network's output, .
The cost function is dependent on the task (what we are trying to model) and our a priori
assumptions (the implicit properties of our model, its parameters and the observed variables).
As a trivial example, consider the model where is a constant and the cost
. Minimizing this cost will give us a value of that is equal to the mean of the
data. The cost function can be much more complicated. Its form depends on the application: for
example, in compression it could be related to the mutual information between and ,
whereas in statistical modeling, it could be related to the posterior probability of the model given
the data. (Note that in both of those examples those quantities would be maximized rather than
minimized).
Tasks that fall within the paradigm of unsupervised learning are in general estimation problems;
the applications include clustering, the estimation of statistical distributions, compression and
filtering.
Reinforcement learning
In reinforcement learning, data are usually not given, but generated by an agent's interactions
with the environment. At each point in time , the agent performs an action and the
environment generates an observation and an instantaneous cost , according to some (usually
unknown) dynamics. The aim is to discover a policy for selecting actions that minimizes some
measure of a long-term cost; i.e., the expected cumulative cost. The environment's dynamics and
the long-term cost for each policy are usually unknown, but can be estimated.
More formally the environment is modelled as a Markov decision process (MDP) with states
and actions with the following probability distributions: the instantaneous
cost distribution , the observation distribution and the transition ,
while a policy is defined as conditional distribution over actions given the observations. Taken
together, the two then define a Markov chain (MC). The aim is to discover the policy that
minimizes the cost; i.e., the MC for which the cost is minimal.
ANNs are frequently used in reinforcement learning as part of the overall algorithm. Dynamic
programming has been coupled with ANNs (Neuro dynamic programming) by Bertsekas and
Tsitsiklis and applied to multi-dimensional nonlinear problems such as those involved in vehicle
routing, natural resources management or medicine because of the ability of ANNs to mitigate
losses of accuracy even when reducing the discretization grid density for numerically
approximating the solution of the original control problems.

Rohan verma Page 46

 Tasks that fall within the paradigm of reinforcement learning are control problems, games and
other sequential decision making tasks.
See also: dynamic programming and stochastic control
Learning algorithms
Training a neural network model essentially means selecting one model from the set of allowed
models (or, in a Bayesian framework, determining a distribution over the set of allowed models)
that minimizes the cost criterion. There are numerous algorithms available for training neural
network models; most of them can be viewed as a straightforward application of optimization
theory and statistical estimation.
Most of the algorithms used in training artificial neural networks employ some form of gradient
descent, using backpropagation to compute the actual gradients. This is done by simply taking
the derivative of the cost function with respect to the network parameters and then changing
those parameters in a gradient-related direction.
Evolutionary methods, gene expression programming, simulated annealing, expectation-
maximization, non-parametric methods and particle swarm optimization are some commonly
used methods for training neural networks.
Employing artificial neural networks
Perhaps the greatest advantage of ANNs is their ability to be used as an arbitrary function
approximation mechanism that 'learns' from observed data. However, using them is not so
straightforward, and a relatively good understanding of the underlying theory is essential.
 Choice of model: This will depend on the data representation and the application. Overly
complex models tend to lead to problems with learning.
 Learning algorithm: There are numerous trade-offs between learning algorithms. Almost
any algorithm will work well with the correct hyperparameters for training on a
particular fixed data set. However, selecting and tuning an algorithm for training on
unseen data requires a significant amount of experimentation.
 Robustness: If the model, cost function and learning algorithm are selected appropriately
the resulting ANN can be extremely robust.
With the correct implementation, ANNs can be used naturally in online learning and large data
set applications. Their simple implementation and the existence of mostly local dependencies
exhibited in the structure allows for fast, parallel implementations in hardware.
Applications
The utility of artificial neural network models lies in the fact that they can be used to infer a
function from observations. This is particularly useful in applications where the complexity of
the data or task makes the design of such a function by hand impractical.
Real-life applications
The tasks artificial neural networks are applied to tend to fall within the following broad
categories:
 Function approximation, or regression analysis, including time series prediction, fitness
approximation and modeling.

