AISC - Techmax (Searchable) With Index - Compressed

Scanned by CamScanner
Introduction to Artificial
Intelligence(Al) and Soft Computing
14.1. Introduction and Definition of Artificial Intelligence.

Nature of Environment, Structure of Agent, types of
1.2 Intelligent Agents : Agents and Environments Rationality,
Agent
soft computing vs. hard computing, various types of soft
1.3 Soft Computing: Introduction of soft computing,
computing techniques.
1.1. Introduction to Artificial Intelligence

in 1956, has defined Al as “the science and
John McCarthy who has coined the word “Artificial Intelligence”
—
r programs.
engineering of making intelligent machines”, especially intelligent compute
the machine needs to take some decision or
— Artificial Intelligence (Al) is relevant to any intellectual task where
act intelligently or rationally. As it has a very
choose the next action based on the current state of the system, in short
wide range of applications, it is truly a universal field.
a Human Brain, where a machine or software shows
— In simple words, Artificial Intelligent System works like
systems or expert systems. You can say
intelligence while performing given tasks; such systems are called intelligent
that these systems can “think” while generating output!!!
variety of application fields. Al applications
— Alis one of the newest fields in science and engineering and has a wide
field, such as writing stories,
range from the general fields like learning, perception and prediction to the specific
, and playing chess.
proving mathematical theorems, driving a bus on a crowded street, diagnosing diseases
Following are the four
— Alis the study of how to make machines do thing which at the moment people do better.
approaches to define Al.
1.2 Foundations and Mathematical Treatments

moment human do
— In general, artificial intelligence is the study of how to make machines do things which at the
better. Following are the four approaches to define Al.
- Historically, all four approaches have been followed by different group of people with different methods.
1.2.1. Acting Humanly : The Turing Test Approach

LOSE As)
- Definition 1: “The art of creating machines that perform functions that requires intelligence when performed by
people.” (Kurzweil, 1990)
Definition 2 : “The study of how to make computers do things at which, at the moment, people are better.” (Rich and
Knight, 1991)
Al&SC (MU-Sem. 7- Comp) Intro. to Artificial Intelligence Al) and Soft Computing
To judge whether the system can act like a human, Sir Alan Turing had designed a test known as Turing test.
As shown in Fig. 1.2.1, in Turing test, a computer needs to interact with a human interrogator by answering his
questions in written format. Computer passes the test if a human interrogator, cannot identify whether the written
responses are from a person or a computer. Turing test is valid even after 60 year of research.
Human
Human
interrogator
Fig.1.2.1 : Turing Test Environment
For this test, the computer would need to possess the following capabilities:
“1: (NLP) : This unit enables computer to interpret the English language and
Natural Language Processing
communicate successfully.
2. Knowledge Representation : This unit is used to store knowledge gathered by the system through input devices.
3. Automated Reasoning: This unit enables to analyze the knowledge stored in the system and makes new
inferences to answer questions.
4, Machine Learning: This unit learns new knowledge by taking current input from the environment and adapts to
new circumstances, thereby enhancing the knowledgebase of the system.
To pass total Turing test, the computer will also need to have computer vision, which is required to perceive objects
from the environment and Robotics, to manipulate those objects.
Capabilities a Computer needs to Possess
_ Natural _ Knowledge Automated. Q ‘Machine —

_ Language | Representation} | Reasoning ~ Learning’.
Processing | | fe 2
Fig. 1.2.2 : Capabilities a Computer needs to possess
Fig. 1.2.2 lists all the capabilities a computer needs to have in order to exhibit artificial intelligence. Mentioned above
are the six disciplines which implement most of the artificial intelligence.
1.2.2 Thinking Humanly : The Cognitive Modelling Approach
— Definition1 : “The exciting new effort to make computers think ... machines with minds, in the full and literal sense”.
(Haugeland, 1985)
Definition 2 : “The automation of activities that we associate with human thinking, activities such as decision making,
problem solving, learning ...” (Hellman, 1978)
— Cognitive science :It is interdisciplinary field which combines computer models from Artificial Intelligence with the
‘techniques from psychology in order to construct precise and testable theories for working of human mind.
: .
; WF Al&SC (MU-Sem. 7-Comp) 1-3 Intro. to Artificial Intelligence(Al) and Soft Computing
— In order to make machines think like human, we need to first understand how human think. Research showed that
there are three ways using which human’s thinking pattern can be caught.
1. Introspection through which human can catch their own thoughts as they go by.
2. Psychological experiments can be carried out by observing a person in action.
3. Brain imaging can be done by observing the brain in action.
- By catching the human thinking pattern, it can be implemented in computer system asa program and if the program's
input output matches with that of human, then it can be claimed that the system can operate like humans.
1.2.3 Thinking Rationally : The “Laws of Thought” Approach
— Definition1 : “The study of mental faculties through the use of computational models”. (Charniak and McDermott,
1985)
— Definition2 ; “The study of the computations that make it possible to perceive, reason, and act”.
their study initiated the field called
— The laws of thought are supposed to implement the operation of the mind and
logic. It provides precise notations to express facts of the real world.
Also computer programs based on
It also includes reasoning and “right thinking” that is irrefutable thinking process.
those logic notations were developed to create intelligent systems.
_ There are two problems in this approach :

for any problem.
1. This approach is not suitable to use when 100% knowledge is not available
ent a simple human reasoning process; practically, all
2. As vast number of computations was required even to implem
because even problems with just a few hundred facts can exhaust the computational
problems were not solvable
resources of any computer.
oach
4.2.4 Acting Rationally : The Rational Agent Appr
of the design of intelligent agents”. (Poole et at, 1998)
— Definition 1 :“Computational Intelligence is the study
ligent behaviour in artifacts”.
— Definition 2:“Al ... is concerned with intel
(Nilsson, 1998)
Rational Agent
period and a dapt to change to create and
through sensors over a prolonged time
— Agents perceive t heir environment
nal agent i s the one that does “right”
ue goal s and tak e actio ns thro ugh actuators to achieve those goals. A ratio
purs
tainty in knowledge.
things and acts rationally s 0 as to achie
ve the best outcome even when there is uncer
advantages over the other approaches

- The rational-agent approach has two
y can be achieved by selecting
ared to other appr oach es this ig the more general approach as, rationalit
1. As comp
ral available.
the correct inference from the seve
genera | and can be used to
has speci fic stan dard s and is math emat ically well defined and completely
2. Rationality subjective and cannot be
that achi eve it. Hum an beha vior , on the other hand, is very
develop agent designs
proved mathematically.
w TechNnowledgé
Pp ubii¢ations
. ,
oF _Al&SC
Al&SC (MU-Ser
(MU-Sem. 7-Comp) 1-4 Intro. to Artificial Intelligence(Al) and Soft Computing
—. The two approaches namely, thinking huma
nly and thinking rationally are based on the
reasoning epee
intelligen syst
t ems while; the other two acting humanly and acting rationally are based on
the intelligent behaviour
expected from them. .
In our syllabus we are going to study acting rationally approach.

1.3 _ Categorization of Intelligent System
s . -
As Al is a very broad concept, there are different types or forms of Al. The critical categories of Al can be based on the
Capacity of intelligent program or what the program is able to do. Under this consideration there are three main
categories: :
1. Artificial Narrow Intelligence/ Weak Al
Weak Al is Al that specializes in one area. It is not a general purpose intelligence. An intelligent agent is built to solve a
particular problem or to perform a specific task is termed as narrow intelligence or weak Al. For example, it took years
of Al development to be able to beat the chess grandmaster, and since then we have not been able to beat the
machines at chess. But that is all it'can do, which is does extremely well.
2. Artificial General Intelligence / Strong Al
Strong Al or general Al refers to intelligence demonstrated by machines in performing any intellectual task that human
can perform. Developing strong Al is much harder than developing weak Al. Using artificial general intelligence |
machines can demonstrate human abilities like reasoning, planning, problem solving, comprehending complex ideas,
learning from self experiences, etc. Many companies, corporations’ are working on developing a general intelligence
but they are yet to complete it.
3. Artificial Super Intelligence
As defined by a leading Al thinker Nick Bostrom, “Super intelligence is an intellect that is much smarter than the best
human brains in practically every field, including scientific creativity, general wisdom and social skills.” Super
” intelligence ranges from a machine which is just a little smarter than a human to a machine that is trillion times
smarter. Artificial super intelligence is the ultimate power of Al.
- :
1.4 Components of Al
Al is a vast field for research and it has got applications in almost all possible domains. By keeping this in mind,
components of Al can be identified as follows: (Refer Fig.1.4.1)
1. Perception
2. Knowledge representation
3. Learning
4. Reasoning
5. Problem solving
Natural language processing (Language-understanding)
Wey TechKnowledge
y-Al&SC (MU-Sem: 7-Gomp) =» and Soft Computing
intro. to Artfcial ntaligence(Al)
Reasoning Linguistic intelligence
8
Learning
isa
Perception
j
Problem solving
Fig. 1.4.1: Components of Al
1. Perception
In order to work in the environment, intelligent agents need to scan the environment and the various objects in it.
Agent scans the environment using various sense organs like camera, temperature sensor, etc. This is called as
perception. After capturing various scenes, perceiver analyses the different objects in it and extracts their features
and relationships among them. 7
2. Knowledge representation
The information obtained from environment through sensors may not be in the format required by the system.
Hence, it need to be represented in standard formats for further processing like learning various patterns, deducing
inference, comparing with past objects, etc. There are various knowledge representation techniques like Prepositional
logic and first order logic.
3. Learning
Learning is a very essential part of Al and it happens in various forms. The simplest form of learning is by trial and
error. In this form the program remembers the action that has given desired output and discards the other trial
actions and learns by itself. It is also called as unsupervised learning. In case of rote learning, the program simply
remembers the problem solution pairs or individual items. In other case, solution to few of the problems is given as
input to the system, basis on which the system or program needs to generate solutions for new problems. This is
known as supervised learning.
4. Reasoning
Reasoning is also called as logic or generating inferences form the given set of facts. Reasoning is carried out based on
strict rule of validity to perform a specified task. Reasoning can be of two types, deductive or inductive. The deductive
reasoning is in which the truth of the premises guarantees the truth of the conclusion while, in case of inductive
reasoning, the truth of the premises supports the conclusion, but it cannot be fully dependent on the premises. In
programming logic generally deductive inferences are used. Reasoning involves drawing inferences that are relevant
to the given problem or situation.
5. Problem-solving
Al addresses huge variety of problems. For example, finding out winning moves on the board games, planning actions
in order to achieve the defined task, identifying various objects from given images, etc. As per the types of problem,
there is variety of problem solving strategies in Al. Problem solving methods are mainly divided into general purpose
methods and special purpose methads. General purpose methods are applicable to wide range of problems while,
special purpose methods are customized to solve particular type of problems.
Intro. to Artificial Intelligence(Al) and Soft Computing
6. Natural language processing
Natural Language Processing, involves machines or robots to understand and process the language that human speak,
and infer knowledge from the speech input. It also involves the active participation from machine in the form of dialog
ie, NLP aims at the text or verbal output from the machine or robot. The input and output of an NLP system can be
speech and written text respectively.
1.4.1. Computational Intelligence vs. Artificial Intelligence
Computational Intelligence is the study of the design of | Artificial Intelligence is study of making
machines which
intelligent agents can do things which at presents human do better.
Cl involves numbers and computations. Al involves designs and symbolic knowledge
representations.
Cl constructs the system starting from the bottom level | Al analyses
the overall structure of an intelligent system
computations, hence follows bottom-up approach.
by following top down approach.
Cl concentrates on low level cognitive function |
Al concentrates of high level cognitive structure design.
implementation.
1.5 History of Artificial Intelligence

— The term Artificial Intelligence {Al) was introduced
by John McCarthy, in 1955. He defined artificial
intelligence as
“The science and engineering of making intelligent machin
es”.
— Mathematician Alan Turing and others presented
a study based on logic driven computational
theories which showed
that any compu ter program can work by simply shuffl
ing “O” and “1” (i.e. electricity off and
electricity on). Also, during
that time period, research was going on
in the areas like Automations, Neurology, Contr
ol theory, Information theory,
etc. :
— This inspired a group of researchers to
think about the possibility of Creating an
electronic brain. In the year 1956 a
conference was conducted at the campus of Dartmouth College wher
e the field of artificial intelligence resea
founded. rch was
-
— This conference was attended by
John McCarthy, Marvin Minsky,
supposed to be the pioneers Allen Newell and Herbert Simon
of
artificial intelligence research etc., who are
for a very long time. During that
Intelligence systems were develo time seed Artificial
ped by these researchers and thei
r students.
— . Let’s see few examples of such artificial intelligent systems
:
1. Game : Checkers : Computer played as an opponent,
‘
2. Education : Algebra : For solving word problems,
7
3. Education : Math :Proving logical theorems,
4. Education : Language : Speaking English, etc,

—___ During that time period these founders
predicted that in few
years machines can do any wor
"they failed to recognize the difficulties which can be faced k that a man can do, but
,
TechKnowledga
Puostcations Useee rEC
Al&SC (MU-Sem.7-Comp) = 1-7 Intro. to Artificial Intelligence(Al) and Soft Computing
Meanwhile we will see the ideas, viewpoints and techniques which Artificial Intelligence has inherited from other
disciplines. They can be given as follows :
1. Philosophy : Theories of reasoning and learning have emerged, along with the viewport that the mind is
constituted by the operation of a physical system.
2. Mathematical : Formal theories of logic, probability, decision making and computation have emerged.
3. Psychology : Psychology has emerged tools to investigate the human mind and a scientific language which are
used to express the resulting theories.
4. Linguistic : Theories of the structure and meaning of language have emerged.
5. Computer science : The tools which can make artificial intelligence a reality has emerged.
1.6 Applications of Artificial Intelligence

we have | Robot, Wall-E, The
You must have seen use of Artificial Intelligence in many SCI-FI movies. To name a few
potential of using Al and sometimes
Matrix Trilogy, Star Wars, etc. movies. Many a times these movies show positive
such movies, which show us many probable
also emphasize the dangers of using Al. Also there are garnes based on
applications of Al.
or/and predicting output for a system. Al can
Artificial Intelligence is commonly used for problem solving by analyzing
in wide range of fields for example in diagnosing
provide solutions for constraint satisfaction problems. It is used
in entertainment field, etc.
diseases, in business, in education, in controlling a robots,
Fig. 1.6.1 : Fields of Al Application

be many fields in which
1.6.1 shows few fields in which we have applications of artificial intelligence. There can
Fig.
be used.
Artificially Intelligent Systems can
Education
children are developed
can be built using artificial intelligence techniques. Software for pre-school
Training simulators nal theory are the current areas
e learn ing with fun games . Auto mate d grading, Interactive tutoring, instructio
to enabl
of application.
Entertainment
as an opponent when human
s, robo ts are desi gned to play asa character. In games they can play
Many movies, game
able.
- player is not available or not desir |
Medical
complex operations of
card iolo gy (CRG ), Neu rol ogy (MRI), E mbryology (Sonography),
Al has applications in th e field of f rotations, store and ret
rieve
in organ izing bed schedules, managing staf
internal organs, etc. It can be also used can provi de with medica
l
expert system s are enable d to predict the decease and
information of patient. Many
prescriptions. py pl{c ations -
4. Military
in Ii :
or - ife Stacking :
Training simulators can be used in military applications. Also areas where human cannot reach
to be made quickly taking into”
conditions, robots can be very well used to do the required jobs. When decisions have
can p provide crucial ~ é
account an enormous amount of information, and when lives are at stake, artificial intelligence
r creatin 3
assistance. From developing - intricate flight plans to implementing complex supply systems 0 8 training
simulation exercises, Al is a natural partner in the modern military.
5. Business and Manufacturing
Latest generation of robots are equipped well with the performance advances, growing integration of vision and an
enlarging capability to transform manufacturing.
6. Automated planning and scheduling
Intelligent planners are available with Al systems, which can process large datasets and can consider all the -
constraints to design plans satisfying all of them.
7. Voice technology
Voice recognition is improved a lot with Al. Systems are designed to take voice inputs which are very much applicable
in case of handicaps. Also scientists are developing an intelligent machine to emulate activities of a skillful musician.
Composition, performance, sound processing, music theory are some of the major areas of research.
8. Heavy industry
Huge machines involve risk in operating and maintaining them. Human robots are better replacing human operators.
These robots are safe and efficient. Robot are proven to be effective as compare to human in the jobs of
repetitive
nature, human may fail due to lack of continuous attention or laziness.
1.7 Sub Areas/ Domains of Artificial Intelligence
Al Applications can be roughly classified based on the type of tools/a

pproaches used for inoculating intelligence i in the
system, forming sub areas of Al. Various sub domains/ areas in intelligent
systems can be given as follows; Natural _
Language Processing, Robotics, Neural Networks and Fuzzy
Logic. Fig. 1.7.1 shows these areas in Intelligent Systems.
Fig. 1.7.1 : Sub-areas in Intelligent

Systems
1. Natural language processing: One of the
application of Al is in field of Natural Langu
age Processing (NLP). NLP £ :
enables interaction between computer
s and human (natural)
machine translation (e.g. Lunar System), language. Practical applications of NLP are in
information retrieval, text categori
zation, etc. Few more appl
extracting 3D information using vision, spee ications are
ch recogn ition,perception, image
formation.
We TechKnowledge
_- 1-9 _ Intro. to Artificial Intelligence(Al) and Soft Computing
¥ Al&SC (MU-Sem..7-Comp)
whose environment is the
2. Robotics : One more major application of Al is in Robotics. Robot is an active agent
field, in military, etc. for
physical world. Robots can be used in manufacturing and handling material, in medical
automating the manual work.
Neural Network is a system that works like
3. Neural networks : Another application of Al is using Neural Networks.
analysis, in character recognition, in image
a human brain/nervous system. It can be useful for stock market
Optical Character Recognition (OCR), etc.
compression, in security, face recognition, handwriting recognition,
the help of Fuzzy Logic. Fuzzy Logic can be useful in
4. Fuzzy logic : Apart from these Al systems are developed with
making appr oximations rather than having a fixed and exact
reasoning for a problem. You must have seen sy stems
fuzzy logic (they call it we sense technology!”).
like AC, fridge, washing machines which are based on
Current Trends in Artificial Intelligence _

1.8
conditioners, to smart
of our life. Fram washing machine, Air
Artificial Intelligence has touche d each and every aspect
as well. Robots are doing the
In industry, Al is doing marvellous work
phones everywhere Al is serving to ease our life.
become a reality. WiFi-enabled Barbie
uses speech-recognition to talk and
sound work in factories . Driverles s cars have
saw significant advances in
e their pro duct and increase sales. Al
listen to children. Companies are using Al to improv
ts.
in which Al is showing significant advancemen
machine learning. Following are the areas
1. Deep learning
area of focus in Artificial *
enabling th e concept of deep learning is the top most
Convolutional Neural Networks
and applications areas of Al like,
natural language and text processing,
problems
intelligence in todays’ era. Many wered by
imodal information processing empo
recognition, computer vision, inf ‘ormation retrieval, and mult
speech
multi-task deep learning.
2, Machine learning
e a given
computers to use example data or past experience to solv
of machine learning is to program
The goal ict
that analyze past sales data to pred
y succ essf ul appl icat ions of machine learning include systems
problem. Man minimum resources, and extract
r, opti mize robo t beha viou r so that a task can be completed using
customer behaviou
data.
knowledge from bioinformatics
3. Alreplacing workers
are getting replaced by
fety haza rds, robo ts are doin g a good job. Human resources
sa
In industry where there are exceedingly well
see that the whit e colo r jobs of data processin g are being done
rie 5 to
robots rapidly. People are wor ght together technologists and
A stud y from The Nati onal Academy of Sciences brou
by intelligent programs.
to happen.
ts to figure out what's going
economists and social scientis
4. Internet of Things (loT)

ntion of internet of
cars and smar ter worl d is evolving rapidly with the inve
smarter
The concepts of smarter homes, sly connected to something in orde
r to pe rform
y obje ct will be wir eles
and ever
things. The future i 5 no far when each how the mined data can pote
nti ally
any huma n instr uctio ns or interference. The worry is
soma smart actions without
be exploited.
TechKnowledye
G Peprications
W_Algsc (MU-Sem. 7-Comp) 4-10 ——
Intro. to Artificial Intelligence(Al) and Soft Computing
esses |
5. Emotional Al
Emotional Al, where Al can detect human emotions, is another upcoming and important area of research. Computers
ability to understand speech will lead to an almost seamless interaction between human and computer. With d
increasingly accurate cameras, voice and facial recognition, computers are better able to detect our emotional state, 3
Researchers are exploring how this new knowledge can be used in education, to treat depression, to accurately
predict medical diagnoses, and to improve customer service and shopping online.
6. Alin shopping and customer service
Using Al, customers’ buying patterns, behavioral patterns can be studied and systems that can predict the purchase or
can help customer to figure out the perfect item. Al cab be used to find out what will make the customer happy or
unhappy. For example, if a customer is shopping online, like a dress pattern but needs dark shades and thick material,
computer understand the need and brings out new set of perfectly matching clothing for him.
7. Ethical Al
With all the evolution happening in technology in every walk of life, ethics must be considered at the forefront of
research. For example, in case of driverless car, while driving, if the decision has to be made between weather to dash
a cat or a lady having both in an uncontrollable distance in front of the car, is an ethical decision. In such cases how
the programming should decide who is more valuable, is a question. These are not the problems to be solved by
computer engineers or research scientists but someone has to come up with an answer.
1.9 Intelligent Agents .
1.9.1 What is an Agent?
Agent is something that perceives its environment through sensors and acts upon that environment through effectors
or actuators. Fig. 1.9.1 shows agent and environment.
Take a simple example of a human agent. It has five senses : Eyes, ears, nose, skin, tongue. These senses sense the
environment are called as sensors. Sensors collect percepts or inputs from environment and passes it to the
processing unit.
Actuators or effectors are the organs or tools using which the agent acts upon the environment. Once the sensor
senses the environment, it gives this information to nervous system which takes appropriate action with the help of
actuators.
— Incase of human agents we have hands, legs as actuators or effectors.
Agent
_ Sensors
ovide inpu
Environment
Fig. 1.9.1 : Agent and Environment
TechKnowledge -
Pubiicatians ;
Al&SC (MU-Sem. 7-Comp) Intro. to Artificial Intelligence(Al) and Soft Computing’ .
Fig.1.9.2 : Generic robotic agent architecture

for a robotic agent, can you think of
After understanding what an agent is, let’s try to figure out sensor and actuator
sensors and actuators in case ofa robotic agent?
, etc. used as sensors, while various types of motors,
The robotic agent has cameras, infrared range finders, scanners
input.
screen, printing devices, etc. used as actuators to perform action on given
ayes, ears, and other
gan
__ s0ffor ss
sensor
Human
hands, legs, mouth, and other.
_ body parts for actuators,
‘ _cameras and Infrared :
Robotic |
various motors ;
_ > for actuators
and robotic agent
Fig.1.9.3 : Sensors and actuators in human
to do. The agent function
funct ion is the descr iptio n of what all functionalities the agent is supposed
The agent
represented as [f: P* > A]
ences to the desired actions. It can be
provides mapping between percept sequ
architecture suitable language. Agent
that implements agent function in an
Agent program is a computer program
That device must have some form of
need s to he insta lled on a devic e in order to run the device accordingly.
programs ion of the architecture
envi ronm ent and actua tors to act upon it. Hence agent is a combinat
sensors to sense the
.
hardware and program software
m
Agent = Architecture + Progra
cleaner agent in “WALL-E” (animated
exam ple of vacu u m clea ner agent. You might have seen vacuum
Take a simple (outputs) used in case of a vacu
um
rsta nd how t o repr esen t the percept’s (input) and actions
movie). Let’s unde
cleaner agent.
: OF Tecthnam letns
= Publications
a Comp) 112 __ Intro. to Artificial Inteligence(Al)
and Soft Computing
OC»)
OOSd 8S |
Fig.1.9.4 : Vacuum cleaner agent
As shown in Fig. 1.9.4, there are two blocks A and B having some dirt. Vacuum cleaner agent supposed to sense the
dirt and collect it, thereby making the room clean. In order to do that the agent must have a camera to see the dirt
and a mechanism to move forward, backward, left and right to reach to the dirt. Also it should absorb the dirt. Based
on the percepts, actions will be performed. For example : Move left, Move right, absorb, No Operation.
Hence the sensor for vacuum cleaner agent can be camera, dirt sensor and the actuator can be motor to make it
move, absorption mechanism. And it can be represented as :
[A, Dirty], [B, Clean], [A, absorb],[B, Nop], etc.
1.9.2 Definitions of Agent
There are various definitions exist for an agent. Let’s see few of them.
IBM states that agents are software entities that carry out some set of operations on behalf of a user or another
program.
— FIPA : Foundation for Intelligent Physical Agents (FIPA) terms that, an agent is a computational process that
implements the autonomous functionality of an application.
— Another definition is given as “An agent is anything that can be viewed as perceiving
its environment through sensors
and acting upon the environment through effectors”.
Agent will perform

all the tasks on
your behalf
Solving
experiments
and
assignments
Washing clothes Cleaning service
Fig.1.9.5 : Interactive Intelligent Agen
t
— By Russell and Norvig, F. Mills and R. Stuff
le beam’s definition says that “An agent
is anything that is capable of acting |
upon information it perceives. An intelligent agent Is an agen
t capable of making decisions about how it acts based on
experience”.
Sa TechKnewledga
Fubtications
— From above definitions we can understand that an agent is: (As per Terziyan, 1993)
o Goal-oriented o Creative
o Adaptive . o Mobile
o Social o — Self-configurable
1.9.3 Intelligent Agent

System” which helps in deciding an
- Inthe ura agent example, we read that there is something called as “Nervous
In robotic agent, we have software’s
sei with the assistance of effectors, based on the input given by sensors.
which demonstrates the functionality of nervous system. ,
its sensors and act upon the
— Intelligent agent is the can take input from the environment through
one which
to achieve a goal.
always directed
environment through its actuators. Its actions are
to be self-governed, responsive, goal-oriented, etc.
— The basic abilities of an intelligent agent are to exist
Generally observed
modules are responsible for exhibiting intelligence.
— Incase of intelligent agents, the software
:
capabilities of an intelligent agent can be given as follows
o Ability to remain autonomous (Self-directed)
o Responsive
o Goal-Oriented
ugh its sensors and act upon the
take input from the environment thro
— Intelligent agent is the one which can a goal.
actions are always directed to achieve
environment through its actuators. Its
1.9.3(A) Structure of Intelligent Agents

cture of an intelligent agent.
— Fig.1.9.6 shows the general stru
Observe the inpu
Scan DB for
corresponding action
Database of Input
Action is
given as output
ucture of intel ligent agent

Fig.1.9.6: General str environment
ironment inter: act with each other. Every time
ed h ow agent and env ns the database of
— From Fig. 1.9.6, it can be obs erv
ough its sensors and get the input, then sca
es the environmen t thr the appropriate
changes the age nt firs t obs erv
In put and lastly sets the internal state to
s for the cor res pon ding action.for given
input and action
ndings is an
action. der you are an agent and your surrou
e. Con si see
wi th a real life exampl and by mistake you tou
ch a hot pan. We will
— Let’s understand this working coo kin g in kit che n e tou che d
a situation W here you are m environment (i.e. you hav
énvironment. Now, take ste p. You r tou ch sensors take input fro = 5
step by =
what happens in this situation = OF renee
Intro. to Artificial Intelligence(Al) and Soft Computing 3 a
WF AlaSc (MU-Sem. 7-Comp) 1-14
should be taken when you go near hot elements?”

some hot element), then it asks your brain if it knows “what action
from the hot element
Now the brain will inform your hands (actuators) that you should immediate ly take it away
take your hand away from the h ot pan.
otherwise it will burn. Once this signal reaches your hand you will
In above example, if
The agent keeps taking input from the environment and goes through these states every times
your action takes more time then in that case your hand will be burnt.
the states which will be followed in
So the new task will be to find solution if the hand is burnt. Now, you think about
and Jennings, “An intelligent agent is one that is capable of taking flexible
this situation. As per Wooldridge
self-governed actions”.
- They say for an intelligent agent to meet design objectives, flexible means three things:
Reactiveness . Pro-activeness
Social ability
1. Reactiveness : It means giving reaction to a situation in a stipulated time frame. An agent can perceive the
environment and respond to the situation in a particular time frame. In case of reactiveness, reaction within
situation time frame is more important. You can understand this with above example, where, if an agent takes
more time to take his hand away from the hot pan then agents hand will be burnt.
2. Pro-activeness : It is controlling a situation rather than just responding to it. Intelligent agent show goal-directed
behavior by taking the initiative. For example : If you are playing chess then winning the game is the main
objective. So here we try to control a situation rather than just responding to one-one action which means that
killing or losing any of the 16 pieces is not important, whether that action can be helpful to checkmate your
opponent is more important.
3. Social ability ; Intelligent agents can interact with other agents (also humans). Take automatic car driver exa mople,
where agent might have to interact with other agent or a human being while driving the car.
- Following are few more features of an intelligent agent.
o Self-Learning : An intelligent agent changes its behaviour based on its previous experience.
This agent keeps
updating its knowledge base all the time.
o Movable/Mobile : An Intelligent agent can move from one machine to another
while performing actions.
o Self-gaverning : An Intelligent agent has control over its own actions.
1.10 Rational Agent
- For problem solving, if an agent makes

a decision based on some logical reasoning,
“Rational Decision”. The way humans have abilit then, the decision is called
y to make right decisions, based on; his/)
reasoning; an agent should also be able to make her exp eri enc
na e e and logical
correct decisions, based on what it kno
and actions which are carried out by that agent from ws from the perc ept sequence
its knowledge.
prolonged time Period and adap
pursue goals and take actions through actuators t to change to create and
to achieve those goals. A rational
things and acts rationally so as to achieve agent is the one that does “right”
the best outcome even when ther
e is uncertainty in knowledge
-— Arational agent is an agent that has clear .
fuinctt . preferences, can model uncertai
eae iainty via expected values of variables or :
ee and always chooses to perform the
action with the optimal expected
among g a easible action
i s. A rational agent can be anything that makes decisions typically
outcome for it 7"
i a person as “ee ve
software program. ? » a machine, o
var TechKnowledga
2 Publications
¥ Al&SC (MU-Sem. 7-Comp) IS iit 10Artificial Intelligence(Al) and Soft Computing
criterion of success for
criteria: First is the performance measure which defines the
Rationality depends on fou r main the agent
an agent, second is the agent's prior knowledge of the environment, and third is the action pe rformed by
to date.
and the last one is agent's percept sequence
measuring success of an agent's performance. Take a
Performance measure is one of the major criteria for s
re of a vacuum-cleaner agent can depend upon variou
vacuum-cleaner agent's example. The performance measu
that dirt, consumption of electricity, etc.
factors like it's dirt cleaning ability, time taken to clean
se it
which is very useful for decision making, becau
_ For every percept sequence a built- in knowledge base is updated, then
consequen ces of performin g some particular action. If the consequences direct to achieve desired goal
stores the state, then we get a
if the consequences do not lead to desired goal
we get a good performance measure factor, else,
poor performance measure factor.
hammer (b) Agent Is using nail and hammer efficiently

(a) Agent's finger is hurt while using nail and
Fig. 1.10.1
it for the next time
see Fig.1 .10.1 . If agent hurts his finge r while using nail and hammer, then, while using
— for example, will be able to use the
be more caref ul and the proba bilit y of not getting hu rt will increase. In short agent
agent will
hammer and nail more efficiently.
nce, experience and knowledge to
agent who m akes use of its percept seque
— Rational agent can be defined as an
s the most feasible action which
of an agent for every pro bable action. It select
maximize the performance measure
will lead to the expected results optimally.
Environmental
experience
oI
‘Input | —> Goal:
127 Alternate possible actions

coy
Ez Selected actions which leads to

Agent optimal expected results
Fig.1.10.2 ; Rational Agent
Agent
1.11 Nature of Environment and PEAS Properties of
/ Task Environment Properties
1.11.1 Environments Types / Nature of Environment
sia
TechXnowledge
Puofigatioas
7-
WF alasc (MU-Sem.h.Z-COMP)
am _intro.to Artificial Inteligence(Al) and Soft Computing .
1. Fully observable vs. Partially observable

can have access to.
The first type of environment is based on the observability. Whether the agent sensors
e ~
complete state of environment at any given time or not, decides if it is a fully observable or partially observabl
environment.
information required to take actions,
In Fully observable environments agents are able to gather all the necessary
example,
Also in case of fully observable environments agents don’t have to keep records of internal states. For
Word-block problem, 8-puzzle problem, Sudoku puzzle, etc. in all these problem worlds, the state is completely
visible at any point of time.
Known vs. unknown
Fig.1.11.1; Environment types
Environments are called partially observable when sensors cannot provide errorless information at any given
time for every internal state, as the environment is not seen completely at any point of time.
Also there can be unobservable environments where the agent sensors fail to provide information about internal
states.
For example, In case of an automated car driver system, automated car cannot predict what the other drivers are
thinking while driving cars. Only because of the sensor’s information gathering expertise it is possible for an
automated car driver to take the actions.
2. Single agent vs. Multl-agent

The second type of an environment is based on the number of agents acting in the environment. Whether the
agent is operating on its own or in collaboration with other agents decides if it is a Single agent or a multi-agent
environment,
For example : An agent playing Tetris by itself can be a single agent environment, whereas we can have an agent _|
playing checkers in a two-agent environment. Or in case of vacuum cleaner world, only one machine is working,
so it’s a single agent while in case of car driving agent, there are multiple agents driving on the road, hence it’s 4 ;
multi-agent environment.
[
Intro. to Artificial Intelligence( 1) and t Computing

1-17
Al&SC (MU-Sem. 7-Comp) _
multi-agent and Competitive multi-agent. pos you
Multi-agent environment is further classified as Co-operative
of agent environment do we have:
might be thinking in case of an automated car driver system which type
car driving system 'X', other car say
Let's understand it with the help of an automated car driving example. For a
measure and the input taken by car 'Y
'y' is considered as an Agent. When 'Y' tries to maximize its performance
system we have a cooperative
depends on the car 'X'. Thus it can be said that for an automated car driving
multi-agent environment.
ts, and trying to maximize their own
Whereas in case of “chess game” when two agents are operating as opponen
performance, they are acting in competitive multi agent environment.
3. Deterministic vs. Stochastic

state of the environment can be completely
An environment is called deterministic environment, when the next
determined by the previous state and the action executed by the agent.
the next state of the environment
For example, in case of vacuum cleaner world, 8-puzzle problem, chess game
solely depends on the current state and the action performed by agent.
of
Stochastic environment generally means that the indecision about the actions is enumerated in terms
probabilities. That means environment changes while agent is taking action, hence the next state of the world
does not merely depends on the current state and agent’s action. And there are few changes happening in the
environment irrespective of the agent’s action. An automated car driving system has a stochastic environment as
the agent cannot control the traffic conditions on the road.
In case of checkers we have a multi-agent environment where an agent might be unable to predict the action of
the other player. In such cases if we have partially observable environment then the environment is considered
to be stochastic.
If the environment is deterministic except for the actions of other agents, then the environment is strategic. That
is, in case of game like chess, the next state of environment does not only depend upon the current action of
agent but it is also influenced by the strategy developed by both the opponents for future moves.
We have one more type of environment in this category. That is when the environment types are not fully
observable or non-deterministic; such type of environment is called as uncertain environment.
4. Episodic vs. Sequential
An episodic task environment is the one where each of the agent's action is divided into an atomic incidents or
episodes. The current incident is different than the previous incident and there is no dependency between the
current and the previous incident. In each incident the agent receives an input from environment and then
performs a corresponding action. :
Generally, classification tasks are considered as episodic. Consider an example -of pick and
place robot agent tl
which is used to detect defective parts from the conveyor belt of an assembly line. Here, every time
agent will
make the decision based on current part, there will not be any dependency between
the current and previous
decision.
In sequential environments, as per the name suggests, the previous

decision can affect all future decisions. The
next action of the agent depends on what action he has taken
previously and what action he is supposed to take
in future.
Sa TechKnowledge
Pudlications
1-18 Intro. to Artificial Intelligence(Al) and Soft Computing
For or example, z in checkers where previous move can affect all the following moves. Also sequential environment
vet
can be understood with the help of an automatic car driving example where, current decision can affect the next
decisions. If agent is initiating breaks, then he has to press clutch and lower down the gear as next consequent
actions.
5. Static vs. Dynamic

You have learnt about static and dynamic terms in previous semesters with respect to web pages. Same way we
have static (vs. dynamic) environments. If an environment remains unchanged while the agent is performing
given tasks then it is called as a static environment. For example, Sudoku puzzle or vacuum cleaner environment
are static in nature.
- If environment is not changing over the time but, an agent's performance is changing then, nt Is elles asa
semi-dynamic environment. That means, there is a timer exist in the environment who is affecting the
performance of the agent.
— For example, In chess game or any puzzle like block word problem or 8-puzzle if we introduce timer, and if agent’s
performance is calculated by time taken to play the move or to solve the puzzle, then it is called as semi-dynamic
environment.
—Lastly, if the environment changes while an agent is performing some task, then it is called dynamic environment.
— _ In this type of environment agent's sensors have to contin uously keep sending signals to
agent about the current
state of the environment so that appropriate action can be taken with immediate effect.
— Automatic car driver example comes under dynamic environment as
the environment keeps changing all the
time.
6. Discrete vs. Continuous
— You have seen discrete and continuous signals in

old semesters. When you have distinct, quantized,
defined values of a signal it is considered as discrete signal. clearly
— Same way, when there are distinct and clearly defin

ed inputs and outputs or precepts and actions,
a discrete environment. For example : chess then it is called
environment has a finite number of distinct
inputs and actions.
— When a continuous input signal is received by
an agent, all the Precepts and actions cannot
beforehand then it is called continuous be defined
environment. For example : An automatic
car driving system
7. Known vs. Unknown
- In a known environment, the output for

all probable actions is given, Obvi
environment, for an agent to make a decision, it has ously, in case of unknown
to gain knowledge about -
how the environment works.
— Table 1.11.1 summarizes few task environm
ent and their characteristics.
Table 1.11.1 : Task env
ironments
| | Part ~ Picking © Cross word aS
Observable Partially
Partlally fully Partially
Agents . M ulti Full
ully
agent|Sin gle agent
- single Multi
— (coope
. rative) ; i
|(Competitive) eee
Deterministic (compe titive) =
tochastici
Stochastic D eterminisitic
Strategic
3
W
Strategici
TachKnowledga
Poolications °
” ~
.
|
+
a7 Al&SC (MU-Sem. 7-Comp)
Episodic Sequential Episodic Sequential Sequential Sequential
Static Dynamic Dynamic Static Dynamic Semi |
Discrete Continuous Discrete Discrete Continuous Discrete
Known and Unknown Known Known Known Known

Unknown
1.11.2 PEAS Properties of Agent
— PEAS : PEAS stands for Performance Measure, Environment, Actuators, and Sensors. It is the short form used for
performance issues grouped under Task Environment.
— You might have seen driverless/ self driving car videos of Audi/ Volvo/ Mercedes, etc. To develop such driverless cars
we need to first define PEAS parameters.
— Performance Measure : It the objective function to judge the performance of the agent. For example, in case of pick
’ and place robot, number of correct parts in a bin can be the performance measure.
— Environment: It the real environment where the agent need to deliberate actions.
— Actuators : These are the tools, equipment or organs using which agent performs actions in the environment. This
works as the output of the agent. -
— Sensors : These are the tools, equipment or organs using which agent captures the state of the environment. This
works as the input to the agent.
— Tounderstand the concept of PEAS, consider following examples.