Rohan verma Page 47

  Classification, including pattern and sequence recognition, novelty detection and
sequential decision making.
 Data processing, including filtering, clustering, blind source separation and compression.
 Robotics, including directing manipulators, prosthesis.
 Control, including Computer numerical control.
Application areas include the system identification and control (vehicle control, process control,
natural resources management), quantum chemistry, game-playing and decision making
(backgammon, chess, poker), pattern recognition (radar systems, face identification, object
recognition and more), sequence recognition (gesture, speech, handwritten text recognition),
medical diagnosis, financial applications (e.g. automated trading systems), data mining (or
knowledge discovery in databases, "KDD"), visualization and e-mail spam filtering.
Artificial neural networks have also been used to diagnose several cancers. An ANN based
hybrid lung cancer detection system named HLND improves the accuracy of diagnosis and the
speed of lung cancer radiology. These networks have also been used to diagnose prostate cancer.
The diagnoses can be used to make specific models taken from a large group of patients
compared to information of one given patient. The models do not depend on assumptions about
correlations of different variables. Colorectal cancer has also been predicted using the neural
networks. Neural networks could predict the outcome for a patient with colorectal cancer with
more accuracy than the current clinical methods. After training, the networks could predict
multiple patient outcomes from unrelated institutions.
Neural networks and neuroscience
Theoretical and computational neuroscience is the field concerned with the theoretical analysis
and the computational modeling of biological neural systems. Since neural systems are
intimately related to cognitive processes and behavior, the field is closely related to cognitive
and behavioral modeling.
The aim of the field is to create models of biological neural systems in order to understand how
biological systems work. To gain this understanding, neuroscientists strive to make a link
between observed biological processes (data), biologically plausible mechanisms for neural
processing and learning (biological neural network models) and theory (statistical learning
theory and information theory).
Types of models
Many models are used in the field, defined at different levels of abstraction and modeling
different aspects of neural systems. They range from models of the short-term behavior of
individual neurons, models of how the dynamics of neural circuitry arise from interactions
between individual neurons and finally to models of how behavior can arise from abstract neural
modules that represent complete subsystems. These include models of the long-term, and short-
term plasticity, of neural systems and their relations to learning and memory from the individual
neuron to the system level.
Neural network software
Main article: Neural network software

Rohan verma Page 48

 Neural network software is used to simulate, research, develop and apply artificial neural
networks, biological neural networks and, in some cases, a wider array of adaptive systems.
Types of artificial neural networks
Main article: Types of artificial neural networks
Artificial neural network types vary from those with only one or two layers of single direction
logic, to complicated multi–input many directional feedback loops and layers. On the whole,
these systems use algorithms in their programming to determine control and organization of their
functions. Most systems use "weights" to change the parameters of the throughput and the
varying connections to the neurons. Artificial neural networks can be autonomous and learn by
input from outside "teachers" or even self-teaching from written-in rules.
Theoretical properties
Computational power
The multi-layer perceptron (MLP) is a universal function approximator, as proven by the
universal approximation theorem. However, the proof is not constructive regarding the number
of neurons required or the settings of the weights.
Work by Hava Siegelmann and Eduardo D. Sontag has provided a proof that a specific recurrent
architecture with rational valued weights (as opposed to full precision real number-valued
weights) has the full power of a Universal Turing Machine using a finite number of neurons and
standard linear connections. They have further shown that the use of irrational values for weights
results in a machine with super-Turing power.
Capacity
Artificial neural network models have a property called 'capacity', which roughly corresponds to
their ability to model any given function. It is related to the amount of information that can be
stored in the network and to the notion of complexity.
Convergence
Nothing can be said in general about convergence since it depends on a number of factors.
Firstly, there may exist many local minima. This depends on the cost function and the model.
Secondly, the optimization method used might not be guaranteed to converge when far away
from a local minimum. Thirdly, for a very large amount of data or parameters, some methods
become impractical. In general, it has been found that theoretical guarantees regarding
convergence are an unreliable guide to practical application.
Generalization and statistics
In applications where the goal is to create a system that generalizes well in unseen examples, the
problem of over-training has emerged. This arises in convoluted or over-specified systems when
the capacity of the network significantly exceeds the needed free parameters. There are two
schools of thought for avoiding this problem: The first is to use cross-validation and similar
techniques to check for the presence of overtraining and optimally select hyperparameters such
as to minimize the generalization error. The second is to use some form of regularization. This is
a concept that emerges naturally in a probabilistic (Bayesian) framework, where the
regularization can be performed by selecting a larger prior probability over simpler models; but

Rohan verma Page 49

 also in statistical learning theory, where the goal is to minimize over two quantities: the
'empirical risk' and the 'structural risk', which roughly corresponds to the error over the training
set and the predicted error in unseen data due to over-fitting.

Confidence analysis of a neural network

Supervised neural networks that use an MSE cost function can use formal statistical methods to
determine the confidence of the trained model. The MSE on a validation set can be used as an
estimate for variance. This value can then be used to calculate the confidence interval of the
output of the network, assuming a normal distribution. A confidence analysis made this way is
statistically valid as long as the output probability distribution stays the same and the network is
not modified.
By assigning a softmax activation function, a generalization of the logistic function, on the
output layer of the neural network (or a softmax component in a component-based neural
network) for categorical target variables, the outputs can be interpreted as posterior probabilities.
This is very useful in classification as it gives a certainty measure on classifications.
The softmax activation function is:

Controversies
Training issues
A common criticism of neural networks, particularly in robotics, is that they require a large
diversity of training for real-world operation. This is not surprising, since any learning machine
needs sufficient representative examples in order to capture the underlying structure that allows
it to generalize to new cases. Dean Pomerleau, in his research presented in the paper
"Knowledge-based Training of Artificial Neural Networks for Autonomous Robot Driving," uses
a neural network to train a robotic vehicle to drive on multiple types of roads (single lane, multi-
lane, dirt, etc.). A large amount of his research is devoted to (1) extrapolating multiple training
scenarios from a single training experience, and (2) preserving past training diversity so that the
system does not become over-trained (if, for example, it is presented with a series of right turns –
it should not learn to always turn right). These issues are common in neural networks that must
decide from amongst a wide variety of responses, but can be dealt with in several ways, for
example by randomly shuffling the training examples, by using a numerical optimization

Rohan verma Page 50

 algorithm that does not take too large steps when changing the network connections following an
example, or by grouping examples in so-called mini-batches.
Hardware issues
To implement large and effective software neural networks, considerable processing and storage
resources need to be committed. While the brain has hardware tailored to the task of processing
signals through a graph of neurons, simulating even a most simplified form on Von Neumann
technology may compel a neural network designer to fill many millions of database rows for its
connections – which can consume vast amounts of computer memory and hard disk space.
Furthermore, the designer of neural network systems will often need to simulate the transmission
of signals through many of these connections and their associated neurons – which must often be
matched with incredible amounts of CPU processing power and time. While neural networks
often yield effective programs, they too often do so at the cost of efficiency (they tend to consume
considerable amounts of time and money).
Computing power continues to grow roughly according to Moore's Law, which may provide
sufficient resources to accomplish new tasks. Neuromorphic engineering addresses the hardware
difficulty directly, by constructing non-Von-Neumann chips with circuits designed to implement
neural nets from the ground up.
Practical counterexamples to criticisms
Arguments against Dewdney's position are that neural nets have been successfully used to solve
many complex and diverse tasks, ranging from autonomously flying aircraft to detecting credit
card fraud .
Technology writer Roger Bridgman commented on Dewdney's statements about neural nets:
Neural networks, for instance, are in the dock not only because they have been hyped to high
heaven, (what hasn't?) but also because you could create a successful net without understanding
how it worked: the bunch of numbers that captures its behaviour would in all probability be "an
opaque, unreadable table...valueless as a scientific resource".
In spite of his emphatic declaration that science is not technology, Dewdney seems here to
pillory neural nets as bad science when most of those devising them are just trying to be good
engineers. An unreadable table that a useful machine could read would still be well worth
having.
Although it is true that analyzing what has been learned by an artificial neural network is
difficult, it is much easier to do so than to analyze what has been learned by a biological neural
network. Furthermore, researchers involved in exploring learning algorithms for neural networks
are gradually uncovering generic principles which allow a learning machine to be successful. For
example, Bengio and LeCun (2007) wrote an article regarding local vs non-local learning, as
well as shallow vs deep architecture.
Hybrid approaches
Some other criticisms came from believers of hybrid models (combining neural networks and
symbolic approaches). They advocate intermix of these two approaches and believe that hybrid
models can better capture the mechanisms of the human mind.

Rohan verma Page 51

 Gallery

A single-layer feedforward artificial neural network. Arrows originating from are

omitted for clarity. There are p inputs to this network and q outputs. In this system, the
value of the qth output, would be calculated as

A two-layer feedforward artificial neural network.

Bee Algorithm (BA)

The Bees Algorithm [9] or BA is a bio-inspired metaheuristic behavior of honey bees and how
they searching for plant to obtain the necessary pollen for honey production.
A colony of bees search in a large territory looking for new sources of food and begins to thrive
on discovering new food sources. When these sources have much pollen are visited by large
numbers of bees and when the pollen decreases the number of bees collected from these sources
decreases too.
When season for recollecting pollen start, the colony sent so many bees, which are called scouts
bees to reconnoiter randomly the territory to inform at the colony where are the best food
sources. Once the harvesting season starts the colony maintains a certain percentage of their
scout bees in order to detect new and better sources of food. When scout bees have returned to
the colony and found a better food source than the currently is using the colony, makes the
Dance by means of which transmits the exact position of the source food and then the colony
began to send more bees to the food source.
An application of the Bee Swarm Optimization BSO to the Knapsack Problem is given below.
Require: population size (ps), neighborhood (n).
Initialize population with random solutions.
Evaluate fitness of the population.
While stopping criterion not met do