(A) Automated car driving agent
1. Performance measures which should be satisfied by the automated car driver:
(i) Safety : Automated system should be able to drive the car safely without dashing anywhere.
(ii) Optimum speed : Automated system should be able to maintain the optimal speed depending upon the
surroundings.
(iii) Comfortable journey : Automated system should be able to give a comfortable journey to the end user, i.e.
depending upon the road it should ensure the comfort of the end user.
of energy
(iv) Maximize profits : Automated system should provide good mileage on various roads, the amount
user is
consumed to automate the system should not be very high, etc. such features ensure that the
benefited with the automated features of the system and it can be useful for maximizing the profits.
2. Environment
(i) Roads : Automated car driver should be able to drive on any kind of a road ranging from city roads to
highway. :
wW TechKnowledyé
Pudiications
Al&SC (MU-Sem. 7-Comp) Intro. to Artificial Intelligence(Al) and Soft Computing :
(ii) Traffic conditions : You will find different set of traffic conditions for different type of roads. Automated
system should be able to drive efficiently in all types of traffic conditions. Sometimes traffic conditions are
formed because of pedestrians, animals, etc.
(iii) Clients : Automated cars are created depending on the client’s environment. For example, in some countries 3
you will see left hand drive and in some countries there is a right hand drive. Every country/state can have
designed.
different weather conditions. Depending upon such constraints automated car driver should be
3. Actuators are responsible for performing actions/providing output to an environment.
In case of car driving agent following are the actuators :
(i) Steering wheel which can be used to direct car in desired direction (i.e. right/left)
(ii) Aaccelerator, gear, etc. can be useful to increase or decrease the speed of the car.
(iil) Brake is used to stop the car.
(iv) Light signal, horn can be very useful as indicators for an automated car.
4. Sensors: To take input from environment in car driving example cameras, sonar system, speedometer, GPS,
engine sensors, etc. are used as sensors.
(B) Part-picking ARM robot
(i) Performance measures : Number of parts in correct container.
(ii) Environment : Conveyor belt used for handling parts, containers used to keep parts, and Parts.
(ili) Actuators : Arm with tooltips, to pick and drop parts from one place to another.
(iv) Sensors : Camera to scan the position from where part should be picked and joint angle sensors which are used
to sense the obstacles and move in appropriate place.
(C) Medical diagnosis system
(i) Performance measures
a. Healthy patient: system should make use of sterilized instruments to ensure the safety (healthiness) of the
patient.
b. Minimize costs : The automated system results should not be very costly otherwise overall expenses of the
patient may increase, Lawsuits. Medical diagnosis system should be legal.
(ii) Environment : Patient, Doctors, Hospital Environment
(iii) Senso rs printer

: Screen,
_ (iv) Actuators : Keyboard and mouse which is useful to make entry of symptoms, findings, patient's answers to given
questions. Scanner to scan the reports, camera to click pictures of patients.
(D) Soccer player robot
(i) Performance measures : Number of goals, speed, legal game.
(ii) Environment: Team players, opponent team players, playing ground, goal net.
(iii) Sensors: Camera, proximity sensors, infrared sensors.

tiv) Actuators : Joint angles. motors.
Al&SC ( MU Sem 7 7 Comp ) 1-2 1 Intri 0. to to Artifici | telli
Icial
ic I Intelligenc
lig e (Al) and
in ft Co 7
So mputing
1.12 Structure of Agents / Types of Agents
Di epending upon the degree of 'intelligence and ability to achieve th
Boal, agents are categorized into five basic types.
e
These five types of agents are depicted in the Fig. 1.12.1
leaming agents
Fig.1.12.1 : Types of agents
Let us understand these agent types one by one.
1.12.1 Simple Reflex Agents

r
)
Simple Reflex Agent
input)
Sensors < pullpercept
output / action
Effectors
_/
Fig.1.12.2 : Simple reflex agents

ous inputs is called as simple
h perf orms actio ns base d on the current input only, by ignoring all the previ
An agent whic
reflex agent. .
the situation and its
lic ate d type of agen t. The sim ple reflex agent's function is based on
It is a totally uncomp matching action is taken without
ocol). If the c ondition is true, then
t
g acti on (con diti on-a ctio n prot

correspondin
considering the percept history.
oaches eye then, you will
refl exes with the hel p of a real life example, say some object appr
You can understand simple
ex.
reflex is called nat ural/innate refl
blink your eye. This type of simple
is based only on whether
agen t. It is a simp le reflex agent, as its decision
vacuu m cleaner
Consider the example of the
in Table 1.12.1.
ains dirt. The a gent function is tabulated det
the current loca tion cont Techhannle
OBSF. publi cations
f\&SC (MU-Sem. 7-Comp) 1-22 Intro. to Artificial Intelligence(Al) and Soft Computin
— Few possible input sequences and outputs for vacuum cleaner world wi th 2 locations are considered for simplicity.
Table 1.12.1
x 5 {A, clean} Right
A B {B, clean} Left
A B {A, dirt} Suck

<4,
A B {B, dirt} Suck

i)
A B Input sequence Output / action
=A, {location, content } Right, left, suck, no-op
A B
{A, clean}{A, clean} Right
A B
—4,] {A, clean}, dirt} Suck
A B
,
i
Sg TechKnawledga
Publications
_Al&SC (MU-Sem. 7- Comp)_
In case of above mentioned vac
uum agent only one sensor i 5
used and that is a dirt se nsor
if there is dirt or not. So the possible . This dirt sensor can detect
inputs are ‘dirt’ and ‘clean’.
Also the agent will have to mai
ntain a database of actions, whi
c h will help to decide what outp
an agent. Database will contain ut should be given by
conditions like : If there is dirt
on the floor to left or right.then
find out if there is dirt
il the entire assigned area is clea
ned then, vacuum cleaner should
suck
instruction.
If the vacuum cleaner agent keeps sear

ching for dirt and clean area, then, it
will surely get trapped in an in finite loop.
Infinite loops are unavoidable for simpl
e reflex agents Operating in partiall y observable
environments. By randomizing
its actions the simple reflex agent can avoid
these infinite loops. For example, on recei
ving {clean} as input, the
vacuum cleaner agent should either go to left
or right direction.
If the performance of an agent is of the right
kind then randomized behaviour can be considered
as rational in few
multi-agent environments.
1.12.2 Model-Based Reflex Agents
= Partially observable environment cannot be handled

well by simple reflex agents because it does not keep
track on
the previous state. So, one more type of agent was created
that is model based reflex agent.
An agent which performs actions based on the current input
and one previous input is called as model-based agent.
Partially observable environment can be handled well by model-
based agent.
From Fig. 1.12.3, it can be seen that once the sensor takes
input from the environment, agent checks for the current
state of the environment. After that, it checks for the previous state
which shows how the world is developing and
how the environment is affected by the action which was taken by the agent
at earlier stage. This is termed as model
of the world.
™
input/percept
Sensors <— =
». What action...
should.be tak
| output / action
Effectors
L Model-based Reflex Agents J
Fig.1.12.3 : Model-based reflex agents

Once this is verified, based on the condition-action protocol an action is decided. Thist decision
7
isPagiven to effectors and a
the effectors give this output to the environment.
wAl&SC (MU-Sem_7-Comp) Intro. to Artificial Intelligence(Al) and Sott Computing
The knowledge about “how the world is changing” is called as a model of the world. Agent which uses such mode| :
while working is called as the “model-based agent”. a4
— Consider a simple example of automated car driver system. Here, the world keeps changing
all the time. You must ‘
have taken a wrong turn while driving on some or the other day of your life. Same thing applies for. an agent. Suppose
if some car “X” is overtaking our automated driver agent “A”, then speed and direction in which “X”
and “A” are i
moving their steering wheels is important. Take a scenario where agent missed a sign
board as it was overtaking other
car. The world around that agent will be different in that case.
— Internal model based on the input history should be maintained by model-b

ased reflex agent, which can reflect at
least some of the unobserved aspects of the current state. Once this is
done it chooses an action in the same Way as
the simple reflex agent.
1.12.3 Goal-Based Agents

f
>
input/percept
Sensors «<—
‘ Whataction — ~
oul ai sya Sones
output / action
Goal-based Agents Effectors
LL
Fig.1.12.4 : Goal-based age

)
nts
— Model-based agents are further deve
loped based on the “goal” informat
ion. This new type of agent is
goal-based agent. As the nam called as
e suggests, Goal information
will illustrate the situations
that is desired. These agents
the specified goals. Goal bas

ed agents can only differentiat
e bet
performance can be 100% or zero
,
TechKaowledga
Pudsications
input/percept
Sensors < _
_
What is the curren
of th
__. What willbe the stateif
ction A’.is.perfo
output /
action
Utility-based Agents Effectors
Fig.1.12.5:Utility-based agents
Utility function is used to map a state to a measure of utility of that state. We can define a measure for determining
how advantageous a particular state is for an agent. To obtain this measure utility function can be used.
The term utility is used to depict how “happy” the agent is to find out a generalized performance measure, various
world states according to exactly how happy they would make an agent és compared.
Take one example; you might have used Google maps to find out a route which can take you from source location to
your destination location in least possible time. Same logic is followed by utility based automatic car driving agent.
Goals utility based automatic car driving agent can be used to reach given location safely within least possible time
and save fuel. So this car driving agent will check the possibte routes and the traffic conditions on these routes and will
select the route which can take the car at destination in least possible time safely and without consuming much fuel.
1.12.5 Learning Agents
Al&SC (MU-Sem. 7-Comp) Intro. to Artificial Intelligence(Al) and Soft Computing
Performance standard
~
input/percept
Feedback
r—______ changes
_ Learning
a element
ees «
Learning g
goals
- Problem
See
generator
IN :
Effectors output
/ action
Learning Agents ‘
a /
Fig. 1.12.6 : Learning agents
Why do you give mock tests ? When you get less marks for some question, you come to know that you have made
some mistake in your answer. Then you learn the correct answer and when you get that same question in further
examinations, you write the correct answer and avoid the mistakes which were made in the mock test. This same
concept is followed by the learning agent.
Learning based agent is advantageous in many cases, because with its basic knowledge it can initially operate in an
unknown environment and then it can gain knowledge from the environment based on few parameters and perform
actions to give better results.
Following are the components of learning agent :
1. Critic
Learning element
Performance element
Problem generator
1. Critic : It is the one who compares sensor’s input specifying effect of agent’s action on the environment with the
performance standards and generate feedback for leaning element.
2. Learning element : This component is responsible to learn from the difference between performance standards
and the feedback from critic. According to the current percept it is supposed to understand the expected -
behavior and enhance its standards
3. Performance element : Based on the current percept received from sensors and the input obtained by the
learning element, performance element is responsible to choose the action to act upon -the external
environment. ;
4. Problem generator : Based on the new goals learnt by learning agent, problem generator suggests new OF
alternate actions which will lead to new and instructive understanding.
1.13 Introduction to Soft Computing

Soft computing is a collection of all the techniques that help us to construct computationally intelligent systems. It has
been now realized that, real world problems are complex, pervasively imprecise and uncertain. To solve such
problems, we require computationally intelligent systems that combine knowledge, techniques and methodologies
from various sources.
Al&SC (MU-Sem. 7- Comp) _ ‘
_1-27— ___Intro. toArtificial Intelligence(Al) and Soft Computing
- There are three main requirements of any
intelligent system.
1. They must possess human like expertise
within a specific domain.
2. They should be able to adapt and learn to do
better in changing environment.
3. Should be capable of making decisions and taking
actions accordingly.
1.14 Soft Computing vs. Hard Computing

Hard computing involves the traditional methods of computing
that require precisely stated analytical models. They
often require more computational time. Examples of hard computi
ng are:
o Solving numerical problems (e.g. roots of polynomial, integration etc.)
o Searching and sorting techniques

0 Solving Computational geometry problem etc.
Unlike hard computing, soft computing techniques are tolerant of imprecision, uncertainty, partial truth and
approximation that are present in the real world problems. Examples of soft computing techniques are Neural
networks, Fuzzy logic, genetic algorithms etc.
- _ Following are differences between hard computing and soft computing.

z=
1. Hard computing is a conventional type of computing that | Soft computing techniques are imprecision,
requires a precisely stated analytic model. , approximation and uncertainty tolerant.
2. Hard computing requires programs to be written. Soft computing techniques are model free.
They can evolve their own models and
programs.
3. Hard computing is deterministic and uses two-valued logic. Soft computing is stochastic and uses
multi-valued logic such as fuzzy logic.
4. Hard computing needs exact data to solve a particular | Soft computing can deal with incomplete,
problem. uncertain and noisy data.
5. Hard computing techniques perform sequential computation. | Soft computing allows parallel
computations. E.g. Neural networks.
6. The solution or output of hard computing is precise. Soft computing can generate approximate
output or solution.
7. Hard computing is based on crisp logic, binary logic and | Soft computing is based on neural
numerical analysis. networks, fuzzy logic, and evolutionary
computations etc.
8. Hard computing techniques are not fault tolerant. The reason | Soft computing techniques are fault
is conventional programs and algorithms are bullt in such a | tolerant due to __ their redundancy,
. way that errors have serious consequences, unless enough | adaptability and reduced precision.
L redundancy Is added Into the system. characteristics.
‘Ey TechKnculedgs
Puolications
= i
ifici Intelligence
Artificial
Intro. toto Artiicial (Al) and | Soft Computi nqa
M 7-Com
m. 7-Comp)
yy _Al&SC (MU-Sem. —_1-28- _Intro.
1.15 Various Types of Soft Computing Techniques —
i
Soft Computing is the fusion of different techniques that were designed ,
to model.and enable solu tions to com plex lex real
world problems.
These real world problems are the problems that are too difficult to model, mathematically.
These problems result from the fact that our world seems to be imprecise,
i i ifficult to categorize.
uncertaini and diffi B
. . ‘
The soft computing techniques are capable of handling such uncertainty, impreciseness and vagueness ess p present in th e
real world data.
we
Most of the Soft computing techniques are based on some biological inspired methodologiesi such as hu mann ervous
systems, genetics, evolution, ant’s behaviors etc.
Soft Computing is the fusion of different techniques that were designed to model and enable solutions to complex real
world problems, which are not modeled or too difficult to model, mathematically.
— Soft computing consist several computing paradigms mainly are:

o Neural Network
Oo Fuzzy Logic
© Evolutionary Algorithms such as Genetic algorithm
Every paradigm of soft computing mentioned above

has its own strength. In order to build a computationally
intelligent system, we may integrate multiple techniques or methodol
ogies to take advantage of the strengths of each
of them. Such systems are called Hybrid soft computing systems
.
— Table 1.15.1 summarizes the soft computing methodologies
and their strengths.
Table 1.15.1 ; Soft computing constituents and their
strengths
1. Neural networks Has capability of learning and adaptation.

2. Fuzzy set theory Handles uncertainty and incorporates human-like
reasoning into the system.
‘
3. Evolutionary algorithms Has capability of finding optimum solution
to a problem.
— The seamless integration of these methodologie
s forms the base of soft computing.
— Neural networks have the capability of
recognizing patterns and adapting them
environments. selves to cope with changing
— The evolutionary algorithms such

as Genetic Algorithms are search
and Optimization techniques base
evolution that help us to optimize certain
parameters in a give d on biological
n problem.
— . Fuzzy logic incorporates human
knowledge and performs inference
and decision making
1.15.1 Introduction to Neu
ral Networks
= An Artificial Neural Network (ANN),
inspired by the biologi
human brain, ‘ Y i
lological nervous system basicall y tries to mimic the working ofa
“neurons work in parallel

to solve a specific proble
Py TechKnowletga m.
Publications ” ° ” °
W_Al&SC (MU-Sem.7- Comp) 1-29 Intro. to Artificial Intelligence(Al) and Soft Computing
An ANN learns by examples the way humans learn by their experiences.
ANN can be designed and configured for a specific application such as data classification, pattern reorganization, data
clustering etc.
Advantages of Neural Networks
Neural networks provide human like artificial intelligence.
A neural network learns and does not need to be reprogrammed.
A neural network can do a task that a linear program cannot do.
Parallel organization of neural networks permits solutions to problems where multiple constraints must be satisfied
simultaneously,
5. Because of its parallel nature, when an element of the neural network fails, it can continue without any problem.
Applications of Neural Networks
Neural networks have been successfully applied to a broad spectrum of data-intensive applications. Few
of them are
listed below.
(a) Forecasting
Neural network can be used very effectively in forecasting exchange rates, predicting stock
values, inflation and cash
forecasting, forecasting weather conditions etc. Researchers have proved that the forecastin
g accuracy of NN systems
tend to excel over that of the linear regression model.
(b) Image compression
— Digital images require a large amount of memory for storage. As a

result, the transmission of image from one
computer to another can be very expensive in terms of time and bandwidth
required.
— With the explosion of Internet, more sites are using images. Image compression is a technique that removes
some of the redundant information present in the image without
affecting its Perceptibility, thus, reducing the
storage size required to store the image.
— NN can be effectively used to compress the image. Several

NN techniques such as Kohonen’s self organizing
maps, Back propagation algorithm, Cellular neural network
etc. can be used for image compression.
(c) Industrial process control
— Neural networks have been applied successfully in industrial process

control of dynamic systems.
— Neural networks (especially multi layer perceptron
s) have been proved to be the best choice for
modelling non-
linear systems and implementing general — purpo
se non-linear controllers, due to their universal
approximation
capabilities. For example control and management of agriculture machi
nery.
(a) Optical character recognition
Well known application using image recognition is the Optical. Character Recognition (OCR)
available with the standard scanning software for the tools that. are
home computer.
Scansoft has had great success in combining NN with
a rule based s ystem for correctly recognizing
characters a nd words, to get a high level of accur both
acy. .
W TechKnowledga.
Puoticarinas
d
Ww Al&SC (MU-Sem. 7-Comp) 1-30 Intro. to Artificial Intelligence(Al) and Soft Computing: :
(e) Customer relationship management
Another popular application for NN is Customer Relationship Management (CRM).
Customer Relationship Management requires key information to be derived from raw data collected for each
individual customer. This can be achieved by building models using historical data information.
— Many companies are now using neural technology to help in their day to day business
Processes. They are doing this to achieve better performance, greater insight, faster development and increased
productivity.
:
— By using Neural Networks for data mining in the databases, patterns, however complex, can be
identified for the
different types of customers, thus giving valuable customer information to the company.
— Also, NN could be useful for important tasks related to CRM, such as forecasting
call centre loading, demand and
sales levels, monitoring and analyzing the market, validating, completing and
enhancing databases, clustering and
profiling client base etc.
— One example is the airline reservation system AMT, which could predict
sales of tickets in relation to destination,
time of year and ticket price.
(f) Medical science
— Medicine is the field that has always taken benefits

from the latest and advanced technologies.
— Artificial Neural Networks (ANN) is currently
the next promising area of interest in medical
science.
— It is believed that neural networks will have
extensive application to biomedical probl
ems in the next few years.
— ANN has already been successfully
applied in medical applications such
as diagnostic systems, bio chemical
analysis, disease detection, image analysis
and drug development.
1.15.2 Introduction to Fuzzy Logic
— Fuzzy logic is an approach to

computing based on "degrees
of truth" rather than the usua
Boolean logic on which the modern com l "true or false" (1 or 0)
puter is based.
present in complex real world problems.
~ Using fuzzy logic, it is now

possible to include vague
human assessment in Com
puting problems
— _ Also, it provides an
effective means for conflict
resolution of multiple crit
eria and better assessmen
—. Fuzzy logic can be used t of options .
in the development of var
ious applications such as
applications, identification pattern recognition, opt
and any intelligent system imization, a control
for decision making.
-
A. Advantages of Fuzzy Log

ic Controllers
— — simplicity and flexibil
ity
— canhandle problems wit
h imprecise and incomp
lete data
W TechKnowledga ° - =
Publications i : ij
ence(Al) and Soft Computing
1-311
1-3 itro. to Artificiat Intellig
Intr
| v Al&SC (MU-Se m.. 77-Comp)
sem
of arbitrary complexity
— can model nonlinear functions
_ cheaper to develop, terms.

omizable in natural language
ra wide r rang e of oper atin g conditions, more readily cust
— cove
p. Applications of Fuzzy Logic

ity to generate precise
ions whe re hum an like decision making with an abil
applicat
— Fuzzy logic can be used in
ate information is required.
solutions from certain or approxim
as washing machine,
used in desi gn of cont roll ers for home appliances such
ely
— Fuzzy logic has been extensiv
etc.
vacuum cleaner, air conditioner
anti-skid braking systems,
for othe r appl icat ions such as facial pattern recognition,
— Fuzzy logic can also be used
helicopters.
subway systems and unmanned
transmission systems, control of
systems for multi objective
of fuzz y logic is dev elo pment of knowledge-based
- Another application area uct pricing or project risk
ems, weat her fore cast ing systems, models for new prod
optimization of power syst
treatment plans, and stack trading.
assessment, medical diagnosis and
image processing,
ssful ly used in nume rous fields such as control systems engineering,
— Fuzzy logic has been succe
and optimization.
on, robotics, consumer electronics
power engineering, industrial automati
orithms
4.15.3 Introduction to Genetic Alg
ry ideas of natural selection and
ithms are adapt ive heuri stic searc h algorithms based on the evolutiona
— Genetic Algor
genetics.
which otherwise would take a
near-optimal solutions to difficult problems
— GAs are often used to find optimal or
lifetime to solve.
solutions.
of candidate designs and find optimum
— GAcan efficiently explore a large space
Computation.
- GAs area subset of an Evolutionary
Michigan,
students and colleagues at the University of
- GAs were developed by John Holland and his
of possible solutions to the given problem.
- InGaAs, we select the initial pool ora population
recombination and mutation which in turn produce new
- These solutions then undergo various GA operations like
children.
— The process is repeated over various generations.

the fitter individuals are given a higher chance to
~ Each individual or candidate solution is assigned a fitness value and
an Theory of “Survival of the Fittest”
mate and yield more “fitter” individuals. This is in line with the Darwini
generations, till it reaches a stopping criterion
— Thus GA keeps “evolving” better individuals or solutions over
local search
nature, but they perform much better than random
— Genetic Algorithms are sufficiently randomized in
A. Advantages of Genetic Algorithms
dge of complex mathematics
— GAsare easy to understand since they do not demand the knowle
— Does not require any derivative information
TachKnowledge
Pootications
— Good for noisy environment.
— Easy to discover global optimum.
— They can solve multimodal, non differentiable, non continuous or even NP-complete problems.
— GAsare inherently parallel and distributed.
— Provides a list of “good” solutions and not just a single solution.

— Flexible in forming building blocks for hybrid applications.
—- Have substantial history and range of use.
— Useful when the search space is very large and there are a large number of parameters involved.
Applications of Genetic Algorithms
1. Automotive design : Genetic algorithms can be used to design composite materials and aerodynamic shapes
for race cars to provide faster, lighter, more fuel efficient and safer vehicles for all the things we use vehicles for.
2. Engineering design : GA are most commonly used to optimize the structural and operational design of
buildings, factories, machines, etc. GAs are used for optimizing the design of robot gripping
arms, satellite booms,
building trusses turbines, , flywheels or any other computer-aided engineering
design application. 3.
3. Robotics : GAs have found applications that span the range of architec
tures for intelligent robotics. GAs can be
used to design the entirely new types of robots that can
perform multiple tasks and have more general
application.
Q. 1 Define artificial intelligence.
Q. 2 Write a short note on: Applications of artificial intelligence.
Q. 3 Explain the various artificial intelligence problems

and artificial intelligence techniques.
Q. 4 What is artificial intelligence ?
Q. 5 What are the components of Al?
Q. 6 What are the various Al techniques ?
Q. 7 Explain various applications of Artificial Intelligence,
Q. 8 Explain PEAS representation with example.
Q. 9 Define agent and give classification of

agents.
Q. 10 What is intelligent agent 7
Q. 11 Write a short note on: Rational agent

.
Q. 12 Write a short note on : Structure
of intelligent agents.
Q. 13 Give types of agents.
Ww TechKnowledga
Publications
=
Al&SC (MU-Sem. 7-Comp) 1-33 Intro. to Artificial Intelligence(Al) and Soft Computing
Q. 14 What are various agent environments ? Give PEAS representation for an agent.
Q. 15 Define in your own words, the following terms :

1. Agent 2. Agent function
8. Agent program 4. Autonomy
Q. 16 Explain various types of intelligent agents, state limitations of each and how it is overcome in other type of agent.
Q.17 What do you mean by PEAS? Explain properties of task environment.
Q. 18 Explain detail architecture of goal based agent.
Q. 19 Explain Simple reflex agent architecture.
Q. 20 Explain learning agent architecture.
Q. 21 What are the constituents of Soft Computing ? Explain each in brief.
Q. 22 Differentiate between hard computing and soft computing.
Q. 23 Explain the basics of genetic algorithm along with its applications.
Q.24 What are the applications of neural networks ?
o00
Problem Solving
seit Bataan Sosy

€
2.1 Problem Solving Agent, Formulating Problems, Example Problems
2.2 Uninformed Search Methods : Depth Limited Search, Depth First Iterative Deepenin
g (DFID), Informed Search |
’ Method : A* Search
, |
2.3 Optimization Problems : Hill climbing Search, Simulated anneali
ng, Genetic algorithm
my
Search is an indivisible part of intelligence. An intelligent agent is the one

who can search and select the Most
appropriate action in the given situation, among the available set of actions.
When we play any game like chess, cards, tic. |
tac-toe, etc.; we know that we have multiple options for next move,
but the intelligent one who searches for the correct :
_ Move will definitely win the game. In case of travelling salesma
n problem, medical diagnosis system or any expert system;
all they required to do is to carry out search which will produce
the optimal path, the shortest path with minimum cost and
_
efforts. Hence, this chapter focuses on the searching
techniques used in Al applications. Those are known
as un-informed |
and informed search techniques.
‘
2.1 Solving Problems by Searching
Now let us see how searching play a vital role in

solving Al problems. Given a problem, we can genera
te all the
possible states it can have in real time, including
start state and end state. To generate solution
for the same is
nothing but searching a path from start state to end state.
;
— Problem solving agent is the one who finds the
goal state from start state in optimal way by follow
ing the shortest
path, thereby saving the memory and time. It’s suppo
sed to maximize its performance by fulfilling all the
performance
measures.
— Searching techniques can be used in game playing like Tic-Ta
c-Toe or navigation problems like Travelling Salesman —
Problem.
— _ First, we will understand the representation of given proble
m so that appropriate searchin
g techniques can be applied
to solve the problem.
2.2 Formulating Problems
- Given a goal to achieve; problem formulation is the Proces

s of deciding what states to be considered and what actions
1
to be taken to achieve the goal. This is the first step to
be taken by any problem solving agent.
s |
— State space : The state space of a problem is the set of
all states reachable fr om the initial state by executing any”
sequence of actions. State is representation of all possible outcomes.
4
. :
- The state space specifies the relation among various problem states thereby, forming a directed network 4
or graph in ;‘
which the nodes are states and the links between nodes represent actions.
:
a
Al&SC (MU-Sem. 7-Comp) 2-2 Problem Solvin
_ State Space Search : Searching in a given space of states pertaining to a problem under consideration is called a state
space search.
- Path: Apath is a sequence of states connected by a sequence of actions, in a given state space.
2.2.1 Components of Problems Formulation

LOREAL ESAs)
Problem can be defined formally using five components as follows :

1. Initial state 2. Actions
Successor function 4. Goal test
Path cost
1. Initial state : The initial state is the one in which the agent starts in.
2. Actions : It is the set of actions that can be executed or applicable in all possible states. A description of what each
action does; the formal name for this is the transition model.
3. Successor function : It is a function that returns a state on executing an action on the current state.
4. Goal test : It is a test to determine whether the current state is a goal state. In some problems the goal test can be
carried out just by comparing current state with the defined goal state, called as explicit goal test. Whereas, in some
of the problems, state cannot be defined explicitly but needs to be generated by carrying out some computations, it is
called as implicit goal test.
state
For example : In Tic-Tac-Toe game making diagonal or vertical or horizontal combination declares the winning
which can be compared explicitly; but in the case of chess game, the goal state cannot be predefined but it’s a
scenario called as “Checkmate”, which has to be evaluated implicitly.
5. Path cost : It is simply the cost associated with each step to be taken to reach to the goal state. To determine the cost
to reach to each state, there is a cost function, which is chosen by the problem solving agent.
and path
Problem solution : A well-defined problem with specification of initial state, goat test, successor function,
the goal state.
cost. It can be represented as a data structure and used to implement a program which can search for
initial state
A solution to a problem is a sequence of actions chosen by the problem solving agent that'leads from the
to a goal state. Solution quality is measured by the path cost function.
solutions.
Optimal solution : An optimal solution is the solution with least path cost among all
with the goal to be
A general sequence fallowed by a simple problem solving agent is, first it formulates the problem
achieved, then it searches for a sequence of actions that would solve the problem, and then executes the actions one at a
time.
2.2.2 Example Problems

MU - Dec. 12
1. Example of 8-puzzle problem

— Fig. 2.2.1 depicts a typical scenario of 8-puzzle problem. It has a 3x3 board with tiles having 1 through 8 numbers on it.
There is a blank tile which can be moved forward, backward, to left and to right. The aim is to arrange all the tiles in
the goal state form by moving the blank tile minimum number of times.
7|/6]5 7) 8) -
Initial State Goal State

Fig. 2.2.1 : A scenario of 8-Puzzle Problem
This problem can be formulate
d as follaws :
States : States can represented by a 3 x 3 matr
ix data structure with blank denoted by 0.
1. Initial state : {{1, 2, 3},{4, 8,
O},{7, 6, 5}}
2. Actions : The blank space can move in Left, Righ
t, Up and Down directions specifying the actions.
3. Successor function : If we apply “Down” operator to
the start state in Fig. 2.2.1, the resulting state has the
the blank switching 5 and
their positions.
4. Goal test : {{1, 2, 3),{4, 5, 6},{7,
8, O}
5. Path cost : Number of steps to reac
h to the final state.
Solution:
{{1, 2, 3},4, 8, 01,47, 6, S)}-> {{1, 2, 31,4, 8, 5},{7, 6, Of}

((1, 2, 3),
{4, 8, 5},{7, 0, 6}} 3{{1, 2, 3}{4, 0, 5},{7, 8, 6}}-9
{{2, 2, 3},{4, 5, O}{7, 8, 6}} >{{1, 2, 3},{4, 5, 6},{7,
Path cost = 5 steps 8, OFF
2. Example of missionaries and cann

ibals problem
The problem statement as discusse
d in the previous section. Let's form
ulate the problem first.
— States : In this problem, state can
be data structure having triplet (i, L
k) representing the number of
cannibals, and canoes on the left bank missionaries,
of the river respectively.
1. Initial state : It is (3, 3, 1), as all missiona
ries, cannibals and Canoes
are on the left bank of the river,
2. Actions : Take x number of missionar
ies and y number of cannibals
3. Successor function : If we take one missionary
, one cannibal the other side of the river will have two
and two cannibals left. missionaries
4. Goal test : Reached state (0, 0, 0)
5.
Path cost : Number of crossings to
attain the goal state,
Solution:
The sequence of actions within the path

:
(3,3,1) > (2,2,0)->(3,2,1) 3(3,0,0) >(3
,1,1) >(1,1,0) >(2,2,1) ~(0,2,0) 3(0
,3,1) (0,1,0) > (0,2,1) 3(0,0 0)
Cost = 11 crossings
3. Vacuum-cleaner problem
States : In vacuum cleaner prob
lem, state can be repr
esented as [<block>,
in one of the two blocks which can clean] or [<block>, dir
be eith er clean or dirty. ty]. The agent can be
Hence there are tot
18 |
1. Initial State : Any state can bec on
a sta tes in the vacuum cleaner
sidered as initial sta world. |
te. For example
2.__ Actions : The possible actions f TA, dirty]
:
TachKnowledga ft, right, absorb,
"PuURLICaALIons idle.
td
Problem Solving
3.
(lee I ath R
BP | AS | RS
| RS
(JA) 7] [adh
V7 LJ$s
$s
Fig. 2.2.2 : The state space for vacuum world

4. Goal Test : The aim of the vacuum cleaner is to clean both the blocks. Hence the goal test if [A, Clean]
and [B, Clean].
5. Path Cost : Assuming that each action/ step costs 1 unit cost. The path cost is number of actions/
steps taken.
. Example of real time problems
There are varieties of real time problems that can be formulated and solved by searching. Robot Navigation,
Rout
Finding Problem, Traveling Salesman Problem (TSP), VLSI design problem, Automatic Assembly Sequencing,
etc. are
few to name.
There are number of applications for route finding algorithms. Web sites, car navigation systems that provide
driving
directions, routing video streams in computer networks, military operations planning, and airline travel-plann
ing
systems are few to name. All these systems involve detailed and complex specifications.
For now, let us consider a problem to be solved by a travel planning web site; the airline travel problem.
State : State is represented by airport location and current date and time. In order to calculate the path cost state may
also record more information about previous segments of flights, their fare bases and their status as domestic or
international.
1. Initial state : This is specified by the user's query, stating initial location, date and time.
2. Actions : Take any flight from the current location, select seat and class, leaving after the current time, leaving
enough time for within airport transfer if needed.
3. Successor function : After taking the action i.e. selecting fight, location, date, time; what is the next location date
and time reached is denoted by the successor function. The location reached is considered as the current location
and the flight's arrival time as the current time.
4. Goal test : Is the current location the destination location?
Path cost : In this case path cost is a function of monetary cost, waiting time, flight time, customs and
immigration procedures, seat quality, time of day, type of airplane, frequent-flyer mileage awards and so on.
Ep 3 Tectiinculedys
Publications
2.3 Measuring Performance of Problem Solving Algorithm / Agent
There are variety of problem solving methods and algorithms available in Al. Before studying any of oe aeons |
in detail, let’s consider the criteria to judge the efficiency of those algorithms. The performance of all these algorithms cap,
be evaluated on the basis of following factors.
1. Completeness : If the algorithm is able to produce the solution if one exists then it satisfies completeness criteria.
2. Optimality : If the solution produced is the minimum cost solution, the algorithm is said to be optimal. . .
3. Time complexity : It depends on the time taken to generate the solution. It is the number of nodes generated during
the search.
Space complexity : Memory required to store the generated nodes while performing the search.
Complexity of algorithms is expressed in terms of three quantities as follows :
1. Bb: Called as branching factor representing maximum number of successors a node can have in the search tree.
2. di: Stands for depth of the shallowest goal node.
3. m:Itis the maximum depth of any path in the search tree.
2.4 _Node Representation in Search Tree

In order to carry out search, first we need to build the search tree. The nodes are the various possible
states in the
State space.
The connectors are the indicators of which all states are directly reachable from current
state, based on the successor.
function.
Thus the parent child relation is build and the search tree can be generated.
Fig. 2.4.1 shows the representation ofa
tree node as a data structure in 8-puzzle problem.
Parent node
State
=) Child node
Fig. 2.4.1 : Node representation of state in

searching
~ Node is the data structure from which the searc
h tree is constructed. Each node has a Parent,
nodes directly reachable from that node. a state, and children
‘
- For each node of the tree, we can have follo
wing structure components :
1. State / Value : The state in the state space
to which the node corresponds or value
2. Parent node : The node in assigned to the node.
the search tree that generated this node.
3. Number of children: Indicating numb
er of actions that can be taken
4. to generate next states (children
Path cost : The cost of the path from nodes).
the initial state to the node.
2.5 Uninformed Search
— _ Whyis it called uninformed search? What i .

Uy Teckinouleasa f S$ not been infor
med about the search?
SF AlBSC(MU-Sem.7-Comp) 8G Problem Solving
state along with
The term “uninformed” means they have only information about what is the start state and the end
the problem definition.
— These techniques can generate successor states and can distinguish a goal state from a non-goal state.
- All these search techniques are distinguished by the order in which nodes are expanded.
- The uninformed search techniques also called as “blind search”.
2.6 Depth First Search (DFS)_
2.6.1 Concept
- Indepth-first search, the search tree is expanded depth wise; i.e. the deepest node in the current branch of the search
tree as expanded. As the leaf node is reached, the search backtracks to previous node. The progress of the search is
illustrated in Fig.2.6.1. :
- The explored nodes are shown in light gray. Explored nodes with no descendants in the fringe are removed from
memory. Nodes at depth three have no successors and M is the only goal node.
Process
7+®
A)
()
OQ) ©
GE) @
Fig. 2.6.1 : Working of Depth first search on a binary tree
AI&SC (MU-Sem. 7-Comp) 2-7 Problem Solving
2.6.2 Implementation 2
DFS uses a LIFO fringe i.e. stack. The most recently generated node, which is on the top in the fringe, is
chosen first for
expansion. As the node is expanded, it is dropped from the fringe and its successors are added. So when there are no more :
successors to add to the fringe, the search “back tracks” to the next deepest node that is still unexplored.DFS can be
implemented in two ways, recursive and non-recursive. Following is the algorithm for the same
2.6.3 Algorithm
(a) Non recursive implementation of DFS
1. Push the root node on a stack
2. while (stack is not empty)
(a) Pop anode from the stack;
(i) if node is a goal node then return success;
(ii) push all children of node onto the stack;
3. return failure ,
(b) Recursive implementation of DFS
DFS(c) :
1. If node is a goal, return success;
2. for each child c of node
(a) if DFS(c) is successful,
(i) return success
3. return failure;
2.6.4 Performance Evaluation
Completeness : Complete, if m is finite.

Optimality : No, as it cannot guarantee the shallowest solution.
Time Complexity: A depthfirst search, may generate all of the O(b™) nodes in the search tree, where m is the
maximum depth of any node; this can be much greater than the size of the state space.
— Space Complexity: For a search tree with branching factor b and maximum depth m, depth first search requires
storage of ronm, ab ") nodes, as at a.time only the branch, which is getting explored, will reside! in memory.
ses pe
MU - May 16, 10 Marks
Problem Solving
Fig. P. 2.6.1
2.7 _ Breadth First Search (BFS)
2.7.1 Concept
- Asthe name suggests, in breadth-first search technique, the tree is expanded breadth wise.
- The root node is expanded first, then all the successors of the root node are expanded, then their successors, and so
on.
— In turn, all the nodes at a particular depth in the search tree are expanded first and then the search will proceed for
the next level node expansion.
- Thus, the shallowest unexpanded node will be chosen for expansion. The search process of BFS is illustrated in
Fig. 2.7.1.
2.7.2 Process
=>O ® ® ©
® ©@
=>O ® =O =O
Fig. 2.7.1 : Working of BFS on binary tree
2.7.3 Implementation
— In BFS we use a FIFO queue for the fringe. Because of which the newly inserted nodes in the fringe will automatically
be placed after their parents.
— Thus, the children nodes, which are deeper than their parents, go to the back of the queue, and old nodes, which are
shallower, get expanded first. Following is the algorithm for the same.
2.7.4 Algorithm
1. Put the root node on a queue
2. while (queue is not empty)
(a) remove a node from the queue
(i) if (node is a goal node) return success;
(ii) put all children of node onto the queue;
3. return failure;
ree &SC-(MU-S8m. 7-COMP) es
Completeness : It is complete, provided the shallowest goal node is at some finite depth.
— Optimality : it is optimal, as it always finds the shallowest solution.
~ Time complexity : O(b°), number of nodes in the fringe.
— Space complexity : O(b*), total number of nodes explored.
2.8 Uniform Cost Search (UCS) _ _ _

2.8.1 Concept
Uniform cost search is a breadth first search with all paths having same cost. To make it work
in real time conditions
we can have a simple extension to the basic implementation of BFS. This results
in an algorithm that is optimal with
any path cost.
— _ In BFS as we always expand the shallowest node first; but in uniform
cost search, instead of expanding the shallowest
node, the node with the lowest path cost will be expanded first. The implementation
details are as follow.
— Uniform cost search can be achievedby implementing the fringe as a priority queue ordered by
path cost. The
algorithm shown below is almost same as BFS; except for the
use of a priority queue and the addition of an extra
check in case a shorter path to any node is discovered.
— The algorithm takes care of nodes which are inserted in
the fringe for exploration, by using a data structure having
priority queue and hash table.
The priority queue used here contains total cost from root
to the node. Uniform cost search gives the minimum path
cost the maximum priority. The algorithm using this priority
queue is the following.
2.8.3 Algorithm
Insert the root node into the queue.
— While the queue is not empty : |

(i) Dequeue the maximum priority node from the queue.
(If priorities are same, alphabetically smaller node is chosen)
(ii) If the node is the goal node, print the path and exit.
Else
— Insert all the children of the dequeued node, with
their total costs as priority.
— The algorithm returns the best cost path
which is encountered first and will never
80 for other possible paths. The
solution path is optimal in terms of cost.
— As the priority queue is maintained
on the basis of the total path cost
of node, the algorithm never expa
which hasa cost greater than the cost
of the shortest nds a node |
path in the tree,
- 7
The nodes in the priority queue have
almost the same costs at a given time,
Search”. and thus the name “Uniform Cost |
f
wer Tarh¥nmnladaa
, Al&SC (MU-Sem. 7-Comp) - 2-10 Problem Solving
aS tw BAO
Completeness : Completeness is guaranteed provided the cost

of every step exceeds some small positive constant.
- Optimality : It produces optimal solution as
nodes are expanded in order of their path cost.
Time complexity : Uniform-cost search considers path costs rather than depths; so its
complexity is does not merely
depend s on b and d. Hence we Consider C* be the cost of the optimal solution, and
assume that every action costs at
least €. Then the algorithm's worst-case time and space complexity is O(b°’*)
, which can be much greater than bd.
Tey» c* oo .
~ Space complexity : O(b“’*), indicating number of node in memory at execution time.
2.9 Depth Limited Search (DLS)
2.9.1 Concept
In order to avoid the infinite loop condition arising in

DES, in depth limited search technique, depth-first search
is
carried out with a predetermined depth limit.
The nodes with the specified depth limit are treated as if they don’t have any successor
s. The depth limit solves the
infinite-path problem. ,
But as the search is carried out only till certain depth in the search tree, it introduces problem
of incompleteness.
Depth-first search can be viewed as a special case of depth-limited search with depth limit equal
to the depth of the
tree. The process of DLS is depicted in Fig. 2.9.1.
2.9.2 Process
lf depth limit is fixed to 2, DLS carries out depth first search till second level in the search tree.
Q @ @ Q
© OO ®©§ © QO GQ 9
©
Fig. 2.9.1 : DIS working with depth limit
— Asincase of DFS in DLS we can use the same fringe implemented as queue.
Additionally the level of each node needs to be calculated to check whether it is within the specified depth limit.
Depth-limited search can terminate with two conditions :
1. If the solution is found.
2. If there is no solution within given depth limit.
2.9.4 Algorithm
Determine the start node andthe search depth.

Check if the current node is the goal node
Set TechKnnwledye
~Pebjicatinns 9)
AI&SC (MU-Sem. 7- Comp)
If not: Do nothing
If yes : return
Check if the current node is with
in the Specified search depth
If not: Do nothing i
If yes : Expand the node and

save all of its successors in a
stack.
Call DLS recursively for all nod
es of the stack and go back
to Step 2.
2.9.5 Pseudo Code
.
Completeness : Its incomp

lete if Shallowest goal is
beyond the depth limit.
Optimality : Non optimal,
as the depth chosen can
be greater than d.
— Time complexity : Same
as DFS, O (b'), where | is the
specified depth limit.
— Space complexity : Same as
DFS, O(b), where | is the spec
ified depth limit.
2.10 _Depth First Iterative
Deepening (DFID)
2.10.1 Concept
-
— It eventually turns out to be the bread

th -first searc
before going on to the next layer.
— It does this by gradually incr

easing the depth limit -fir
st 0, then 1,
then 2, and-
guarantees the op
timal solution. It
erative deepenin & Co mb .
ines the
so on
“until a goal is found; and
— The search process is depicted in Fig. 2.10.1.
benefits of dep
th thus
‘first and breadadth-f
th- irst search.
wW TethKnowledge
Publicatians
WF _Al&SC (MU-Sem. 7-Com p) 2-12
Problem Solving =
eee
Limit =0 +@)
ee oo on oe
Limit = 2 >) :
o © © ©
O ©
(A)
Fig.2.10.1 : Search process In DFID (contd...)
Tech Knowledge
Peblicarions
Fig. 2.10.1 : Search process in DFID
— Fig. 2.10.1 shows four iterations of ona binary search tree, where the solution is found on the fourth iteration.
2.10.2 Process
function iterative : Depending search (problem) returns a solution, or failure
for depth = 0 to odo
result <—- Depth — Limited — Search (problem, depth)
if result + cutoff then return result
Fig. 2.10.1 the iterative depending search algorithm, which repeatedly applies depth limited search with increasing
limits. It terminates when a solution is found or if the depth limited search returns failure, meaning that no solution exists.
It has exactly the same implementation as that of DLS. Additionally, iterations are required to increment the depth
limit by one in every recursive call of DLS.
2.10.4 Algorithm
— Initialize depth limit to zero.
— Repeat Until the goal node is found.
(a) Call Depth limited search with new depth limit.
(b) Increment depth limit to next level.
2.10.5 Pseudo Code
DFID()
Bee
: limit = 0;
found = false;
while (not found)
found = DLS(root, limit, 0);

limit = limit + 1;
oh
— Completeness : DFID is complete when the branching factor b is finite.
- Optimality : It is optimal when the path cost is a non-decreasing function of the depth of the node.
- Time complexity : .
o Do you think in DFID there is a lot of wastage of time and memory in regenerating the same set of nodes again
and again ?
o It may appear to be waste of memory and time, but it’s not so. The reason is that, in a search tree with almost
same branching factor at each level, most of the nodes are in the bottom level which are explores very few times
as compared to those on upper level.
o The nodes on the bottom level that is level ‘d’ are generated only once, those on the next to bottom level are
generated twice, and so on, up to the children of the root, which are generated d times. Hence the time
complexity is O(b’).
- Space somplenity Memory requirements of DFID are modest, i.e. O(b* ).
8 performance evaluai uite satisfactory on all the four paramete

arch method when the search space is large and the depth of the solution isi not knov
2.11 Bidirectional Search
2.11.1 Concept
In bidirectional search, two simultaneous searches are run. One search starts from the initial state, called forward
search and the other starts from the goal state, called backward search. The search process terminates when the searches
meet at a common node of the search tree. Fig. 2.11.1 shows the general search process in bidirectional search.
2.11.2 Process
Fig. 2.11.1 : Search process in bidirectional search
— In Bidirectional search instead of checking for goal node, one need to check whether the fringes of the two searches
intersect; as they do, a solution has been found.
SS (uC Sain.
7 -Comp)
— When each node is generated or selected for expansion, the check can be done. It can be implemented: with a hash
table, to guarantee constant time.
— For example, consider a problem which has solution at depth d= 6. If we run breadth first search in each direction,
then in the worst case the two searches meet when they have generated all of the nodes at depth 3. If b= 10.
— This requires a total of 2,220 node generations, as compared with 1,111,110 for a standard breadth-first search.
- Completeness : Yes, if branching factor b is finite and both directions use breadth first
search.
— Optimality : Yes, if all costs are identical and both directions use breadth first
search.
— Time complexity : Time complexity of bidirectional search using breadth-first searches
in both directions is O(b”),
— Space complexity: As at least one of the two fringes need to kept in memory
to check for the common fede, the
space complexity is o(b y,
2.11.5 Pros of Bidirectional Search
— Itis much more efficient.
— Reduces space and time requirements as, we perform two b? searches

, instead of one b* search.
- Example: ,
Oo Suppose b = 10, d = 6. Breadth first search will examine10° = 1, 000,
000 nodes.
© Bidirectional search will examine2 x 10° = 2,000 nodes.
— One can combine different search strategies in different directio
ns to avail better performance.
2.11.6 Cons of Bidirectional Search
— The search requires generating predecessors of states.
— Overhead of checking whether each new node appears in

the other search is involved.
— For large d, is still impractical!
— For two bi-directional breadth-first searches, with

branching factor b and depth of the solution d
requirement of pb?” for each
we have memory
search. |
2.12 Comparing Different Techniques
CURSE
Table 2.12.1 depicts the comparison of all
uninformed search techni ques basis
on their Performance evalua
know, the algorithms are evaluated on four tion. As We
criteria viz, completeness,
optimality, time complexity and space complexity.
The notations used are as follows :
- b: Branching factor
- d: Depth of the shallowest solution

— m:Maximum depth of the search tree
- {: Depth limit
Table 2.12.1 : Comparison of tree-search strategies basis on performance Evaluation
Completeness Yes Yes No No Yes Yes

Optimality Yes Yes No No Yes Yes
Time Complexity O(b*) O(b°") O(b") - o(b) ob’) o(b”)
Space Complexity O(b*) o(b°"*) O(b") o(b) o(b’) o(b%”)

2.12.1 Difference between Unidirectional
and Bidirectional Search
idirectio
1. | These methods use search tree, start node and goal node as input | These methods have additional information
for starting search. -about the search tree nodes, along with the
start and goal node.
2. | They use only the information from the problem definition. They incorporate additional measure of a
potential of a specific state to reach the goal.
3. | Sometimes these methods use past explorations, e.g. cost of the | All these methods use a potential of a state
path generated so far. (node) to reach a goal is measured through
heuristic function. :
4. | All unidirectional techniques are based on the pattern of | All bidirectional search techniques totally
exploration of nodes in the search tree. depend on the evaluated value of each node
generated by heuristic function.
5. | In real time problems uninformed search techniques can be | In real time problems informed search
costly with respect to time and space. techniques are cost effective with respect to
time and space.
6, | Comparatively more number of nodes will be explored in these | As compared to uninformed techniques less
methods. number of nodes are explored in this case.
7. | Example : Breadth First Search Example : Hill Climbing search

Depth First search Best First search, A* Search, IDA* search, |
Uniform Cost search, SMA* search
Depth Limited search,

Iterative Deepening DFS
2.12.2 Difference between BFS and DFS
BFS Stands for “Breadth First Search”. DFS stands for “Depth First Search”.
. BES traverses the tree level wise.i.e. each node near DFS traverses tree depth wise. i.e. nodes in particular
to root will be visited first. The nodes are explored branch are visited till the leaf node and then search
left to right. continues branch by branch from left to right in the
tree.
TechKnowledgs
Pubiteations
Al&SC (MU-Sem. 7-Comp)
3. Breadth First Search is implemented using | Depth First Search is implemented using Stack which
queue which is FIFO list. is LIFO list. _ a
— OO mV;
4. This is a single step algorithm, wherein the visited | This is two step algorithm. In first stage, the visiteg
vertices are removed from the queue and then | vertices are pushed onto the stack and later on when
displayed at once. ; ‘) there is no vertex further to visit those are popped
out.
5. BFS requires more memory compare to DFS. DFS require less memory compare to BFS.
6. Applications of BFS : Applications of DFS :

To find Shortest path | Useful in Cycle detection
Single Source & All pairs shortest paths In Connectivity testing
In Spanning tree Finding a path between V and Win the graph.
In Connectivity Useful in finding spanning trees and forest.
7. BFS always provides the shallowest path solution. DFS does not guarantee the shallowest path solution.
8. No backtracking is required in BFS. Backtracking is implemented in DFS.
9. BFS is optimal and complete if branching factor is | DFS is neither complete nor optimal even in case of
finite. finite branching factor.
10. BFS can never get trapped into infinite loops. DFS generally gets trapped into infinite loops, as |
search trees are dense.
11. Example : Example :
A A
J\ > /\
B C
B Cc
/ /\ / /N
DE F D E—E F
A,B, C,D,E,F A,B,
D, C, E, F
2.13 Informed Search Techniques
Informed searching techniques is a further extension of basic un-informed search techniques. The main idea is to
generate additional information about the search state space usin
g the knowledge of problem domain, so that the.
search becomes more intelligent and efficient. The evalua
tion func tion is developed for each state, which quantities
the desirability of expanding that state in order to reach
the goal.
All the strategies use this evaluation function in
order to select the nex t state under consideration,
“Informed hence the name)
Search”. These techniques are very much efficient with respect to time and space requirements as
compared to uninformed search techniques.
er TechKnowledga : 7
2.14 Heuristic Function
- A heuristic ur ctior is ane

1 eV: alua iI on unction,
i to which
i the search
r state is give nas p t and
i inpu it
i generates the tangible
i
- It maps the problem sta te de seri pti ‘O| n to m easures

r O f desirability, usual ly rep ese ted as number weigh
of a I euristic functio 1ata given ode in the Ss earch process gives a good estimate o tk at node 6
— individu srineatin
enanonmay
It evaluates
One nr an determines how much promising the state is. Heuristic functions are the
\ ing additional knowledge of the problem states t o the search i . i Fig.
_ the general representation of heuristic function. “wee ee

Gs to!
hae
State / Node 'n'
Eee! ‘Heuristic Function of node n, 'e
Value
= h(n) a
function
Fig. 2.14.1 : General representation of Heuristic
red to
the goal node or number of hopes requi
The representation may be the approximate cost of the path from
reach to the goal node, etc. cheapest

a node n is, h(n) = estimated cost of the
are considering in this sy \labus, for
The heuristic function that we
n to a goal state.
path from the state at node
far can be a simple heuristic
ng Sal esm an Pro ble m, the sum of the distances traveled so
— Example : For the Travelli
function. Maximization Function or
of two ty pes depe ndin g on the problem domain. It can be a
be
Heuristic function can
o f the path cost.
Minimization function le; in case of minimization
rist ic, gre ate r the cost of the node, better is the node whi
of heu l applicability as well as domain
— In maximization types a re heuristics of every genera
the node. There
the cost, better is
heuristic, lower is pose heuristics.
strategies are gen eral pur on, even though there is no
specific. The search alwa ys lea d to faster and better soluti
ti c will
in general, a heuris
- It is believed that ong direction in the sea
rch tree.
in the wr
it will never lead
guarantee tha t of search.
pla ys a vita l role in perfor mance among all that are
- Design of heuristic ide the sear ch proc ess in the most profitable path
tion is to gui r a path is good or bad.
ah euristic func ly good estimate of whethe
- As the purpose of c functions ca n pr o vide s a fair
signe’ d heuristi uristic function would be
more than the effort
available; a well de ng th e va lu e of a he
stof com puti luating a heuristic function
ve r in ma ny problems, the co re is a tr ad e- off betw een the cost of eva
~ Ho we Hence generally
the
arch process.
saved in the se nc tion provides. the
search that the fu
how
word of caution. See
and the savings in n heuristi c fu nc ti on def i nitions? Here is the
ink of your ow
th
— So, are you ready to of searching
ple tely alters th e scenario
function definition impact. how design of he uri sti c fun cti on com
e the examples dem onstrate

. e
- Following ar
process. SH TechKnewl edgt
ativas
Pubiic
2-49 : Pribiem ov
W Alasc (MU-Sem. 7-Comp)
2.14.1 Example of 8-puzzle Problem
ata cdoae Naa

- Remember 8-puzzle problem? Can we estimate the number of steps required to solve an 8-puzzle from a given sate
What about designing a heuristic function for it?
715 )4 om 1 |2
5 6 3 °)4]5
8 {3 |1 6 |7 |8
Start state Goal State
Fig. 2.14.2 : A scenario of 8-puzzle problem
- Two simple heuristic functions are:
© h, =the number of misplaced tiles. This is also known as the Hamming Distance. In the Fig. 2.14.2 example, the
start state has h, = 8. Clearly, h, is an atceptable heuristic because any tile that is out of place will have to be
moved at least once, quite logical. Isn’t it? a
© h,=the sum of the distances of the tiles from their goal positions. Because tiles cannot be moved diagonally, the
distance counted is the sum of horizontal and vertical distances. This is also known as the Manhattan Distance. In
the Fig. 2.14.2, the start state hash, =3+1+2+2+ 2+3+3+42= 18. Clearly, h, is also an admissible heuristic.
because any move can, at best, move one tile one step closer to the goal.
As expected, neither heuristic overestimates the true number of moves required to solve the puzzle, which is
26 (h,+ h,). Additionally, it is easy to see from the definitions of the heuristic functions that for any given state, h, will
always be greater than or equal to h,. Thus, we can say that h, dominates h,.
2.14.2 Example of Block World Problem
Start Goal
Fig. 2.14.3': Block problem
— Fig. 2.14.3 depicts a block problem world, where the A, B, C,D letter bricks are piled u
P On one another and required
to be arranged as shown in goal state, by moving one brick at a time. As shown, th
€ goal state with the particular
arrangement of blocks need to be attain from the given start state. Now it’ s time to scratch your head and
define a
heuristic function that will distinguish start state from goal state. Confused??
- Let’s design a function which assign $+ 1 for the brick at right position and —1 for
the one which i
Consider Fig. 2.14.4 is at wrong positionition. «|
(a) Local heuristic :

— +1 for each block that is resting on the thing it is supposed to be resting on.
- - 1 for each block that is resting on a wrong thing.
GF Technamletg
Publications
as M
y
Start | Goal
Blocks world
Fig. 2.14.4 : Definition
of heuristic function “hy”
AR flee
Fig. 2.14.5 : State evaluations using Heuristic function “hy”
Fig. 2.14.5 shows the heuristic values generated by heuristic function “h,” for various different states in the state
space. Please observe that, this heuristic is generating same value for different states.
promising
Due to this kind of heuristic the search may end up in limitless iterations as the state showing most
state as the state
heuristic value may not hold true or search may end up in finding an undesirable goal
evaluation may lead to wrong direction in the search tree.
a new heuristic function “hz”
Let’s have another heuristic design for the same problem. Fig. 2.14.6 is depicting
+1 for each brick in the support structure.
definition, in which the correct support structure of each brick is given
for each brick in the wrong support structure.
And the one not having correct support structure, —1
Start | Goal
-B 6
Blocks world
function “he”
Fig. 2.14.6: Definition of heuristic
Fig. 2.14.5, but this time using h2,
states are considered again as that of
As we observe in Fig. 2.14.7, the same ion h2.
value generate according to heuristic funct
each one of the state is assigned a unique
ch will be carried
easi ly unde rsta nd that, in the second part of the example, sear
Observing this example one can to it.
state is getting a unique value assigned
out smoothly as each unique
le search is
vital role in search process, as the who
that, the design of heuristic plays a
This example makes it clear
cting the next state to be explored.
ied out by con sid eri ng the heuristic values as basis for sele
carr prior candidate for
promising the goal state will be the first
value to reach to
The state having the most
till we find the goal state.
exploration, this continues
° ° ” tys
Tettanaleions
OF -Puplicar
: : - :
(b) Global heuristic :
Fig. 2.14.7 : State evaluations using heuristic function “h2”

— For each block that has the correct support structure : + 1 to every block in the support structure.
- For each block that has the wrong support structure :— 1 to every block in the support structure.
— This leads to a discussion of a better heuristic function definition. ,
- Is there any particular way of defining a heuristic function that will guarantee a better performance in search.
process??
2.14.3 Properties of Good Heuristic Function
1 It should generate a unique value for each unique state in search space.
2 The values should be a logical indicator of the profitability of the state in order to reach the goal state.
3. It may not guarantee to find the best solution, but almost always should find a very good solution.
4 It should reduce the search time; specifically for hard problems like travelling salesman problem where the time
required is exponential.
The main objective of a heuristic is to produce a solution in a reasonable time frame that is good enough for solving
the problem, as it’s an extra task added to the basic search process.
The solution produced by using heuristic may not be the best of all the actual solutions to this problem, or it may
simply approximate the exact solution. But it is still valuable because finding the solution does not require a
prohibitively long time. So we are investing some amount of time in generating heuristic values for each state in.
search space but reducing the total time involved in actual searching process.
Do we require to design heuristic for every problem in real world? There is a trade-off criterion for deciding whether
to use a heuristic for solving a given problem. It is as follows.
o Optimality : Does the problem require tofind the optimal solution, if there exist multiple solutions for the same? ’
o Completeness : In case of multiple existing solution of a problem, is there a need to find all of them? As many.
heuristics are only meant to find one solution.
o Accuracy and precision : Can the heuristic guarantee to find the solution within the precision limits? Is the error
bar on the solution unreasonably large?
o Execution time : Is it going to affect the time required to find the solution? Some heuristics converge faster than.
others. Whereas, some are only marginally quicker than classic methods.
In many Al problems, it is often hard to measure precisely the goodness of a particular solution. But still it is important
to keep performance question in mind while designing algorithm. For real world problems, it is often useful to
introduce heuristics based on relatively unstructured knowledge. It is impossible to define this knowledge in such a
way that mathematical analysis can be performed. 3
Scanned by CamScanner 7
_ Problem S
LY UM
gy alasc (MU-Sem. 7-ComP)
2.15 Best First Search
apped on
2.15.1 Concept And br ea dt h fir st se arch never gets tr
getting expanded. a 5! ingle path at a ti
me, but
all co mp et in g branches are not and BFS, it wou ld be “fol low
st sear ch es of both DFS Best First
— Indepth frpath
dead end s. If we combine these properti ing than the curr ent one”. This is what the
s whe nev er som e com pet ing path look more promis
switch path
node chosen
search is..!!
the sear ch tree by exp anding the most promising
algorithm which explores mise of node
_ Best-first search is a search l described best-first sear
ch as estimating the pro
value of nodes. Judea Pear description of the
according to the heuristic gene ral, may dep end on the description of n, the
function f(n) which, in
nbya “heuristic evaluation a knowledge about the
up to that poin t, and most important, on any extr
hered by the sear ch
goal, the information gat
problem domain”. 4 priority queue.
ate for exte nsio n is typi cally implemented using
current best candid noted below the
Efficient selection of the first search on an example
search tree. The values
ch process of Best
Fig. 2.15.1 depicts the sear
heuristic values of nodes.
nodes are the estimated
a fl fo
3 5
[E|
4 6
ion scenario
t search tree expans
Fig. 2.15.1: Best firs
2.15.2 Implementation se are namely OPEN list and

CLOSED list for
rec ord the pat h. The
two lists in order to
— Best first search uses in
implementation pu
rpose. organized as a priority queue,
n gen era ted , but have not examined. This is imi zat ion
bee lementing max
nodes that have der of thei r heuristi
c value, assuming we are imp
= OPEN list stores in cr ea si ng or ns io n.
e stored with
the
candidate for exte
which nodes ar le ct io n of the current best that have been
heuristic. It provid
es effi ci en t se
ed. Thi s CLO SED list contains all nodes
been exa min
es th at hav e already it has been generated
CLOSED list stores nod
ll
generated, check pa
value is better than
~~ a new nod e is
Whenever
and will not be looked at again. val ue and cha nge the parent. if this new
ted its recorded ps.
evalua
visited before, check ck into an infinite loo
before. If it is already eva lua ted twi ce, and will never get stu
avoid any node being
previous one. This will
Best First Search
2.15.3 Algorithm :
OPEN = [initial state]
CLOSED = [J
-
while OPEN is not empty
do
PWN repr Remove the best node from OPEN, call it n, add it to CLOSED.
If n is the goal state, backtrack path to n through recorded parents and return path.
Create n's successors.
For each successor do :
a. If it is not in CLOSED and it is not in OPEN : evaluate it, add it to OPEN, and record its parent.
b. Otherwise, if it is already present in OPEN with different parent node and this new path is better than )
previous one, change its recorded parent.
i. Ifit is not in OPEN add it to OPEN.
ii. Otherwise, adjust its priority in OPEN using this new evaluation.
done
This algorithm of Best First Search algorithm just terminates when no path is found. An actual implementation would
of course require special handling of this case.
2.15.4 Performance Measures for Best first search
Completeness : Not complete, may follow infinite path if heuristic rates each state on such a path as the best option.
Most reasonable heuristics will not cause this problem however.
Optimality : Not optimal; may not produce optimal solution always.
Time Complexity : Worst case time complexity is still O(b "where m is the maximum depth.
m
Space Complexity : Since must maintain a queue of all unexpanded states, space-complexity is also O(b ).
2.15.5 Greedy Best First Search
A greedy algorithm is an algorithm that follows the heuristic of making the locally optimal choice at each stage with
the hope of finding a global optimum.
When Best First Search uses a heuristic that leads to goal node, so that nodes which seems to be more promising are
expanded first. This particular type of search is called greedy best-first search.
In greedy best first search algorithm, first successor of the parent is expanded. For the successor node, check the
following :
1. If the successor node's heuristic is better than its parent, the successor is set at the front of the queue, with the
parent reinserted directly behind it, and the loop restarts.
2. Else, the successor is inserted into the queue, in a location determined by its heuristic value. The procedure will
evaluate the remaining successors, if any of the parent.
In many cases, greedy best first search may not always produce an optimal solution, but the solution will be locally
optimal, as it will be generated in comparatively less amount of time. In mathematical optimization, greedy algorithms:
solve combinatorial problems.
For example, consider the traveling salesman problem, which is ofa high computational complexity, works well with -
greedy strategy as follows. Refer to Fig. 2.15.2. The values written on the links are the straight line distances from the ¥
nodes. Aim is to visit all the cities A through F with the shortest distance travelled. ‘a
Let us apply a greedy strategy for this problem with a heuristic as, “At each stage visit an unvisited city nearest to the’
current city”. Simple logic... isn’t it? This heuristic need not find a best solution, but terminates in a reasonable number
of steps by finding an optimal solution which typically requires unreasonably many steps. Let’s verify.
Fig. 2.15.2 : Travelling Salesmen Problem example
— As greedy algorithm, it will always make a local optimal choice. Hence it will select node C first as it found to be the
one with less distance from the next non-visited node from node A, and then the path generated will be
A>C>D->B-E-F with the total cost = 10 + 18 +5 + 25 + 15 = 73. While by observing the graph one can find the
optimal path and optimal distance the salesman needs to travel. It turns out to be, A->~B—>D—>E—»F—9C where the cost
comes out to be 18+5+15+15+4+18=68.
2.15.6 Properties of Greedy Best-First Search

1. Completeness : It’s not complete as, it can get stuck in loops, also is susceptible to wrong start and quality of heuristic
function. ,
2. Optimality : It’s not optimal; as it goes on selecting a single path and never checks for other possibilities.
m
3. Time Complexity : O(b ), but a good heuristic can give dramatic improvement.
4. Space Complexity : O(b") , It needs to keep all nodes in memory.
2.16 A* Search
MU=May 13) Dec: 13; May 14; Dec. 14, May 15
At isoptimally efficient.
@ At algori
‘At algorithm with example.
2.16.1 Concept
= A* pronounced as “Aystar” (Hart, 1972) search method is a combination of branch and bound and best first search,
‘
combined with the dynamic programming principle.
— It’s a variation of Best First search where the evaluation of a state or a node not only depends on the heuristic value of
the node but also considers its distance from the start state. It’s the most widely known form of best-first search.
A* algorithm is also called as OR graph / tree search algorithm.
— In A* search, the value of a node n, represented as f(n) is a combination of g(n), which is the cost of cheapest path to
reach to the node from the root node, and h(n), which Is the cost of cheapest path to reach from the node to the goal
node. Hence f(n) = g(n) + h(n).
~ As the heuristic can provide only the estimated cost from the nade to the goal we can represent h(n) as h*(n);
similarly g*(n) can represent approximation of g(r) which Is the distance from the root node observed by A* and the
algorithm A* will have,
f*(n) = g*(n)
+ h*(n)
o. 7
BF Al&SC (MU-Sem. 7-Comp) 2-25 Problem Solving 4
— As we observe the difference between the A* and Best first search is that; in Best first search only the heuristic ’
estimation of h(n) is considered while A* counts for both, the distance travelled till a particular node and the
estimation of distance need to travel more to reach to the goal node, it always finds the cheapest solution. 4
4
— Areasonable thing to try first is the node with the lowest value of g*(n) +h*(n). It turns out that this strategy is more
than just reasonable, provided that the heuristic function h*(n) satisfies certain conditions which are discussed further —
in the chapter. A* search is both complete and optimal.
A* does also use both OPEN and CLOSED list.
2.16.3 Algorithm (A*)

Initialization OPEN list with initial node; CLOSED=6; g = 0, f = h, Found = false;
2. While (OPEN # 9 and Found = false )
t
i. Remove the node with the lowest value of f from OPEN to CLOSED and call it as a Best_Node.
ii If Best_Node = Goal state then Found = true
iii. else
C
a. Generate the SuccofBest_Node

b. For each Succ do
-
i. Compute g(Succ) = g(Best_Node) + cost of getting from

Best_Node to Succ.
ii. IfSucce OPEN then /* already being generated but not

processed */ : ; .
{ 4
a. Callthe matched node as OLD and add it in the list of Best_Node successors.
b. Ignore the Succ node and change the parent of OLD, if required.
- If g(Succ) < g(OLD) then make parent of OLD to beBest_Node and change the values of g and f for OLD
- If g(Succ) >= g(OLD) then ignore
y
a. If Succe CLOSED then /* already processed */

Fe
i. Call the matched node as OLD and add it in the list of Best_Node successors.
ii. Ignore the Suce node and change the parent of OLD, if required
- If g(Succ) < g(OLD) then make parent of OLD to be Best_Node and change the values of g and f for OLD.
- Propogate the change to OLD’s children using depth first search
- If g(Succ) >= g(OLD) then do nothing
~
a. If Succé OPEN or CLOSED

°
we
i. Add it to the list of Best_Node’s successors
ii. Compute f(Succ) = g(Succ) + h(Succ)
iii,
. Put Suce on OPEN list
ist with
with its
i f value
} /* for loop*/
}¥ /* else if */
} /* End while */.
ve
If Found = true then re port the b est path else report failure
i
2.16.4 Behaviour of A* Algorithm

1 of heu istic function a id how we
A* . tota ly depen ds upon the desig
As sta ted already the uccess of
succ
node. Let us unde stand the effect o heu

by estimating its distance fram the goal
eval yate each node
optimality gets affected by it

execution of the algorithm and how the
A. Underestimation
ent to goal that is, the
i s t y func i
tion o n ‘ ‘h’ neve r over estimates actual value from curr
wecan
Ifvelae antese ane he
guared atne the
— oe erat of hopes required to reach to
his " ealwa ys lesse r than the actual cost or actual number
voal state On y exists.
one
to find an optimal path to a goal, if
_In this case, A* algorithm is guaranteed
Example :
mated.
f = g+h,Herehis underesti A
Underestimated
(1+4)C (145)D
(1+3)B
goal
3 moves away trom
(2+3)E
from goal
3 moves away
(3+3)F
Fig. 2.16.1 uted. B is
d D. ‘f va lu es fo r each node is comp
an
expanded to B, C e path currently we
of all ar cs to be 1. Ais PP OS e we re solve in favor of E, th
st f(E) = £(C) = 5. SU
— {If we consider co E, We notice that F)=6 so we will now
expand node C.
ex pa nd ed to F is st op pe d as f(
chosen to be nsion © fa node was farther away
E is ex pa nd ed to F. Expa t ev en tu al ly discovered that B
bu
are expanding. e d some effort
in g h( B) , We have wast ll find optimal
path.
de re st im at th , an d wi
- Hence by un and try anothe
r pa
th ou gh t. Th en Ww e go back
than we
Wie
; _Al&SC> (MU- -Sem. _7-Comp) )
B. Overestimation
generated for each nade is greate
r than the actual number of step
Here h is overestimated that is, the value
required to reach to the goal node.
Example :
~—\
\. B (1+4)C (1+5)D
(2+2) E
(3+1)F
(440) G
Fig. 2.16.2
— As shown in the example, A is expanded to B, C and D. Now B is expanded to E, E to F and F to G for a solution
a path off
path of length 4. Consider a scenario when there a direct path from D to G with a solution giving
length 2.This path will never be found because of overestimating h(D).
Thus, some other worse solution might be found without ever expanding D. So by overestimating h, one cannot
guarantee to find the cheaper path solution.
2.16.5 Admissibility of A*
MU - Dec. 12, May 14
— Asearch algorithm is admissible, if for any graph, it always terminates in an optimal path from initial state to goal
state, if path exists. A heuristic is admissible if it never overestimates the actual cost from current state to goal state.
Alternatively, we can say that A* always terminates with the optimal path in case h(n) is an admissible heuristic.
function.
— Aheuristic h(n) is admissible if for every node n, if h(n) < h*(n), where h*(n) is the true cost to reach the goal state
from n. An admissible heuristic never overestimates the cost to reach the goal. Admissible heuristics are by nature
optimistic because they think the cost of solving the problem is less than it actually is
— Anobvious example of an admissible heuristic is the straight line distance. Straight line distance is admissible because /
the shortest path between any two points is a straight line, so the straight line cannot overestimate the actual road,
distance.
o Theorem : If h(n) is admissible, tree search using A* is optimal.
o ~~ Proof: Optimality of A*with admissible heuristic. { ‘ 4
4
— Suppose some suboptimal goal G, has been generated and is in the fringe, Let n be an unexpanded node in the fringe
such that nis on a shortest path to an optimal goal G. ‘
ww TechKnowledge
Publications
— if aS
Start a
Fig. 2.16.3 : Optimality of A*

(G2) = g(G,) since h(G.) = 0
g(G2) > g{G) since G, is suboptimal
f(G) = g(G) since h(G)
=0
(G2) > f(G) from above
h(n) s h*(n) since h is admissible
g(n)+h(n) < g(n)+h*(n)
fin) < f(G)
Hence f(G2) > f(n), and A* will never select G, for expansion.
2.16.6 Monotonicity
— Aheuristic function h is monotone or consistent if,

—- VWstates X, and X; such that X, is successor of Xi,
h(Xi)— h(X)) $ cost (Xi, X)
where, cost (Xj, X))actual cost of going from X; to X; and h (goal) = 0

- In this case, heuristic is locally admissible i.e., consistently finds the minimal path to each state they encounter in the
search. The monotone property in other words is that search space which is everywhere locally consistent with
heuristic function employed i.e., reaching each state along the shortest path from its ancestors: With monotonic
heuristic, if a state is rediscovered, it is not necessary to check whether the new path is shorter. Each monotonic
heuristic is admissible.
-. Acost function f(n) is monotone if f(n) <f(succ(n)), Vn.
~ For any admissible cost function f, we can construct a monotone admissible function.
~ Alternatively, the monotone property : that search space which is everywhere locally consistent with heuristic
function employed i.e., reaching each state along the shortest path from its ancestors.
~ With monotonic heuristic, if a state is rediscovered, it is not necessary to check whether the new path is
shorter.
2.16.7 Properties of A*
1, Completeness : It is complete, as it will always find solution if one exist.
2. Optimality : Yes, it is Optimal.

3. Time Complexity : O(b”), as the number of nodes grows exponentially with solution cost.
4, Space Complexity : 0(b™),as it keeps all nodes In memory.
= =
TechKnowledyé
Puolications
- ee py o-cy Eroplem Solving!
2.16.8 Example : 8 Puzzle Problem using A*
Algorithm
Start stat Goal state
Evaluation function -
f for EPP
Start state
Search tree
1-044
up
(1+3)
5807 6
502
up
left
(243) (243) |
306
3°7°6 : 3°76
57 2 O52
240158 §20
Ac de Bs fold
left right
(342) | (3+4)
O36 3.61
5 72: B57 2)
4.408 413°
down
£6. 3h 6" 5 3° 6°
O72 7O2 an
A otyk. Ae jces Se
Fig. 2.16.4: Solution
of 8- puzzle usin
The choice of evaluatio g A*
n function critically de
termines search re
Consider Evaluation fun sults,
ction
f(X) = g(X) +(x)
h(X) = the number of til
es'not in their 80al
position in a given
state X
a(X) = depth of node X in the sea
rch tree.
For Initial node f(init
ial_node) =4
WF TochKnowtedga
Publications
—=y
MM
i Al&SC (MU-Sem. 7-Comp) _
Problem Solving
(s)15
3 4
14G) (4) 12
Open : 4(12 +4), 1(14 + 3)

Closed : $(15)
Open : 5(10 + 6), 1(14 + 3)

Closed : S(15), 4 (12 + 4)
Open : 2(10 + 7) 6(10+10)

Open : 1(14 +3) 6(10 + 10) 2(10 + 11)
Closed : S(5) 4(12 + 4) 5(10
+ 6) 1(14 +3)
Closed : S(15) 4(12 + 4) 5(10 + 6)
TachKnowl
*Pedwicar
Open : 3(8 +11), 6(10 + 10)
Open : 6(10 + 10)
Closed : S(15),4(12+4),5(10+6),
Closed : s(15), 4(12 + 4), 5(10 + 6),
/1(14+ 3),2(10+7) 1(14 +3) 2(10 + 7) 3(8 + 11)
Open: 7(0 + 13}

+ 10)
Closed: $(15), 4(12 + 4), 5(10 + 6), 1(14 + 3), 2(10 + 7), 5(8- + 4), 6(10
2.16.9 Caparison among Best First Search, A* search and Greedy Best First Search
y | A* search | Best
Best First Search
"Algorithm _| Greed First Sear
Completeness Not complete complete | Not complete
Optimality Not optimal optimal Not optimal

Time Complexity O(b™) O(b™) O(bm)
“Space Complexity | O(b™) O(b™) O(bm)
TechKrowledga
puosicakions
(D) 0+12=12
10 8
E G
Q) 815213
lorsa18
10 10 8 16
DC; 4 1
Cc
1642=18 2)! 24+0=24
2015=25 Ci 20+0=20
10 10 3/ ©
J K
E F
3045235 © 30+0=30 2440224 [© (2) 244+5=29
A A. 4 A
wo 13(15) ©
20
3 B G G
=o
150 Ors = 13 py
150 rd
H
4
8 A
8 A
SA
10 20(24)@
ie B
B GS
Oo 20()Q)
160
D
O20
nodes
* with memory size of 3
Fig. 2.16.5 : Process of SMA
Se ar ch Al go ri th ms an d Op timization Problems
2.17 Local
2.17.1 Hill Climbing MTU Ey 2s scoe ez cee May 16
:
dback is used here to decid eon

pth first with generate and test where a fee
ination of de
~ Hill climbing is simply @ comb
search space. onally hard problems, which’
the direction of motion in the rt if ic ia l in te ll igence, to solve solving computati
used widely in a
- Hill climbing technique is = = ~ Wig Tech Knowledge
‘has multiple possible solutions: Puptications ;
= a
ee Ee Eroblen'Salam
AlgSC (MU-Sem. 7-Comp) 2-33
tion. But in hill climbi ng the test function
the test func tion will mer ely accept or reject a solu
In the depth-first search, e is to goal state.
which provides an esti mate o f how close a given stat the heuristic
is provided with a heuristic function d to fin d th e solution, i.e.
ion n ee de
ided with the additional informat
— In Hill climbing, each state is prov the c complete search tree.
Rather, it looks only a
ient since it does not mal intai n
value. The algorithm is memory effic
s.
the current state and immediate level state |
from your current lo cation. There are
n possible paths with different
For example, if you want to find a mall mall, so
ion will just give you the distance of each
path which is reaching to the
to reach to the mall. The heuristic funct
ent for you to reach to the mall.
that it becomes very simple and time effici
ider all the
nt state by means of an evaluation function. “Cons
— Hill climbing attempts to iteratively improve the curre point on the landscape corresponds
to thei
of a landscape. The height of any
possible states laid out on the surface
. Fig. 2.17.1 depicts the typical hill climbing
evaluation functionof the state at that point” (Russell and Norvig, 2003) 4
:
reach to the hill top from ground level.
scenario, where multiple paths are available to
Goal (Hill top)
Start
point
Fig. 2.17.1 : Hill Climbing Scenario
- = Hill climbing always attempts to make changes that improve the current state. In other words, hill climbing can onl
advance if there is a higher point in the adjacent landscape. “_ 4
t In many cases where state spac ia
— — Hill climbing is a type of local search technique. It is relatively simple to implemen space is)
"
of moderate size, hill climbing works even better than many advanced techniques
;7
— For example, , hill climbing when applied to travelli initially iit produc
elling salesman problem; ; initially produces random combi + age ofae
..
Then .it selects the better rout by switching the ination
. . ..
solut : - ord
7 ions having all the cities visited.
g the order, which visits all the cities inj
minimum cost..
- There are two variations of hill climbing as discussed follow
2.17.1(A) Simple Hill Climbing
It is the simplest way to i il climbi ;

is the algorithm for simple hill climbi i |
the procedure han ler ts tater
at of generate and Following
hil climbing, test but, the main di So
he two eneeiae ong
is use of heuristio
function for state evaluation which is used in hill climbing The erence between-t
Y state is decided by the heuristic value
that state. It can be either incremental heuristic or detrimental anes ofr
Algorithm
1. Evaluate the initial state. If it Is a goal state, then return and quit; otherwis
2. Loop until a solution is found or there are no new Operators
rs left
le = mate fravcurvent state and g0 tose
a
explored).) top @ applied
ied (i.e.
(i no new children nodes left to be5@
ii a. Select and apply a new operator (i.e, generate new child ;
_ b; Evaluate the new state : pon!
(i) Ifitisa goal state, then return and quit
Soe Sa ~ peda hia a eek,
= ss bored
‘7 AlaSC (MU-Sem. 7-Com 2-34 Problem Solving
(ii) If itis better than curren t state then make i
€ it a new current state.
If it is
(iii)
iii
not better than the current state then continue the loop, go to Step 2
is better than the current state is
.
As we study the algorithm, we obse
in every pass the first node / state that
considered for further exploration Tie ee the problem, but may
optimal solution to
egy may not guarantee that most
save upon the execution time.
2.17.1(B) Steepest Ascent Hill Climbing

ee
As the name tes
suggests,* stee pest hillil climbing
climbi always finds the steepest path to hill top. It does so by selecting the best
a erent of this ata current node / state. All the states are evaluated using heuristic function. Obviously, the
a egy is: more as compared to the previous one. The algorithm for steepest ascent hill climbing
i lows
Algorithm
t; otherwise make it as a current state.
1. Evaluate the initial state, if it is a goal state, return and qui
ent state :
a solu tion is foun d or a comp lete iteration produces no change to curr
2. Loop until will be better than SUCC.
of the current state
a. SUCC=a state such that any possible successor
e the new state :
b. For each operator that applies to the current state, evaluat
(i) If itis goal; then return and quit
(ii) If it is better than SUCC then set SUCC to this state.
set the current state to SUCC.
c. SUCC is better than the current state —
ti me requirement and
with stee pest ascent, we find that there is a tradeoff for the
As we compare simple hill climbing
solution.
the accuracy or optimality of the successors are
go for first bette r succ es sor, the time is saved as all the
g technique as we
— Incase of simple hill climbin ches ge tting explored, in
turn the solution found may
of node s and bran
to more num ber
not evaluated but it may lead
not be the optimal one. the bes t among all the successors is
selected for
que , as eve ry tim e
ascent hill climbing techni ution found will be
"= While in case of steepest eva lua tin g all the suc ces sor s at earlier stages, but the sol
involves more time in makes it clear that the
further expansion, it tes lea din g to hill top are explored. This also
sta
solutio n, as only the performance of the algorithm.
always the optimal def ini tio n pla ys a vit al role in deciding t he
on
ie. the heuristic functi
evaluation function
ng
Limitation s of Hill Climbi
2.17.1(C) MU -May 13, May 14, Dec.
re MEW ae
~ (May
13, Dec.14, )
May 14 May =
ction on the hill climbing techniques.
co rr ec t de si gn of heuristic fun
e impact © f in
algorithms may lead to
a position,
what can be th bing strateey- Some
t imes the
Nowlet’s see n hill cl im further
— may ar is e |
d to a better place on hill i. e. no
Following are the
probl ems that is no mo ve possible which will lea states.
- which th er e
the following three
so on, but from
lu ti if we ve reached one of
ha
which is not a the solution. This
will happen hill but is not
g cl o se r to ic h is at he ig ht fr om other parts of the
state that is goin wh
mu m’” is a location in hill s, but there is not
next better state
“c al ma xi all its ne ig hb or
1. Local Maximu
m : A te better than solution. In such
e se ar ch tree, it Is a sta so me ti me s oc cu r within sight of a
p. In th Local maximum
the actual hill to rt he r expansion.
osen for fu
which can be ch “goothills”.
lled
cases they are ca
¥ ‘Tech Knowles
Puoticatlt
Fig. 2.17.2 : Local Maximum
- Inthe search tree local maximum can be seen as follows :
Fig. 2.17.3 : Local maxima in search tree

2. Plateau : A “plateau” is a flat area at some
height in hilly region. There is a large area of same
the search space, plat eau situation occurs
height in plateau. In
when all the neighbouring states have the
same value. On a plateau, it
is not possible to dete rmine the best direct
ion in which to move by making local comparison
s.
Plateau
Fig. 2.17.4 : Plateau in hill climbing

In the search tree plateau can be identified
as follows :
()
YO @ ©
Fig. 2.17.5
3. Ridge: A "ridge" is an area in
the hill such that, it is hi gher
uphill path from ridge. In the than the Surroundin
search tree it Is the situation g areas, but there is no further
lesser, it’s a ridge condition , where all SuCCesso
. The sultable successor c ann rs are eithe r of same value or
ot be searched in a
simple move,
a , Ridgo
Fig. 2.17.6 : Ridge In hill

climbing
W TechKaowledga
Publications
=
Se
In the search tree ridge can be identified as follows :
|
Fig. 2.17.7
. |
Fig. 2.17.8 depicts all the different situations together in hill climbing
global
maxima
Objective function
Plateau
Local
maxima
Local maxima
State space
Fig. 2.17.8 : Hill climbing problems scenario
2.17.1(D) Solutions on Problems in Hill Climbing

will
ng techniques. At times combination of two techniques
— Inorder to overcome these problems we can try followi
provide a better solution.
direction.
back track to some earlier nodes and try a different
1. Agood way to deal with local maximum, we can
This can be done by
in some direction to a new area in the search.
2. Incase of plateau and ridges, make a big jump strategy is dealing with
ing two more rules of the same rule several times, before testing. This is a good
apply
plateau and ridges. -
mediate” consequences of its
what to do next by looking only at the “im
- Hill climbing is a local method. It decides bing becomes inefficient in large proble
m
encoded in heuristic functions. Hill clim
choices. Global information might be ly a constant amount, as it
comb inat oria l expl osio n occur s. But it uses very little memory, usual
spaces, and when
doesn’t retain the path. h on the shape of the
d with othe r met hod s. The succ ess of hill climbing depends very muc
|tisa useful when combine a good solution very
—
few loca l max ima an d plateau, random-restart hill climbing will find
state space. If there are
quickly.
2.17.2 Simulated Annealing

an analogy to
hill clim bing . Simu late d anne eal
alin g technique can . be explained by
2 V ariation of ting point and then cooled.
- Simulated annealing is lids, a solid is heated past mel
process in case of so
annealing in solids. In the annealing es. If the liquid is cooled slowly, it gets
transformed in
ges its prop erti
With the changing rate of cool
ing, the solid chan will not get enough time and
~ quickly, the crystal formation
crystals. While, if it is cooled
steady frozen state and forms
stals.
it produces imperfect cry duc e a min ima l energy final state after rais
ing the substance to high
| s ' to pro
ealing process c function is a minimal
- The aim of physical ann we are actually going downhill and the heuristi
lev el. He nc e in simulated annealing rath er than climbing up in this case
we are descending
energy ni mu m v alue , and
is the one with mi
heuristic. The final state i. Ww TechKnowledge
the valley. ; Publications
Problem So}Vin
.
— _ The idea is to use simulated annealing to search for feasible .
solution s and converge to an optii mal solution
tion.. In Order to,
achieve that, at the beginning of the process, some downhill moves may
be made. These em moves ae Made
purposely, to do enough exploration of the whole space early on, so that the final
solution is relatively insensitive ta
. :
the starting state. It reduces the chances of getting ;
caught at a local maximu idge.
m, or plateau, or a ridg
Algorithm
1. Evaluate the initial state.
2. Loop until a solution is found or there are no
new operators left to be applied :
‘Set T according to an annealing sche
dule
Select and apply a new operator
Evaluate the new state :
goal quit
AE = Val(current state) —Val(new
state)
AE < 0 > new current state
else new current state with probability elk
— We observe in the algorithm that,
if the next state is better than the curre
nt, it readily accepts it as a new curre
State. But in case when the next nt
5 tate is not having the desirable value
probability, e“*"" where AE even then it accepts that state with some
is the p Ositive change in the energy level,
T is temperature and k is Boltzmann’s cons
tant.
the small one. Also, the
temperature decrease. Hence uphil
beginning of the annealing process, l mov @s are more likely
when the temperature is high. As in the
the cooling process starts, tem
down, in turn the uphill moves.
Downhill moves are allow ed any per atu re comes
time in the whole Process. In this
very small upward moves are allo
wed till finally, the process con
way , comparatively
verges to a local minimu
desired low point destination in m configuration, i.e. the
the valley.
2.17.2(A) Comparing Simulated Annealing
with Hill Climbing
A hill climbing algorithm never mak
es “downhill” moves toward
states with lower value
because it can get stuck on a loca
l maximum. and it can be incomplete,
Hill climbi
' Ng Procedure chooses the
best state from those av
€xpansion. Unlike ailable or at least bett
hill climb er than the current stat
. State turned out ng, simulated annealing chooses e for further.
a random move from the neighb
Hastie Cessor
to be bett t : ourhood. If the successor c
. ah han
state is worse, then it WillitsbeCurr ent state then simulated annealing will acce
acce pted based on some probability pt it for further expansion, If the.
a
W TechKnowle i ge
S c anned by CamS canner
Al&SC (MU-Sem. 7-Com 4
2 2-38 Problem Solving
2.17.3 Local Beam Search
| n all the variations of hill climbing
climbing till¢ now, we have considered only one node getting selected
at a time for further
search p process. These algori gorithms are memory efficient
Ses in that
sense. But when an unfruitful branch gets explored
even for some amount of time it 's a
Complete waste of time and memory. Also the solution produce may not be the
optimal one. ;
The local beam sea i

rch algorithm keeps track of k best states by performing parallel k searches. At each step it
gene rates successor nodes and selects k best nodes for next level of search. Thus rather than focusing on only one
b ranch it concent rates on k paths which
; seems to be promising.
ea
If any of the successors found to be the goal, search
process stops.
In parallel local beam search, the parallel threads communicate to each other,
hence useful information is passed
among the parallel search threads.
In turn, the states that generate the best successors say to the others, “Come
over here, the grass is greener!” The
algorithm quickly terminates unfruitful branches exploration and moves its resources to where
the path seems most
promising. In stochastic beam search the maintained successor states are chosen with a
probability based on.their
goodness.
Algorithm : Local Beam search
Step1: Found = false;
Step2: NODE = Root_node;