Rohan verma Page 52

 select the sites for neighborhood search
recruit bees for n selected sites and evaluate their fitness
select the fittest bee from each site
assign remaining bees to search randomly and evaluate their fitness
end while
Particle Swarm Optimization (PSO)
The Particle Swarm Optimization [4][7] or PSO is a Bio-inspired metaheuristic in flocks of birds
or schools of fish. It was developed by J. Kennedy and R. Eberthart based on a concept called
social metaphor, this metaheuristics simulates a society where all individuals contribute their
knowledge to obtain a better solution, there are three factors that influence for change in status or
behavior of an individual:
• The Knowledge of the environment or adaptation which speaks of the importance given to the
experience of the individual.
• His Experience or local memory is the importance given to the best result found by the
individual.
• The Experience of their neighbors or Global memory referred to how important it is the best
result I have obtained their neighbors or other individuals.
In this metaheuristic each individual is called particle and moves through a multidimensional
space that represents the social space or search space depends on the dimension of space which
depends on the variables used to represent the problem.
For the update of each particle using something called velocity vector which tells them how fast
it will move the particle in each of the dimensions, the method for updating the speed of PSO is
given by equation (10), and it is updating by the equation (11).
where:
• is the velocity of the i-th particle
• is adjustment factor to the environment.
• is the memory coefficient in the neighborhood.
• is the coefficient memory.
• is the position of the i-th particle.
• is the best position found so far by all particles.
• is the best position found by the i-th particle
The Algorithm of the PSO applied to the Knapsack Problem is given below.
Require: adaptation to environment coefficient, local memory coefficient, neighborhood memory
coefficient, swarm size
Start the swarm particles.
Start the velocity vector for each particle in the swarm.
While stopping criterion not met do
for i=1 to n do
if the i-particle's fitness is better than the local best then replace the local best with the iparticle

Rohan verma Page 53

 if the i-particle's fitness is better than the global best then replace the global best with the
iparticle.
Update the velocity vector
Update the particle's position with the velocity vector
end for
end while
Bee Swarm Optimization (BSO)
The Bee Swarm Optimization or BSO is a hybrid metaheuristic population between the PSO and
the BA. The main idea of BSO is based on taking the best of each metaheuristics to obtain better
results than they would obtain.
The BSO use the velocity vector and the way to updating it, equation (10), and applies the social
metaphor to get better results from the PSO and use the exploration and a search radius from the
BA to indicate which is where they look for a better result.
The first thing that the BSO do is update the position of the particles through the velocity vector
and then select a specified number of particles, in this case it is proposed to select the best as
being the new food supply that the scout bees discovered,
and conducted a search of the area enclosed by the radius search and if there are a better solution
in this area than the same particle around which are looking for then the particle is replaced with
the position of the best solution found in the area.
For this metaheuristic to work properly you need a way to determine what the next and earlier
particle position, if we cannot determine this then it is impossible to apply this meta-heuristic
because you would not know what the next and previous particle to search in that area.
We defined the next and before elements like binary operations, adding and subtracting one
number to the solution vector. For example if we have a Knapsack Problem with 5 elements the
solution vector will have 5 spaces and in each space can be 0 or 1, we can see the example of the
next and before solution vector in the Figure 1.

Fig. 1. Example of the previous and next solution vector

The algorithm of the BSO applied to the Knapsack Problem implemented in the present paper is
the following.
Require: adaptation to environment coefficient, local memory coefficient, neighborhood memory
coefficient, swarm size, scout bees, search radio
Start the swarm particles.
Start the velocity vector for each.
While stopping criterion not met do
for i=1 to n do
if the i-particle's fitness is better than the local best then replace the local best with the iparticle
if the i-particle's fitness is better than the global best then replace the global best with the
iparticle.

Rohan verma Page 54

 Update the velocity vector
Update the particle's position with the velocity vector
Choose the best particles for all best particle
Search if there are some better particle in the search radio and if exist it replace the particle
with the best particle in the search radio
end for all
end for
end while

GENETIC ALGORITHM

BASIC CONCEPT:

GA encodes the decision variables of a search problem into finite-length strings of

alphabets of certain cardinality. The strings which are candidate solutions to the search problem
are referred to as chromosomes, the alphabets are referred to as genes and the values of genes are
called alleles. For example, in a problem such as the traveling salesman problem, a chromosome
represents a route, and a gene may represent a city. In contrast to traditional optimization
techniques, GAs work with coding of parameters, rather than the parameters themselves. To
evolve good solutions and to implement natural selection, we need a measure for distinguishing
good solutions from bad solutions. The measure could be an objective function that is a
mathematical model or a computer simulation, or it can be a subjective function where humans
choose better solutions over worse ones. In essence, the fitness measure must determine a
candidate solution‟s relative fitness, which will subsequently be used by the GA to guide the
evolution of good solutions. Another important concept of GAs is the notion of population.
Unlike traditional search methods, genetic algorithms rely on a population of candidate solutions.

Rohan verma Page 55

 The population size, which is usually a user-specified parameter, is one of the important factors
affecting the scalability and performance of genetic algorithms. For example, small population
sizes might lead to premature convergence and yield substandard solutions. On the other hand,
large population sizes lead to unnecessary expenditure of valuable computational time. Once the
problem is encoded in a chromosomal manner and a fitness measure for discriminating good
solutions from bad ones has been chosen, we can start to evolve solutions to the search problem
using the following steps:
1. Initialization: The initial population of candidate solutions is usually generated randomly
across the search space. However, domain-specific knowledge or other information can be easily
incorporated.
2. Evaluation: Once the population is initialized or an offspring population is created, the fitness
values of the candidate solutions are evaluated.
3. Selection: Selection allocates more copies of those solutions with higher fitness values and
thus imposes the survival-of-the-fittest mechanism on the candidate solutions. The main idea of
selection is to prefer better solutions to worse ones, and many selection procedures have been
proposed to accomplish this idea, including roulette-wheel selection, stochastic universal
selection, ranking selection and tournament selection, some of which are described in the next
section.
4. Recombination: Recombination combines parts of two or more parental solutions to create
new, possibly better solutions (i.e. offspring). There are many ways of accomplishing this (some
of which are discussed in the next section), and competent performance depends on a properly
designed recombination mechanism. The offspring under recombination will not be identical to
any particular parent and will instead combine parental traits in a novel manner
5. Mutation: While recombination operates on two or more parental chromosomes, mutation
locally but randomly modifies a solution. Again, there are many variations of mutation, but it
usually involves one or more changes being made to an individual‟s trait or traits. In other words,
mutation performs a random walk in the vicinity of a candidate solution.
6. Replacement. The offspring population created by selection, recombination, and mutation
replaces the original parental population. Many replacement techniques such as elitist
replacement, generation-wise replacement and steady-state replacement methods are used in
GAs.
7. Repeat steps 2–6 until a terminating condition is met. Goldberg has likened GAs to
mechanistic versions of certain modes of human innovation and has shown that these operators
when analyzed individually are ineffective, but when combined together they can work well.

ENCODING:
As for any search and learning method, the way in which candidate solutions are encoded is a
central, if not the central, factor in the success of a genetic algorithm. Most GA applications use
fixed−length, fixed−order bit strings to encode candidate solutions. However, in recent years,
there have been many experiments with other kinds of encodings,Common approaches used are:
Binary Encoding: Every chromosome is a string of 0 or 1.Suppose we have a knapsack
of capacity C and N items, then we can encode this problem as follows Chromosome, in this case
is a string of 0s and 1s with N bits Represent item i of problem with bit in the chromosome
bit is 1 iff item has been selected, 0 otherwise. The set of all such chromosomes ( ) is
the solution space of the problem.

Rohan verma Page 56

 Chromosome 1: 1 0 1 1 0 0 1 0 1 1 0 0 1 0 1 0 1 1 1 0 0 1 0 1
Chromosome 2: 1 1 1 1 1 1 1 0 0 0 0 0 1 1 0 0 0 0 0 1 1 1 1 1

The example shown above has 24 items (and therefore 24 bits) with item1 selected in both
chromosome 1 and 2 whereas item2 is selected in chromosome 2 but not in chromosome 1.
Permutation Encoding (Travelling Salesman Problem): Every chromosome is a string
of numbers that represent position in a sequence.Problem description: There are cities and given
distances between them. Travelling salesman has to visit all of them, but he doesn't want to
travel more than necessary. Find a sequence of cities with a minimal travelled distance.

Chromosome A: 1 5 3 2 6 4 7 9 8
Chromosome B: 8 5 6 7 2 3 1 4 9

Encoding: Here, encoded chromosomes describe the order of cities the salesman visits. For
example, in chromosome A, the salesman visits city-1 followed by city-5 followed by city-3 and
so on.
Tree Encoding : (Genetic Programming) In tree encoding, every chromosome is a tree of
some objects, such as functions or commands in programming language. Tree encoding is useful
for evolving programs or any other structures that can be encoded in trees.

Value Encoding: Every chromosome is a sequence of some values (real numbers, characters
or objects).Direct value encoding can be used in problems, where some complicated values, such
as real numbers, are used. Use of binary encoding for this type of problems would be very
difficult. In value encoding, every chromosome is a string of some values. Values can be
anything connected to problem, form numbers, real numbers or chars to some complicated
objects.
Chromosome A 1.2324 5.3243 0.4556 2.3293 2.4545
Chromosome B ABDJEIFJDHDIERJFDLDFLFEGT
Chromosome C (back), (back), (right), (forward), (left)

Example of Problem: Finding weights for neural network

The problem: There is some neural network with given architecture. Find weights for inputs
of neurons to train the network for wanted output.
Encoding: Real values in chromosomes represent corresponding weights for inputs.