Step3: If NODE is the goal node, then Found = true else find SUCCs of NODE, if any with its estimated cost and store in
OPEN list;
Step4: While (Found = false and not able to proceed further)
{
Sort OPEN list;
Select top W elements from OPEN list and put it in W_OPEN list and empty OPEN list;
While (W_OPEN #¢ and Found = false)
{ .
Get NODE from W_OPEN;
if NODE = Goal state then Found = true else
{
Find SUCCs of NODE, if any with its estimated cost
store in OPEN list;
}
}// end inner while
}// end outer while

Stop
Step5: If Found = true then return Yes otherwise return No and
, BF Tectknamedgs a
Pudlications
¥ AI&SC (MU-Sem. 7-Comp)
k=2
I Continue till goal stateis |

| found or not able to |
v proceed further v
Fig. 2.17.9 : Process of local Beam Search

— As shown in Fig. 2.17.9, here k = 2, hence two better successo
rs are selected for expansion at first level of search and |
at each next level, two better successors will be selected by both searches
. They do exchange their information with
each other throughout the search process. The search will continue
till goal state is found or no further search is
possible.
— _ It may seem to be that local beam search is same as running
k searches in parallel. But it is not so. In case of parallel
searches, all search run independent of each other. While
in case of local beam search, the parallel running threads
continuously coordinate with one another to decide the
fruitful region of the search tree.
— Local beam search can suffer from a lack of diversity among
the k states by quickly concentrating to small region of
“the state space.
— While it is possible to get excellent fits to training
data, the application of back Propagation is fraught
and pitfalls for the prediction of the performance with difficulties
on independent test data. Unlike most other learnin
have been previously discussed, there are far more g systems that
choices to be made in applying the gradient descent
method.
— The key variations of these choices are : The
learning rate and local minima - the selection
critical importance in finding the true global minimum of a learning rate is of
of the error distance. ,
— Back propagation training with too small a
learning rate will make agonizingly slow progr
will proceed much faster, but may simply produ ess. Too large a learning rate
ce oscillations between relatively poor solutions.
:
— Both of these conditions are generally
detectable through experimentation and
number of training epochs. sampling of results after a fixed
:
— Typical values for the learning rate paramete
r are numbers between Oand1:0.05<h<
0.75.
— - One would like to use the largest learning
rate that still conver; ges to the minimu
evidence shows that the use of a term called m solution momentum - empirical
momentum in the
back propagation algorithm can be helpful in spee
the convergence and avoiding local minima. ding
- The idea about using a moment
um is to stabilize the weight change
combination of the gradient decr by making non radical revisions using a:
easing term with a fraction of
the Previous weight change :
Aw(t) = -— OEe/Ow(t) + 4 A w(t y
— 1)
where a is taken O£ a £ 0.9,
and t is the index of the curren
t weight change.
— — This gives the system
a certain amount of inertia
since the weight vector wil
direction unless opposed by l tend to continue moving
the gradient term. in the same |
— The momentum has the followin
g effects :
Td TethXnowledga
Pubfications
7 a
Alxsc (MU-Sem.7-Comp)
It smooths the weight changes ang Suppresses ae
error valey; Cross-stitching, t
‘9 When all weight changes are all in the same
directi .
convergence, ection the momentum amplifies the learning rate causing a faster
o Enables to escape from small local mini Ma

on the erro T surfac
e.
e hope is that the momentu i
the other hand, a eoraing
nieta, otOn én-COnsiehgeneess
mr cane nest learning rate and that this will speed convergence and avoid local
encounteea. 1 with no Momentum will be much faster when no problem with toc
for training,ticular’
and the length ofth training oft €n* isSequential
cxamole, measured or inrandom presentation - the epoch is the fundamental Mie
terms of epochs. During a training epoch with revision
after a Fare pie, the examples can be presented in the same sequential order, or the examples could be
presented in a different random order for
each epoch. T he random representation usually yields better results.
:
The randomness has advantages and disadvantages :
o aaa : It gives the algorithm some stochastic search properties. The weight state
tends to jitter around its
equilibrium, and may visit occasionally nearby points. Thus it may escape trapping in
suboptimal weight
configurations. The on-line learning may have a better chance of finding a global minimum than the true gradient
descent technique.
Disadvantages : The weight vector never settles to a stable configuration. Having found a good minimum it may
then continue to wander around it.
Random initial state - unlike many other learning systems, the neural network begin in a random state. The network
weights are initialized to some choice of random numbers with a range typically between -0.5 and 0.5 (The inputs are
usually normalized to numbers between 0 and 1 ). Even with identical learning conditions, the random initial weights
can lead to results that differ from one training session to another.
The training sessions may be repeated till getting the best results.
2.17.4 Genetic Algorithms

MU — Dec. 15, May 16)
Gas are adaptive heuristic search algorithms based on the evolutionary ideas of natural selection and genetics. As such
they represent an intelligent exploitation of a random search used to solve optimization problems. Although
randomized, Gas are by no means random, instead they exploit historical information to direct the search for better
performance within the search space. The basic techniques of the Gas are designed to simulate processes in natural
systems necessary for evolution, especially those following the principles of “survival of the fittest” laid down by .
Charles Darwin.
Genetic algorithms are implemented as a computer simulation in which a population oikabstract representations
(called chromosomes or the genotype oF the genome) of a solutions (called individuals, creatures, or
as
better solutions. The solutions are bgbiesanted in binary
Phenotypes) to an optimization problem evolves towlatus
strings of Os and 1s, but other encodings are also possible. The phon usually starts trom a population of randomly
generated individual and occurs in generations. In each generation, the fitness of every individual in the population is
evaluated Itiple individuals are stochastically selected from the current population (based on their fitness), and
ated, multiple in
used in the next iteration of the algorithm.
Modifieledd toto f form a new population : The new population is then
Ee Tech knawleays }
Pubiications
Scanned by CamScanner |
W_Alasc (Mu-Sem. 7-Comp) _
2.17.4(A) Terminologies of GA
- Gene
in the problem
Gene is the smallest unit in genetic algorithm. The gene represents the smallest unit of information
the problem context were, for:
domain and can be thought of as the basic building block for a possible solution. If
number of shares of a |
example, the creation of a well-balanced investment portfolio, a gene might represent the
‘particular security to purchase.
-— Chromosome
Chromosome is.a series of genes that represent the components of one possible solution to the problem. The
chromosome is represented in computer memory as a bit string of binary digits that can be “decoded” by the genetic |
algorithm to determine how good a particular chromosome’s-gene pool solution is for a given problem. The decoding
process simply informs the genetic algorithm what the various genes within the chromosome represent.
- Encoding
the type of 1
Encoding of chromosomes is one of the problems, to start solving problem with GA. Encoding depends on
the problem. There are various types of encoding techniques like binary encoding, permutation encoding, value
encoding, etc.
— Population
A population is.a pool of individuals (chromosomes) that will be sampled for selection and evaluation. The
performance of each individual will be computed and a new population will be reproduced using standard genetic
operators.
— Reproduction
Reproduction is the process of creating new individuals called off-springs from: the parents population. This new
population will be evaluated again to select the desired results. Reproduction is done basically using two genetic
operators : crossover and mutation. However, the genetic operators used can vary from model to model, there are a
few standard or canonical operators : crossover and recombination of genetic material contained in different parent —
chromosomes, random mutation of data in individual chromosomes, and domain specific operations, such as
migration of genes.
2.17.4(B) Genetic Operators
—. Selection
In this process, chromosomes are selected from the population to be the parents for the crossover. The problem is
how to select these chromosomes. According to Darwin’s evolution theory the best ones should survive and create ©
new offspring. There are many methods how to select the best chromosomes, for example roulette wheel selection,
Boltzman selection, tournament selection, rank selection, steady state selection, etc.
— Crossover
Crossover involves the exchange of gene information between two. selected chromosomes. The purpose of the
crossover operation is to allow the genetic algorithm to create new chromosomes that shares positive characteristics el
while simultaneously reducing the prevalence of negative characteristics in an otherwise reasonably fit solution. Types .”
of crossover techniques include single-point crossover, two-point crossover, uniform crossover, mathematical
crossover, tree crossover, etc. ©
a
Saal
m Solving
ED
Al&sC (MU-Sem. 7-Comp) .
2-42 Problem Solving
mutation
is another re finement step th
Mutation a
: rity of 7 a ndomly chan setting to 4
completely differen one. The majo
t
Mutations f or Bes the value
of a gene from its current
the case in nature, less
problem
fit than more so. Occasional
ly, however,
a
.
hi ht med by this process are, as is often the
ighly superio Mutation provides
were, for
netic algorithm with t he Opportunity to Create rand beneficial mutation will occur.
previously
ares of a Be
uncharted areas of the solution 5 Pace, thus in chromosomes and information genes that can explore
;
creasing th ; i i
‘ous typ types of mutati .
various utation techniques like bit inversion y siasrching
changing, adage
value encoding. ea
2.17.4(C) The Basic Genetic Algorithm
em. The
po :
genetic Start] Generate random
1. ions for the problem) -
ecoding Population of n chromosomes (suitable solut
2. [Fitness] Evaluate the fitness f(x) of each chromoso me x foll
in the population
populati
repeating
Create a new population by
;
3, [New population] two parent chromosomes from. ing i
following g ste steps untili the new population is complete
1 ;
(selection] select
(the better fitness, the bigger
type of population according to their fitness
chance to be selected)
, value no crossover was
5, , 4 With
[Crossover] of a crossover probabilit y cross over the | parents to form a new offspring (children). If
performed, offspring is an exact copy of parents.
in chromosome).
6. [Mutation] With a mutation probability mutate new offspring at each locus (position
1. The
7, [Accepting] Place new offspring in a new population
enetic
run of algorithm
8, [Replace] Use new generated population for a further
return the best solution in current population
9. [Test} If the end condition is satisfied, stop, and
new 10. [Loop] Go to step 2
netic 2.17.4(D) Example of Genetic Algorithm

technique and are undergoing the
area
g pair of chr omo som e s encoded using permutation encoding
arent Let’s consider followin selection method and will be
applied, arithmetic
uming that they are selected using rank
h as complete process of GA. Ass
ng mutation techniques.
crossover and value encodi
a. Parent chromosome
Chromosome A
B
Chromosome
m is h chromosomes.
cro sso Vv er: i.e adding bits of bot
sate after arithmetic
Child chromosome
ion,
b. Child chromosome
909 9 8 7837
Chromosome c . subtracting 1
ing 2 sma ll val ue to selecte d values. E.g
su btract
- adding or
the mutation i.€
After applying valu e encoding
tics 7
to3™ and 4" bit.
pes
ical
" 908 e
c sT
8 8 7837
Se is mu c h better than
both parents.
- Techkanaledys
Oe Podrtications
uc ed
pr od
It can be observed that the child
: Al&SC (MU-Sem. 7-Comp) 2-43 Problem Solving
2.18 Adversarial Search ——___3

4
and it is played according to ;
— Adversarlal Search Problem is having competitive activity which involves ‘n’ players
j
certain set of protocols.
the surrounding environment in a game 1
— Gameis called adversarial because there are agents with conflicting goals and
is competitive as there are ‘n’ players or agents participating.
wants to win the game.
— Wesay that goals are conflicting and environment is competitive because every participant
agent environment.
— From above explanation it is understandable that we are dealing with a competitive multi
- Asthe actions of every agent are unpredictable there are many possible moves/actions.
analyse the action of other agents and
—- In order to play a game successfully every agent in environment has to first
agent executes.
how other agents are contributing in its own wining or loosing. After performing this analysis
2.18.1 Environment Types |
Fig. 2.18.1 : Environment types
There can be two types of environments in case of multi-agent : Competitive and cooperative.
1. Competitive environment
— In this type of environment every agent makes an effort to win the game by defeating or by creating superior it~
over other agents who are also trying to win the game.
— Chess is an example of a competitive environment.
2. - Cooperative environment
— — Inthis type of environment all the agents jointly perform activities in order to achieve same goal.
— Car driving agent is an example of a cooperative environment.
2.18.2 Al Game — Features

make the game j
Under artificial intelligence category, there are few special features as shown in Table 2.18.1, which e
.
more interesting.
Table 2.18.1 : Al game features 4
Two player When there are two opponents/agents playing the game
it is called 2- play game. To increase difficulty of the
game Intelligence is added to agents in Al games.
Note that, in case of Al Games we must have at least two
players. (i.e. single player games don’t come under Al
games category). ,
7 TechKnowledge
USF punsicarions
SS
_—
Multi-agent When there are two o Monopoly
T more opponents/agents playing
the
naa game veit j's called multi agent environment where
ery agents affects the action of other agents.
Non-cooperative When th @ surrounding
i environment is not helpful for | / Card games
environment winni
ing the game it is called as non-cooperative or
competitive.
Turn taking In - a multi

ulti- agent environment
i when the agent/ player | Any board game,
p or ms a move and has to i for the next player to | Chess, carom,
wait
make the next move. etc
Time limit One more constrainti can be incorporated in a : game i.e. Time bound
f
keeping a limitation on time. Every player will get a finite chess games
_ amount of time to take an action.
Unpredictable In Al games action agent is fuzzy which | Card game with

of opponent
makes the game challenging and unpredictable. multiplayer
opponent
Players are called unpredictable when the next step
upon an input set which is generated by the

depends
other player.
9,18.2(A) Zero Sum Game

ance of the game
gned to each player when the inst
ept is asso ciated with payoffs which are assi
“Zera sum game ” conc state. (Le. agents winning
ances when the game is in a neutral
tation of circumst
is over. It is a mathematical represen agents).
the winning and losing of other
or losing is even handed with the loss of player 2 is
ches s game it is mark ed with say +1 point and at the same time
For example, if player 1 wins 2 are
- ; in that case players 1 and
thus sum is zero. Anot her condition is when game is draw
marked with =~ 1 point ,
d payoffs).
(Here +1,-1 and 0 are calle
marked with zero points.
2.18,2(B) Non-Zero Sum Game

player winning the game
alge brai c sum of payo ffs as zero. In this type of games one
have
Non-zero sum game’s don t
er has lost the game.
does not necessarily mean that the other play
sum games :
There are two types of non-zero
1. Positive sum game
2. Negative sum game
2.18,2(C) Positive Sum Game have same goal and they contribu
te together to play the game.
ayers
It is also called as cooperative Bame- Here, all play
For example, educational games-
2.18.2(D) Negative Sum Game every player has a different goal

so no one really wins the game
,
,
It is also called as competitive game Here the most. 2
everybody loses. Real world example of a war suits TechKnowledye
Puotications
8 Alasc (MU-Sem. 7-Comp) 2-45 Problem Solving
2.19 Relevant Aspects of Al Game ; 4
To understand game playing, we will first take look at all appropriate aspects of a game which give overview of the
stages in a game play. See Fig. 2.19.1.
- Accessible environments : Games with accessible environments have all the necessary information handy. For
example : Chess.
|
— Search : Also there are games which require search functionality which illustrates how players have-to search through
possible game positions to play a game. For example : minesweeper, battleships.
— Unpredictable opponent : In Al games opponents can be unpredictable, this introduces uncertainty in game play and
thus game-playing has to deal with contingency/ probability problems. For example : Scrabble.
Accessible
environments,
"Relevant aspects ©
Fig. 2.19.1 : Relevant aspects of Al game
2.20 Game Playing
— Fig. 2.20.1shows examples of two main varieties of problems faced in artificial intelligence games. First type is “Toy
Problems” and the other type is “Real World Problems”.
"Real World
B-queen/n-queen “(NP hard)
Vacuum World - oe Robot Navigation
Missionaries and cannibals _ Assembly Sequencin

Fig. 2.20.1 : Example Problems
— Game play follows some strategies in order to mathematically analyse the game and genera
te possible outcomes. A ”
two player strategy table can be seen in Table 2.20.1.
4
~
y Alas (MU-Sem. 7-Comp)
= 2-46
Problem Solvin
a
le 2.20.1 ; Tw g _
.
o Player Stra
tegy table
By2 ae
Player1 Wins ; Player 2 Wins
Draw - Player 2 Wins

Player 2 Wins 7 Draw
oon ~ tooo 7
Bame is the strategic

form, mentioned in the introduction. For a
two-person zero-sum game, the payoff functio
i : n of Player Il is the negative of the payoff of Player I, so we may restrict
attention to the single payoff function of Player |, which we call here A.
The strategic form, rg
or normal form, of a two-person zero-sum game is given by a triplet (X, ¥,A), where
~ (1) Xis a nonempty set, the set of stra
tegies of Player |
(2) Yisanonempty set, the set of Strat
egies of Player Il
_ 3) Aisa real-valued function defined on X *Y,
(Thus, A(x, y) is a real number for everyx €X and every y eY,)
The interpretation is as follows. Simultaneously, Player | chooses
x €X and Player Il chooses y € Y, each unaware of
the choice of the other. Then their choices are made known and | wins
the amount A(x, y) from Il.
This is a very simple definition of a game; yet it is broad enough
to encompass the finite combinatorial games and
games such as tic-tac-toe and chess.
On the basis of how many times Player | or Player Il is winning the game, following strategies
can be discussed.
Equalizing Strategy : A strategy that produces the same average winnings no matter what the
opponent does is called
an equalizing strategy.
Optimal Strategy : If | has a procedure that guarantees him at least A(x, y) amount on the
average, and Il has a
procedure that keeps her average loss to at most A(x, y). Then A(x, y) is called the value of the game,
and the
Procedure each uses to insure this return is called an optimal strategy or a minimax strategy. -
Pure Strategies and Mixed Strategies : It is useful to make a distinction between a pure strategy
and a mixed strategy.
We refer to elements of X or Y as pure strategies. The more complex entity that chooses
among the pure strategies at
random in various proportions is called a mixed strategy.-
2.20.2 Type of Games

Game can be classified under deterministic or probabilistic category. Let’s see what we
mean by deterministic and
Probabilistic,
(®) Deterministic
It is a fully ob ble environment. When there are two agents playing the game alternatively and
is a fully observa the final
. ae
results of the game are equal and opposite then the game is called deterministic.
Tak le of ticctac-toe where two players play a game alternatively
and when one player wins a game then
© example of tic-tac- :
Other player losses game.
(b) Probabilistic
; erministic games, where you can ha
Probabilistic is also called as non-deterministic type. It is apposite of tet
q
multiple players and you cannot determine the next action of the player. salilietic
type sbutaritake exaniil :
You can only predict the probability of the next action. To understand pro
a
card games.
another way of classification for games can be based on exact/per : i ed on j
approxima
fect information or based on exact “/
te information. Now, let us understand these ter
ms. ‘ layer is called as a 3
1. Exact/perfect information : Games in
which all the actions are known to A ae
exact or perfect information. For example tic-tac-toe or boa of
rd games like chess, chec . , a 4
2. Inexact / approximate information : Game in which all the action
s are not known to omer p
er (or action
are unp
redictable) is called game of inexact or approximate inform
ation. In this type a game, player's next
action depends upon who played last, who won last hand, etc.
For example card games like hearts.
Consider following games and see how they
are classified into various types of games
we have learnt in above sections based on the parameters Which
:
Deterministic |-— Chess — Battleships

= Checkers - Card Game (Hearts)
Probabilistic - Ludo
— Scrabble
_ — Monopoly
- Poker
Now, let us try to learn about
few games mention in Table
2.20.2.
2.20.2(A) Chess
In chess both players can

see chess board positions
play one after the Other 5
in an order,
(a) Chess bo
ard
(b) Deep Blue
ll vs Garry Ka
sparov
final Position
‘Tech Knowledga
game 1
* Pubricagy ons Fig. 2.20.2
2
) _ problem Solving
(MU-Serm. 7-Comp
ie] «+e
‘ts
Medien bee
8 a
:
d et Se prehsted
F F
sid
tree
Fig. 2.20.3 : Chess game
rs
9.20.2(B) Checke y. This ga me is a two pers
on game where
and exa ct/ per fect information ¢ ategor er.
Checkers comes under det
erm ini sti c the other in an ord
y and players Pp lay one after world
both players can see board
positions, so ther e is no sec rec
dev e loped which de
feated human
call ed dra ugh ts) was
m name Chinook (also
In 1990's a comput er progra
- Tinsley.
champion Marion
rs board
Fig. 2.20.4 : Checke
omer
japan ered
repehaeias)
Foi obi)
ie}
ie)
ie}
OsO<0%08
ey
po orosc
aka e]
o=0
ome}
rome o pie!
o_0-0.0
Merce e(er ome!
kaemmo
e emoO) ne
|
B
omy je p. °8 8
[o] |
o.58 @
eer oil eet
ee ipn-o. O20
om om ome! TethKnowledge
So Puotications
We saw game tree of chess and checkers in earlier section. |
Game tree is defined as a directed graph with nodes and edge. Here nodes indicate positions in a game and edges |
indicate next actions.
Let us try to understand what a game tree is with the help of
Tic-Tac-Toe example. Tic-Tac-Toe is a 2-player game, it is.
deterministic and every player plays when his/her turn come. Game tree
has a Root Node which give the starting
position of the game board, which is a blank 3 x 3 grid.
Say in given example player 1 takes 'X' sign. Then
MAX(X) indicates the board for the best single
indicates that it is player 1's turn. (Remember next move. Also it.
that initial move is always indicated with a
MAX).
If the node is labelled as MIN then it means
that it is opponents turn (i.e. player 2's turn)
single next move is shown. and the board for the best
,
Possible moves are represented with the help
of lines. Last level shows terminal board posi
tion, which illustrates end
of the game instance, Here, we get zero as sum
of all the play offs. Terminal state gives winning
indicates the play off points gained by the board position. Utility
player (— 1, 0, + 1).
In a similar way we can draw a game tree for
any artificial intelligence based games.
MAX (X)
X x x
MIN (0) i
x BS xX _
_ | x x XxX
Ox x]o [Ix
MAX (X)
O eae
x][xTo x[o x]o

MIN (0)
X X| cee
X1X]O} [x]x]o] [x]xJo wove

TERMINAL OfO|X} Jolx]x x
o}x xX/O|o} fJololx
Utility 1 0 +1
Fig. 2.20.3 : Tic-tac-toe gam
e tree
We can formulate a game as a sear
ch problem as follows :
Initial state | ft gives the starting Position

of the game board.
Operators They define the next possible acti
ons.
wy TechKnowledge
tntiane
ee Se
gC (MU-Sem. 7-Comp)_ __
Problem Solvin
Al
Which
Terminal state hich ind :
indicates that all in
stance of game are
over
utility It displays a nu m ' .
ber which indicates if the game was won or lost or it was draw.
; '
From Tic-Tac-Toe
atively small (it game's example nodes),
has 91 terminal you can understan d that for a 3 x 33 grid, player game, where the ga me tree is
pri two os
rel
how O' difficult itit wou would b stillwe cannot draw it completely on one single page.
i
games with bigger grid size.
h © '0 creat e a game
; tree for multi-player games or for the
h a v e have huge
Many game s
7_ ive ge search space complexity. Games have limitation over the time and the amount of memory
space i: can consume.
ume. Finding
Findi :
an optimal .
solution is not feasible most of the times, so there is a need for
approximation.
- Therefore mere 5 a need for an algorithm which will reduce the tree size and eventually will help in reducing the
processing time and in saving memory space of the machine.
and
- One method is pruning where, only the required parts, which improve quality of output, of the tree are kept
,
reaming parts are removed.
- Another method is heuristic method (it makes use of an evaluation function) it does not require exhaustive research.
This method depends upon readily available information which can be used to control problem solving.
2.21 MiniMax Algorithm

- Minimax algorithm evaluates decision based on the present status of the game. This algorithm needs deterministic
environment with perfect/exact information.
- Minimax algorithm directly implements the defining equation. Every time based on the successor state minimax value
is calculated with the help of simple recursive computation.
- In case of minimax algorithm the selected action with highest minimax value should be equal to the best possible
payoff (outcome) against best play.
2.21.1 Minimax Algorithm

stage.
Take example of tic-tac toe game to understand minimax algorithm. We will take a random
terminal states.
Step1: Create an entire game tree including all the
Start action :'0!

X|X]O
o| |x
oO
O'sTum
x|xlo KIx|o X|X]O

0 X Oo xX 0 X
0 Oo 0
Fig, 2.21.1
Next action : 'x'
X{X}O XIX}/o xI{xIo X]x]o xXx]x]|o X|X|O

OJO|X Of/0O/}K O;X}x of |x Of;x}|x oO x
X O X}O Oo Oo o|x jo o|o O|x]o
Fig. 2.21.2
Next action : 'O'
x|x|o
o}| |x
Oo
x|xlo x|x lo x|xlo

O;O]X Oo x Oo xX
° Co Oo olo
X|/xJo_ _x{xlo x|xlo _x|xJo xlxlo

Ofo}x xIxlo
ofolx ofx[x “of [x “olxix “ol Tx
x! fo xlo of fo “oalxlo colo §=xfofo
O'sTum
x|xfo _x{xlo x|x]o x/xlo

ofo|x x[xlo xlxlg
olo|x olx[x “ololx “olxlx
x]O]O ~olotx
olxlo ofolo alxfo- ololo xlolo
Fig. 2.21.3
Step 2: For every terminal state find out utility (playoff points gained 7
b Y every terminal
1 means win and 0 means draw.
state). Terminal position whet
Problem Solving
¥ Al&SC. (MU-Sem. 7-Comp)
x
me)
ase x|x Jo x|x}o

o| |x o| |x
al Bs
° Oo; |90 ojo
ren
_X1x xIxlo x|x x|x]o

X{x]O _X{xXjoO of |x _o|x o| |x
ojo[x ofolx ofx{[x
jo ojxlo ° xlolo
x1 10 xlo o|
f 1 0
1 1
_ O'sTum
_x|x|m_ _x{xjo _x|% xix x|x|o
_x|xjo o|o|xX_
olgix ofx|x o|m|x oO|x
ololx
StH x|o x{o}o
x|o}o x|O
Fig. 2.21.4
e the utility values upward in
oper ator s on the node s of the present stage and propagat
step 3: Apply MIN and MAX
the three.
x|x|o
ol |x
oO
x|x}|o
O° xX
O°
AN
Xx xX|O x|x}O
x | Oo x
oO
O° x|oO
x
oO
Fig. 2.21.5
g minimax decision.
(pa yof f yalu e) sele ct the action at the root node usin
utility value
Step 4: : WitWith h the the m max(of the min)
¥ TechKnowledys
Pubtleakions
7 _Al&SC (MU-Sem.7-Comp) _
Oo 1 1 1 1 oO
x|x]o xx x|xlo x{x x|xlo xIxlo
10x ol@lx olx|x “olg|x o1x|x olfo]x
x/o X]}O -erete g)xlo -ete x}olo
Fig. 2.21.6 °
x|xlo
o| |x
Best move
Oo
oO
Fig. 2.21.7
(Ir CasSie of Step: Ss 2
a id 3 we
are assuming
that the Oppone t will play
pe t fectly
as p er ou re xp ecta
t: ti
tion )
Wy : Tethkhowteaga
Pucitcations ja
l
!
¥
&
ic
f
‘ 2 i
a ae ¥ Al&SC (MU-Sem. 7-Comp)_
2-54 Problem Solving=
2.21.2 Properties of Minimax Algorithm
eS
It is considered as Complete if the game tree size is fin;

S finite,
.
is considered Opti
" er r
Primal when itis played against an optimal numbmbDe o of opponents.
Time complexity of minimax algorithm is indicated asas O(b i ).
Space complexity of minima i oa , ‘
x algorithm is indicated by O(b")(using depth-first exploration approach).
For chess, b = 35, m ~100 for “reasonable” games
Exact solution is completely infeasible in most of the gamiés
2.22 Alpha Beta Pruning
Off. In game search iti resembles to clipping a branch in the search tree, probably which is not
Pruningital means cutting g off.
so fruitful.
choice found i.e.,
At any choice point along the path for max, a is considered as the value of the best possible
then, MAX will avoid it. Similarly we can define
highest-value. For each “x”, if "X" is worse i.e. lesser value than a value
B value for MIN.
making process need not consider each and every
a-B pruning is an extension to minimax algorithm where, decision
node of the game tree.
in making the search
are considered in decision making. Pruning helps
Only the important nodes for quality output
more efficient.
the result remaining parts of
which contribute in improving the quality of
Pruning keeps only those parts of the tree
the tree are removed.
e tree :
Consider the following gam
MAX
MIN
Fig. 2.22.1
for root node.

cal cul ate mi nimax values
to
For which we have ), min(13,7,2))
), min(3,A,B
Max(min(4,10,6
e
Minimax value of root nod
(C <=3)=4
Max(4, C 2)
step.
this step by
’ Let us see how to ch eck
TechKnowledye
. Pupriicatio
as
A [00,00]
MAX
MIN
MAX
MIN
MAX
MIN
MAX
MIN
MAX
MIN
W Tachn a a i
" SC (MU-Sem, 7-Com .
AlaSCs ) 2-56 Problem Solvin
ee
MAX e-0] (>) (4, 13]
MIN
[-0o, 13]
MAX
MIN
MAX » BX] >) 3] Ba] 12, 4]
MIN
[~», 13] [-29, 7] [2, 2]
— Soin this example we have pruned 2 B and 0 a branches. As the tree is very small, you may not appreciate
the effect
of branch pruning; but as we consider any real game tree, pruning creates a significant impact on search as far as the
time and space is concern.
2.22.1 Example of a-B Pruning

Ex. 2.22.1: Explain Min-Max and Alpha Beta pruning algorithm with following example. .
22 1 134437-+11096542656
Fig. P. 2.22.1 : Game Tree
Soln. : Using min-max algorithm
Al&ZSC (MU-Sem. 7-Comp)_ a ns ip OT
1
2211394437-110354256
(a) Min-max solution with optimal path shown in bold
Tota
Ex.
j
q
s
- Fig. P. 2.22.1
|_AI&SC (MU-Sem. 7-Com 58 _ en Problem Solving:
221134437 410354256
(g) Solution by o-f pruning

Fig. P. 2.22.1
Total pruned branches
Q&-cuts=2
B- cuts =3
Ex. 2.22.2: Perform o-® cutoff on the following.
so |-——_,
Jo oh CaSe
Fig. P. 2.22.2
Soln. :
.
No. of &- cuts = 1
No. of B - cuts = 2
min
max
min
10° 5 7 11 (12 8 5 12 11 9 8 7
yd @F _Al&SC (MU-Sem. 7-Comp) : ‘
— 2-60 Problem Solving
2.22.3 Properties of a-B

_- Final results are not affected by prunin
. e
— Ordering ofcede
Fis Good actitio ns helps inini improving effectiveness of pruning technique
perfect ordering then we can get time complexity as O(b™”) .
— Depth of search is doubled with pruning
ce Review Questions. .
Q.1 Differentiate between BFS and DFS
Q.2 How the drawbacks of DFS are overcome by DLS and DFID?
ing techniques.
O38 Compare and contrast.all the un-informed search
h.
Q.4 Write short note on bidirectional searc
orm cost search.
Q.5 Write anote on BEA and Unif
FLS and DFID.
Q.6 Compare and contrast DFS,
A* with example.
var iou s inf orm ed sea rch techniques? Explain
Q.7 What are
mple.
rch and A* with an exa
Q.8 Compare Be st First Sea
e it with simple h ill climbing.
of ste epe st asc e' nt hill climbing. And compar
Q.9 Write algorithm
we solve them?
lim ita tio n s of hill climbing? How can
Q. 10 What are the
s| pecify its propertie
s.
thm for Bes t first search and
Q.11 Write algori with example.
t and gre edy best firs t search? Explain
firs
difference between
bes t
Q.12 What is the heuristic? —
are the qu alities of a good
functio n? What
Q.13 What is heuristic .
local beam search
te d an nealing a nd
Q.14 Writeshort note on simu la
ng.
Hill climbi
Si mu la te d an nealing with
Q.15 Compare and contrast 5? explain with suitab
le example.
ch pr oc es
s the sear
uristic affect tic.
Q.16 How the definitio n of he d overesti mating Heuris
of u nd erestimating an
on behavior of A* in case
Q.17 Write short note
0ptimality.
of A* in case of en options?
Q.18 Discuss admissibility choose SMA* giv
When sho uld we
ge
‘YechKnowledons
ple. ¥ pupiicati
with exam
Q.19 Explain SMA“ alg orithm
WF _Alasc (MU-Sem.7-Comp)
Q. 20 Write a short notes on:
(a) Game types
(b) . Zero-sum game
(c) Relevant aspects of Al games
(d) Features of Al game
Q. 21 Explain minimax algorithm with an example and give its properties
Q. 22 Give a-B pruning algorithm with an example and it’s properties, also explain why is it called a-B pruning.
Q. 23 What is adversarial search?
Q. 24 Apply alpha-beta pruning on example given in Fig. Q. 24 considering first node as max.
Fig. Q. 24
Ol
Comme
Knowledge, Reasoning and Planning
Unit Ill j=
34 Knowledge based agents
3.2: First
et zi ord ler logic:
ic: syntax and Semantic,i Knowledge Engineering in FOL Infere
nce in FOL : Unification, Forward
Chaining, Backward Chaining and Resolution
3.3 Planning Agent, Types of Planning: Partial Order, Hierarchical Order, Conditional Order
- Understanding theoretical or practical aspects of a subject is called as knowledge. We can gain knowledge through
experience acquired based on the facts, information, etc. about the subject.
- After gaining knowledge about some subject we can apply that knowledge to derive conclusions about various
problems related to that subject based on some reasoning.
- We have studied various types of agents in chapter 1. In this chapter we are going to see what is “knowledge based
agent”, with a very interesting game example.
- Weare also going to study how do they store knowledge, how do they infer next level of knowledge from the existing
set. In turn, we are studying various knowledge representation and inference methods in this chapter.
3.1 A Knowledge Based Agent

- As shown Fig. 3.1.1, a knowledge based agents can be described at different levels : Knowledge Base (KB) and an
Inference Engine.
Domain-independent algorithms
“Knowledge base = Domain-specific content
Fig. 3.1.1 : Levels of knowledge base
1. Knowledge level:
specific content.
— Knowledge level is a base level of an agent, which consists of domain-
f ‘ormation about the surrounding environment in which they are working, it does
— In this level agent has facts/in
lementation.
not consider the actual imp
2. Implementation level: .
the data
domain independent algorithms. At this level, agents can recognize
— Implementation level consists of
structures used in knowledge
base and algorithms which use them. For example, propositional logic and
t logic and resolution in this chapter)
resolution. (We will be learning abou
. e choosing any action,
- Knowl
Knowl edge based
edge bas agents are crucial to use in partially observable environments Befor
knowledge along with the current inputs
: from the environment
F
base agents make use of the existing
ledge e based
ki nowledg
in order to infer hidden aspects of the current state.
WF _Alasc (MU-Sem.7-Comp) Knowledge, Reasoning and Plannin
As we have learnt that knowledge base is a set of representations of facts/information about the surroUundin,
environment (real world). Every single representation in the set is called as a sentence and sentences a,
which is
expresses with the help of formal representation language. We can say that sentence is a statement
on language, ?
set of words that express some truth about the real world with the help of knowledge representati
Declarative approach of building an agent makes use of TELL and ASK mechanism. |
o TELL the agent, about surrounding environment (what it needs to know in order to perform some action
TELL mechanism is similar to taking input for a system.
© . Then the agent can ASK itself what action should be carried out to get desired output. ASK mechanism|
similar to producing output for a system. However, ASK mechanism makes use of the knowledge base t
‘ decide what it should do.
TELL and ASK mechanism involve inference. When you run ASK function, the answer is generated with the help,
; knowledge base, based on the knowledge which was added with TELL function previously.
o TELL(K) :Is a function that adds knowledge K to the knowledge base.
© _ASK(K) :Is a function that queries the agent about the truth of K.
An agent carries out following operations: First, it TELLs the knowledge base about facts/information it perceive
with the help of sensors. Then, it ASKs the knowledge base what action should be carried out based on the inpy
it has received. Lastly, it performs the selected action with the help of effectors.
3.1.1 Architecture of a KB Ageni
Knowledge based agents can be implemented at three levels namely, knowledge level, logical level an
implementation level.
Knowledge level . Logical level
Implementation level
See ae oS
Knowledge level :
It is the most abstract level of agent implementation. The knowledge level describes agent by saying what
knows. That is what knowledge the agent has as the initial knowledge.
Basic data structures and procedures to access that knowledge are defined in his level. Initial knowledge
knowledge base is called as background knowledge.
Agents at the knowledge level can be viewed as an agent for which one only need to specify what the aget
knows and what its goals are in order to specify its behaviour, regardless of how
it is to be implemented.
For example :A taxi driving agent might know that the Golden Gate Bridge connects San Francisco with
the Mati
county.
Logical level :
At the logical level, the knowledge is encoded into sentences. This level uses some formal language to represef
the knowledge the agent has. The two types of representations we have are propositional logic and first order¢
predicate logic. :
Both these representation techniques are discus
sed in detail in the further sections.
For example: Links(Golden Gate Bridge, San Francisco, Marin County) “tte
Imptementation level :
In implementation level, the physical
representation of | Ogical level sentences is done. This level also describ
data structures used in knowledge base and algorithms t
hat used for data manipulation: : £4 Dene
i 74
AlaSC. (MU Som Comp) :
3-3 Knowledge, Reasoning and Planning
- Forexa mpl
: Links(G e
olden
Sr Sn
ee el a Gate Brid
“Rin etion KB mney Be, San Francisco , Marin Co unty)
~ Agent (percept) retu ts ani action! | Sarge eens
“static: KB, a knowledge base
_ , @ counter, initially 0,
indicating time
oe Tew (KB. MAKE — Penceyy
“SENTENCE (perce
pt, t)) .
ee action ASk (KB, MAKE-AcTi
; on-Quey ty(t))
LL(KB, MAKE-ACTION “SENTENC
pers
E (actio n,t))
pe ted
rem
returns action
RST
Fig. 3.1.2 : General function of

knowledge based agent
- Fig. 3.1.2 is the generalimplementati On of knowledge
" implemented to perform based agent. TELL and ASK are the sub procedur
the respective actions, es
The knowledge base agent
must he able to perform fol
lowing tasks :
° Represent States, actions,
etc,
° Incorporate new Precepts
, —
° Update internal represent
ations
of the world.
° Deduce hidden properties of
the world.
o Deduce appropriate actions.
3.2 The WUMPUS World Environment
You have learnt about vacuum world problem,

block world problem so far. Similarly we
have WUMPUS world
problem. Fig.3.2.1 shows the WUMPUS world.
WUMPUS is an early computer game also known as “Hunt the
Wumpus” - WUMPUS was developed by
1972/1973. It was originally written in BASIC (Beginner's All-purpose
Gregory Yob in
Sy mbolic Instruction Code).
WUMPUS is a map-based game. Let's understand the game :
© WUMPUS world is like a cave, which represents number of rooms,
rooms, which are connected by Passage ways.
We will take a 4 x 4 grid to understand the game.
© WUMPUS is a monster who lives in one of the rooms of the cave. WUMPUS eats the player (agent) if player
(agent) comes in the same room. Fig. 3.2.1 shows that room (3, 1) where
WUMPUS js Staying.
0 Player (agent) starts from any random position in cave and has to explore the c ave, We are sta
rting from (1, 1)
position.
There are various sprites in the game like pit, stench, breeze
, gold, and arrow. Every sprite has some feature, Let'
Understan this s
d one-by-one :
9 Few rooms have bottomless pits, which trap the player (agent)
if he comes to that room. You can see in the
Fig. 3.2.1 that room (1,3), (3,3) and (4,4) have bottomless pit. Note that
even WUMPUS can fall into a pit.
© Stench experienced In a room, which has a WUMPUS in its neighb
ourhood room, See the Fig. 3.2.1, here room
(2,1), (3,2) and (4,1) have Stench.
° Breeze js experienced in a room, which has a pit In Its
nelghbourhood room, Fig. 3.2.1 shows that room (1,2),
(1,4), (2,3), (3,2), (3,4) and (4,3) consists of Breeze,
° Player (Agent) has arrows and he can shoot these arrows in straight
line to kill WUMPUS.
“— ‘.
3-4
WF _Algsc (MU-Sem. 7-Comp) ws that room
ists of gold , this roo m glitters. Fig.3.2.1 sho
° One of the rooms cons ch are; Bump and scream. A bump jig}
ept two types of percepts whi
s player (agent) can acc
Apart from above feature: a sad scream ¢
walks into a wall. While
generated if player (agent )
is killed.
‘sO s
DAA AAEMA Breeze
AStenchws tt
NY C
TT a
cBreeze" yA i
I< xan
DPAAAAM Breeze
ree
3 Ve “Stench COS
Nyy
7
Gold)
DAAABAMA yu »
AStenchw
2 wStenchys cBreeze"
% san yu
1 " Breeze CER Breeze
Start
1 2 3 4
Fig. 3.2.1 : The WUMPUS World
3.2.1. Description of the WUMPUS World
— Anagent receives percepts while exploring the rooms of cave. Every percepts can be represented with the help of five
element list, which is {stench, breeze, glitter, bump, scream]. Here player (agent) cannot perceive its own location.
— If the player (agent) gets percept as [Stench, Breeze, None, None, None]. Then it means that there is a stench anda
breeze, but no glitter, na bump, and no scream in the WUMPUS world at that position in the game. .
— Let's take a look at the actions which can be performed by the player(agent) in WUMPUS World :
‘9 Move: To move in forward direction,
o Turn: To turn right by 90 degrees or left by 90 degrees,
oO Grab: To pick up gold if it is in the same room as the player(agent),
o Shoot : To Shoot an arrow in a straight line in the direction faced
by the player (agent).
- These actions are repeated till the player (agent) kills the
WUMPUS or if the player (agent) is killed. If the WUMP
US is
killed then it is a winning condition, else if the player(agent
) is killed then itis a losing condition and the game is over.
| have some effect. If this shoot action kills the.