Rohan verma Page 57

 FITNESS FUNCTION:
A fitness function is a particular type of objective function that is used to summaries, as a
single figure of merit, how close a given design solution is to achieving the set aims. In
particular, in the fields of genetic programming and genetic algorithms, each design solution is
represented as a string of numbers (referred to as a chromosome). After each round of testing, or
simulation, the idea is to delete the 'n' worst design solutions, and to breed 'n' new ones from the
best design solutions. Each design solution, therefore, needs to be awarded a figure of merit, to
indicate how close it came to meeting the overall specification, and this is generated by applying
the fitness function to the test, or simulation, results obtained from that solution. The reason that
genetic algorithms cannot be considered to be a lazy way of performing design work is precisely
because of the effort involved in designing a workable fitness function. Even though it is no
longer the human designer, but the computer, that comes up with the final design, it is the human
designer who has to design the fitness function. If this is designed badly, the algorithm will either
converge on an inappropriate solution, or will have difficulty converging at all. Moreover, the
fitness function must not only correlate closely with the designer's goal, it must also be computed
quickly. Speed of execution is very important, as a typical genetic algorithm must be iterated
many times in order to produce a usable result for a non-trivial problem. Fitness approximation
may be appropriate, especially in the following cases:

 Fitness computation time of a single solution is extremely high

 Precise model for fitness computation is missing
 The fitness function is uncertain or noisy.

Two main classes of fitness functions exist: one where the fitness function does not change, as in
optimizing a fixed function or testing with a fixed set of test cases; and one where the fitness
function is mutable, as in niche differentiation or co-evolving the set of test cases. Another way
of looking at fitness functions is in terms of a fitness landscape, which shows the fitness for each
possible chromosome. Definition of the fitness function is not straightforward in many cases and
often is performed iteratively if the fittest solutions produced by GA are not what is desired. In
some cases, it is very hard or impossible to come up even with a guess of what fitness function
definition might be. Interactive genetic algorithms address this difficulty by outsourcing
evaluation to external agents (normally humans).GAs are naturally suitable for solving
maximization problems. Maximization problems are usually transformed into maximization
problem by suitable transformation. In general, a fitness function F(i) is first derived from the
objective function and used in successive genetic operations. Fitness in biological sense is a
quality value which is a measure of the reproductive efficiency of chromosomes. In genetic
algorithm, fitness is used to allocate reproductive traits to the individuals in the population and
thus act as some measure of goodness to be maximized. This means that individuals with higher
fitness value will have higher probability of being selected as candidates for further examination.
Certain genetic operators require that the fitness function be non-negative, although certain
operators need not have this requirement. For maximization problems, the fitness function can be
considered to be the same as the objective function or F(i)=O(i). For minimization problems, to
generate non-negative values in all the cases and to reflect the relative fitness of individual
string, it is necessary to map the underlying natural objective function to fitness function form. A
number of such transformations is possible. Two commonly adopted fitness mappings are
presented below.

Rohan verma Page 58

 F(x)

This transformation does not alter the location of the minimum, but converts a minimization
problem to an equivalent maximization problem. An alternate function to transform the objective
function to get the fitness value as below.

where, O(i) is the objective function value of i th individual, P is the population size and V is a
large value to ensure non-negative fitness values. The value of V adopted in this work is the
maximum value of the second term of equation so that the fitness value corresponding to
maximum value of the objective function is zero. This transformation also does not alter the
location of the solution, but converts a minimization problem to an equivalent maximization
problem. The fitness function value of a string is known as the string fitness.

SELECTION:
Chromosomes are selected from the population to be parents to crossover. The problem is how to
select these chromosomes. According to Darwin's evolution theory the best ones should survive
and create new offspring. In principle, a population of individuals selected from the search space
, often in a random manner, serves as candidate solutions to optimize the problem.The
individuals in this population are evaluated through ( "fitness" ) adaptation function. A selection
mechanism is then used to select individuals to be used as parents to those of the next generation.
These individuals will then be crossed and mutated to form the new offspring. The next
generation is finally formed by an alternative mechanism between parents and their offspring [4].
This process is repeated until a certain satisfaction condition. There are many methods how to
select the best chromosomes, for example roulette wheel selection, Boltzman selection,
tournament selection, rank selection, steady state selection and some others.
Roulette Wheel Selection :
Parents are selected according to their fitness. The better the chromosomes are, the more chances
to be selected they have. Imagine a roulette wheel where are placed all chromosomes in the
population, every has its place big accordingly to its fitness function, like on the following
picture.

Rohan verma Page 59

 Then a marble is thrown there and selects the chromosome. Chromosome with bigger fitness will
be selected more times.The conspicuous characteristic of this selection method is the fact that it
gives to each individual i of the current population a probability of p(i) being selected,
proportional to its fitness f(i).