WUMPUS then you win the game,
otherwise it reduces the probability
of winning the game.
- Lastly there is a die Q
action : It takes places auto
maticall
: room with WUMPUS. Die acti
on is irreversible,
Goal of the game
Main aim of the gam game e isi that player

(agent) should grab the gold and retu
being killed by the monste rn to starting room (here its (1,1)) with
r (WUMPUS) out
: 4 TechKnawledga
Publications
ward and punishment Points are assigned t ~ Knowledge, Reasoning and Planning
0 a player (A °
as follows: Bent) based on the actions it performs. Points can be given
o 100 points are awarded ifplayer (ag
ent ) come
wae S Out of the cave with the gold.
o 1 pointis taken away for every act
onta ken
o 10 points are taken away Y j if the
arrow Is used
0 (agent) getseelss k killed

200 points are taken away if the playerWorld
WUMPUS
3.2.2 PEAS Properties of
WIE EN enePL skorea
4, Performance measure
- +100 for grabbing the gold and coming back to the starting position
- -—200if the player(agent) is killed. .
— -1 per action,
— -—10 for using the arrow.
2. Environment
Empty Rooms.
— Room with WUMPUS.

are smelly.
— Rooms neighbouring to WUMBPUS which
— Rooms with bottomless pits

.
tomless pits which are breezy
— Rooms neighbouring to bot
is glittery.
— Room with gold which
WUMPUS.
- Arrow to shoot the
otic agent)
3. Sensors (assuming a rob
—~ Camera to get the view
l the stench
— Odour sensor to s mel
listen to th e scre
am and bump
Au di o se ns or to
>
(a ss um in g 4 robotic agent)
4. Effectors
left, right
— Motor to move
gold
~ Robot arm to grab the
arrow
m to shoot the g characteristics
+ .:
Robot mechan!
~ is
owin
The WUMPUS world agent 2 has f°pe! terministic; Episodic .
.
ate g. _ Single agen
1. Fully observable Discr'
4. ‘St ati 5. ,
ete
Static Pu
SF Al&SC (MU-Sem.7-Comp) ___ 3-6 _ Knowledge, Reasoning and Planni
3.2.3 Exploring a WUMPUS World

Fig. 3.2.2 as a reference figure..
Let's try to understand the WUMPUS world problem in step by step manner. Keep
4.1 4.2 4.3 4.4 - Agent
B -—Breeze
3.1 3.2 3.3 3.4 G - Glitter, Gold
, -| OK - Safe square
' P - Pit
24 2.2 2.3 2.4
S$ - Stench
V - Visited
11 1.2 1.3 1.4 W -Wumpus
OK OK
Fig. 3.2.2(a) : WUMPUS world with player in room (1,1)
— The knowledge base initially contains only the rules (facts) of the WUMPUS world environment.
Step 1: Initially the player(agent) is in the room (1,1). See Fig. 3.2.2(a).
The first percept received by the player is [none, none, none, none, none]. (remember percept consists of
[stench, breeze, glitter, bump, scream])
Player can move to room(1,2) or (2,1) as they are safe cells.
Step2: Let us move to room (1,2). See Fig.3.2.2(b).
Al - t
41 4.2 4.3 4.4 ae
B - Breeze
3.1 3.2 3.3 3.4 G - Glitter, Gold
- OK - Safe square
2.1 2.2 2.3 2.4
P - Pit
11 1.2 1.3 1.4 S - Stench

v P?
oK V - Visited
B
OK W -Wumpus
Fig. 3.2.2(b) : WUMPUS world with player in room (1,2)
- Asroom (1,1) is visited you can see “V” mark in that room. The player receives following percept :[none, breeze, none
none, none]. ;
- _ As breeze percept is received room (1,2) is marked with “B” and it can be predicted that there is a bottomless pit!
the neighboring room. ‘ ‘ ’
— You can see that room (1,3) and room (2,2) is marked with “P?”. So raom (1,3) and (2,2) is not safe to move in Th
player should return to room (1,1)and try to find other, safe room to move to.
W_AI&SC (MU-Sem. 7-Comp)
3-7 Knowledge, Reasoning and Planning
—<——
figure, Step 3:
ee
41 =
42 43 44
Ls
3. 1 =
3.2 33 34
i 22 |ag 2A
14 -
teem
Vv
12 —
B
13
[|
OK OK
Fig. 3.2.2(c) WUMPUS world with player moving back to room (1,1) and then moves to other safe room (2,1).
As seer: in Fig.3.2.2(c). Player in now in room (2,1), where it receives a percept as follows : [stench, none, none, none,
none] which means that there is a WUMPUS in neighboring room (i.e. either room (2,2) or (3,1) has WUMPUS).
NSists of
As we did not get breeze percept in this room, we can understand that room (2,2) cannot have any pit and from step 2
~ we can understand that room (2,2) cannot have WUMPUS because room (1,2) did not show stench percept.
Thus room(2,2) is safe to move in.
Step4: Player receives [none, none, none, none, none] percept when it comes to room (2,2). From Fig. 3.2.2(d) you
can understand that room (2,3) and room (3,2) are safe #0 move in.
41 42 4.3 4.4
3.1 32. (33 34
2.41 22. |23 24

[AK + 4]
Vv
OK a
1.3 1.4
14 12 "PP
Vv B
OK OK
r moving to room (2,2)
Fig 3.2.2(d) : WUMPUS world with playe
ze, none 3.2.2(e).
to room (3,2). Here, pla rec eiv es [ste nch, bre eze , glit ter, none, none} percept. See Fig.
Step5: Let's move yer cept shows
m (3,1 ), (3, 3) and ) canFieldhav3e ofWUM
(4,2pit. PUS. tField 2wsof the per
jess pit if
c apt
Field 1 of the ma . (2,2 sho ws
) andthat roo
(4,2 ) can hav e bot tom les s the per cep sho that roo m (3,2) has
that room (3,1), of this game is to grab the gold and go back to the startin
g
the pla yer gra bs the gold first. As +the aim
gold. So,
3 in. Thus WUMPUS.
being killed by the
position, without
= ‘ ~~ WF Tettnoateays
Publications
Knowledge, Reasoning and Plann,
WF algsc (MU-Sem. 7-Comp)
41 4.2 4.3 4.4
P?
3.1 3.2 OK] 3.3 3.4

; P?
P?
2.4 2.2 23 2.4
1.1 1.2 1.3 1.4

P?
V B
OK OK
Fig. 3.2.2(e) : WUMPUS world with player moving to room (3,2)

Now, we have to go back to the starting position i.e. room (1,1) without getting killed by WUMPUS. From nad
2, 3 and 4 We know that room (1,1), (1,2), (2,1) and (2,2) are safe rooms. so, we can go back to room (1,1) by
following any of the two paths: i.e. (2,2), (2,1), (1,1) or (2,2), (1,2), (1,1).
Step 6: As can be seen in Fig. 3.2.2(f). We will go from room (2,2) to room (2,1) and from room (2,1) to room (1,1).
Thus we won the WUMPUS World game!!!
41 4.2 4.3 4.4
Ww?
P?
3.1 3.2 3.3 3.4
WwW? W?
P? P?
2.1 2.2 2.3 24
A
vy AFA A
OK V
Yh 1.2 134. [14
V B
OK OK V
Fig.3.2.2(f) ; WUMPUS world with player moving back
to room (1,1) with gold
3.3 Logic
Logic can be called as reasoning which is carrled out oriti
sar eview based on strict rule of validity to perform a
specified task, a
In case of intelligent systems we say that any of logic's particu
lar form cannot bind logical representation and
_, feasoning, they are independent of any
particular form of logic.
Make a note that logic Is beneficlal only If the
knowledge Is represented in small extent and
~ Fepresented
in large quantity the logic when knowledge
Is not considered valuable.
Fig. 3.3.1 depicts that sentences
are physical confi gurations of an agent, also It is shows
This means that reasoning is a proc that sentences need sentence:
ess of forming new physical configurations
Wi TechK nowledge
ontliatians
from old ones.
—~—_=—=_——=_{_{_ {ee eeeerrreteete
ee |
___ Representation 2 J
Real world Bfn- - <
8 B
1
Featites I the ss _. Features of the

orld Follows real world .
Fig. 3.3.1 : Corres
pondence between real world and its representation
Logical
8 reasoning should make sure that the new configurations e world that actually follow
worl d that the old confi gurat ions repre sented represent features of th
the features of the ente
3,3.1 Role of Reasoning in Al

_ Fig.3.3.2 shows how logic can tion janguage. There are various
levels to the logic
t fund : be seen as a knowledge representa
and most fundamental type of logic is propositional logic
.
Multi-valued logic t’ | Non-monotonic logic
Probabilistic logic
Fuzzy logic ee trey YF Cli eric
Fig. 3.3.2 : Logic as Knowledge Representation language

1. Next level
level, where rules are values between range of 0 and
- Propositional logic can be considere d at fuzzy logic
whic h first order predicate logic is implemented.
is also called as probabilistic logic level using
non- monotonic logic
higher order logic which are multi-valued and
- Inthis Fig. 3.3.2, there ar e two more levels above building blocks
and tempor al logic respectively. All these types: of logic are basic
levels and they consist of modal logic es. Hence reason ing plays a very
intelli gent system s and they all use reasonin g in order to represent sentenc
of
important role in Al.
e using Rules
3.4 Representation of Knowledg
ally two levels’

- knowledge can be considered to be represented at gener
the facts.
ledg is level describe s
(i) Know e le ve l :Th
g the symbols for representing the objects, which can be manipula
ted in
s with usin
(ii) Symbol level- ‘This level deal
programs. rules:
e following
nted using th
Knowledge can be represe
Logical representations
ions
(b) Production rule representat
(c) Semantic networks — :
(d) Frame representations et
Al&SC (MU-Sem. 7-Comp) _ Knowledge, Reasoning and Plannin,
(a) Logical representation
— The logical representations are mostly concerned with truth of statements regarding the world.
These statements are f
most generally represented using statements like TRUE or FALSE.
.
— Logic is successfully used to define ways to infer new sentences from the existing ones. There
are certain logics that
are used for the representation of information, and range in terms of their expressiveness.
There are logic that are 4
more expressive and are more useful in translation of sentences from natural languages into the logical ones.
There
are several logics that are widely used :-
1. Propositional logic : These are restricted kinds that make use of propositions (sentences that are either
true or
false but not both) which can be either true or false. Proposition logic is also known as propositio
nal calculus;
sentential calculus or boolean algebra.
All propositions are either true or false, For example: .
(i) Leavesare green (ii) Violets are blue.
Sky is blue true yes
Roses are red | true yes
2+2=5 false yes
2. First Order Predicate Logic : These are much more expressive and make use of variables, constants, predicates,
functions and quantifiers along with the connectives explained already in previous section.
3. Higher Order PredicateLogic : Higher order predicate logic is distinguished from first order predicate logic by
using additional quantifiers and stronger semantics.
4. Fuzzy Logic : These indicate the existence of in between TRUE and FALSE or fuzziness in all logics.
5. Other Logic : These include multiple valued logic, modal logics and temporal logics.
(b) Production rule representation
One of the widest used methods to represent knowledge is to use production rules, it is also known as IF-THEN rules.
Syntax:
Example :
IF pressure is high, THEN volume is small.

IF the road is slippery, THEN driving is dangerous.
— Some of the benefits of IF-THEN rules are that they are modular, each defining a relatively small and, at least in
principle, independent piece of knowledge. New rules may be added and old ones deleted usually independently of
other rules. :
— Production rules are simple but powerful forms of representing knowledge, they provide flexibility for combining
procedural and declarative representations in a unified manner. The major advantage of production rules are that
they are modular, independent of other rules with the provision for addition new rules and deleting older ones.
By TechKnowledgi
wantie PIONS
AlgsC MU-Sem. 7-Comp) ' nin
-
a Knowledge, Reasoning and FIST
mantic networks
( <)
qhese represent knowledge in the form of gra
: Phical
,
;
since graphs are easy to be
ar
stored inside programs as
e networks
they are concisely represented by nodes and dges.
.
a semantic network basically comprises and
of nodesthat are named and represent concepts,
jabelled links representing; relations between conc pts. Nodes re S.
ror example, the semantic network in Fig, 3.4.1. expresses th e knowle
Sates acta
dge to represe nt the following data :
Tomis a cat.
8
Tom caught a fish.

8
Tom is grey in color.

8
Tom is owned by Sam.

8
Tom is a Mammal.
Gc
Fish is an Animal.
G0
Cats love Milk.

oo
All mammals are animals.
Fig. 3.4.1
has a finite,
network, introduced by John Sowa,
Graph : It is a rec ent scheme used fo r semantic
Conceptual previous
- cep ts or conceptual relations. It differs from the
bipart ite graph. The no des represent either con cat color is grey can
connected, Brothers or
labell e d arcs . For exa mpl e: Ram, Laxman an d Bharat are
method that it does not use
be represented as shown.
Fig. 3.4.2
(d) ) F Frame representation Minsky in 1975 They are mostly used when the task becomes quite complex
7 :
irement of
This concept was introduced by Marvin jon. More structured the system becomes more would be the requ of
es that consists of a collection
and needs more structured representatio | Generall y frames are record like structur
using frames which would prove beneficial. t values.
names and values (subfields) called as facets. Facets can have names
Slots or attributes and the corresponding _
he slots have rson Ram,
Slots can be of any size and type
: Tsho i n In the Fig . 3.4.2 for a pe Ww TechKnowledgé
or numbers too. A simple frame is wn Pudlications
Al&SC (MU-Sem. 7-Comp) satire wodse:. Reasoning and Plannif,
Oo (Ram)
( PROFESSION (VALUE professor)
0
(AGE(VALUE 50))
0
(WIFE(VALUE sita))
o
(CHILDREN(VALUE luv kush))

06
(ADDRESS (STREET(VALUE 4C gb road)))

0
CITY(VALUE banaras))
QQ
(STATE(VALUE mh))
O09
(ZIP(VALUE400615))
o
3.4.1 Ontology
Ontology is study about what kind of things or entities exist in the universe. In Al, ontology is the specification of
conceptualizations, used to help programs and humans to share knowledge about a particular domain. In turn,
ontology is a set of concepts, like entity, relationships among entities, events that are expressed in a uniform way in
order to create a vocabulary for information exchange.
An ontology should also enable a person to verify what a symbol means. That is, given a concept, they want to be able
_ to find the symbol, and, given the symbol, they want to be able to determine what it means.
— Typically, it specifies what types of individuals will be modelled, specifies what properties will be used,
and gives some
axioms that restrict the use of that vocabulary. Ontologies are usually written independently
of a particular
application and often involve a community to agree on the meanings of symbols
~ For example: Consider a map showing hotels, railway station, buildings, schools, hospitals
in a particular locality. In
this map the symbols used to indicate these entities are enough to describe them.
Hence the community who knows
the meaning of these symbols can easily recognize it. Hence that becomes
ontology of that map. In this ontology, it
may define a building as human-constructed artifacts.
It may give some restriction on the size of buildings so that shoeboxes cannot be
buildings or that cities cannot be
buildings. It may also state that a building cannot be at two geographically
dispersed locations at the same time.
3.5 Propositional Logic (PL)

Propositional Logic(PL) is simple but powerful for some artificial intelligence problems. You have learnt simple
mathematical logic in which uses atomic formulas are used. (Atomic formula is a formula that has no strict
sub-formulas). Atomic logic formulas are called propositions.
— In case of artificial intelligence propositional logic is not
categorized as the study of truth values, but it is based
relativi on
ty of truth values. (i.e. The relationship between the truth value of
one statement to that of the truth value of
other statement) ”
:
3.5.1 Syntax
Basic syntax followed by the propositional logic can be given as follows:

meses: |
— Propositional symbols are denoted with capital letters like: A, B, C, etc.
4
—_ Propositional logic constants have a truth value
generally truth values have a crisp nature (i.e. O
(false) and 1 (true). - 4 1
But for fuzzy logic truth values can vary in the
range of 0 and 1.
_"
- Propositional logic make use of wrapping parenthesis
while writing atomic sentence. It is denoted as '(....}'.
—_. Literal is an atomic sentence or it can be negation
of atomic sentence. (A, ~A)
Wy TechKnowtedge Soman ee i dst +
MU-Sem.7-Comp) a
Asscl smitefico thenrshisetbenish
;
3-13 Knowledge,
oo
Reasoning and Planning
ifA isa , e.
propositional logic makes use of relationships

. between ti aoe
Pro Positions and it is denoted by connectives, if A and B a re
propositions. Connectives used in Proposition |
Tabl
able 3.5.1: Bic can be seen in the Table 3.5.1. .
Connectives used in Propositional logic
A !
_ _ AB Conjunction
ee
eee _
Vv Or 7
- - AvB Disjunction
———
= = — = _
a Not
- “A Negation
> Implies A>B Implication/ conditional
oe is equivalent/ if and only if ASB Biconditional
- To define logical connectives truth tables are used. Truth table 3.5.2 shows five logical
connectives.
Table 3.5.2
Bo AB
False false false False true true true
False true false True true true false
True false false True false false false
True true _ true True false true true
- . Jake an example, where A a B, i.e. Find the value of A A B where A is true and B is false. Third row of the Table 3.5.2
shows this condition, now see third row of the third column where, A A B shows result as false. Similarly other logical
connectives can be mapped in the truth table.
35.2 Semantics
~ World is set of facts which we want to represent to form propositional logic. In order to represent these facts
Propositional symbols can be used where each propositional symbol’s interpretation can be mapped to the real world
feature,
~ Semantics of a sentence is meaning
of a sentence. Semantics determine the interpretation of a sentence. For
in following manner:
example :You can define semantics of each propositional symbol
d;
Ameans “It is hot”
2 B means “It is humid”, etc.
Sentence , idered true when its interpretation in the real world is true. Every sentence results from a finite
S considered tr
Number of y f the rules. For example, if A and B are sentences then (A B), (A v B), (B > A) and (A <> B) are
sages of the . (
UF Techknowiedys
Pudticati
.
ons —
- If truth values of all symbols in a sentence are given then it can be evaluated for determining its truth value (ie, we.
can say if it is true or false). 3
3.5.3 What is Propositional Logic ?
- AABand BAA should have same meaning but in natural language words and sentences may have different Meanings
Say for an example,
1. — Radha started feeling feverish and Radha went to the doctor.
2. Radha went to the doctor and Radha started feeling feverish.
— Here, sentence 1 and sentence 2 have different meanings.
— In artificial intelligence propositional logic is a relationship betwee
n the truth value of one statement to that of the
truth value of other statement.
3.5.4 PL Sentence - Example

Take example of a weather problem.
— Semantics of each propositional symbol can be defin
ed as follows:
© Symbol Ais a sentence, which means “It is hot”.
© Symbol! B is a sentence, which means “It is humid
”,
9 Symbol Cis a sentence, which means “It is rainin
g”.
- We can also choose symbols which are easy to
understand, like:
oO HT for “It is hot”.
Oo HM for“It is humid”.
oO RN for“It is raining”.
- Ifyou have HM->HT, then that means “If it is humid
, then it is hot”.
— Ifyou have(HT A HM) > RN then it means “If
it is hot and humid, then it is raining” and so on.
assignments of true or false values for

Sentence A, B and C. Then verify the truth
table for the validity. There can be
total 8 possibilities as shown below:
Sentence HT Sentence HM Sentence RN. Validity -

False - False False Valid
__ False False True Valid
__ False True _ False Not Valid
False True - True Not Valid
True ; False False Valid
True False True Valid
True - True False Not Valid
True True True Valid
— Now, if the knowledge base is [HM, HM-
>HT, (HTAHM) — RNJ(i.e. [“It is humi
hot and humid, then it is raining”) ), then d”, “If it is humid, then it is hot” , “If
“True -True - True” is the only possible it is 4
valid model.
Publications
7-Comp)
3.$C (Mu-Sem.
ion —
y and Contradict :
qautolod ions. For example
means valid sentence. It t isi a sentence which is true for all the interpretat
autolo
“it is hot or It is not hot” ions. For
i nterpretat
-
(“Aor not A”): for all the

tradiction means: an inconsistent sentence. . ItIt isi a sentence whic h is false
ample : if,
con
AAA (“A and not A”): “itis hot and it is not hot.”
example : be Tr ue- For Ex
ence X is True, sentence Y
wil l
Y. It
_ Xx entails Y,
is shown as X | |= means that whenever sent Y).
Y = Priya is Pooja's Aunty. . Then en X |= = ¥ ( X entails
_ xe Priya is Pooja's Mother's Sister and
les
3.5.5 inference Ru en A= C. You must
have come
th
can infer that
n
nce. For example : If A= Ba dp =C
ed with the logical infere = C” then we
_ New sentences are form e
ledg ba se ha s “A = 8B ” an d “B
times it implies that if know
across this example many
~
ces of the
“N= Cc’.
o wi ng th e se t of senten
can be create by logically foll
says that new sentence
jn short inference rule
se.
knowledge ba
Table 3.5.3 : Inference Rules
Modus Ponens
Substitution
Chain rule
AND introduction
Transposition
ented as : KB |-Q.
as: KB |= Qa nd Derivati on is repres
ented
Entailment is repres
of inference rules :
There are two types
1, Sound inference
2. Complete inference
cols
dge base ” using given set of proto
Sound inference s de ri ved from the knowle
_
that, if “X i ented as : “lf KB |- X

:
pro p er ty of in ference say5 s e”. So un dn es s prope rty can be repres

Soundness owledge ba
, the n “x js ent ailed by kn
of inference that
A— BI], from this we can conclude
then KB |[=X”. k no wl ed ge base has [A,
(MP) rule we
assume that
Po ne ns e:
- For Mo du s g tru th tabl
can have B. see followin
knowledge base
TechKnowledge
Pusticarions
Al&SC(MU-Sem.7-Comp)
In general,
For atomic sentences pj, p;', and q, where there is a substitution © such that «
SUBST (@,p) = SUBST (©, p;) for alli,
, PrP (p,A pA pA. es, a P. > q)

P, P,
SUBST (©, a)
N + 1 premises = N atomic sentences + one implication.
Example :
A : itis rainy.
B_: I will stay at home.

A-B : If it is rainy, | will stay at home.
Modus Tollens
When Bis known to be false, and if there is a rule “if A, then B,” it is valid to conclude that A is also false.
2. Complete inference
— Complete inference is converse of soundness. Completeness property of inference says that, if “X is entailed by
knowledge base” then “X can be derived from the knowledge base” using the inference protocols.
— Completeness property can be represented as: “ If KB |= Q then KB |- Q”.
3.5.6 Horn Clause
Clauses are generally written as sets of literals. Horn clause is also called as horn sentence. In a horn clause a
conjunction of 0 or more symbols is to the left of “>” and 0 or 1 symbols to the right. See following formula :
A,AA,AA; «.. A,B, where n>=0 and m is in range{0,1}
There can be following special cases in horn clause in the above mentioned formula :
o Forn=0 and m=1: A (This condition shows that assert A is true)
o Forn>0 and m=0:AA Bo (This constraint shows that both A and B cannot be true)
o Forn=0and m=0: (This condition shows empty clause)
Conjunctive normal form is a conjunction of clauses and by its set of clauses it is determined up to equivalence. For a.
horn clause conjunctive normal form can be used where, each sentence is a disjunction of literals with at most one
non-negative literal as shown in the following formula : —-A,v3A,V—A; ... VAA,VB
This can also be represented as : (A > B)= (—A vB)
Significance of horn logic
Horn sentences can be used in first order logic. Reasoning processes is simpler with horn clauses. Satisfiability ofa
propositional knowledge base is NP complete. (Satisfiability means the process of finding values for symbols which will
make it true).
For restricting knowledge base to horn sentences, satisfiability is in A. Due to this reason, first order logic horn ©
sentences are the basis for prolog and data log languages.
Let's take one example which gives entailment for horn formulas. —
Find out if following horn formula is satisfiable?
(true—X) A (XAY—Z) A (Z |W) A (ZAW-> false) A (true Y)
reer
eer
;
mp)
AlgSC (MU-Sem. 7-Co nnit
: :
3-1
Knowledge, Reasoning and Pla
rom the above equY,ation,
we entail if that
E true s : if the que TY atom j et Equation shows that there are clauses which state
-
assign X and Y to true y
tne w ane r Sowe mican alue (i.e. . true 3X AY).
that all premis e. After that we
Then we can say pre es of XAY—»z a orm ation we can assign Z to tru
re true , bas ed on this inf
vansee all premises of ZW are true, so truwee can assissign W tot rue,
of ZAW —false are horn
As now all premises ry atom is false. Therefore, the
fro .
Mm this we can entail that the que
,
aimee
.
iable.
fo mula is not satisf
45.7 propositional Theorem Proving
A
sequenc e of sentences form a “Proof”, se mis e or it can be a sen ten ce d erived from earlier
the inf entence can be pre to prove is called as a query OF
@ goal.
sentences in the proof based on ere nce rul e. Wh at ev er we wan t
query/goal is the last sentence of the theorem in the proof OT.
Take Example of the “weather problem” which we have seen ab ove.
oO HT for “It is hot”.
o HM for “It is humid”.

o RN for “It is raining”.
“It’s humid” . :
| HM Premise (initial sentence)
" “If it’s humid, it’s hot”

2, | HM HT Premise(initial sentence)
1 and 2) “It’s hot” : "

3, | HT Modus ponens(1,2) (sentence derived from
|
“If it’s hot and humid, it’s raining”
4. | (HT AHM)>RN Premise(initial sentence)
“It’s hot and humid”

5. | HT AHM And introduction(1,3)
5) “It’s raining”
te n ce derived from 4 and
6. | RN Modusponens(4,5)(sen
ositional Logic
35.8 Advantages of Prop language.
||
ple kn ow le dge representation
Propositional log ic is a sim based problems.
artificial intelligence
- |
olving some
a nd ef fi ci en t technique for s r Logic (FOL), etc
.
- Itis suffic ient
hi gh er log ics like First O rde
the foundation for
~ Propositional logic forms decidable.
and reas' oning is
~ Propositional logic is NP complete
d by PL.
can be illustra te
~The process of inference
ositional Logic
45.9 Disadvantages of Prop ence Pr oblems.
rt if icial intellig
express complex a t WUMPUS hunter problem.
Propositional logic is cannot
~
ma ll wo rl ds , think abou
for even $5 problems, it can be very
impr actical complex artificial intelligence
~
Propositional logic can be i c to ex pr es s
try to ma ke use Of P ropositional log .
~ Even if we .
Wordy and lengthy. janguaBe
because :
l e : If there are entiti
es like : Priya,
t a t i o n ’. Fo r e x a m p
PLis a weak knowledge repre se n
d entity is “indivi du al
if the use
With PL it is hard to ide nt TechKnowledge
ify
-© Publications
Mumbai, 123, etc.
AI&SC (MU-Sem. 7-Comp) 3-18 pen itoige, Reasoning anid Plann
© PL cannot directly represent properties of individual entities or relations between individual eNtitieg Fat
example, Pooja is tall.
© PL cannot express specialization, generalizations, or patterns, etc. For example: All rectangles have 4 Sides.
3.6 First Order Predicate Logic
Because of the inadequacy of PL discussed above there was a need

for more expressive type of loale: Thus First-Order
Logic (FOL) was developed. FOL is more expressive than
PL, it can represent information using relations, variables
and
quantifiers, e.g., which was not possible with
propositional logic.
° §6“Gorilla is Black” can be represented
as :
Gorilla(x) > Black(x)
°o 6 6(“It is Sunday today” can be represente
d as :
today(Sunday)
— First Order Logic(FOL)is also called as First
Order Predicate Logic (FOPL). Since FOPL
is much more expressive as3
knowledge representation language than
PL it is more commonly used in artificial
intelligence.
3.6.1 Syntactic Elements, Semantic and
Syntax
= FOL symbol can be a constant term, a
variable term or a function.
— Assuming that “xX” is a domain of value
s, we can define a term with followin
g rules :
1. Constant term: It is a term with fixed
value which belongs to the domain.
2. Variable term: It is a term, which can
be assigned values in the domain.
3. Function : Say “f” is a function of
“n” arguments. If we assume that
t vty .t, are terms then f(t,t,, .., t,) is also
called as a term.
— Allthe terms are generated by appl
ying the above three Protocols.
are quantified is called as a “well-formed

formula”.
oO Every ground term is mapped with
an object,
© Every condition (predicate) is
mapped to a relation,
© Aground atom is considered as
true if the predicate’
°
— Example: The sentence “If

the bag is of blue colour,
| will buy it.”
Quantifiers
Apart from these connec

tives FOPL make S use 4
of quantifiers. As the name
variables taking part in suggests they quantify the
the relation or obeying
the rule,
number of
aigsc (MU-Sem. ..
‘y’ Reasoning and Planning
uantifier Knowledge,
universal Q
pronounced as “for all” and it is applicable
to all the Variable
wyx A” means A is true for every replacem s in the predicate
ent of x,
Example: “Every Gorilla iIs Black”
can be repr
esented as
“/x (Gorilla(x) > Black(x)) :
existential Quantifier ‘5°

= Pronounced as “there
exists”
“ax A” means A is tru
e for at least one re
placement of x.
— Example: “There is a white dog” can ber
3 x (Dog(X) presented as,
* white(x))
ly => isthe main connective with

:
ality : term; == termais true under a given interpretation if and only if tennene termyrefer tot the same obje
1. | PLeannot represent small worlds like vacuum cleaner | FOL can very well represent small worlds’ problems.
world.
= 2
Ps2 | PLis a weak knowledge representation
. language — | F OL is a strong way of representing language.
3. | Propositi I La e uses propositions in which FOL uses predicated which involve constants, variables,
Positiona nguag ions, relations.
|__| the complete sentence is denoted by a symbol. Functi
«| PLcanne erties of individual FOL can directly represent properties of individual entities
-| gp cannot directly represent oh idual entities. e.g. | or relations between individual entities using individual
“ntities or relations between indivi predicates using functions. E.g. Short(Meera)
| Meera is short. . ; hat nace
3 ly A fla eralizations, or | FOL can express specialization, generalizations, or
"| PL cannot express specialization,gen patterns, etc. Using relations.
Patterns, etc
:a . f sides(rectangle,
si 4
&B. Allrectangles have 4 sides. E.g. NO_Of_ wee : )
6 icherens FOL is a higher level logic. .
: isa
tationj i
level logic. esent complex .
FOL can represent complex statements.
'S Not sufficiently expressive to repr
| Statements.
ne er :
oo“ — Wines
‘ 3-20 Knowled ge, Reasoning and Plann
‘Propositional logic (PL).

wees
objects, rela tion:

8. | PL assumes the world contains facts FOL assumes the world contains s
functions like natural language.
: anda
9. | In PL Meaning of the facts is context-independent | In FOL Meaning of the sentences is context dependent |i,
. e
unlike natural language. natural language. - ——_|

10. | PLis declarative in nature. derivative e inin natunature.
FOL isis derivativ yl
3.8 InferenceinFOL —_— at — ——__
3.8.1 Forward Chaining
— For any type of inference there should be a path from start to goal. When based on the avallable data a decision
js
taken, then the process is called as the forward chaining. Forward chaining or data-driven inference works from an
initial state, and by looking at the premises of the rules (IF-part), perform the actions (THEN-part), possibly updating
the knowledge base or working memory. This continues until no more rules can be applied or some cycle limit is met, —
— For example, “If it is raining then, we will take umbrella”. Here, “it is raining” is the data and “we will take umbrella’ is
a decision. This means it was already known that it's raining that’s why it was decided to take umbrella. This process
js
forward chaining.
Fig. 3.8.1 : Forward chaining,

— “Forward chaining’ is called as a data-driven inference
technique.
Example
Given :
— Rule: human(A) > mortal(A)
— Data : human(Mandela)
— To prove : mortal(Mandela)
Forward Chaining Solution

— Human (Mandela) matches Left Hand
Side of the Rule. So, we can get A = Man
dela
— based on the rule statement we can get
: mortal(Mandela)
— Forward chaining
star twihe-en dl is used by the “design expert systems” , as it performs
operatioion n iin a forward dire
irectio (ie. fr fromm
ction n (ice.
Example
— Given facts are as follows:

1. It ls acrime for an Americ
an to sell weapons to
t he enemy of Americ
2. Country Nono Is an enemy of Americ a.
a.
3. Nono has some missiles,
W TachKnowledga
Fublications
—_—4, issi were sold to Nono by Col
Allthe missiles 3.2 1 Knowledge: Peascniniy ani Plann
Missile is a Weapon. Nel West, ,
5.
g, Colonel West is American,
we have to prove that West is a criminal,
et’s see how to represent these facts by FOL.
4, Itisa crime for an American to sell we
ponst
American(x) A Weapon(y) A sell (x y 0 the enemy nations.
)A enemy(z, Amer
2. Country Nono is an enemy of America, ica) => Criminal
(x)
Enemy (Nono, America)
3, Nono has some missiles,
oO Owns (Nono, x)
o ~~ Missile(x)
4, Allthe missiles were sold to Nong
by Colonel West
Missile(x)A owns(Nono, x)=> Sell(
West, x, Nono)
5. Missile is a weapon.
Missile(x)=> weapon(x)
6. Colonel West is American.
American (West)
Proof by forward chaining

The proof will start from the given facts. And as we can derive other facts from those,
it will lead us to the solution.
Please refer to Fig. 3.8.2 As we observe from the given facts we can reach to the predicate Criminal
(West).
rican Wea
ning
Fig. 3.8.2 : Proof by forward chai
88.2 Backward Chaining
fhased on the decision the initial ei is te, andd, by then

fetche it is called as backward chaining. Backward chaining or goal-
looking at the working memory to see if goal already there. If not
driven inference works towards a final sta hat will establish goal, and set up sub-goals
for achieving premises of the
look at the actions (THEN- -parts) of rules t e
be applied, appl y to achiev goal state. +
rules (IF-part).This continues until some rule can TH Teatoratay
Publications
Al&SC (MU-Sem. 7-Comp) 3-22 Knowledge, Reasoning and Plann
For example, If while Boing out one has taken umbrella. Then based on this decision it can be guessed that it
is Fainin
Here, “taking umbrella” is a decision based on which the data is generated that
“it's raining”. This process is backwarg
chaining.“Backward chaining” is called as a decision-driven or goal-driven inference technique.
j 4
Fig. 3.8.3 : Backward chaining

- Given:
o Rule: human(A)
— mortal(A)
o Data : human(Mandela)
- Toprove: mortal(Mandela)
Backward Chaining Solution

— mortal (Mandela) will be matched with mortal (A) which gives human (A) i.e. human (Mandela) which is also a given
fact. Hence proved.
— It makes use of right hand side matching. backward chaining is used by the “diagnostic expert systems”, because it
performs operations in a backward direction (i.e. from end to start).
Example
Let us understand how the same example used in forward chaining can be solved using backward chaining.
Proof by backward Chaining

The proof will start from the fact to be proved. And as we can map it with given facts, it will lead
us to the solution.
Please refer to Fig. 3.8.4. As we observe, all leaf nodes of the proof are given facts that means “West is Criminal”.
Weapon Sal (westx
issile (x) | | Missile (x).
True True True

Fig. 3.8.4 : Proof by backward chaining
3.8.3 Differentiate between Forward Chaining and Backward Chaining
Also known as: Goal-driven , Data-driven

Starts from Possible conclusion Pei New data
Processing Efficient ; Somewhat wasteful _
Aims for Necessary data Any Conclusion(s)
TechKnowledga ~ 7 -
Pubdsicatians
——
—
——= =: ————————
_z—
: Knowledge, Reasoning and Planning
2
90" ch Conservative/Cautious
vacticl if Number of possible final answ i torial explosion creates an ‘infinite
set of known alternatives is Sidiatie reasonable or a | Combina
number of possible right answers. —|
aoe Diagnostic — —
? Tiptio ~
pen and debugging application Planning, monitoring, control and
ApP
- - Interpretation applicatio 7
reasoning qopo™ i reasoning
~ Botto m rees e ”
Type of Search — Depth-first search= ” Breadth-first search
who determine | Consequents determine search
reach Antecedents determine search
Flow _ Consequent to_— antecedent . - .| AntAntecede : uen t

d nt tt to conseq
ex.381: Using predicate logic find the course of Anish’s liking for the followin :
(i) Anish only likes easy courses. .
(i) | Computer courses are hard.
(iii) All electronics courses are easy
(iv) DSP is an electronics course.
Soln. :
Step1: Converting given facts to FOL
h, x)
(i) Wx: course (x) A easy (x) — likes (Anis
(x)
(ii) Wx: course (x) A computes (x) —> hard
(x) — easy (x)
(iii) Vx : course (x) A electronics
liv) Electronics (DSP) |
(vy) course (DSP)

:
Step2: Proof by backward chaining
proof from the same fact.
ch cou rse Ani s h likes. So we will start the
As we have to find out whi
likes (Anish, X)
nae (x)
course (x)
x / DSP
electronics (x)
wil veo (DSP) course (x)
| x18
| x / DSP
electronics (DSP)
course (DSP)
True
True
True
Fig. P- 3.8.1
rse.
Hence proved that Anish likes DSP cou
WF TechMnawledga
PuBli¢akiotes
BF Al&SC (MU-Sem. 7-Comp) 3-24 ___ Knowledge, Reasoning and Planning -
3.9 Knowledge Engineering in First Order Logic | —_

Knowledge engineering is a process of knowledgebase construction. It requires a knowledge engineer to investigate a
particular domain, learn the important concepts in that domain, and create a formal representation and logical relations
among objects in that domain. ‘
3.9.1 Knowledge Engineering Process

Following is the general knowledge engineering process which can be applied to problem of any domain.
1. Identify the task: This step is analogues to PEAS process while designing an agent. While identifying task, the
knowledge. engineer must define the scope of knowledgebase and the range of questions that can be answered
through the database. He also need to specify the type of facts that will be available for each specific problem
instance. ‘
2. Assemble the relevant knowledge: Assembling the relevant knowledge of that particular domain is called the process
- of knowledge acquisition. In this the knowledge engineer needs to extract the domain knowledge either by himself
provided he is the domain expert or needs to work with the real experts of the domain. In this process knowledge
engineer learns how the domain actually works and can determine the scope of the knowledgebase as per the
identified tasks.
3. Defining vocabulary: Defining a complete vocabulary including predicates, functions and constants is a very important
step of knowledge engineering. This process transforms the domain level concepts to logic level symbols. It should be
exhaustive and precise. This vocabulary is called as ontology of the domain. And once the ontology is defined, it
means, the existence of the domain is defined. That is, what kind of things exist in the domain is been decided.
4. Encoding of general knowledge about the domain: In this step the knowledge engineer defines axioms
for all the
vocabulary terms by define meaning of each term. This enables expert to cross check the vocabulary
and the contents.
if he finds any misinterpretations or gaps, it can be fixed at this point by redoing step 3.
5. Encode the problem: In this step, the specific problem instance is encoded using the
defined ontology. This step will
be very easy if the ontology is defined properly. Encoding means writing atomic
sentences about problem instances
which are already part of ontology. It can be analogues to input data for a computer
program.
6. Query the Knowledgebase: Once all the above steps are done, all input
for the system is set and now is a time to
generate some output from the system. So, in order to get some
interested facts inferred from the provided
knowledge, we can query the knowledgebase. The inference procedure will operate
on the axioms and facts to derive
the new inferences. This lessens the task of a programmer to write apptication
specific programs.
7. Debug the knowledgebase: This is the step in which one can prove or check the
toughness of the knowledgebase. In |
the sense, if the inference procedure is able to give appropriate answers to
all the queries asked or it stops in between
because of the incomplete axioms; will be easily identifi
ed by debugging process. If one observes the reasoni
stopping in between or some of the queries could not be answere ng chain
d then, it is an indication of a missing or a weak —
axioms. Then the corrective measures can be taken by repeating
the required steps and system can be claimed
have a complete and precise knowledgebase. to
3.10 Unification and Lifting

3.10.1 Unification —
The processes of finding legal substitutions that make different logical
algorithm is a recursive algorithm; the Problem of u expressions look identical. The unification
; i
nification is: given ind i :
return an MGU (Most General Unifier) of them. : ES
3 "wo atoms, to find if they unify, and, if they “i 4
We TethKnowledge - : ” : ~
‘ Fupiications
put, ae
on “inl - atoms Out
: _ most general ome of t; and tsif it eixiota' or éthorwise
B a‘set of saiiality statements

:‘ Ss ‘substitunon
Ee {h3 =t;}
S20
i)
hile (BEL)
. select and remove x= =y ‘from E
if (vis not identical to xyt then eae
oS if (xi isa variable) then -
< ee swith y even inin» Band s
a ‘else iffoe a variable) ‘then’

: ‘replace ywith x everywhere i in Band s
He Styx} US ee
. - else if (xi is M(x,,...,x,) and vis LV ».- ) hee
2 Poluere: 4X =y,} ee
Unification algorithm for Data log
Example:“x. King(x) A Brave(x) => Noble(x)
King(Ram)
Brave(Ram)
We get an ‘x’ where, ‘x’ is a king and ‘x’ is brave (Then x is noble) then ideally what we want is O= {substitution set}
ie. O= {x/ Ram}
Hence, Ram Unifies x.
3.10.2 Lifting
The process of encapsulating inference rule is called as Generalized Modus Ponens.
Generalized Modus Ponens
For atomic sentences pi, p', and q, where there isa substitution @ such that
SUBST (©,p)) = SUBST (©, pi) for all ;
oo os PL q)
P, iP, 1 Py Py (DA BAA PaAee
SUBST (Oo, =
+ one implication,
Ne Premises = N atomic sentences
the conclusion we seek.
“plying SUBST(8 , q) produces
Ps King(Ram) pie Bravely)
Te
Ww, PschKnowledge
ations
-3- 26 -___ , Reasoning and Pig,
edge,e, Reason
KnowlKnowledg
Pp,= King(x) P= Brave(x)
© = {x/Ram, y/ Ram} q = Noble(x)