Where n denotes the population size in terms of the number of individuals. A well-known
drawback of this technique is the risk of premature convergence of the GA to a local optimum,
due to the possible presence of a dominant individual that always wins the competition and is
selected as a parent.
Linear Rank Selection (LRS):
LRS is also a variant of RWS that tries to overcome the drawback of premature convergence of
the GA to a local optimum. It is based on the rank of individuals rather than on their fitness. The
rank n is accorded to the best individual whilst the worst individual gets the rank 1. Thus, based
on its rank, each individual i has the probability of being selected given by the expression

Exponential Rank Selection (ERS) :

The ERS is based on the same principle as LRS, but it differs from LRS by the probability of
selecting each individual. For ERS, this probability is given by the expression:

Tournament Selection (TOS)

Tournament selection is a variant of rank-based selection methods. Its principle consists in
randomly selecting a set of k individuals. These individuals are then ranked according to their
relative fitness and the fittest individual is selected for reproduction. The whole process is

Rohan verma Page 60

 repeated n times for the entire population. Hence, the probability of each individual to be
selected is given by the expression:

Steady-State Selection:

This is not particular method of selecting parents. Main idea of this selection is that big part of
chromosomes should survive to next generation. GA then works in a following way. In every
generation are selected a few (good - with high fitness) chromosomes for creating a new
offspring. Then some (bad - with low fitness) chromosomes are removed and the new offspring
is placed in their place. The rest of population survives to new generation.

Elitism

Idea of elitism has been already introduced. When creating new population by crossover and
mutation, we have a big chance, that we will loose the best chromosome. Elitism is name of
method, which first copies the best chromosome (or a few best chromosomes) to new population.
The rest is done in classical way. Elitism can very rapidly increase performance of GA, because
it prevents losing the best found solution.

REPRODUCTION:
After selection, individuals from the mating pool are recombined (or crossed over) to
create new, hopefully better, offspring. In the GA literature, many crossover methods have been
designed and some of them are described in this section. In most recombination operators, two
individuals are randomly selected and are recombined with a probability pc, called the crossover
probability. That is, a uniform random number, r, is generated and if r ≤ pc, the two randomly
selected individuals undergo recombination. Otherwise, that is, if r > pc, the two offspring are
simply copies of their parents. The value of pc can either be set experimentally, or can be set
based on schema-theorem principles.
k-point Crossover One-point, and two-point crossovers are the simplest and most widely
applied crossover methods. In one-point crossover, illustrated in Figure a crossover site is
selected at random over the string length, and the alleles on one side of the site are exchanged
between the individuals. In two-point crossover, two crossover sites are randomly selected. The
alleles between the two sites are exchanged between the two randomly paired individuals. Two-
point crossover is also illustrated in Figure The concept of one-point crossover can be extended
to k-point crossover, where k crossover points are used, rather than just one or two.

Rohan verma Page 61

 Uniform Crossover Another common recombination operator is uniform crossover. In uniform
crossover, illustrated in Figure every allele is exchanged between the a pair of randomly selected
chromosomes with a certain probability, pe, known as the swapping probability. Usually the
swapping probability value is taken to be 0.5.
Uniform Order-Based Crossover : The k-point and uniform crossover methods described
above are not well suited for search problems with permutation codes such as the ones used in
the traveling salesman problem. They often create offspring that represent invalid solutions for
the search problem. Therefore, when solving search problems with permutation codes, a
problem-specific repair mechanism is often required (and used) in conjunction with the above
recombination methods to always create valid candidate solutions. Another alternative is to use
recombination methods developed specifically for permutation codes, which always generate
valid candidate solutions. Several such crossover techniques are described in the following
paragraphs starting with the uniform order-based crossover. In uniform order-based crossover,
two parents (say P1 and P2) are randomly selected and a random binary template is generated
.Some of the genes for offspring C1 are filled by taking the genes from parent P1 where there is a
one in the template. At this point we have C1 partially filled, but it has some “gaps”. The genes
of parent P1 in the positions corresponding to zeros in the template are taken and sorted in the

Rohan verma Page 62

 same order as they appear in parent P2. The sorted list is used to fill the gaps in C1. Offspring C2
is created by using a similar process

Order-Based Crossover The order-based crossover operator (Davis, 1985) is a variation of the
uniform order-based crossover in which two parents are randomly selected and two random
crossover sites are generated. The genes between the cut points are copied to the children.
Starting from the second crossover site copy the genes that are not already present in the
offspring from the alternative parent (the parent other than the one whose genes are copied by the
offspring in the initial phase) in the order they appear. For example, as shown in Figure, for
offspring C1, since alleles C, D, and E are copied from the parent P1, we get alleles B, G, F, and
A from the parent P2. Starting from the second crossover site, which is the sixth gene, we copy
alleles B and G as the sixth and seventh genes respectively. We then wrap around and copy
alleles F and A as the first and second genes.

Rohan verma Page 63

 Partially Matched Crossover (PMX): Apart from always generating valid offspring, the PMX
operator also preserves orderings within the chromosome. In PMX, two parents are randomly
selected and two random crossover sites are generated. Alleles within the two crossover sites of
a parent are exchanged with the alleles corresponding to those mapped by the other parent. For
example, as illustrated in Figure, looking at parent P1, the first gene within the two crossover
sites, 5, maps to 2 in P2. Therefore, genes 5 and 2 are swapped in P1. Similarly we swap 6 and 3,
and 10 and 7 to create the offspring C1. After all exchanges it can be seen that we have achieved
a duplication of the ordering of one of the genes in between the crossover point within the
opposite chromosome, and vice versa.