SUBST(0,q) is Noble(Ram) 5 from ground (variaige
Generalized Modus Ponens is a lifted version of Modus Ponens. It raises Modus Ponen g 7 fee)
propositional logic to first-order logic. Hence it is called as lifting.
— Here “lifted” indicates transformed from. —
The major advantage of lifted inference rules over :
propositional ind
logic is that only those su bstitutions are mad that
are required so as particular inferences are allowed to proceed. 4
Ex. 3.10.1: Represent following sentences in FOL using a consistence vocabulary.
(i) Every person who buys a policy is smart.
(ii) | No person buys an expensive policy.
(iii) There is an agent who sells policies only to people who are not insured.
(iv) There is a barber who shaves all men in town who do not save themselves.
Soin. :
(i) Wx Vy: person (x) A policy (y) A buys (x, y) > smart (x)
(ii) Wx, Vy: person (x) A policy (y) A expensive (y) 3~ buys (x,
y)
(iii) Wx: person (x) A~ insured (x)
Vy : 3x A policy (y) ~ agent (x) — sells (x, y, x)
(iv)Sx Vy: barber (x) A person (y) A~ shaves (y, y)

shaves (x, y)
Ex. 3.10.2: - Represent following statements in
FOPL.
1. Anyone who kills an animal is loved by no one.
2. A square is breezy if there is a pit in the neigh
bouring squares.
Soln. :
(i) Vx Vy: kills (x, animal) ~ loves (y, x)
(ii) Wx Vy: pit (x, y) Sbreez (x, ¥~ 1) Abree
z (x, y + 1) Abreez (x-1, y) Abreez
(x + 1, y)
Ex. 3.10.3: Write first order logic statements for
following statements :
(i) Ifa perfect square is divisible by
a prime p then it is also divisible
(ii) Every perfect Square is divisible by by square of p.
some prime.
(iii) Alice does not likechemistry and hist
ory.
(iv) Ifitis Saturday and warm, then
sam is in the park,
(v) Anything anyone eats and is not
killed by is food..
Soln. :
(i) Vx: square(x) A prime (y)
A divides (p, x) > (Az: Squ
are_of(z, p) divides(z, x)]
(ii) Vx3y; square(x) A divides
(p, x)
(iii) ~ likes(Alice, History)
A ~ likes (Alice, Chemis
try)
(iv) day(Saturday) A weathe
r(warm) in_park(Sam)
(v) Vx: Vy: Person(x) A
eats(x, y) A “killed (x)
> food(y)
TechKnowtedg
G7 Techinewtedgie
Scanned by CamScanner:
mee? Knowleage, Neaouiiey —_
3B
jon is! a valid inference rulule. Resolution prod
resolution ich ic implied by two clauses containining
. : uces an ew clause which is implied by
t a r y .
liter TI
als. : vered by Alan Robins on in the nid 1960' s .
compleme n ary his resol ution rule was disco .
0
;
we have see n that a lititeral isi an atomic symbol or a Negation of the atomic symbol (i.e.A, ~A)-
ce
resolution is the only interference rule you need, jIn order to bui means that every senten
» i
produced by a procedure will; be “true” ) and com CUE eat aR
“rue ” sentence can be
produced
P ene
plete (complet "tri!
ss means every
bya procedure) theorem proof maker.
Take an example where we are given that :
o Aclause X containing the literal : Z
o Aclause Y containing the literal : =Z
gased on resolution and the information given above we can conclude :

(X-{Z) U (Y - {-2})
Take a generalized version of the above problem :
Given:
- Aclause X containing the literal: Z
- Aclause Y containing the literal: -Y
- Amost general unifier G of Z and -Y
We can conclude that :((X = {Z}) U (Y -{-Y})) | G
3.11.1 The Resolution Procedure

that we
do not have any contradictions, and Z be a sentence
= Let knowledge base be a set of true sentences which
want to prove.
ion (You must
we should assume -Z and then try to find a contradict
- The Idea is based on the proof by negation. So,
if all the knowledge
etry proofs). Then based on the Intuition that,
have followed such methods while solving geom red from knowledge base. Then
createsa contradiction then Z must be infer
base sentences are true, and assuming 4Z
form.
dge base U {-Z} to clause
we need to convert knowle
e th the process after that.
Terminatate
d. Termin
Lae
If there .is a contradict ion i knowledge base, that means Zis proved.
n
we do not find any resolvable
r res olv ent s to the current knowledge base. If
~ Otherwise select two clauses and add thei there is a contradiction in
we termin ate. Else, we have to start finding if
clauses then the procedure fails and then
knowledge base, and so on.
sal Normal Form (CNF)
3.11.2 Conversion from FOL Clau
with =P VO
Replace P 30
Elimination of implication i.e. Eliminat eall‘’:
) with =P A=Q and so on.
Distribute negations:Replace ~-P with P, (PV Q
TechKi on
=
ee
¥ Pub en
;
es ESC IMU-Som.7-Comp) Knowledge , Reasoning and P| lanning
3. Eliminate existential quantifiers by replacing with Skolem constants or Skolemfunctions:
e.g. V XA Y(PA(X,Y) v(P2(X,¥)) VX (P4(X,f(X)) V_ (P2(X,F(X)))
4 Rename variables to avoid duplicate quantifiers.
5. Drop all universal quantifiers
6 Place expression into Cconjunctive NormalForm

.
7. Convert to clauses i.e. separates all conjunctions as
separate clause.
8. Rename variables to avoid duplicate claus
es.
Soln. ::
FOL: A (Bc)
Normalizing the given statement.
(i) A (B>CAc>B)
(ii) (A—(B—>C)) A (A (C— B))
Converting to CNF.
Applying Rule, a8 = ~avB
~Av (~BvC)A~Av (~C vB)
i.e.~ Av ((~B VC) a (~CvB))
3.11.3 Facts Representation
To show how facts can be represen

ted let's take a simple problem:
° “Heads X wins, Tails Y loses,”
°o Our goal is to show that X alw

ays wins with the help of reso
lution,
— Solution can be given as follows:
1. H=> Win(X)
2. T= Loose(Y)
3. =H=> T
4. Loose(Y) > Win(X)
Thus we have : Win(x)

We can write a proof
for this problem as fol
lws:
1. {H, Win(x)}
2. {HT,Loose(Y)}
3, {H,T}
{—Loose(Y), Win(X)}
_ Knowledge, Reasoning and Plentirs
¢ aise (MU-Sem.7EOMP)
—
twin
6 ih Wwin(X)}
(From 2 and 4)
1. {T, Win)
— (From 1 and 3)
5 win(X)}
tease (From 6 and 7)
9. tee (From 5 and 8)
411.4 Example
the same exam learn how to write proofs for resolution.
let stake ple of forward and backward chaining to
step 1:
The given facts are :

tis acrime for an American to sell weapons to the enemy nations.
American(x) A Weapon(y) A sell (x,y,z) A enemy(z, America) => Criminal (x)

7, Country Nono is an enemy of America.
3, Nono has some missiles.
o Owns (Nono, x)
o Missile(x)
Colonel West.
4. Allthe missiles were sold to Nono by
x, Nono)
Missile(x)A owns(Nono, x)=> Sell(West,
5. Missile is a weapon.
Missile(x)=> weapon(x)
6 Colonel West is American.
American (West)
Step
2;
Lets convert them to CNF

erica)
sell (X,Y , z)V ~enemy(2, Am
) V™
1 American(x) V ~Weaponly
V Criminal (x)

Owns (Nono, x)
Missile(x)
l(West, % N ono)
. Missile(x) V~ owns(Nono, x) V Sel
“Missile(x) V weapon(x)
.
=.
Ameri C‘an : Tech Knowledge
Publications
(West)
meen SC (MU-Sem.7-Comp)
Step 3:
To prove that West is criminal using resolution.
~ criminal (West) ~ American (x) v ~ weapon (y) v ~ sell (x, y, Z) V

~ enemy (z, America) v criminal (x)
«/West
~ American (West) v ~ weapon (y) v enemy (Nono, America)
~ Sell (West, y, z) v ~ enemy (z, America)
z/Nono
~ American (West) v ~ weapon (y) American (West)

~ sell (West, y, a
~ weapon (y) v ~ sell (West, y, Nono) ~ Missile (x) v~ owns (Nono, x) —

Vv Sell (West, x, Nono)
~ weapon (y) v ~ Missile (y) v ~ owns (Nono, y) Missile (x)
et
~ weapon (y) v owns (Nono, y) Weapon (x) v ~ Missile (x)
~ Missile (y) v ~ owns (Nono, y)

Missile (x)
~ owns (Nono, y) owns (Nono, x)
\
Hence our assumption was » Wrong. Hence
proved that West iis criminal.
Ex. 3.11.2: Consider following statements :
(a) Ravi Likes all kind of food.
(b) Apple and Chicken are food
(c) Anything anyone eats and is not killed
is food.
(d) Ajay eats peanuts and still alive.
(e) Rita eats everything that Ajay eats.
Prove that Ravi Likes Peanuts using
resolution. What food does Rita
Soln. : eat?
. {A) Proof by Resolution

Step 1 : Negate the statement to be proved.
~likes (Ravi, Peanuts)

Step 2 : Convert given facts to FOL
(a) vx, food (x) => likes (Ravi, x)
(b) food (Apple)

W TechKnowledge
Pubifcatians
Scanned by CamScanner _
GC
anning
oning and Pl
(U-Sem. 7-Gomp) : Knowledge, e, Reas
F food (Chicken) °
my sg xv yreats (x, y) A~ killed (x) —> food (y)
uts) 4 alive (Ajay)

fe) eats (Ajay, Pean
xe eats (Ajay, x) — eats (Rita, x)
in thisis case we have to add few common sense predicate which are always true
(x)
(e) yf x:~killed (x) — alive
(h) ig x alive (x) >~ killed (x)
CNF
Step 3: Converting FOLs to
food (x) v likes (Ravi, x)

(a)
(b) food (Apple)
(c)_ food (Chicken)
food (y)
(d) ~eats (x, y) ¥ killed (x) v
(e) eats (Ajay, Peanuts)
(f) alive (Ajay)

(Rita, x)
(g) ~eats (Ajay, x) v eats
(h) killed (x) v alive (x) —
(i) ~alive (x) v killed (x)

Step 4: Proof by Resolution i, x)
- ~ food (x) V likes (Rav
~likes (Ravi, Peanuts)
x/Peanuts
(y)
killed (x) V food
~ eats (x, ¥) ¥
~food (Peanuts) " y/Peanuts
eats (Ajay, Peanuts)

)
¥ V(xo
~eats (x, Peanuts) x/Ajay
|
~ alive (x) v killed (x)
Kille Co
nn |
tou .
alive (Ajay) |
a
e
~alive e
s out assumption Is
wr on g. He nc e pr ov ed that “Ravi likes
Peanuts”.
|
As the result of thi s re so lu ti on is NIL, it mean
Wy pometent
|
|
|
~ Knowledge, Reasoning and Planni
To answer : What food Rita eats ?
(B) Proof by backward chaining : (Referring to FOLs of Step 2)

eats(Rita, x)
eats(Ajay, x)
x PPoanutiy
eats(Ajay, Peanuts) _alive(Ajay)
True True
Hence the answer is Rita eats peanuts.
Ex. 3.11.3: Using a predicate logic convert the following sentences to predicates and prove that the statement “Ram did
not jump” is false.
(a) Ram went to temple.
(b) . The way to temple is, walk till post box and take left or right road.
(c) The left road has a ditch. .
(d) Way to cross the ditch is to jump
(e) A logis across the right road.
(f) | One needs to jump across the log to go ahead.
Soln.:
Step 1 : Negate the statement to be proved.
~jump (Ram)
Step 2 : Converting given statement to FOL
(a) At (Ram, temple)
(b,) > x: At (x, temple) ——sAt (x, PostBox) a take left (x)
(b,) *% x: At (x, temple) ——At (x, PostBox) A take right (x)
(c) °* x: take left (x) —> cross (x, ditch)
(d) *¥ xcross (x, ditch)

——> jump (x)
(e) *¥ x: take right (x) at (x, log)
(f) *% x: at (x, log) —> jump (x)
Step 3: Converting FOLs to CNF
(a) At (Ram, temple)
(b,,) ~ At (x, temple) v At (x, PostBox)
(b,,) ~ At (x, temple) v take left (x)
(b,,) ~ At (x, temple) v At (x, PostBox)
(b,,) ~ At (x, temple) v take right (x)
7-Comp)
a7 _Al&SC (MU-Sem. _ Knowledge, Reasoning and Planning
(c) ~take left (x) v cross (x ditch)
(d) ~ cross (x, ditch) v jump (x)
(e) ~take right (x) v at (x, log)
(f)~at (4 log) v jump (x)

Step 4 : Proof by Resolution
~jump (Ram) at (x, log) vjump (x)
ie
~At (Ram, log) ~ take right (x) Vv at (x, log)
te
~take right (Ram) ~ At (x, temple) v take right (x)
\
~ At (x, temple) At (Ram, temple)
g
Hence proved.
_
Ex. 3.11.4: Consider following statements.
1. Rimiis hungry.
2. If Rimi is hungry she barks.
Raja is angry.
3. _ If Rimi is barking then
CNF form.
pr edicate logic. Convert them into
Explain statements in
ry using resolution.
Prove that Raja is ang
Soln. :
to FOL.
Step 1: Converting given facts
1. Hungry (Rimi)
2. Hungry (Rini) — barks (Rimi)
3. Barks (Rimi) > angry (Raja)
to CNF.
Step2: Converting FOL statements
1. Hungry (Rimi)
(Rimi)
2. ~ hungry (Rimi) v barks
3. ~barkes (Rimi) V angry (Raja)
Step3: Negate the stmt to be proved
T.P.T. Angry (Raja)
YechKnowledge
Negation : ~ Angry (Raja)
Publications —
AI&SC (MU-Sem. 7-Comp) Knowledge, Reasoning and Plannin.
Step 4: Proof by resolution

~ Angry (Raja)
~ Barks (Rimi)
(Raja)
Vv Angry
a
~ Barks (Rimi)
Barks (Rimi)
ry (Rimi)
v~ hung
a
~ hungry (Rimi)
hungry
a (Rimi)
9
Me
This shows that our assumption is Wrong. Hence proved that Raja is Angry.
Ex. 3.11.5: Consider following facts.
1. If maid stole the jewellery then butler was not guilty.
2. Either maid stole the jewellery or she milked the cow.
3. If maid milked the cow then butler got the cream.
4. Therefor if butler was guilty then he got the cream.
Prove the conclusion (step 4) is valid using resolution.
Soln. :
Step 1 :Converting given facts to FOL.
1. steal (maid, jwellary) >~ guilty (butler)
2. steal (mail, jwellary) v milk (maid, cow)
3. silk (maid, cow) >got_Cream (butter)
To prove that
4. guilty (butler) >got_cream (butter)
Step 2 :Converting FOL to CNF.
1. ~steal (maid, jwellary) v~ guilty (butler)
2. steal (maid, jwellary) v milk (maid, cow)
3, ~ milk (maid, cow) vgot_cream (butler)
4, ~guilty (butler) vgot_cream (butler)
Step 3 :Negate the proof sentence

As sentence 4 is the one to be proved. -
guilty (butler) A~ got_ cream (butler)
35C (MU-Sem. 7-Comp) =. 3
Al&S
sa Knowledge, Reasoning and Planning
git aiproof by resolution Avgot_cream (butter)
guilty (butler)
~ milk (maid, cow)
Y Bot_cream (butler)
guilty (butler) V~ ll cow)
Steal (maid jwellery)

v~ milk (mail, cow)
guilty (butler) v steal (maid, jewellery)

~ guilty (butler)
a V~ steal (maid jwellery)
4
Hence proved.
All happy people smile. :
jepresent these axioms in FOL.

rmul to.CNF. .
TUEUISS GRP
Soln. :
Step 1: Converting axioms to FOL.
i) Vx: graduating (x) > happy (x).

(il) Vx: happy (x) -» smile (x)
tli) 3x : graduating (x)
Step 2: Converting FOL to CNE.
{) ~ graduating (x) v happy (x)
(i) ~ happy (x,) v smile (x,)
i) Braduating (x,)
Sep3: TPT, x, smile (x3)

Negating the stmt
~ Smile (x;)
Proof by resolution
Se : WienPublicatipnas —
_Al&SC (MU-Sem. 7-Comp) et
Knowledge, Reasoning and Plannin
~ smile (x3)
~ happy (x,) v
smile (R,)
X3 |X
~ happy (x,)
~ graduating (x)
V happy(x)
x, | x
~ graduating (x) graduating | (x,) .
x |x,
b
Hence our assumption is wrong.
Hence proved.
3.12 Planning
— Planning various tasks is a part of day to day activities in real

world. Say, you have tests of two different subjects of
one day then, you will plan your study timetable as per your
strengths and weaknesses in those two subjects.
— Also you must have learnt about various scheduling algori
thms (e.g. first in first out) in operating systems subject
anc
how printers plan/schedule their printing tasks based on
tasks importance while printing.
= These examples illustrate how planning is important. We
have seen that artificially intelligent systems are rationa
system s. So, devising a plan to perform actions becom
es a part of creating an artificially intelligent
think about giving intelligence to a system or device agent. When we
, we have to make sure that it priorities between
given activities o
tasks.
— In this we are going to learn about how the machines can
become more intelli gent by using planning
while performin,
various actions.
3:12.1 Introduction to Planning
Planning in Artificial Intelligent can be defined

as a problem that ne eds decision making
by intelligent systems t¢
accomplish the given target.
f
— The intelligent system can be a robot ora computer
program.
- Take example of a driver who has to
pick up and drop people from one place
people from two different places then he has to to another, Say he has to pick up tw?
follow some sequence, he cannot pick both Passengers at same time.
— There is one more definition of planning
which says that, Planning is a n activity
sequence of actions to accomplish target. where agent has to come up wit
h#
- Now, let us see what information Is aval
lable while formulating a plannin
- We have information about the initial statu
s of the agent
Ge TechKnowledya
¥ Al&SC (MU-Sem. 7-Comp) _
- Aim ofan agent is to find 3-37 Knowledge, Reasoning and Planning
. J
the Proper se uen H :
—<—= a as ST
an efficient solution. Guence of actions which will lead from starting state to goal state and produce
3.12.2 Simple Planning Agent

- Take example of an agent which can be a
3 people who have access to this Spene ;
Coffee maker, a printer and a mailing system, also assume that there are
- Suppose at same time if all 3 Users

and sending a mail. of ana gent give a command to execute 3 different tasks of coffee making, printing :
- Thenas per definition of planning, agent has

to decide the sequence of these actio
ns.
Fig. 3.12.1 : Example ofa Planning Problem
Fig. 3.12.2 depicts a general diagrammatic representation of a planning agent that interacts with environment with its
sensors and effectors/actuators. When a task comes to this agent it has to decide the sequence of actions to be taken
and then accordingly execute these actions.
Sensors
af
“~)
Environment
Effectors
C-E)-S
nt
Fig. 3.12.2 : Planning age
3.13 Planning Problem
mning problem
BN erage? torrmuat
information Is avallable while formulating a planning problem and what results
tin
We have seen in above section pia ble here that, states of an agent correspond to the probable surrounding
are expected. Also it Is understanc gent are specified based on logical formalization.
Is ofan a W TechKnowledge
environments while the actions and goa Publications
_ 3-38
Al&SG (MU-Sem. 7-Comp)__ in chapter 1. Whic! h sh“hows that, to achi
eve any Boal an
various types of intellige nt age nts ow it will affect the UPCOMing a
Also we have learnt about its act ion s ”
be t he effect of
agent has to answer few questions like “what will reasoning about it s
future actions, states.
be able to provide a proper
actions”, etc. This illustrates that an agent must
of surrounding environments, etc. ofactions
ste p, he /s he has to follow sequence
in one
Consider simple Tic-Tac-Toe game. A Player cannot win a game
er old steps and has to imagine the probable future -
to win the game. While taking every next step he/she has to consid time he/ she should also consider the
move and at the same !
actions of an opponent and accordingly make the next
consequences of his/her actions. ~
’ A classical planning has the following assumptions about the task envir'
‘onment:
environment.
o Fully Observable :Agent can observe the current state of the
.
© Deterministic: Agent can determine the consequences of its actions
Finite: There are finite set of actions which can be carried o ut by the agent at every
state in order to achieve the
o
goal.
© Static : Events are steady. External event which cannot be handled by agent is not considered.
of time.
© Discrete :Events of the agent are distinct from starting step to the ending(goal) state in terms
So, basically a planning problem finds the sequence of actions to accomplish the goal based on the above
assumptions.
Also goal can be specified as a union of sub-goals.
Take example of ping pong game where, points are assigned to opponent player when a player fails to return the ball
within the rules of ping-pong game. There can a best 3 of 5 matches where, to win a match you have to win 3 games
and in every game you have to win with a minimum margin of 2 points.
3.13.1 Problem Solving and Planning
Generally problem solving and planning methodologies can solve similar type of problems. Main difference between
problem solving and planning is that planning is a more open process and agents follow logic-based
representation.
Planning is supposed to be more powerful than problem solving because of these two reasons
|
Planning agent has situations (i.e. states), goals (i.e. target end conditions) and operations (i.e. actions performed) .All
these parameters are decomposed into sets of sentences and fi urther i ” a
in sets words dependi the need of the |
system. . Pp ding on
Planning agents can deal with situations/states more efficiently because of its explicit reasoning capability
communicate with the world. Agents can reflect on their targets and we can minimize the compl P t Y
also it caf
wg
.
problem by independently planning for sub-goals of an age plexity of the plann
nt. Agents have informati
actions and the important point is that it i
can predict the eff ect of actions by inspecting the
ithe op
opeer
ratat
ions. re
Planning is a logical representation, based on situation, goals and operations, of
problem solvi
’ olving.
Planning = Problem solving + Logical representati
on
3.14 Goal of Planning

—- We 4
ba planne activities in order to achiev e@ some m
goals. Main goala ca n be diviivi ded
into su b goal
into sub- Oa s t Oo make
ke ng mo!
planning x:i
Ww TechKnowledge ~
Poblications ” ’
° : ~
tg algsc (MU-Sem. 7-Comp)
3-39
at super Knowledge, Reasoning and Planning
take example of a grocery shopping rksi
et,st» er
su SUPPose You want to buy milk, brea an eg
then your initial state will be — “at home” a 1. i d d g from supermarket,
nd g a will be ~ “get milk, bread an
d ege”.
and that branch i fact
ing or can be enormous dependin upon the se
of actions, for e.g. Watch TV, read book etc » aVailable at th ; g t
at .
go
: io Kear; point of time.
attend lecture
change channel
increase/decrease.
eos Volume
Fig. 3.14.1 : Supermarket examples to understand need of planning

as in the
e actio ns at any state, set can be very large, such
Thus branching factor can be defined as a set of all probabl will also
of probab le actions increases, the branching factor
supermarket example or block problem. If the domain space.
increas e as they are directly proportional
to each other, this will result in the increase of search
ions, then you have to
state you have to follow many steps, if you consider using heuristic funct
To reach the goal h of
will be helpful only for guiding the searc
that it will not be able to eliminate states; these functions
remember
states. to make sure that all three
to supermarket we need
s diffi cult to choo se b est actions. (i.e. even if we go
- So it bec ome
then goal state can be ach
ieved).
listed items are picked, only re can be combined goals (as seen
is dif fic ult to describ e every state, and the
- Asthere are many possible a ctions
and it
6 oals efficiently. In order to be mo
re efficient planning is
ua te to ach iev e
is inadeq
in supermarket ex ample) searching
required. of planning we need to
quir es expl icit knowledge that means in case
d that planning re
In above sections, we discusse
l.
~ o rder to achieve the goa
l be useful in
ons wil
be same. For example, you
know the exact sequence of acti
WwW! hi ch
and th e order of execution need not
nning
that, the ord er of pla
Advantage of planning is
ket.
- planning to go to supermar
s for Broce ry before Pp olicy by dividing /decomposing the
can plan how to pay bill us e of di vi de and conquer
at you can make
~ Another advantage of planning is th
goal into sub goals.
3.14.1 Major Approaches

: few major approaches used for planning :
are W
pro blems. Following
solve planning
There are many approaches to
space search.
QO Planning with state
g)-
Partial ordered planning. s i t i o n ( HTN plannin
h ical decompo
Hierarchical planning/ hierare
o oO
on calculus/ plan
ng situati
Planni l anning.
Conditiona pl e TechKnowledge
Oo
Publications
Planning with operators.

o
WF Alas (MU-Sem.7-Comp) ___Knowledge,Reasoning
and Plann
o Planning with graphs.
o Planning with propositional logic.
o —- Planning reactive.
- Out of these major approaches we will be learning about following approaches in detail :
© Planning with state space search
Partial ordered planning and
0
Hierarchical planning / Hierarchical decomposition (HTN planning).

0
Conditional planning.
0
Planning with operators.

o
3.15 Planning Graphs - |
— Planning graph is a special data structure which is used to get better accuracy. It is a directed graph : and is usefultp
Also
accomplish improved heuristic estimates. Any of the search technique can make use of planning graphs.
GRAPHPLAN can be used to extract a solution directly.
- Planning graphs work only for propositional problems without variables. You have learnt Gantt charts. Similarly,jn
case of planning graphs there are series of levels which match to time ladder in the plan. Every level has set of literals
and a set of actions. Level 0 is the initial state of planning graph.
Set of literals
e All Literals which are true at that point of time
Depends. on executed actions at previous time step:
Set of actions
e Ail actions which meet prerequisite
at that point of time
Jepends upon which literals are true
Fig. 3.15.1 : Planning graph level
Example
= Init(Have(Apple))
— Goal(Have(Apple) AAte(Apple))
Action(Eat(Apple), PRECOND: Have(Apple)
— EFFECT: -Have(Apple) AAte(Apple))
Action(Cut(Apple), PRECOND: - Have(Apple)
— EFFECT: Have(Apple))
Ly : Ao Ly Ay 35
Cut (Apple)
Have (Apple) ————---}—- Hava (Apple) Have (Apple)
— Have (Apple) — Have (Apple)
Ate (Apple)
- Ate (Apple).
Ate (Apple) Ate (Apple)
— Ate (Apple) ———-—}+——— + Ato (Apple) ~———__{-} ____ — ate (Apple)

Fig. 3.15.2 ae
Ww TethKnowledgé
Punsications
E———=—————
yr
OO—COCLLz———————————————
Al&SC (MU-Sem. 7-Comp) ning

Knowledge, Reasoning and Plan
eee!
start at level LO and determine action level AO and next leve HLL.
AO >> all actions who Se prerequisite
is Satisfied at pr
evious | level.
90008C~"dDhC~«sd
Connect precondition and effect of action Lo-

s >L1,
persistence acti
Inaction is represented by
: possible actions ions.s
Level AO contains the
Conflicts b etween actior
i sare shown by mutual exclusi fon links.
link:
Level L1 contains all lite 7 als that could result from picking any subset of actions in Level AO.
0
Conflicts
exclusion between
links. lite rals which can n ot occur together (as a effect of selection action) are represented by m utual
o°
L1 de ines iultiple stat es and the mutual j k s are th e constr ain ts h that d aetin
exclusision lin fine
i th i sseto states .
| ‘ |
oO
i
Continue until two
_ consecutive levels are the same or contain the same amount of literals.
omrer
NO Ee ae tion of the a
effect of — fone literal is the nega
One action cancels out the
-—
another action. literal OR
- one of the effects of action is negation — |f each possible action pair that could
0 preconditions of other action. achieve the literal is mutually exclusive.
— One of the preconditions of one action is

mutually exclusion with the precondition |
of the other action.
GRAPH PLAN algorithm of following algorithm :

from planning graph with the help
GRAPHPLAN can \ directly extract soluti on
1 lition or ae
nso
Deak all nonemutex i

4: ogee
solution«EXTRACT Sceuhirh
G Hégraph))
n
e then alai
ifsolution# failur
P O S S I B B l e a ! then retum f
N-
Properties of planning graph achieved!.

t he n go al cannot be
last level
~ If goal is absent from t in the last
level.
goal is presen
~ Ifthere exists a path to goal then may not ex ist any
path.
then there
-
If goal is prese nt i
in last Jevel
,
3.16 i ate-Space Search | ks of pri nti ng, sending a mail and mak
ing coffee namely
per for m thr ee tas
#5 S : tcan
Planning
~ fan agent tha e different tasks then, let us
We have seen example 0
. at a same time to perform these.e thre ;
~ lets! call this agent as office 2 gent | ook if we have
a finite spac
So TechKnowledg s
- order fro me prossiem witl
ace seare ca
When this office agent Bet
statee Sp’
Public ations
see how planning with
. :
:
eo
¥ Al&SC (MU-Sem. 7-Comp) 3-42 Knowledge, Reasoning and Planning =

You can understand from Fig. 3.16.1 that the office agent is at location 250 on the state space grid. When he gets a
task he has to decide which task can be performed more efficiently in lesser time. ! a |
Agent's location is 250

-Shraddha wants to print
-Raj wants to send mail to
John throught mailing server
-Gerry wants coffee
Fig. 3.16.1 : Office agent example with finite state space
If it finds some input and output locations nearer on state space grid for example in case of printing task then the
probability of performing that task will increase.
But to do this it should be aware of its own current location, the locations of people who are assigning tasks and the
locations of the required devices.
State space search is unfavourable for solving real-world problems because, it requires complete description of every
searched state, also search should be carried out locally.
" There can be two ways of representations for a state:

1. Complete world description
2. Path from an initial state
Complete world description
— Description is available in terms of an assignment of a value to each previous suggestion.

- Orwecan say that description is available as a suggestion that defines the state.
—. Drawback of this types is that it requires a large amount of space.
~Path from an initial state
woes Fak neledee natin
AlaSC (MU-Sem. 7-Comp) Planning
Knowledge, Reasoning and can be a
=
oe
. Because of this it
— Drawback of these types is that it does
ether two statesnot ar explic i ly
plicit specify eae “What holds in every state”
difficult to determine wh i e same.
Complete World
Fig. 3.16.2 : Representation

of states
3.16.1 Example of State Space Search
Let us take an example of a water jug problem

that
i r has any measuring markers on it. There is a pump
a 4- - gallon one and a 3- gallon one. Neithe
- ee ve twoj
Weha “\ a jug?
get exact 2 gallons of water into the 4- gallon
used to fi l the jugs with water how can you
such that x = 0,1,2,3 or 4,
this prob lem can be descr ibed as of ordered pairs of integers (x, y),
The state space for quantity of water
_— or 3, representing the
of water in the 4- gallon jug and y = 0,1,2
representing the number of gallons (2, n) for any value of n, since the
problem does not
state is (0, 0). The goal state is
in the 3- gallon jug. The start
to be in the 3- gallon jug.
specify how many gallons need bellow. They are represented
as rules
to solv e the pro ble m can be described as shown resu lts from
The operators to be used ribe the new state that
—
against the current state and whose right sides desc
whose left side are matched
applying the rule.
Rule set
jug
(xy) — (4,y) fill the 4- gallon
1.
If x<4
jug
2. (xy) (x,3) fill the 3-gallon .
lf x<3 4- gallon jug

po ur so me water out of the
(x-d,y)
3. (x,y) —_
jug
If x>0 f the 3- gallon
pour some wate rout 0
4. (xy) —P-d)
If y>O
llon ju gon the ground
(y)em pty the 4- ga
5. (x,y) ——>
ound
If x>0
3- ga ll on jug on the gr
empty the
6. (xy) ——-»{x,0)
e 4- gallon
If y>0 th e 3- ga ll on jug into th
water from
“x )) pour
on
jug is full n jus it the 3-gall
(x,y) > (4,y-(4
7.
i : the 4-galoon = a alto
d y7 0 a p e
if x+y>=4 an th
)) P r water from
v) ,3
(x,y) ——> (x-(3-
go
the 3-gall t
s until
neneteons
Wy eaiiic
=3 an d x2 0 ju
If xty>
OF punarl
$
an
(x+y,0) pour all the water from the 3 -gallon jug into.
9. ~ (x,y) ——
If xty<=4 and y>0 the 3-gallon jug

10. (x,y) ——> (0,x+y) pour all the water from the 4 -gallon jug into
If xty<=3 and x>0 the 3-gallon jug

11. (0,2) ——® (2,0) pour the 2-gallon from the 3 -gallon jug into
the 4-gallon jug
12. (2,y) ——»(0,x) empty the 2 gallon in the 4 gallon on the ground.
Production for the water jug problem
0 0
0 3 2
3 0 9
3 3 2
4 2 7
0 2 Sor 12
2 0 Sor 11 ©
One solution to the water jug problem.

(x, y)
(0, 0)
(4, 3) (1, 3)
(4, 3) (0,3) (4,0) (4, 0)
4,0 43) 0) 1)
40)
—7\™.
0,3) (0,0) 1)
—{/\~
(4,3) (4) OR
43) 0.3) @o) 41)

Fig. 3.16.3
ig_AlSC (MU-Sem.7-Comp) ae 345 Knowledge, Reasoning and Planni
ng
3.17 Classification of Planning with State Space Search

Asthe name suggests state spa .
_ on the spatial searching.
pace search planning techniques is based
Planning with State
_ Forward State Space

_ Search/progression |
ig. 3.17.1: Classification of Planning with State Space Search

techniques.
_— Planning with state space
pace search can be done by both forward and backward state-space search
3.18 Progression Planners we

planning technique, as
search” is also called as “progression planner ” It is a ‘deterministic
- “
Forward state-space
state in order to attain the goal.
plan sequence of actions starting from the initial
d we start with the initial state and go to final goal state. While doing this
_ With forward state space searching metho every state.
ts of the actions taken at
we need to consider the probable effec available actions
ial wor ld stat e inf ormation, details of the
of planning is to ha ve init
Thus the prerequisite for this type
of the goal state.
of the agent, and description that action.
prec on ditions and effects of
the available actions include flight 1 is at location A
— Remember that detail s of for a simple example where
ress ion plan ner
e-space grap h of prog only flight 1 moves
-— See Fig. 3.18.1 it gives a stat are movi ng from locat ion A to location B. In 1st case
A. these flights flight 1 is at location B
and flight 2 is also at location show s that after performing that action B
from location A to location
state
tion B, so the resulting
from location A to loca A. simil arly In 2nd case only flight 2 moves is at its origi nal
its original location — B while flight 1
where as flight 2 is at that actio n flight 2 is at location
shows that after performing
and the resulting state
location — A.
Take Flight1 from
a at Location A
“Flight
Take Flight 2 from

Location A to B
rs
ne
Progression plan
e -space graph of a
Fig. 3.18.1 :St at ation), and lines
sh ow sta te of the flig hts (i.e. their current loc
ctan gles to other location).
he Fig 3.18.1 that, re . mo ve from one location
an he r (i.e if the agent is
- It can be observe tions f rom
one state to ot
bl e ac ti on s wh ic! h can be accepted
d Froing
m e permis si
give the correspon tches to all of th
every state ma
out of
- Here the lines coming
in that state.
rithm
Progression planner algo iti ve, the liti erals whicich don
RF ear
't app
ch problem : whi ch has a set of positi
te space sear PF
£ the planning oblem
;
Formulize the st
1.
e first state © Ww TechKnowledg é
— Initial state is th . Puslicart ons
lse.
ered as fa
are consid
Knowledge, Reasoning and Plannine
~ _ If preconditions are satisfied then the actions are favoured i.e. if the preconditions are satisfied then Positie
effect literals are added for that action else the negative effect literals are deleted for that action. aS
— Perform goal testing by checking if the state will satisfy the goal.
— Lastly keep the step cost for each action as 1.
Consider example of A* algorithm. A complete graph search is considered as a complete planning

algorithm. Functions
are not used.
Progression planner algorithm is supposed to be inefficient because of the irrelevant action problem and requirement
of good heuristics for efficient search.
3.19 Regression Planners

——$__
“Backward state-space search” is also called as “regression planner” from the name of this method you can make out
that the processing will start from the finishing state and then you will go backwards to the initial state.
So basically we try to backtrack the scenario and find out the best possibility, in-order to achieve the goal to achieve
this we have to see what might have been correct action at previous state. 3
In forward state space search we used to need information about the successors of the current state now, for
backward state-space search we will need information about the predecessors of the current state.
Here the problem is that there can be many possible goal states which are equally acceptable. That is why this
approach is not considered as a practical approach when there are large numbers of states which satisfy the goal.
Let us see flight example, here you can see that the goal state is flight 1 is at location B and flight 2 is also at location B,
We can see in Fig. 3.19.1. If this state is checked backwards we have two acceptable states in one state only flight 2is
at location B, but flight 1 is at location A and similarly in 2nd possible state flight 1 is already at location B, but flight2
is at location A.
As we search backwards from goal state to initial state, we have to deal with partial information about the state, since
we do not yet know what actions will get us to goal. This method is complex because we have to achieve a
conjunction of goals.
at Location A Take Flight1 from

-Flight atL nB
Flight
i
|
Location:B
j Flight 1 at
\f
Flight.2 at: Location;A: Take Flight 2 from

OIE Location B toA
Fig. 3.19.1 : State-space graph of a regression planners
In this Fig. 3.19.1, rectangles are goals that must be achieved and lines shows the corresponding actions.
Regression algorithm
1. Firstly predecessors should be determined:
- To do this we need to find out which states will lead to the goal state after applying some actions on it.
— We take conjunction of all such states and choose one action to achieve the goal state.
- If we say that “X” action is relevant action for first conjunct then, only if pre-conditions are satisfied itworks.
- Previous state is checked to see if the sub-goals are achieved, — |
Knowledge, Reasoning and Planning
b eee 7
Actions must be consistent it sh ould r
not undo preferred literals. If there are positive effects of actions which appea
in goal then they are deleted. o
ed. Ot ; dy appears.
herwise Each precondition literal of action is added, except it alrea
Main advant age
geo this me thod l
is only I Y r elevant actio ns
i ns are taken i into co side
i ;
backward search method has much lowe bra cl ing factor
3.19.1 Heuristics for State-Space Search

“hue, tee
Progression :
or Regression 2 isi not very efficient
i with complex problems. They need good heuristic to achieve better
. olution is NP Hard (NP stands for Non-deterministic
and Polynomial-time).
There are two ways to make state space search efficient:

: on their immediate successors or predecessors.
: build
Use line ar method ::Add the steps which
o
ed on
o i
parti l planning
Use partia method .
:As per the requi. rement at execution time ordering constraints are impos
agent.
3.20 Total Order Planning (TOP)

at all stages of
regression planners impose a total ordering on actions
We have seen in above section that forward and
the planning process.
at once and to do this
Total Order Plann ing (TOP) , we have to follow sequence of actions for the entire task
In case of
important thing which
ple comb inat ions of requi red actio ns. Here, we need to remember one most
we can have multi
w hile creating the sequence of actions.
should take care of preconditions
should be taken care of, is that TOP
the left so ck and we cannot wear the
right shoe without
we C ann ot wea r left shoe without wearing
For exam ple,
actio ns in total ordered planning, wea
ring left sock action
t soc k. So whil e crea ting the sequence of
wearing the righ k action should be executed before wearing
and wearing the right soc
fore wearing the left shoe
should be executed be
can see in Fig.3.20.1.
the right shoe. As you
Wearing Shoe :
0.1 Total order planning of . :
Fig, 3.2
" Tt
° Puolications
o
A\&SC .(MU-Sem. 7-Comp) 3-48 Knowledge, Reasoning and Planning

fail in non-cooperative.
— If there is a cycle of constraints then, total ordered planning cannot give good results. TOP can
environments. So we have Partial Ordered Planning method.
. : - :
3.21 Partial Order Planning
7 MU = May 13, May 14, Dec. 14, May 15, Dec. 15
-
planning does not

Plan ning (POP ),or deri ng of the actions is partial. Also partial ordered
= Incase of P artial Ordered
actions which are placed in plan.
specify which a ction will come first out of the two
the environment is
planning, problem can be dec compo sed, so it can work well in case
With partial ordered
non-cooperative.
partial ordered planning.
Take same example of wearing shoe to understand
LeftSockOn
LeftSockOn,
RightShoeOn
Fig. 3.21.1 : Partial order planning of Wearing Shoe
- Apartial order planning combines two action sequences
o First branch covers left-sock and left-shoe.

o _Inthis case to wear a left shoe, wearing left sock is the precondition, similarly.
o Second branch covers right-sock and right-shoe.
o Here, wearing a right sock is the precondition for wearing the right shoe.
— Once these actions are taken we achieve our goal and reach the finish state.
3.21.1 POP as a Search Problem
TUS CCORES Ree
TechKnowledgé 2
“PF Publications | - 6 sea
——— _ _ Knowledge, Reasoning atand Planning
as
-If we auisieree POP a search Problem, deen We say that states are
small I pla
pl ns.
,
States are generally unfinished actions. If of only starting and finishing.
we take an empty plan then, it will consist
actions.
Every
very p plan has fo ur maini components, which can be given
as follows:
Set of actions
- T hese are the steps of a plan. Acti on s which

i ‘i r med
can be perto i n order i t oO achieve
i g goal | are stored i nse i
component.
. -
- . For example : Set ions= Lefts hoe, Finis h}
4 P of Actions={ Start, Rightsock, Rightshoe, Leftsock,
ileft s : shoe are set of actions.
ock, wearing left shoe, wearing right sock, wearing right
< ere, wearing
8
Set of ordering constraints/ preconditions
- constraints. (i.e. without performing action "y" we cannot perform
i ed as ordering
-— Prec " tion s are consider
seaondi
< Left-shoe} that is In order to wear shoe, first

— For example : Set of Ordering ={Right-sock < Right-shoe; Left-sack
we should wear a sock. :
Left-shoe (Wearing Left-sock aetion should be, taken
- So the ordering constraint can beWear Left-sock < wear
before
right-shoe(Wearing right-sock action should be taken
before wearing Left-shoe)Or Wear right-sock < wear
wearing right-shoe).
inconsistency.
— If constraints are cyclic then it represents
should not be any cycle of preconditions.
— | we want to have a consistent pl an then there
Set of causal links

"E" for action B
— Action A achieves effect Effect E
ActionB
Eat Apple
Lin k Par tia l Ord er Pla nni ng (b): Causal Link Example
Fig.3. 91.2: (a) Causal an apple and the
bu y an apple it’s effect can be eating
ca n un de rs ta nd that if you
you
- From Fig.3.21.2(b)
cutting apple.
an app le is
precondition of eating
an effe ct -E and, according| to the ordering constraints it
ion C that has
flict |if there is an act
— There can be con tion can be
action B.
s af te r ac tion A and before wa nt to ma ke a de co rative apple swan. This ac
come d of that we
’t wan t t o eat an apple instea ehintkon 3
-— Say we don "E". ct , Leftsock
t have effe _» Right-shoe
n d It does no
between A and Ba nt se ck ei n
- -sock- scig
Links = {Right }.
Seet of Ca us al
, leftshoe > leftshoeon — Finish
o For example: Rightshoeon _5 Finish
e > links.
conflicts with the causal
ts ho
Leftshoe, Righ sh ou ld not be any
the re
nsistent plan
o Tohavea co
ditions in the plan. Least commitment strategy
Set of open precon cannot beachieved by some actions
It
led open if
~ Preconditions are © al choice dur
ing search
.
he precondition
can be used by delaying ¢
pen
5 fouls not be an
y © TechKaouledy’
nt plan there
euotic ations
To have a consiste
Scanned by CamScanner -
3-500 Knowledge, Reasoning and Planning -
3.21.2 ) Consistent Planis a Solution for POP > Problem

links and does not have
As consistent plan does nott have cycle of constraints; it does not have conflicts in the causal
.
open preconditions so it can provide a solution for POP problem
open preconditions in order tg
While solving POP problem operators can add links and steps from existing plans to
the potential
fulfil the open preconditions and then steps can be ordered with respect to the other steps for removing
the steps and try solving the problem with POP.
conflicts. If the open precondition is unattainable, then backtrack
help of POP we can progress from vague plan to
Partial ordered planning is a more efficient method, because with the
plan in less number of steps, this is
complete and correct solution in a faster way. Also we can solve a huge state space
because search takes place only when sub-plans interact.
3.22 Hierarchical Planning

organized ina hierarchical format.
Hierarchical planning is also called as plan decomposition. Generally plans are
of links between
Complex actions can be decomposed into more primitive actions and it can be denoted with the help
various states at different levels of the hierarchy. This is called as operator expansion.
For example :
move (X,y,Z)
pickup (x, y) putdown (x, Z)

Fig. 3.22.1 : Operator expansion
can observe,
Fig. 3.22.2 shows, how to create a hierarchical plan to travel from some source to a destination. Also you
at every level how we follow some sequence of actions.
Travel (source, dest.)
Take-Fiight Take-trein Take-Bus
YONCatch (Train) Leave (Train, dest.)