Rohan verma Page 64

 Cycle Crossover (CX): We describe cycle crossover with help of a simple illustration
(reproduced from Goldberg (1989b) with permission). Consider two randomly selected parents
P1 and P2 as shown in Figure that are solutions to a traveling salesman problem. The offspring
C1 receives the first variable representing city 9) from P1. We then choose the variable that maps
onto the same position in P2. Since city 9 is chosen from P1 which maps to city 1 in P2, we
choose city 1 and place it into C1 in the same position as it appears in P1 (fourth gene), as shown
in Figure 4.5. City 1 in P1 now maps to city 4 in P2, so we place city 4 in C1 in the same
position it occupies in P1 (sixth gene). We continue this process once more and copy city 6 to the
ninth gene of C1 from P1. At this point, since city 6 in P1 maps to city 9 in P2, we should take
city 9 and place it in C1, but this has already been done, so we have completed a cycle; which is
where this operator gets its name. The missing cities in offspring C1 is filled from P2. Offspring
C2 is created in the same way by starting with the first city of parent P2

Rohan verma Page 65

Rohan verma Page 66
 DIFFERENCES BETWEEN GA'S AND TRADITIONAL METHODS:

The brief list, based on Goldberg (1989), of the essential differences between GAs and other
forms of optimization is the following:.

 Genetic algorithms a coded form of the function values (parameter set), rather than with
the actual values them. So, for example, if we want to find the minimum of the function
f(x) =x3+x2+5, the GA would not deal directly with x or y values, but with strings that
encode these values. For this case, strings representing the binary x values should be
used.
 Genetic algorithms use a set, or population, of points to conduct a search, not just a single
point on the problem space. This gives GAs the power to search noisy spaces littered with
local optimum points. Instead of relying on a single point to search through the space, the
GAs looks at many different areas of the problem space at once, and uses all of this
information to guide it.
 Genetic algorithms use only payoff information to guide themselves through the problem
space. Many search techniques need a variety of information to guide themselves. Hill
climbing methods require derivatives, for example. The only information a GA needs is
some measure of fitness about a point in the space (sometimes known as an objective
function value). Once the GA knows the current measure of "goodness" about a point, it
can use this to continue searching for the optimum.
 GAs are probabilistic in nature, not deterministic. This is a direct result of the
randomization techniques used by GAs.
 GAs are inherently parallel. Here lies one of the most powerful features of genetic
algorithms. GAs, by their nature, are very parallel, dealing with a large number of points
(strings) simultaneously. Holland has estimated that a GA processing n strings at each
generation, the GA in reality processes n3 useful substrings.

* GA's work with string coding of variables instead of variables so that coding
discrediting the search space even though the function is continuous.

* GA's work with population of points instead of single point.

* In GA's previously found good information is emphasized using reproduction operator

and propagated adaptively through crossover and mutation operators.

* GA does not require any auxiliary information except the objective function values.
* GA uses the probabilities in their operators.
This nature of narrowing the search spaces the search progresses is adaptive and is the
unique characteristic of Genetic Algorithms.

SOME APPLICATIONS OF GENETIC ALGORITHMS:

Rohan verma Page 67

 The algorithm described above is very simple, but variations on this basic theme have been
used in a large number of scientific and engineering problems and models, including the
following:

 Optimization: GAs have been used in a wide variety of optimization tasks, including
numerical optimization as well as combinatorial optimization problems such as circuit
layout and job-shop scheduling.
 Automatic Programming: GAs have been used to evolve computer programs for specific
tasks, and to design other computational structures, such as cellular automata and sorting
networks.
 Machine learning: GAs have been used for many machine-learning applications,
including classification and prediction tasks such as the prediction of weather or protein
structure. GAs have also been used to evolve aspects of particular machine-learning
systems, such as weights for neural networks, rules for learning classifier systems or
symbolic production systems, and sensors for robots.
 Economic models: GAs have been used to model processes of innovation, the
development of bidding strategies, and the emergence of economic markets.
 Immune system models: GAs have been used to model various aspects of the natural
immune system including somatic mutation during an individual's lifetime and the
discovery of multi-gene families during evolutionary time.
 Ecological models: GAs have been used to model ecological phenomena such as
biological arms races, host-parasite co-evolution, symbiosis, and resource flow in
ecologies.
Genetic Algorithm Application Areas:

 Dynamic process control

 Induction of rule optimization
 Discovering new connectivity topologies
 Simulating biological models of behavior and evolution
 Complex design of engineering structures
 Pattern recognition
 Scheduling
 Transportation
 Layout and circuit design
 Telecommunication
 Graph-based problems

Rohan verma Page 68