Goto (trein, Source) Buy-Ticket (Tran)
Goto (counter) Request

aNPay (ticket)
(ticket)
Fig. 3.22.2 : Hierarchical planning example
3.22.1 POP One Level Planner
if you are planning to take a trip, then first you have to decide the location. To decide location we can search for
various good locations from internet based on, weather conditions, travelling expenses, etc.
Say we select Rajasthan, with one level planner, first we switch on PC, then we open browser, after that we open
Indian Railways website booking site for ticket booking, then we enter the date, time, etc details to book the railway
ticket. After that we will have to do hotel's ticket booking and so on. ,
This type of planning is called one level planning. If the type of problem is simple then we can make use of one level
planner. For complex problems, one level planner cannot provide good solution.
— ee
ing and Planning”
AlaSC (MU-Sem. 7-Comp
) _Knowledge, Reason
nning example
Fig. 3.22.3 : One Level pla
ons
3.22.2 Hierarchy of Acti would cover more precis
e
chy of act ion s can be decided. Minor activities
or actions, hierar king, Hot el
In terms of major and min have railway Ticket Boo
acti viti es. In case of above example, we can
the major es.
activities to a ccomplish oying there, coming
back are the major activiti
g and enj
Booking, Reaching Raj
asthan, Sta yin are the Minor activities.
ligh t din ner in palace, Take photos, etc
t ion, Have candle
to reach railway sta
While take a taxi
.
complex problems of a test match(180 overs).Numbe
r of
In real wo rid there can be of 4 bowlers i n.2 days
am plans the order
am pl e : A ca pt ai n of a cricket te
For ex
. try out a large
probabilities gi? = 16° uce the size of sea rch spa ce. For plan ordering we have to
primitive
be hi nd thi s plannin g is to red hav e lim ite d way s in whi ch we can select and order
Motivati on chie s we
ible plans. with plan hierar
number of poss empt to
steps are decided we att
operators.
giv en mor e imp ortance. Once the major
s are
ing major step
In hierarchical plann d to return
the mi no r det ail ed actions. at a min or ste p of plan. In such case we nee
solve n may run into difficult ies
th at ma jo r steps of pla ordered sequence
to devise the plan.
It is possible oduce a pp ro pr ia te ly
ep agai" to pr
to the major st
3.22.3 Planner
nditions.
ti fy a hi er ar ch y of maj or co ps), so we postpone the
details to next level.
First iden mi no r ste
teps then
wr Pr
str uct a pla n i n le vels (Major §

Con ecome visible.
s b
s detall action
Patch major levels a
Finally demonstrate.
P
Ep Techtn ealedes
lic ations
a a
Poo
ee
ee
~
; La)
| EXTERNAL
| |NTERFACE |
“Criticality" I
1
!
_, Preconditions
of...
PLANNING y
EXECUTIVE |
I
I
|
!|
4|
I
!
| |
| I
; y |
I
Ber < First step of |
_ "Skeleton Plan’ Aertically™ > I &
I
I < minimum
I
I
1
|| | Plan to Achieve a State !
| in which Preconditions of
2 T |
I
| I
| I
| I
I |
! I
J I
|
| I
!
|
|I
| I
I I
7 |
|
| I
!
|
Fig. 3.22.4 : Planner
Example:
Actions required for “Travelling to Rajasthan” can be given as follows :
Opening yatra.com (1)
Finding train (2)
Buy Ticket (3) : : a
ler TachMnowledaa
AI&SC (MU-Sem. 7-Comp)_
eee Reasoning and Planning_
Get taxi(2)
— Reach railway station(3)
— Pay-driver(1)
— Check in(1)
— Boarding train(2)
- Reach Rajasthan (3)
1* level plan
Buy Ticket (3), Reach Railway Station(3),

Reach Rajasthan (3)
2™ level plan
Finding train (2), Buy ticket (3), Get taxi(2), Reach railway station (3)
Boarding train(2), Reach Rajasthan (3).
3" level plan (final)
Opening yat
pening yatra.com (1), Finding train (2), Buy ticket (3), Get taxi(2), Reach Railway station (3), Pay-driver(1), Check
in(1), Boarding train(2), Reach Rajasthan (3).
3.23 Planning Languages

MU - May 13, Dec. 14
: Explain STRIPS repr esentation of planning problem. ioeos oe 2 oe
Language should be expressive enough to explain a wide variety of problems and restrictive enough to allow efficient
algorithms to operate on it.
- Planning languages are known as action languages.
Stanford Research Institute Problem Solver (STRIPS)

- Richard Fikes and Nils Nilsson developed an automated planner called STRIPS (Stanford Research Institute Problem
_ Solver) in 1971.
— Later on this name was given to a formal planning language. STRIPS is foundation for most of the langiiaeesttin order to
express automated planning problem instances in current use.
Action Description Language (ADL)
ADL is an advancement of STRIPS. Pednault proposed ADL in 1987.
Comparison between STRIPS and ADL
j Only allows positive literals in the states, Can support both positive and negative literals.
in STRIPS is expressed | For example : Same sentence is expressed
For example: A valid sentence -Stupid A -
as => Intelligent* Beautiful. é as => -Stupi Ugly _
Makes use of Open World Assumption (i.e
Makes use Oe care -worldfalse)
assumption (i.e. e
unmentioned literals are unknown)
Unmentioned litera
We can find quantified variables in goals.
in goals.
We only can find ground literals
For example :3x At (P1, x) A At(P2, x) is the goal of
nt A Beautiful
For example: Intellige having P1 and P2 in the same place in the example of
the blocks
ing and Planning
Knowledge, Reason
Al&SC (MU-Som. 7-Comp)
disjunctions
| Goals are conjunctions Goals may involve conjunctions and
4.
For exam| ple: (Intelligent A (Beauti
ful V Rich)).
For example : (Intelligent A Beautiful).
allowe d: when P:E meansE is
5. Effects are conjunctions Conditional effects are
an effect only if P is satisfied
built in.
Does not support equality.
Equality predicate (x =y ) is
6.
Supported for types
7. | Does not have support for types
For example : The variable
;
p: Person
3.23.1 Example of Block World Puzzle

Tool
— rz rh
Start
Fig. 3.23.1
Standard sequence of actions is
Grab Z and Pickup Z

bP
Then Place Z on the table

u WN
Grab Y and Pickup Y
Then Stack Y onZ

aF
Grab X and Pickup X

Stack X on Y
piece of knowledge or
Elementary problem is that framing problem in Al is concerned with the question of what
information is pertinent to the situation.
To solve this problem we have tom make an Elementary Assumption which is a Closed world assumption. (i.e. If
something is not asserted in the knowledge base then it is assumed to be false, this is also called as “Negation by
failure”)
— Standard sequence of actions can be given as for the block world problem :
on(Y, table) on(Z, table)
on(X, table) on(Y, Z)
on(Z, X) on(X, Y)
hand empty hand empty
clear(Z) clear(X)
clear(Y)
4
Tool
Start
Fig. 3.23.2
Knowledge, Reasoning and Planni
Rule 1
hand empty, on(X,table), holding(X)
: clear(Xx)
Rulmie putdown(x) . x)
holding( hand empty, on(X, table),
- .
clear(X)
Rule 3 stack(X,Y) holding(x), on(X,Y),
clear(Y) ; _clear(X)
Rule 4 unstack(X,Y) on(X,¥), “holding(X),
_ . _| Clear(X)
; clear(Y) .
- Based on the above rules, plan for the block world prob
lem:Start > goalcan be specified as follows:
1. unstack(Z,x) 2. putdown(Z)
3. pickup(Y) 4. stack(Y,Z)
5. pickup(x) 6. stack(X,Y)
Execution of this plan can be done by making use of a data struc
ture called "Triangular Table".
1 on (C, A) clear (C) unstuck
handempty (C, A)
2 holding (C) putdown (C)
3 on (B, table) handempty | pickup
__ (B)
4 clear (C) Hoiding | stack
(B) (B,C)
on (A, table) clear (A) Handem | pickup
5
ply (A)
6 clear (B) | holding | stack

1A {ap
on (C, table) on (B, C) on (A, B)
7
é clear (A)
0 1 2 3 4 5 6
Fig. 3.23.3
In a triangular table there are N + 1 rows and columns. It can be seen from the Fig. 3.23.4 that rows have 1 — n+1
condition and for columns 0 — n condition is followed. The first column of the triangular table indicates the starting
table indicates the goal state.
state and the last row of the triangular
With the help of triangular table at ree is formed as shown below to achieve the goal state :
Fig. 3.23.4
c arm) can ha ve some amount of fault tolerance. Fig. 3.23.5 shows one such example.
An agent (in this case roboti
TEN
ee
_Knowledge, Reasoning and Planning
Not allowed
Tool : | |
ry os : Ey
Start Wrong
. move
Fig. 3.23.5
3.23.2 Example of the Spare Tire Problem
— Consider the problem of changing a flat tire. More precisely, the goal is to have a good spare tire properly mounted
onto the car's axle, where the initial state has a flat tire on the axle and a good spare tire in the trunk. To keep it
simple, our version of the problem is a very abstract one, with no sticky lug nuts or other complications.
There are just four actions: removing the spare from the trunk, removing the flat tire from the axle, putting the spare
on the axle, and leaving the car unattended overnight. We assume that the car is in a particularly bad neighborhood,
so that the effect of leaving it overnight is that the tires disappear.
— The ADL description of the problem is shown. Notice that it is purely propositional. It goes beyond STRIPS in that it
uses a negated precondition, -At(Flat, Axle), for the PutOn(Spare, Axle) action. This could be avoided by using Clear
(Axle) instead, as we will see in the next example.
Solution using STRIPS
— — Init(At(Flat, Axle) A At(Spare, Trunk ))

— Goal(At(Spare, Axle)) Action(Remove(Spare, Trunk ),
— PRECOND: At(Spare, Trunk )
— EFFECT: - At(Spare, Trunk ) A At(Spare, Ground))
- Action(Remove(Flat, Axle),
— PRECOND: At(Flat, Axle)
= EFFECT: - At(Flat, Axle) A At(Flat, Ground))
— Action(PutOn(Spare, Axle),
— PRECOND: At(Spare, Ground) A - At(Flat, Axle)
— EFFECT: - At(Spare, Ground) A At(Spare, Axle))
- Action(LeaveOvernight)
— PRECOND
—.. EFFECT: - At(Spare, Ground) A = At(Spare, Axle) A - At(Spare, Trunk } A - At(Flat, Ground) A- At(Flat, Axle))
3.24 Planning and Acting in the Real World

In earlier sections, we have discussed planning that is observable and deterministic. Real world problems however are
not predictable and completely unobservable. The planning and acting on real world problems require more
sophisticated approach.
One of the most important characteristic of real world is uncertainty: In an uncertain environme
nt, it is very important
for agent to rely upon its percepts (series of past experience)
Whenever some unexpected condition encountered, agent should refer its percept and should change
course of
action accordingly. In other words agent should be able to cancel or replace the currently executing
plan with some
other more suitable and reliable plan if something unexpected happens.
:
aigssc (MU-Sem. 7-Comp)
it should be not
eS Knowledge, Reasonin 2fanning
ed that real world
itself is not uncertain 5 uncertain. In
ion related to world i
and hence machine als o receivgin
artificial intelligence, we try to give human Perception ability to machine, ond es the
tke
ity vs icine,
perception of uncertainty about the real world. So machine has to deal with incomplete and incorrect infornis
human does. 7
| they Iways
Determining the condition of state depends on availabl kno
e wle dge . In real worl d, kno wle dge avai lability son ’
limited, so most of the time , conditions are non deterministi
c. ‘
The amount or degree of indeterminacy depends upon the knowledge available. The inter determinacy 's. calle
“pounded indeterminacy” when actions
can have unpredictable effects.
Four planning strategies are there for handling indeterminacy :
(i) Sensorless planning
(ii) Conditional planning
(iii) Execution monitoring and replanning
(iv) Continuous planning
(i) Sensorless planning
Sensorless planning is also known as conformant planning. These kinds of planning are not based
on any perception.
The algorithm ensures that plan should reach its goal at any cost.
(ti) Conditional planning
Conditional plannings are sometimes termed as contingency planning and deals with bounded indeterminacy
discussed earlier. Agent makes a plan, evaluate the plan and then execute it fully or partly depending on the
condition.
(ill) Execution monitoring and replanning

In this kind of planning agent can employ any of the strategy of plannin
g discussed earlier. Additionally it observes the
plan execution and if needed, replan it and again executes and observes.
()) Continuous planning

Continuous planning does not stop after performing action. It persist over time and
keeps on Planning on
Predefined events. These events include any type unexpected circumsta
some
nce in environment.
8.25 Multi-Agent Planning

-~
Whatever planning we have discussed so far, belongs
to single user environment. Agent acts alone in a single
€nviro
user
nment.
When the environment consists of multiple agent
to, then the way a single agent plan its action
get changed
We have a glimpse of environment where multiple: agents have to take actions ba sed on
. current st
£NViro nment could be co-operative or competitiave. In both the cases agent , S action ate. The
influe nces each other. .
Few of the multi agent planning strategies are listed below :
i Co-operation (ii) | Multi body planning
lili)” Ep. Ordination
mechanisms (iv) Competition
Knowledge, aNd Plann
(i) Co-operation
In co-operation Strategy
agents have joint goals and plans. Goals
combined to ach
can be divided into sub goals but ultimate
ieve ultimate goal.
(il) Multibody planning
Multi body planning is the strategy
of implementing correct joint plan.
(ili) Co-ordination mechan
isms
These strategies specify the co-ordination
between co-operating agents. Co-ordinatio
Co-operating plannings. n mechanism is used in Several
(iv) Competition
Competition strategies are used when agents
are not co-operating but competing with each
to achieve the goal first. other. Every agent Wants
3.26 Conditional Planning

Conditional planning has to work regardless
of the outcome of an action.
Conditional Planning can take place in Fully
Observable Environme nts (FOE) where the
known environment is fully observable. current state of the agent is
The out come of actions cannot be determined
be nondeterministic. so the environment is said to
:
8
Left Suck
we
1 {2 e[ |= >| [2
GOAL Right Suck LOOP Left Suck
4) &
<A)
2 lee [as
ad) 7 ad) 5
é
af)
és | as}
1 lt)
§
TA
GOAL
LOOP
Fig. 3.26.1 : Conditional Planning
- vacuum world example
— In conditional planning we can check what is
happening in the environment at predetermined
deal with ambiguous actions. points of the plan t
’
It can be observed from vacuum world example ,Cond : ;
itional Planning needs to take some actions at every state a. d
must be able to handle every outcome for the action
it takes. A State node is represented with a squar
node is represente e and chan¢
d with a circles.
For a state node we have an option of choosing
some actions. For a chance node agent has to
handle every outcome:
Conditional Planning can also take place in the Partia
lly Observable Environments a
track on every state. Actions can be Uncertain because o f (POE) where, we cannot keeP
the Imperfect sensors,
EF TechKnewietss
Publications
ase Ay
|
2 = Knowedge, Reasoning and Pih stcasea s
e if the dirt rtisi i Then in suc
In vacuum agent exampl ut Righ t, but not about Left.
th abo
pirt might be left behind when
ws ef
t and agent
agent, nenleav
kno
called as a state set OF a beli
|
e , a clean square. Initial state is also
€s
state. :
sing can be
environments. Automatic sen
role in Con diti ‘
sensors play important onditi
ng for part iall y obs erv abl e
agent Pla nni
useful; with automatic sensing an
d is Active
-
gets all the avai cdleea at every. SIP: Another

metho
,
nsinging,
sens with which percepts are obtained only by svete oeecific sensory actions.
a
ee see,
-
sy
en eee
wee Soo
CleanL 7 CleanL
em Em
cate ie ET 7s.
ee
sh tae
gah
.~ ee
7 nae
~ és.
és, “s eee
ween
(condition 2)
Pl an ni ng - va cu um world example
onal
Fig. 3.26.2 :Conditi
ent?
Q.1 What is Knowledge Based Ag nt for the same.
s pec ify PEA S pro perties and type of environme
wo RLD Environment.
Q2 Describe WUMPUS ques.
esen tation techni
c? Ex pl ai n va ri ou s knowledg e repr
Q3 What is Logi and exa mpl e sentences for propositiona
l logic.
ax and sem ant ics
Q4 positiona | logic? Writ
e synt examples.
What is pro nal logic with suitab
le
opositio
process i ncase of pr
Q5 Explain the inference
with example.
Q6
Explain Hom Clause ics of FOL with
example.
Q7 What is first order logic‘

lifting? aining.
Q8
What is unification and Forward C
hainin g.and Backward Ch
using
Qg9 s in FOL
Explain inference proc es
.
Q, .
10 Write a short
note on : Resoluti
on. Vee
Al&SC (MU-Sem. 7-Comp) 3-60 ____ Knowledge, Reasoning and Plann
Q. 11. What is planning in Al? :
Q.12 Explain planning problem,
Q. 13 Explain goal of planning with supermarket example.
Q. 14 What are the major approaches of planning? Explain conditional planning with example.
Q.15 Write short note on planning graphs.
Q. 16 How planning strategies are classified? Explain regression planning.
Q.17 €xplain progression planners with example.
Q. 18 Explain regression planners with example.
Q. 19 Explain total order planning with example.
Q. 20 Explain partial order planning with example.
Q.21 Explain process of generating solution to partial order planning problem.
Q. 22 Write short note hierarchical planning.
Q. 23 What is one level planner?
Q. 24 What are different languages used for implementation of planning?
Q. 25 Compare and contrast planning languages.
Q. 26 What are various planning strategies used to handle indeterminacy?
Q. 27 Write a short note on multi-agent planning.
Q. 28 Explain conditional planning with example.
Fuzzy Logic
Unit IV
i
Introducti
to Fuzzy Set:
.
Fuzzy set theory, Fuzzy set versus crisp set, Crisp relation & fuzzy relations,
uction
membership functions,
4,2 Fuzzy Logic: Fuzzy Logic basics, Fuzzy Rules and Fuzzy Reasoning
4.3 Fuzzy inference systems: Fuzzification of input variables, defuzzification and fuzzy controllers.
4.1__ Introduction to Fuzzy Set
= Fuzzy logic was introduced by Prof. Lofti A. Zadeh in 1965.
-— The word fuzzy means “Vagueness”:

— Fuzzyness occurs when a boundary of a piece of information is not clear.
- He proposed a mathematical way of looking at the vagueness of the human natural language.
Why fuzzy set is required?

are crisp, deterministic and precise.
— Most of our traditional tools for formal modelling, reasoning and computing
the structures and parameters of the model are
While designing the system using classical set, we assume that
or their occurrence.
definitely known and there are no doubts about their values
knowledge that is vague, imprecise, uncertain, ambiguous,
— But in real world there exists much fuzzy knowledge;
inexact or probabilistic in nature.
= There are two facts ;
precisely.
deterministic and they cannot be described
1. Real situations are very often not crisp and
a human being could ever
often would require more detailed data than
2. The complete description of areal system
process and understand.
recognize simultaneously,
n do not reflect the nature of human
elling the real system using classical sets ofte
- Because of these facts, mod
abstract, imprect ise and ambiguous.
concepts a nd thoughts which are
incomplete information.
to cope with such unreliable and
- The classical (crisp) sets are unable and give expert opinion.
We want our systems should also be able to cope with unreliable and incomplete information
-
unreliable, incomplete, vague and imprecise information.
~ Fuzzy set theory has been introduced to deal with such
element have degree of membership.
~ Fuzzy set theory is an extension to classical set theory where
g.
the whol e inte rval betw een 0 (fals e) and 1 (true) to describe human reasonin
~ Fuzzy logic u ses
4.2 Fuzzy Set vs. Crisp Set

crisp boundary. ”
crisp set) Is a set with a
~
Aclassical set ( or conventional or
as
For example, a classical set A of real numbers greater than 6 can-be expressed
A = {xIx>6}
number, then x belongs to the set
Where there is a clear, unambiguous boundary ‘6’ such that if x is greater than this
eae
A, otherwise x does not belong to the set.
Although classical sets are suitable for various approximations and have proven to be an important tool for
mathematics and computer science, they do not reflect the nature of human concepts and thoughts, which are
Saas
abstract, imprecise and ambiguous.
igen
height is more
For example, mathematically we can express the set of all tall persons as a collection of persons whose
than 6 ft.
A = {x|lx>6}
Where A = “tail person” and x = “height”.
The problem with the classical set is that it would classify a person 6.001 ft. tall as a tall person, but a person 5.999 ft.
tall as “not tall”. This distinction is intuitively unreasonable.
The flaw comes from the sharp transition between inclusion and exclusion ina set.(Fig. 4.2.1)
1.0 i
i
‘
t———->tall (= 1.0) :
Degree 2
of j
membership
“a a
a
| Not tall (1 = 0.0) . 44 a
> Height (X) 4

|
Fig. 4.2.1 : Sharp edged membership function for TALL
Fuzzy logic uses the "degrees of truth" rather than the usual "true or false" (1 or 0) Boolean logic.
Fuzzy logic includes 0 and 1 as extreme cases of truth but also includes the various states of truth in between so that, 3
for example, the result of a comparison between two things could be not "tall" or "short" but "0.38 of tallness."
1.0 Continuous __, definitely a tall
membership person (p = 0.95)
- Degree of function for
membership, p TALL -
_» leally not very
0.0 tall at all (1 = 0.30)
Height .
Fig 4.2.2 : Fuzzy membership function TALL
&SC (MU-Sem. 7-Comp)
4-3 2, Fuzzy Logic
as shown in Fig 4.2.2 the fuzzy logic defines a aly A person’ss height
hei may now
Smooth transition from ‘not tall’ to ‘tall’.
op to both the groups “taly’ and
pelong Not all’ but now it will have the degree of membership associated with it for each
group.
A person has 0.30 membership in ‘Not tall’ group and ‘0.95 ip in ‘tall’ group, so definitely the per:
categorized as a tall person. membership in ‘tall’ group, s : son is
key differences between Fuzzy Set

ang Crisp Set
1. | A fuzzy set follows the infinite

-valued logic Acrisp set is based on bi-valued logic.
2. | A fuzzy set is determined by its indeterminate
A crisp set is defined by crisp boundaries, and
boundaries, there exists an uncertainty about the set | contain’the precise location’ of the set
boundaries - boundaries.
3. | Fuzzy set elements are permitted to be partly | crisp set elements can have a_ total
accommodated by the set (exhibiting gradual | member
or non-membership.
ship
membership degrees).
4, | Fuzzy sets are capable of handling uncertainty and | Crisp set requires precise , complete and finite
vagueness present in the data data.
43 Fuzzy Set Theory

43.1 Fuzzy Set : Definition
- IfXisa collection of objects denoted generally by x, then a fuzzy set A in Xis defined as a set of ordered pairs :
A = {Pg
Ixe X)
Where [1g(x) is called the Membership Function (MF) for the fuzzy set A.
and 1.
The MF maps each element of X to a membership grade between 0
Ifthe value of membership function [;' (x) is restricted to either 0 or 1, then A is reduced to a classical set.
Here, X is referred to as the Universe of discourse or simply the Universe and it may consist of discrete objects or
, eferre
Continuous space.
*82 Types of Universe of Discourse

non ordered universe
Fuzzy sets with a discrete
par discrete non -ordered
ord objects.
Universe of discourse may contain
YF AI&SC (MU-Sem. 7-Comp)
For example,
Let X = {San Francisco, Boston, Los Angeles} be the set of cities one may choose to live in.
The fuzzy set “desirable city to live in” may be described as follows:
A = {(San Francisco, 0.9), (Boston, 0.8), (Los Angeles, 0.6)}
— Here, the universe of discourse X is discrete and it contains non-ordered objects, in
this case three big cities h
United States. 7
2. Fuzzy sets with a discrete ordered universe
— LetX={0, 1, 2,3, 4, 5, 6} be the set of number of children a family may choose to

have.
— Then, a fuzzy set “Sensible number of children in family” may be described as
A = {(0,0.1), (1, 0.3), (2, 07), (3, 1), (4, 0.7), (5, 0.3), (6, 0.1)}
Here we have a discrete ordered universe X.
The MF is shown in Fig. 4.3.1.
1.0. f-----------+- ®
0.9
0.8
0.7 }-------- @---+---
0.6
- .
ig
0.3 ;---@---4----------}---
0.1 Q---f----nnnpnn
nnn -f
fone Ai
0 1 2 3 4 5 6G
X =no of children
Fig. 4.3.1 : MF on a discrete universe
3. Fuzzy sets with a continuous universe

Let X = R” be the set of possible ages for human beings (Real numbers - continuous). Then the
fuzzy set B = “about50
years old” may be expressed as, “ar ee
B { (%, Hg(x)) Ix © X)
Where, Ha(x) = 1+ |
A&SC (MU-Sem. 7-Comp)
2.
This is illustrated in Fig. 4.3.
>X
0 10 20 30 40 50 60 70 80 90 100
Fig. 4.3.2 : MF for “about 50 years old”
Table 4.3.1 shows [5(x) for some value of x.

old”
Table 4.3.1: x and corresponding ba (x) for “about 50 years
55 56 58 | 60
0.94 | 0.89 | 0.71 | 0.5
discourse and
fication of a ‘suitable universe
te membership function.
Specification of an appropria membership functions.
ber shi p fun cti on is sub jec tive, which means that the
The “specification |of mem . Therefore, the ‘subjec vity-
sa me co nc ep t by ‘different persons may vary considerably
; ified for ihe a izzy sets and probe ty
sets is theprimary, aifference., between the
resenting Fuzzy Sets

43.3 Different Notations for Rep
1) Using ordered pairs :
XK = (pg) xe X)
= (1, 2, 3s ++ 6)
Bg. LetX
um be r of be dr oo ms in a ho us e” then comfortable house for four
cribed by X = “N
Xis available types of houses des dered palr as,
Persons family is described using OF
A = {(1,0.2), (2,05). 3, 0.8), (4, 1), (5, 0.7), ©, 0.3)}
2 function :
) Using membership function.
Its membership
statin‘ing
A fuzzy set can be represented by 10”.
ly Jarger than
“re al nu mb er s o© nsiderab
Eg. To repr es en t
“Public malate
F_Al&SC (MU-Sem.7-Comp)
We define,
A = ((x, g(x) !xe X},

0 xs10
x(x) = —— ; x>10
1+q= 10"
3) Using + Notation :
as,
Fuzzy set for “comfortable type of house for a four person family” may be descr ibed
0.2 05 08 1,07 03}
Pi = (Pepe stat 5 + 6
i.e. we define A as
A = pg Ou) /x1 + Mg Oa) ate

n
= > Mg! x
i=l
4) Using Venn diagrams :

represents membership function.
Sometimes it is more convenient to give the graph that
wax)
0 = x
Fig. 4.3.3 : Representation of fuzzy set using Venn diagram

5) Other notations :
{(3, 0.1) + (4, 0.3)} + .....}

>
a~ _= [01{803,96} 41
or
03 06
4.3.4. Linguistic Variables and Linguistic Values
aged” and “old” that are characterized by &
Suppose that X = “age”. Then, we can define fuzzy sets “young”, “middle de
MFS Hyoung (x), Hmiddle aged (x) and Hoig (x).
aged” and “gid” in this
A linguistic variable (“age”) can assume different linguistic values such as “young”, “middle
case.
old) and
Note that, the universe of discourse is totally covered by these MFs (MFs for young, middle aged and
a
transition from one MF to another is smooth and gradual.
a support of a fuzzy y s set Ai
A is the set of all points xin X
such th at, x(x) > 0.
Support (A(A)) == {x1 g(x) >0}

2. Core / Nucleus :
Th e core of a fu zzy set A Ai is the set of all points x in X such that [x (x) = 1
AMAT Ss
Core A = ~(y\)=
(x1 HAG) =1}
3, Normality:
that
must be at least one point x € X such
A fuzzy set A is normal if its core iis non-empty. In other words there
bg(x) =1.
4, Crossover points :
isa point x € Xat which [,(x) = 0.5.
Across over point of a fuzzy set A
Crossover (A ) = {xl px@= 0.5}
5, Fuzzy singleton : .
1is called a fuzzy singleton
e sup por t is a sin gle poi nt in X with L(x) =
A fuzzy set whos
45 year old
Member
ship
Singleton
grade
Age
singleton
Fig.4.3.4: A fuzzy
of a fuzzy set.
po rt an d crossover points)
sup
ee parameters (core, Middle aged
Fig. 4.3.5 shows thr
1.0
Member
ship
0s
grade
g
Age
— Core"
N Crossover
point
Support =
fuzzy set
points of a
and Crossover
4.3-9
4.3.5: : Core ' Support
Fig.
& o-cut : defined by,

a crisp set
t ofaf uzzy set A Is
The c-cut or - Jevel se TechKnowledgé
Ag = {xt HA 7%
Peblications
WE _Alasc (MU-Sem. 7-Comp)
7.: Strong a-cut/ strong a-level set :
Strong c-cut is defined by
Al = (xIPg(x)>0}
support and core of a fuzzy set Aas,
Using the above notations, we can express
Support (A) = Ay Here a=0
Core (A) =A, Here a=1
8. Convexity :
ANAL
Two convex membership
functions
Non convex
membership functions
Fig. 4.3.6 : Convex and Non Convex MFs
A fuzzy set A is convex if and only if for any x, and x,¢ X and any Ae [0, 1].
Mx (Ax; +(1 —A) x2) = min {Hx (x), UR (x2) }
or
A is convex if all its a-level sets are convex.
9. Fuzzy numbers:
A fuzzy number A isa fuzzy set in the real line (R) that satisfies the conditions for normality and convexity.
10. Bandwidth of normal and convex fuzzy set :

For a normal and convex fuzzy set, the bandwidth or width is defined as the distance between two unique crossover
points.
Width (A) = tx,—xjl

Where [a (x1) = Wx (Xa) = 0.5
11. Symmetry:
A fuzzy set A is symmetric if its MF is symmetric around a certain point x = c,
ba (c+x) =Px(c-x) forallxe X
Pr
Al&SCL (MU-Sem. 7-Comp
)
Fuzzy Logic
42. open left, Open right and cleéed ars :
left if,
A fuzzy set A is open
lim
Ux (x) = 1 and
xO-2
Bx (x) =0
_
A(x) *
Young
= —» x age
-
Fig. 4.3.7 : Open Left MF
A fuzzy set A is open right if,

lim
3-0 HAM = 0 and
lim
x $00 HK OD =1
Old
BAX)
—> x age
a
MF
Fig. 4.3.8: Open right
A sa seti is closed if,

and
an 400M ACO = °
X>-poo Ha @) =° Middle aged

Ha)
x age
Closed MF
Fig. 4.3.9:
"8. Cardinality :
A is defined as
Cardinality of a fuzzy set
Al = D pA
TechKaeuledge
Puptications
WY _Alasc (MU-Sem. 7-Comp)
14. Relative cardinality :
Relative cardinality of a fuzzy set A is defined as,
~
WAM =
Al
15. Height of a fuzzy set :
degree Um
The height of a fuzzy setA in X, is equal to the largest membership
~ Sup
hgt(A) = ye x ha)
If hgt (A )=1 then, A is normal.
If het (A ) <1 then, A is subnormal.
hgt
Vv
Fig. 4.3.10: Height of a fuzzy set
4.3.6 Properties of Fuzzy Sets
Fuzzy sets follow the same properties as crisp set except for the law of excluded middle and law of contradiction.
That is, for fuzzy set A
AUA #U ; AnA#6
The following are the properties of fuzzy sets,
1. Commutativity :
ANB = BNA
2. Associativity:
AU(BUC) = (A UB) UC
AN(BNC) = (ANB)nC
7 ag gigSC (MU-Sem. 7-Comp)
ZY Logig q
a4
= pistributivity : Fuzzy Log jic
t
ie ?
‘
AUBnS ~ R n AUC)
f
~ on
i ~ Jd: d= (AUB)
=(,ano ~ on
AN Buc)
tity :
* AUd =A : Auuseu
A = ~ ~
=o 5 ANU=A
5, Involution:
A =A
6. Transitivity :
ACB CC, thenAC Cc

7, De Morgan’s law :
ANB = AUB
43.7 Operations on Fuzzy Sets

1. Containment or Subset :
for alll x.
in fuzzy set B ifand only if Pax) S [g(x)
Fuzzy set A is contained
Kk cBeRx@SPa®
ne
en t or subset
Fig. 4.3.1 1: Containm
= at whose FS
2 Union (Disjunction) :
a
Aunion of two fuzzy sets
publicatigas
IESE (MU-Sen
7 one)
—p X
>X ‘ »
Fig. 4.3.12 : Union of two fuzzy sets

3. Intersection (Conjunction) :
The intersection of two fuzzy sets A andB isa fuzzy set C, such that whose MF is defined as
Mes) = min (Ug(%), Hg)

A B
AnB
ne
Fig. 4.3.13 : Intersection of two fuzzy sets
4. Complement (Negation) :
The complement of a fuzzy set A , denoted by A is defined as,
Me(x) = 1-Hx()
Fig. 4.3.14 : Complement of a fuzzy set

More operations of fuzzy sets
1. Algebralc sum :
Mx+ 50) = Ug(X) + Wg(x) — Wa(X) g(x)

2. Algebraic product:
Mx.a(x) = Wg(X)-Mg(X)
3. Bounded sum :
Mxe@ g(x) = min [1, Wx(x) + wp(x)]

4. Bounded difference :
Lx @ H(X) = max [0, M(x) — Hg(x)]
aaa
| | winty vole
pgs Mu-Sem. 7COm)
a
4_crisP Relation and Fuzzy Relations
i crisp Relation
4
ann-ary relation over My, Mz, Ms, ... My is a subset of the Cartesian product Mx M2 --- Mp, where n= 2, the relation
ear We 2 Mp
M 1X Mp. This is called a int, nelotion
isa subset of the Cartesian product
aa
tet Xand Y be two universe and X x Y be their Cartesian product
Y can be defined as, Prone
Then XX
1et XxY = {@ y)Ixe X, ye Y)
every element in X is related to every element in Y
X and
_ wecan define characteristic function f that gives the strength of the relationship between the each element of
Y.
1, &yWEXxY
fxxy Oy) = | &y)
0 ? (x, y) éXxXY
- We can represent the relation in the form of matrix.

Ann-dimensional relation matrix represents on n-ary relation.
- $0, binary relation is represented by 2 dimensional matrices.
fy: Consider the following two universe,
X={a,b, ch, Y={1, 2, 3}
- The Cartesian product - X x Y is,
Xx¥={(a, 1), (a, 2), (a, 3), (b, 2), (b, 2), (b, 3), (c, 2), (c, 2), (c, 3}
From the above set, we may select a subset R, such that
R = {(a,1), ©, 2), (6, 3), D, (, 3)}

Then R can be represented in matrix form as,
The relation between set X and Y can also be represented as coordinate diagram as shown in Fig. 4.4.1.
m of a relation
Fig. 4.4.1: Co-ordinate diagra
a | : TW Tekaeaieaes
Fig. 4.4.2: Mapping representation of a relation
- A characteristic function is used to assign values of relationship in the mapping of X x Y to the binary values and jg
given by, /
fan R(x, y) =| 0
1,@%yeR
; (x, y)ER
4.4.1(A) | Cardinality of Classical Relation
— Let X and Y be two universe and n elements of X are related to m elements of y.

— Let the Cardinality of X is n, and cardinality of Y is ny, then the cardinality of relation R between X and Y is,
Nxxy = 1x*Ny
- The Cardinality of the power set P (X x Y) is given as, -
(xny)
Npxxy = 2 xy
4.4.1(B) Operations on Classical Relations
— Let Aand B be two separate relations defined on the Cartesian universe X x Y.

— Then the null relation defined as,
ds =],
oa
And complete relation is defined as,

1117
Ea 111
111
The following operations can be performed on two relations A and B.
1. Union
AU B-> fixup (XY): fue (% y) = Max [fs (x, v), fa (XY)

2. Intersection
ANB fang (% ¥) + fan (% ¥) = min [fy (x, Y), fa (x, y)]
- 3. Complement
A—> Fi (x,y): fa(X, ¥) = 1— fe (x, Y)

4. Containment
“ACB fa (x y): fa (% Y) S fa (%, y)

2 eee
(N&SC (MU-Sem. 7- Comp)
5. Identity
goa and XE,
44.4(C) Properties of Crisp (Classical) Relati ons

idempotency hold
The properties of classical set such as co ty, ass oci ati vit y, involution, distributivity and
mmutativi
good for classical relation also
Also De Morgan’s
: law and excluded middle | aws hold good |for crisp relations.
4,4.1(D) omposition of Crisp Relations
_ Composition ombining two compatible binary relations to get a single relation.

eee aclis a process of combini
_ te = that maps elements from universe X to universe Y.
Let B be a relation that maps elements from universe Y to universe Z.
_ The two binary relations A and B are said to be compatible if
ACXXY and BCYXZ

A and B can be denoted as A 0 B.
The composition between the two relations
Be:
Let X = {aj,a-4s)
Y = {b,, be, bs}
Z = {c1,C2,¢sh
Let the relation A and B as, b2)}

A= X x Y = {(@y, by), (a1 ba), (@2»
5 C1)» Or» Cs)» (b3 »C0)}
B = YXZ= {by
A and B
4. 3: II tu st ra tion of rel ations
Fig. 4.
written a5,
Then Ao B can be A oO B = {(ay ’ Cy)» (a2 ?
C4)s (a2 ’c3)}
nat
is given as,+ by bs
i matrixi form |S
Bin
i c
‘on of° Aand
: Th erepresentason bh, 7100
a, f110
b 001
Be
9 10|;
A =
0 1 0
b3
0 0 0
a3
WW alasc(mu-Sem.7-Comp) 4G
Then, composition A 0 B is represented as,
Cy C2 C3
af1o0l
AoB = »|001
a, L.0 0 0
There are two types of composition operations:

Max-min composition
Max-product composition
1. The max-min composition is defined as,

. T = AoB_
fr(x,z) = Vv [fia y) Af fy, 2]
ye Y
‘2. . The max-product composition is defined as,
T = AoB
fr(x,z) = v_ [fa Gy): fe (y, 2]
ye Y
Properties of composition operation

1. Associative : (AOB)OC=Ao (BoC)
2. Commutative : AOB#BOA
3. Inverse : (AoB) =A OB
4.4.2 Fuzzy Relations

fuzzy relation Is
In general, a relation can be considered as a set of tuples, where a tuple is an ordered pair. Similarly, a
a fuzzy set of tuples i.e. each tuple has a membership degree between 0 and 1.
Definition:
Let U and V be continuous universe, and pz : U x V [0, 1] then,
R= J yp Cusvy Ay)
uxv
Is a binary fuzzy relation on UX V.

If U and V are discrete universe, then
R= 2 pate, v)Mu,v)
UxV
i
4-17 ney “Fuzzy Logic,
we can expr ess fuzzy relatio NR=UXVin
VR= a
Matrix form as
HR (u,, V1) MR (uy, V2) coe PR (uy, Vn)
R =| OR (U2, V1) Hr (Ua, V2) ve HR (U2, Vo)
BR (Ups V1) Hp (Uppy V2) BR (Un Vn)

where U = (%1 U2, U3, ...Um}) and
V_ = (Vis V2, V35.-Vn} are universe of discourse.
a . > = pes
ex. 4.4.1 = Given universe of discourse '

U = {1,2,3} ;form a relation R where “ x is approximately equal to y”
Then Relation R can be defined as,
_ ”
R== J- ! 1 1 08 0 308 | 0. .
{an *@2) *B3) *11.2) *G3ay* er +t 23) * ei BD)
goin. : The membership function [1g of this relation can be described as,
1. when x=y
Ur (x, y) = 0.8 when |x-yl=1
0.3 when Ix-yl=2
The matrix notation is,

Y
1 2 3
F
2o |
oo
o
p
x
nN
0.3 |} 08] 1
Ww
N - ary fuzzy relation

relation of pairs.
It is possible to defi ine n-ary fuzzy relations as fuzz
y set of n-tuples. In general it is a
Up (Xqyeeeee Xa)/(Xppeeee Xn)
4.4.2(A) Operations on Fuzzy Relation

E
EEE.
OQ
yyy
interaction between variables.
Fuzzy relations are very impo ant in fuzzy controller because they can describe
relation.
on fuzzy
Four types of operations can be performed |
(2) Union
(1) Intersection
ion
Cylindrical extens
(3) Projection
Intersection
x x Y. The interse ction of R and S is defined by, :
lations defined on
S be binary re
L R and
, ~ Let
p(X, y)) Yb
5 (x, y)= min (palX
Vv :
n be used.
wr
Ky) EX: Han any T- Norm ca
~ > Instead of the minimum,
Al&SC (MU-Sem,. 7- Comp)
2. Union
— The union of Rand Sis define
d as,
V(% V)EXXY: Meus (x, y) = max (Hal, Y),
Hel y))-
Instead of maximum, any S—norm can be used.
Given two relations R andS
Y2 =Y2~—sY3 Y1 Y2 °Y3
x: | 0.3 | 0.2] 0.1 . x, [04] 0 | 0.1

R=» /o4]o6lo1 S= x] 1 [02/08
x, | 0.2103] 05 x3 | 0.3 | 0.2.] 0.4
— Then, using max operation.
RUS =
0.4 | 0.2 | 0.1

1 | 0.6 | 0.8
0.3 | 0.3 | 0.5
— Suppose a simples - norm
S (a,b) = a+b-—a-:
bis used then,
RUS =
0.58 0.2 0.19
1 | 0.68 | 0.84
0.44 | 0.44 | O7
— This operation is more optimistic than the max operation. All the membership degrees are at least as high asin
the max operation. ;
Now, using min operation
RNS =
0.3); 0 | 0.1
0.4 | 0.2 | 0.1
0.2 | 0.2 | 0.4
b
— Suppose a simple T-norm T (a, b) = at+b-ab is used then,
RNS
0.20} O 0.1
0.4 | 0.17 | 0.10
0.13 | 0.13 | 0.28
— The above operation is more optimistic than the min operation. All the membership degrees are less than in the mi
operation.:
Fuzzy Logic
“Sem. 7-Comp)
projection set. ora

re la ti on to a fuzzy
binary relation, or a binary
The projection relation brings a ternary relation back to a
fuzzy set to a single crisp value.
w.
EX. consider the relation R as given belo
Yr Yo Y3 Ya
X,{ 0.8] 1 | 0.1] 0.7
| 0/08) 0} 0
x,|09| 1 | 0.7] 08
X means that
Then the projection on
of the first row.
X, is assigned the maximum
the second row.
x, is assigned the maximum of
m of the third row.
- .%3is assigned the maximu
Thus,
1 08 1
Proj. RonX = Z-+y + X3
Xy X2
Similarly
09 tT 1,07 08
j- Yous y4
Ran Y1 y2 * ¥3
Proj.
_ Cylindrical Extension ndrical extension.

ost alw ays use d in combi nation with cyli
n is alm
— Theprojection op eratio jection. It converts fuzzy set toa
relation.
op po si te of pro ,
more or less
Cylindrical exte nsion is
Ex: Consider a fuzzy set, X = Ix, + 0.8/x2+ 1/X3.
A = proj.ofR on
n on the domai nXxYis

= Itscylindrical extensio
ce(A) =
Ce (B)
WW
TechNanuledga i.
Publications
¥ Al&SC (MU-Sem. 7-Comp) 4-20
4.4.2(B) Properties of Fuzzy Relations

relations are statey a
the universe X x Y. Then, the properties of fuzzy
Let R, S and T be fuzzy relations defined on
below:
1. Commutativity
RUS SUR
RNS SOR
Associativity
RU(S UT) (RUS)UT

RN(GSNAT) (RNS)NT
Distributivity
RU(SNT) (RUS)
N (RUT)
RNO(SUT) (RNS) U(RNT)
Idempotency .
RUR R
ROAR R
Identity
R Udp R , R Nop
= op
RUER ER» ROER=R
y.
Where 0, and Ep are null relation (null matrix) and complete relation (unit matrix of all 1s) respectivel
Involution
R
De-Morgan’s law
| Ras RUS
RUS RNAS
are not satisfied.
Law of excluded middle and law of contradiction
i.e. RUR x Er
and RAR # Op
4.4. 2(C) Fuzzy Com position
Composition operation can be used to combine two fuzzy relations in different product spaces.
There are two compositions that are used commonly.

1. Max-min composition
Max - product composition
Fuzzy Logic
Al& gc (MU-Sem_Z-Comp) 4-21 ————
LX x - Min Composition
1. etMA Ri be? fuzzy relation defined on X x Y,
and Rz be a fuzzy relation defined on Y x.
then the max - min composition of two fuzzy relations R, and R, is denoted by R; 0 R2 and defined as,
RoR, = {[(x, Dry vey canin (Hpi(x, Y)» PR2C¥> pyle X ye¥.2€ 2}
OR Prior2 (XZ) = yet y {min (Hpi % y), poly, 2))}
Max - Product Composition
2.
The max - product composition is defined as,
RioRy = (16D. yey (el ¥) Be Os mike Xye ¥.z€ Z)

OR Pri oR2(X,Z) = ye ¥ {Uris y) “Mraly> z)}
binary relations defined onXxY¥,YxZ ;
The following are the properties of fuzzy composition. Assuming R, S and T are
and ZX W respectively.
1, Associativity > Ro(SoT)=>(ROS)oT
2. Monotonicity >SCT=>RoOSEROT
3. Distributivity > R o (S U T) > (ROS) U (ROT)

4, Inverse
— (RO sy '=S oR”
45 Membership Functions
MU)=May.12;.Dec.12,Dec:.13, Dec.14
One way to represent a fuzzy set is by stating its Membership Function (MF). MFs can be represented using any
mathematical equation as per requirement or using one of the standard MFs available.
There are several different standard MFs available.
Types of Membership Functions
4.|. Inoreasing MFs (T Function) |
Hi
GEIST:
2.Decreasing MP (L ibviction),
3.3, Triangular MF. M function)
4. Trapezoid MFs (x tution)
5. Gaussian. MFs
6. Generalized bell MF / Cauchy MF
7. Sigmoial MES ,
my
' TechKnowleags
ual
Pubticaria
se
Increasing MFs (T Function)
An increasing MF is specified by two parameters (a, b) as follows:
0 : x<a_
T(x;a,b) = 4(x-aj/(b-a) aSxSb
: 1 ; xz2b
palx)
1
Fig. 4.5.1: Increasing MF

2. Decreasing MF (L function)
A decreasing MF is specified by two parameters (a, b) as follows : Bax) 4
1
1 ; xSa 1
1
1
!
L(x;a,b) =. (b-x)/(b-a) ; aSx<Sb '
'
{
0 ; x
a b.
Fig. 4.5.2 : Decreasing MF
3. Triangular MF (A function)
— Atriangular MF is specified by three parameters (a, b, c) as follows:

0 3 xSa
(x-—a)/(b-—a) ; aSxSb
ACABO=) OO oe-b) ; bSxK<e
Fig. 4.5.3 : Triangular MF
0 ; x2c
4. Trapezoidal MFs (zx function) .

— A trapezoidal MF is specified by four parameters (a, b, c, d) as follows :
0 ; xSa
(x-a)/(b-a) ; asSx<b
(x ; a, b,c, d)=%) 1
Trapezoid » bsxSc
(d-x)/(d—c) ; cSxSd
0 ; x2d
a(x)
1
a b cq ™*
Fig. 4.5.4 : Trapezold MF

— Analternative expression using min and max can be given as,
E ‘a gC (MU-Sem. 7-
Comp)
( 4-23 Fuzzy Logic
vapen0id (X35 & Dr Cr Y= MAX(min X-a
baal ld
a? 4)
rd—g)>0
The parameters {a, b, c, d} (with a <be .
-
. the x coordinates of the four corners of the trapezoidal
c <d) determine
MF.
Here, if b =c then trapezoidal Me reduces tot

Otria ngle MF,
; <
efficiency,
ers. However due to the simple formulae and computational
Paramet
they are used extensively.
~ Some smooth and non -line ‘ar MFs (Gaussian
i and Generalized Bell) 1.0
are discussed below :
0.8
5 Gaussian MFs 0.6
;
_ AGaussian MF is specified by two parameters {c,0}
; -1/2 (S)
Gaussian (x;¢, 6)=e 9G
crepresents MFs center and

o determines MFs width.
3. Generalized bell MF / Cauchy MF

- Ageneralized bell MF (or bell MF) is specified by three parameters
(a, b, c).
1
bell (x; a,b,c) x-C¢ /®
=—7{_.
—_—
1+) 5
be obtained by a proper selection

- Adesired generalized bell MF can
of the parameters a, b, c.
fi the cen ter ofa bell MF
c specifies 20 30 4050607080 100.
bell MF
a specifies the width of a Fig. 4.5.6: Bell (x ; 20, 4, 50)
and b determines th e slope at
the crossover points.
pe of the curve.
ct o f cha ngi ng these parameters on the sha
~ Fig. 4.5.7 illustrates the effe
Changing ‘b’ (slope)

dth)
Changing ‘a’ (w!
Changing ‘c’ (centre) WY fedtentatst

~ Publications:
1.0 Dibgsigae
Slope= — b/2a ae
0.5
Fig. 4.5.7 : Effect of change of different parameters in Bell MF
The bell MF is direct generalization of Cauchy distribution used in probability theory; so it is also referred to as the
Cauchy MF. The bell MF has more parameter than Gaussian MF, so it has more degree of freedom to adjust the
steepness at the crossover point.
Although Gaussian and bell MFs achieves smoothness, they are unable to specify asymmetric MFs.
Asymmetric MFs
Asymmetric and close MFs can be achieved by using either the absolute difference or the product of two sigmoidal
functions.
Sigmoidal MFs
- Asigmoidal MF is defined by,

1
sig(Xs&C)= Trexpacx-ol
Where, a controls the slope at the crossover point x = c.
Depending on the sign of the parameter a, a sigmoidal MF is open right or open left and thus is. appropriate for
representing concepts such as “very large” or “very negative”.
— They are widely used as the activation function in artificial neural networks.
4.6 Fuzzy Logic

4.6.1 Fuzzy Logic Basics
— Fuzzy logic is an extension of Boolean logic.
-— In Boolean logic we express everything in the form of 1 or Oi.e. true or false respectively
Fuzzy logic handles the concept of partial truth, where the range of truth value is in between completely true and
completely false, that is in between 0 and 1. In other words , Fuzzy logic can be considered as multi-valued logic
In other words, fuzzy logic replaces Boolean truth-values with some degree of truth.
This degree of truth is used to capture the imprecise modes of reasoning.
The basic elements of fuzzy logic are fuzzy sets, linguistic variables and fuzzy rules.
Usually in mathematics, variables take numerical values whereas fuzzy logic alinWs the non-numeric Lingus
variables to be used to form the expression of rules and facts.
WF TechKnawledga
Pobricatians
aigsc (MU-Sem_7-Comp) _. 1 Fuzzy Logic
4-25 ——
aq;
the linguistic variables are Words, specifically
————
—
so on. A fuzzy set
-
v adjectives like “small,” “little,” “medium,” “high, "and
isa collection of couples of elements
ynguistic Variables and Lingulstic Values
wh
Alinguistic variable is a variable +x
Ose values a Te words or sentences in: a natural or artificial language.
consider the variable X = “age ”
Alinguistic variable (“age”) can assume diff aged”, “Mature” and “old”
ifferent linguistic values such as “young”, “Middle
in this case.
wt
Then, ‘age’ can be considered iddle aged” “Mature” and “old”
tneen variable whose values can be “young”, “M
and these values can be characte young (x), Umiddle aged (x) » Umature (xjand Hold (x).
. by MFs
;
and old) and transition from
The universe of discourse is t Otally covered by these MFs (MFs for young, middle aged
one MF to another is smooth and gradual
Grade
1 Young Middle Age Mature old
35 5 35 65
ns
Fig 4.6.1 : Linguistic variable “age” as membership functio
; |
4.6.2 Fuzzy Rules and Fuzzy Reasoning
new knowledge from an existing knowledge.
Fuzzy inference is the process of obtaining a
be repres' ented in some form.
= Toperform inference, knowledge must
represent its knowledgebase.
- Fuzzy logic uses IF -THEN rules to
according to W hich, we can infer the truth
infe renc e in tradi tiona | two-valued logic is modus ponens,
- The basic rule of — B.
h of Aan d the implication A
ofa proposition B from the trut
ato is ripe” then if it is true that
A is iden tifi ed with - “the tomato is re d” and B with “the tom
Ex. If
is ripe”.
true that “the tomato
“the tomato is red” it is also
XisA .
ie. Premise 1 (fact)
if X is A then Y is B
Premise 2 (rule)
clusion) :YisB
Consequence (con an approximate manner.
easoning, modus ponens is employed in
However, in most of the huma n r
~ is red, then it is ripe” and we know that,
“if the tomato
n rule,
. : if we ha ve th e same implicatio
For e.g infer that
red” then we may!
the tomato is more or less
a
. “a”
“the tomato is more or less ripe

Premise 1 (fact) X is A’
B
if Xis A then Yis
Premise 2 (rule) . ‘
Yis B
Conclusion
Where A’ is close toA and
B’ is close to B ’ r TechKnowledga
Publications
Fuzzy Lo Gig :
re t&SC (MU-Sem. 7-COMP)
: ; cedure is called “a
- When A, B, A’ and B’ are fuzzy sets of appropriate universes, the inference pro PProximat,
(GMP).
reasoning” or fuzzy reasoning, it is also called Generalized Modus Ponens
Definition : Approximate reasoning / fuzzy reasoning ; ey
Let A, A’ and B be fuzzy Y sets of X, X and Y respectively. e that the fuzzy implication A B is expressed as a fy,,,
P Y Assum “ifx is A theny
wit age
is B” is defined by,
relation R on X x Y. Then the fuzzy set B induced by “x is A” (fact) and the fuzzy rule
| pip (y) = max min [p1y? 0), Hr Ok y)] = Vx Ear () Adie O 91
or ff = A’o(A>B)
ifxis A, yis B
a. Single rule with single antecedent

min
Fig. 4.6.2 : Graphic interpretation of GMP using Mamdani’s fuzzy implication and max-min composition
Here Wg can be defined as ;
p(y) [Vy (Mar (x) Aba (X))] A Bay)

WA p(y)
Thus, we first find the degree of match w which is the maximum of ly (x) A Ha (x).
Then the MF of resulting B’ is equal to the MF of B clipped by w.
Single rule with multiple antecedent

A fuzzy if then rule with two antecedents can be written as,
“ifx is Aand y is B then zis C”
Premise 1 (fact) : xis A’ andyisB’
Premise 2 (rule) : lfxisAandyis BthenzisC
‘ Consequence : zisC’
min
A Cc
Fig. 4.6.3 : Approximate reasoning for multiple antecedents
Ho(z) = {Vx [Ha (x) Ada (X)]) A
wy
| Scanned by CamScanner
' tp 6) 1 on and G2 Arte onan woman nee ean ys oe
W2
= (WA Ww) Alc (2)
firing strength
when en wW, Wy and w 2 are the maxima
i of the MFs of AQ, A’ and BB’ respectivel
? ectively.
us, W, denotes t .
Th 1 he degree of Compatibility between A and A’, similarly for w2
Since the antecedent parts parts of the fuzzy rule ‘is constructed using and connective, WiAW; is called firing strength or
degree of fulfilment of the fuzzy rule. .
The firing strength represents the degree to which the antecedent part of the rule is satisfied.
The MF of the resulting C’ is equal to the MF of clipped by the firing strength w (when w = W1AW2)
C. Multiple rules with multiple antecedents

as,
The GMP problem for multiple rules with multiple antecedents can be written
Premise 1 (fact) : xis A’ and y is B’
Premise 2 (rule 1) > Ifxis Az and y is By then zis Cy
Premise 3 (rule 2) . Ifxis Az and y is B2 then zis C2
Consequence : zisC’
s
ltiple antecedent
ple rules with mu
ms Ep etneeletes.
Puotications
|
7 _ABSC(MU-Sem.7 -Comp) Fuzzy Logie
Lg
— Here C,’ and C,’ are the inferred fuzzy sets for rule 1 and rule 2 respectively.
.
When a given fuzzy rule assumes the form “if x is A or y is B” then firing strength isis given
of as the maximum of degr;
Bree of
match on the antecedent part for a given condition.
Ex.
If xis Ay or y is By then zis Cy.
min
By By
i rooo nee
7 ‘
Fig. 4.6.5
— In the above example, because two antecedents are connected using or, we take
maximum of w, and wy as a firing
strength.
— Since w.> wy, we take W? as a firing strength and then we apply min implication operator
on the output MF C;.
4.7 Fuzzy Inference Systems

MU!- May 12, May1 ec. 15
Fuzzy Inference System is the key unit of a fuzzy logic

system. Fuzzy inference (reasoning) is the actua
l process of
mapping from a given input to an output using fuzzy logic.
It uses the “IF...THEN” rules along with conne
ctors “OR” or “AND” for drawing essential
decision rules.
4.7.1 Construction and Working Principle of FIS
— Fig. 4.7.1(a) shows the block diagram of general fuzzy
inference system.
Crisp
input
. Fig. 4 .7.1 (a) : Block diagra

m : Fuzzy Inference sys
tem
PF"Y Publicat
leataeutions.
engs : . ; ~
_ Fuzzy Logic
Output of «
rule n
(Fuzzy)
r
Final output y
Fig 4.7.1(b) : Fuzzy Inference using If-Then rules
As shown in Fig. 4.7.1(a), FIS involves five important modules.

Fuzzification Inference Unit (FU)
Decision making / Inferencing Unit

Rule Base
Data Base
Defuzzification Inference Unit (DU)
Fuzzification Inference Unit

a fuzzy set. Here we need to decide the proper
This block performs a fuzzification which converts a crisp input in to
fuzzification strategy.
Decision making/Inferencing Unit

of the control output variable based on the
The basic function of the inference unit is to compute the overall value
rule base.
individual contribution of each rule in the
input is matched to each rule-antecedent.
The output of the fuzzification module representing the crisp
control output
The degree of match of each rule is established. Based on this degree of match, the value of the
control
is, we get the “clipped” fuzzy set representing the
Variable in the rule-consequent Is modified. The result
Output variable.
esent the overall fuzzy value of control output.
The set of all clipped control output values of the matched rules repr
Atule base
rules.
Itcontains a number of fuzzy IF-THEN
A
A databa se
used in the fuzzy rules.
mem ber shi p func tions of the fuzzy sets
~ Data Base def ine s the
Tech m
Fupticattane
|
— The information in the database includes:
© Fuzzy Membership Functions for the input and output control variables
zed values alon g with the scaling facto
rs.
i o The physical domains of the actual problems and their normali
5. Defuzzification Unit -
output into a single crisp value.
— It performs defuzzification which converts the overall control
— The rule base and the database are jointly referred to as the knowledge base.
Working
The input to the FIS may be a Fuzzy or crisp value.
any of the fuzzification methods.
1. Fuzzification Unit converts the crisp input into fuzzy input by using
knowledgebase.
2. The next, rule base is formed. Database and rule base are collectively called
3. Finally, defuzzification process is carried out to produce crisp output.
Methods of FIS
— The most important two types of fuzzy inference method are :
1) Mamdani FIS
2) Sugeno FIS
Mamdani
h
Mamdani fuzzy inference is the most commonly seen inference method. This method was introduced by
and Assilian (1975).
Another well-known inference method is the so- called Sugeno or Takagi-Sugeno—Kang method of fuzzy inference
process. This method was introduced by Sugeno (1985). This method is also called as TS method.
The main difference between the two methods lies in the consequent of fuzzy rules.
.
1. Mamdani FIS
ation.
Mamdani FIS was proposed by Ebahim Mamdani in the year 1975 to control a steam engine and boiler combin
|| ‘ To compute the output of this FIS given the inputs, six steps has to be followed.
1. Determining a set of fuzzy rules.

2. Fuzzifying the inputs using the input membership functions.
3. Combining the fuzzified inputs according to the fuzzy rules to establish a rule strength (Fuzzy Operations).
of the rule by combining the rule strength and the output membership function
4. Finding the consequence
i f ,
(implication).
5. Combining the consequences to get an output distribution (aggregation).
6. Defuzzifying the output distribution (this step is only if a crisp output (class) is needed).
1 Fuzzy Rule Composition In Mamdani Model

}
|
1 — In Mamdani FIS, The fuzzy rules are formed using IF-THEN statements and AND/OR connectives.
'i
— The consequent of the rule can be obtained in two steps.
i
4
o By computing the strength of each rule
By clipping the output membership function at the rule strength. . : vet
| °
wie
ist
AI&SC (MU-Sem. 7-Comp) _ ,
The outputs of all the fuzzy rules are then combine d to obtain the aggregated fuzzy output. Finally, defuzzification is
applied on to the aggregated fuzzy output to obtain a crisp output value.
consider two inputs, two rule
Mamdani fuzzy inference syst
em
Assume two inputs are crisp value x and y
assume the following two rules :

Rule 1: ifx is Ay and y is B, then zis C
Rule 2 : if x is A, and y is B, then z is Cc
- Fig. 4.7.2 (a) shows Mamdani fuzzy inference system using min — max decomposition.
x and y. In the
Fig. 4.7.2 (a) illustrates a procedure of deriving overall output z when presented with two crisp inputs
above Mamdani inference system, we have used min as T - norm and max as T - conorm operators.
The T-norm operator is used for inferencing antecedent part of the rule. And co-norm operator used to aggregate
outputs resulting form of each rule.
Mamdani model also supports max - product composition to derive overall output z. Here the algebraic product is
used as T-norm operator and max is used as T-conorm operator. ,
Ay
: z
ZMOM
omposition
» sy:systems using max - min dec
|
Inference
Fig. 4.7.2 (a) :7 Mam
dani fuzzy
Be , soPMR
Me eine
Zz
N—>
Fig. 4.7.2 (b): Mamdani fuzzy inference systems using max - product decomposition
2. Takagi-Sugeno-Kang (TSK) FIS
— Takagi - Sugeno FIS was proposed by Takagi, Sugeno and Kang in the year 1985.
— Atypical fuzzy rule in TSK model has the form, .
IF xis A and y is B then z = f(x,y)
Where,
x, y and zare linguistic variables.
A and Bare fuzzy sets in the antecedent part of the rule.
Z = f(x, y) is a crisp function in the consequent part of the rule.
Usually f(x, y) is a polynomial in the input variables x and y.
Fuzzy Inference Process

The fuzzy inference process under Takagi-Sugeno Fuzzy Model (TS Method) works in the following way :
Step1: Fuzzifying the inputs
Here, the inputs of the system are made fuzzy.
Step2: Applying the fuzzy operator ,
In this step, the fuzzy operators must be applied to get the output.
First order Sugeno fuzzy model
When f(x,y) is a first order polynomial (e.g. z = ax + by + c) the resulting FIS is called , first order Sugeno fuzzy model.
Zero Order fuzzy model
In zero order fuzzy model, the output z is a constant (i.e. a = b = 0).
The typical form of the rule in zero order FIS is
IF xis Aand y is B then z=c
a e. uzzy Logic
gpscmuemtoms
7-Comp)
Al&SC (Mu-Sem.
average
nt
where cisa consta
|
is obtained vi?! welentod
in this case the output of each fuzzy rules Si i and hence the overall output
s a constant
method-
p level 2,0,of each rule isi weighted
the ou tput i by the firing strength w; of the rule . _
wa min or
product
Ww, 24 = Pix tan *
pA
Wo Zp = Pox + Gay + 12
average
J ]peones
wy421 + W222
wy +Wo
Z=
.
no FIS
oning in Suge
Fig. 4.7.3 : Reas
Model ip function.
i Sy s te m an d the Sugeno outpu t membersh
th e Ma md an he ba si s of
bet ween them is on tl
Comparison T he main d iffe
rence betwe en
:
ship Function er linear 0 r co
nstant. of fuzzy rules
Output Member ip fu nc tions are eith lie s i n th e consequence
tput member sh em al so
The Sugeno ou e between th
oc ed ur e : The difference
ation Pr e also differs.
at io n an d Defuzzific de fu zz if ic ation pr ocedur
- Ag gr eg tion an d mdani rule.
e sa me their agereBe th e Su ge no rule tha n the Ma
and due to th exi st fo r co ntroller.
at he ma ti cal rules me ter s th an the Mamdani
m ra
Rules : More re adjustable
pa
- Mathema tical no co nt roller has mo
The Su ge
ameters :
~ Adjustable P ar s
ic at io n © f In put Variable t.
47.2 Fuzzif into a fuzzy se
nv er ti ng @ crisp set
s of co e uncertain,
i 5 the proces uistic variab
les. te Tr ather they ar
- . Fuzzification in to li ng prec is e an d ac cu ra
transfo rmed are not very
He re th e crisp value is th e in put values
- mes be
many 4 ti cases, variable may
rd problems,
Inreal wo ness of data. In such
~
un kn ow n. gu en es s an d incomplete
imprecise and may arise du
e to the va
membership
f unction.
The un ce rt ai nt y as fuz zy
ted
n be represen
-
a5 fu zzy and ca
nt
represented
p yalue of assignme
membershi
Methods of
W TePuotbttaivonatleiodenss
_AI&SC (MU-Sem. 7-Comp)
1. Intuition
— As the name suggest, this method is based upon the common intelligence of human. The human develops
membership functions based on their own understanding capability.
Cool Warm Hot
010 20 40 60
—> Temperature
Fig. 4.7.5 : Membership functions for fuzzy variable “temperature”

— As shown in Fig. 4.7.5 each triangular curve is a membership function corresponding to various fuzzy
(linguistic)
variables such as cold, coal, warm etc.
2. Inference
— In inference method we use knowledge to perform deductive reasoning. To deduce or infer a conclusion, we use the
facts and knowledge on that particular problem. Let us consider the example of Geometric shapes for the
identification of a triangle.
Let A, B, C be the interior angles of a triangle such that,
A2B>C>0° and A+B+C=180°

For this purpose, we define five types of triangles.
R = Approximately Right-angle triangle
p
| = Approximately Isosceles triangle

wN
E = Approximately Equilateral triangle

= Isosceles Right-angle triangle
“oP
T = Other type of triangle
Now, we can infer: membership values for all those types of triangles through t the method of inference because we
possess the knowledge about the geometry of their shapes.
The membership values for five types of triangle can be defined as,
pwR(A,B,C) = =e 1A-90°|
HI(A,B,C) = a ia-ma-o)
pE (A,B,C) = 1-3 551A-Cl
ul-R(A,B,C) = pINR(A,B,C©)
= min {WI (A, B, ©), uR (A, B, C)}
wT (A,B,C) = (RUIUe) =RAINE
- algSC (MU-Sem..7-Comp)
; gx
H(A, B,C)
:
= (80, 65,35}
Mr (A,B,C) ) == 121495 180-901=3
mi 1 (A, (A,B,B,CC)
) = 1-61-4 minmi ( 15, 45} =5
Mp (A, B,C BC) == 1-755 l451=4z

Hin (A,B,C) = min (4 Ha} =
wr = RNIN ES
min} i i
= Ale
9°4°4).
|
4, Rank ordering
nion methods
es are ass ign ed by a sin gle ind ivi dual, committee, a poll and other opi
_ |nrank ordering method, preferenc
ues to fuzzy variables.
can be used to assign membership val determine ordering of the
isons and they are used to
by pair wise compar
Here the preferences are determined
membership.
Example: orange,
e and thei r pair wise pre fer ences among the colors red,
pond to a questionnair
Let’s suppose 1000 people res
yellow and blue is given as below.
525 545 661

Red - 517
477 576
- 891
Orange | 483
534 614
159 -
Yellow 474
- 643
523 466
Green 455
357 -
424 386
Blue 339
4 Angular fuzzy sets ts on ly in th eir coordinate

description.
fuzzy se .
An i from no rmal s; henc e they are of repeating shapes
for every 27 cycles.
angle
Bular fuzzy sets differ universe of
as “truth
on @ s, which are known
Angular fuzzy sets are defined h
quantitative desc
ription of the linguistic variable
used int e
Angular fuzzy sets are
Values”,
nateded pond.pond.
aminat
a contam
Wi We know that,
¢ water samples are taken from
lev:
rample
°
et’s consider that pH values
cod
i | solution.
.
a a
FP, value is 7 means it’s 2 neutra
F _Al&SC (MU-Sem. 7-Comp)
- Levels of P, between 14 and 7 are labelled as Absolute Basic (AB), Very Basic (VB), Basic (B), Fairly pai (én
Neutral (N) drawn from 6 = F to 6==
- _ Levels of P, between 7 to 0 are called neutral, Fairly Acidic (FA), Acidic (A), Very Acidic (VA), Absolutely Acidic (AA), arg
drawn from 6 = 0 to a=—F.
— Linguistic values vary with 8 and their membership values are given by equation,
Eg = ttan@
Here ‘t’ is the horizontal projection of the radial vector.
+ (0)
AB VB B FB N FA A VA AA
v
&
O= + 1/2
a AB §VB= 3n/8
4.7.3 Defuzzification
— Defuzzification is the process of converting a fuzzy set into a crisp value.

— The output of a fuzzy process may be union of two or more fuzzy membership functions. In that case we need tonfag 2
crisp value as a representative of the entire fuzzy MF. . d
— Different methods of defuzzification are listed below:
geet mate
etn
AN&SC (MU-Sem. 7-Comp) :
a
Methods of de
fuzzificatio n
~| 1. Max membersh
; ip principle — |
; oe
2. Centre of Qravity or cen
troid
3. Weighted average |
4. Mean of maximum (MoM)
5..Centre of sums (CoS) : —
=> 6. Centre of largest area
“>| 7. First (or last) maxima
—> 8, Bisector
Fig. 4.7.7 : Defuzzitication methods

1. Max-membership principle / Height method
- This method is limited to peak output functions. It uses the individual clipped or scaled central outputs.
Fig. 4.7.8 : Max-membership
The algebraic expression is:

wg(x*) 2 pig(x) for all x EX
Centre of Area / Gravity (Centroid) Method

This method is the most preferred and physically appealing of all the defuzzification methods.
- is method i —
function. (i.e. it takes uni f
below the com bined membership |
tre of the area
~ This method determines the cen
all output fuzzy sets)
ts). lapping. area will be considered only once. Thus overlapping
;
areas are not reflected.
So, if there exist an ove rlap ,
lex and therefore results In slow inference cycle.
This operation is computationally comp
Algebraic expression /s
TForeontinuous | ForDiscrete
pete xd 5 be (Xi) +X
Juco dx |, tek
2X be (4)
Ist
SS CLC py TechK nomiea

Publications
Taken
only once
Fig. 4.7.9 : Centroid method

3. Centre of Sum (COS)
This is faster than many defuzzification methods that are presently in use.
This method involves the algebraic sum of individual output fuzzy sets, instead of their union.
The idea is to consider the contribution of the area of each output membership curve.
In contrast, the centre of area/gravity method considers the union of all output fuzzy sets.
In COS method, we take overlapping areas. If such overlapping areas exist, they are reflected more than once.
0 2 4 6]
8 10 oh
ke
x
Fig. 4.7.10 : Centre of Sum method
b x Hcy (x) dx
k=1
x* = x*=
N
kz He; (x) dx
k=1
Advantage : It can be implemented easily and leads to a faster computation.
4. Weighted average method
— This method is only valid for symmetrical output membership functions.

Algebraic expression is :
n
DL Wg (Xj)
° xi
i=l
x* =
n
DX pe (x)
i=l
mp)
‘a Alasc (MU-Sem. 7-Co .
Fuzzy Logic’
H 4-39
:
{1
\'
{|
by,
b x
method
Fig. 4.7.11: Weighted average .
put by its respective
ighted a
.
ing eac h me mb er sh ip function in the out
verage method Is formed by weight
m_— ue.
maximum membership val
ral form of defuzzification.
fun cti ons sho wn in Fig. 4.7. 11 would result in the following gene
_ The two
ye (ax 0.5) + (b x 0.9)
0.5 +0.9
maxima)
5, Mean-max membership (Middle of ; except that the
max -me mbe rsh ip prin cipl e (height defuzzification) method
d to the
This method is closely relate be non-unique (can be more than
one).
bership can
locations of the maximum mem mizing MF.
havi ng max imu m membership value of Maxi
rage of th e elements
In that case we take the ave
mum method
Fig. 4.7.1 2: Mean of maxi
oe
on ,
- Algebraic expressi is
x* = 72
works for convex:

This method only
y t is non-convex i.e. it consists of at
when the combined output fuzz se _
6 Centre of largest area .
used
a r e a m ethod is |
t rea an d defi to
~ The centre of! a r g e s t
. n v e x f u z z y s u b s et with the larges a crisp output value x*
s o
ex fuzzy § ub 5 the c een
s e t
least two conv ‘
d e t ermine a oc
met h o d t fuz z y
- Then the ea of th
e larges
tr f ar
e the cen tre o in
b
=
dX aa
he, (x) +
x
,
a.
rgest are
h a t h as the la
t
fuzzy subset
convex
:
atipas
, C m !5 t h e . Babli
c
Where c_ i s
Al&SC (MU-Sem. 7-Comp) 4-40
> MF with
largest area
Fig. 4.7.13 : Centre of largest area

7. First (or last) of maxima
— This method uses the overall output (i.e. union of all individual output MF).
First of maxima is determined by taking the smallest value of the domain with maximized membership degree.
Last of maxima is determined by taking the greatest value of the domain with maximized membership degree.
Fig. 4.7.14: First (or last) of maxima

8. Bisector method
This method uses the vertical line that divides the region into two equal areas
as shown in Fig. 4.7.15 This line is called
bisector.
x" | Bisector
Fig. 4.7.15 : Bisector method of defuzzificat

ion
4.8 Fuzzy Controllers
- Most commercial fuzzy products use Fuzzy Know
ledge-Based Controllers (FKBC).
— The principal structure ofa FKBC Is shown
in Fig. 4.8.1.
UP Techtnowlengé : oe
*PUDTICatians ns
, * ‘ Eu
J
Ale gc (MU-Sem. 7-Comp) Logic

ik Fuzzy
441
Crisp Process state
.
value
: Crisp control
output values
-
Option
obligato:
Fig. 4.8.1 : The structure of FKBC
three important modules.

_ Asshown in Fig.4.8.1, FKBC involves
4, Fuzzification module
le
2. Decision making or Inferencing modu
3, Defuzzification module
more components
_ Inaddition to this, it uses two
= Data base and
} Knowledge base
—Rule base
Fuzzification module :
1.
functions.
forms the following two
Fuzzification module per
(a) Normalization normalized

of input variables in to a
on wh ic h ma p: 5 the physical values qu ir ed .
le transformati used then this block is no
t re
This block performs a sca n-norma lized domain is
optional. If ano
universe of discourse. This block is
the proper
(b) Fuzzification a crisp input in to @ fuzzy
set. Here we need to decide
n which converts
This block performs a fuzzificatio
fuzzification strategy-
ule:
ng mod variable based on the
2 Decision making/Inferenci pute the overall value of the control output
, ina |sIs to com
on of the inf erence engine
The basic functi | e inthe
rule base- rule-antecedent.
oe
he crisp input is matched to each
individual contribution of each ie ule re e value of the control outp
ut
ific n this degree of match, th
The output of the fuzz ation mod y set representing the con
we get the “clipped” fuzz
trol
of eac
The degree of match
Variable in the rule-consequ fuzzy value of control outpu
t.
Output variable.
vari the matched F ules represent the overall ;
yy . ut values of Ween
Reape
| he set of all clipped control outp
Defuzzification module :
Defuzzification module performs two tasks :
(a) Defuzzification
It performs defuzzification which converts the overall control output into a single crisp value.
(b) Denormalization module
This block maps the crisp value of the control output into the physical domain. This block is optional. It is used only if
normalization is performed during the fuzzification phase.
The knowledge base basically consists of a database and a rule base.

— The database provides the necessary information for proper functioning of the fuzzification module, the rule base
and the defuzzification module.
— The information in the database includes :
° Fuzzy MFs for the input and output control variables .
° The physical domains of the actual problems and their normalized values along with the scaling factors.
4.8.1 Steps in Designing FLC
Following are the steps involved in designing FLC
Steps In Designing FLC
|. 1.| Nentification of ieee.
apy! 2._ Fuzy subset configuration
em} 3. Obtaining pemean. function E
4. Euzay turule base configuration = ’
me 5S Euzeification
mp 6. Combining fuzzy outputs -
to a, Defuzzification.
THT
Fig. 4.8.2 Steps in designing FLC
Identification of variables : Here, the input, output and state variables must be identified of the plant which is under
consideration.
Fuzzy subset configuration : The universe of information is divided into number of fuzzy subsets and each subset Is .
assigned a linguistic label. Always make sure that these fuzzy subsets include all the elements of universe.
3. Obtaining membership function : Now obtain the membership function for each fuzzy subset that we get in the above
step. |
Fuzzy rule base configuration : Now formulate the fuzzy rule base by assigning relationship between fuzzy input and
output.
Fuzzification :The fuzzification process Is initiated in this step.
m. 7-Comp) 4-43
ass, (MU Se : =
l ying meh fuzz = . =
g fuz:zy ou tp uts : By app
ppl
im at e rea son ing , locate the fuzzy output and merge the
m, =. <=
mbinin zy. approx
?é
ly, initi iate defuzzification
jon : Finally, i Process to fi orm a crisp output
é pefuazificat
Ss
advantages of FL
48.
for reasoni ning.
“It uses VEY simple Mathematical concepts
j
an fis can be modified by y just addini g or deleting rules due to flexibility ofof fuz y fuzzy logi
logic.
orted, noisy input inf rmation.
y logic systems can take imprecise, dist , ; info
Fuzz
struct and understand
FLSS are easy to con . it resembles.
human re as
oning,
ne, 2s
solution to co mplex problems in all fields of life, including medici
ruzzy !0! g ic is a
making.
and decision
ntages of FLSs
4 8.3 Disadva
ch to fuzzy system designing.
There is no systematic approa
le only when simple.
They are understandab
h do not need high accuracy.
They are suitable for the p roblems whic
g Solved Problems
SEONG ENCES
of all integer numbers.

s
un iv er se of di scourse be the set

Soin.: Let
X = Integers
defined as
set “N um be rs close to 6” can be
Then fuzzy 1
px &) = 44(x-6)
0s
ip of eac element. pa (x)
s th e pl ot of de gr ee of membersh Table P. 4.9.1 - x
and corresponding
Fig, P. 4.9.1 show
pa
1.0
i
09
0.8
07
08
0.5
1
0.4
0.3
0.2
Teyy Tech inewletes

Publications
‘ Fuzzy Lj .
r close to 10". °=! ss
Ex. 4.9.2: Model the following fuzzy set using the suitable fuzzy membership function “Numbe
; i
Soln. :
Let X be the universe of discourse.
X = Integers .
Then fuzzy set “Number close to 10” can be defined as,

1
MRO) = TE K-10)
In 7
: Fig. P. 4.9.2 shows the plot of degree of membership for each element. :
Table P. 4.9.2 : x and corresponding By (x)
pa 5. | 0.03
6 | 0.05
0.9 ry
i 8 | 0.2
0.8 I ‘
0.7 { 1 9 | 0.5
i
0.6 i i 10 | 1
Dela ce aman
aga ee oO 11 | 0.5
é
o4
0.3 !/ ‘\ 7 12 | 0.2
--- 2-2 nnn nnn @ Jenne nena e nn nnn nn @------
0.2}------------------ 13 | 0.1
0.1: _® , Sa 14 | 0.05 _
0 1 2 8 4 5 67 8 9 10 1 12 13 14 18 |15]0.03 |
Fig. P. 4.9.2 : Plot of x 5 (x)
Ex. 4.9.3: Model the following fuzzy set using suitable membership function. “Integer number considerably larger
than 6”. :
Soln. :
Here universe of discourse is set of all integer numbers.
xX = Integers
Then fuzzy set for “Number considerably larger than 6” can be defined as.
1
Wy @) = 1 +*q_oy
~6)
Fig. P.4.9.3 shows the plot of x —>j14(x)

So membership function for “Number considerably larger than 6” is defined as
0 »x<6
baw =) —L—
1
»X>6
+o o>
-Sem. 7-Comp)
0.94 6 {0
4'
0.8
7 105
i.
0.7 hu
0.6 i 8 | 08
0.5
9 | 0.90
0.4
0.3 | 10 | 0.94
1
0.2
i } 11 | 0.96
0.1 '
dee 12 | 0.97
Soln. :
The following are a-level sets
(1,2,34,5.6)
{2, 3,45,,6}
{2, 3,45}:
{3, 4,5}
3.4) "
(4) .
level sets.
Following are strong a- (23,4, 5,6)
(2, 3,4,5}
(3,45) tan
(3, 4) hy
@eeid bieladuidads
set
fo r ih e flowing fuzzy
levelsets
and strong©ol
Find ou t all a-level
set
8) (1 0, 1) (12, 0.8), (14, 0.6)}
F.4.9.5: (7,0 6) , (8 ,, 0.
(5 , 0 - 3 ) , (6: 0.4)
-2):
A=((3, 0.1), (4, 0
Soln, :
G-level sets
Al&SC (MU-Sem. 7-Com
Ags = (5,6, 7, 8, 10, 12, 14}

Ags -= {6,7,8, 10, 12, 14}
Aog = (7/810, 12,14}
Ags = {8,10,12)
A, = {10}
Strong o-level sets
Aor = (4,5, 6,7, 8, 10, 12, 14} -
Aox = (5,6, 7,8, 10, 12, 14}
Agy = {6,7,8, 10, 12, 14}
Aga = (7,8, 10, 12, 14}
Age = {8,10,12)
Ags = {10}
Ay = 6
Ex. 4.9.6. —A realtor wants to classify the houses he offers to his clients. One indicator of comfort of these houses is the
number of bedrooms in them. Let the available types of houses be represented by the following set.
U={1, 2, 3, 4, 5, 6, 7, 8, 9, 10} a 7
The houses in this set are specified by the number of bedrooms in a house. Describe comfortable house for
4-person family’ using a fuzzy set.
Solin. :
The fuzzy set for “comfortable type of house for a 4-person family” may be described as,
A = {(1, 0.2), (2, 0.5), (3, 0.8), (4, 1), (5, 0.7), (6, 0.3)}
yt bo. .
1.0 +
0.9 +
0.8 7
0.7 T
0.6 +
0.5 +
0.4 7
0.3 +
0.2 + 2 A
0.1-+ a
—> U
12345687
91
Fig. P. 4.9.6: Plot of up, (u)
1 l&SC (MU-Sem. 7-Comp)
x 49.7 : Assume A=" .

considerably larger than 10”and =“
B X approximately 11” characterized by
= {x Ba) |X € X} Draw the plot§or both the
Sets and show AUB and A 0B in a plot.
soll:
Fuzzy set Acan be defined as,
O . ,x10
Wy (x) TT, x> 10
1+——
(x- 10)
set B can be defined as, ©

1
He) = 75-1:
Then,
min (1 + (x10), +@-11y'] »x> 10
Hx ng &) 0 ,xs10
larger than 10 and

y set Aan d B repr esen ts a new fuzzy set “y considerably
ion on fuzz
That is, intersection operat
approximately 11”.
and
= {max((1+@- ior +(x-11)y'] x€ X
ppg
1.0 7
0.9 7
0.8 7
0.7 7
0.6 T
0.5 7
0.4 T
0.3 7
0.2 7
0.1 4
ere
Al&SC (MU-Sem..7-Comp)
pL a
17 = 4
0.9 Prrrnn nner nnn nents nnn 0g 1
nn nnnc cen ft dann ——
0.8 pannnnnnnennnnnnnn 1
O77
0.6
0.5 [n nnn nanan nn nnn :
0.4 i
0.3 i 4
0.2 4--------------------¢
--4 choo
D1 fannaannnonnnays plintenn enn pann fanaa ' 1
p<} 4 —— j}—}-—+ >X _.

5 6 7 8 9 10 11 12 13 14 15 16 1
Fig. P. 4.9.7(a) : Plot of AUB

Ex. 4.9.8 : Model the following as fuzzy set using trapezoidal membership function “Number close to 10”.
Soin. : ,
“Number close to 10” can be represented by,
IG Pa : 0 nt oi VX s a
(x=a)/(b=a) , aSx<b
Trapezoid (x ; a, b,c,d) = 1 » bsxSc

(d-x//(d—c) , csx<d
0 » x>I15
=12
We have selected a =5, b=8,candd=15
“O os, x<5
(x=5)/3, 5<x<B8
id 12,15) = 4 >. 16,
(x 35,8,
Trapezo 8Sx<12_
(I5—xy/3 , 12$x<15
0 > x>15
1
'
'
'
1
I
Il
'
t
'
1.
!
'
'
!
1
'
'
t +
4567 8 9 10 111219141516 |
Fig. P. 4.9.8 : Trapezoldal MF for “Number close to 10”
AlgSC (MU-Sem. 7-Comp)
bee 4-49
ex. 4.9.9: LetA={a,, @},B = (b,,b,,
a} C = fo, C}. Let R be a relation from Ato B defined by matrix.
[Tey bz | bs |
04/0510
a | 0.2 | 0.8 |0.2
Let S be a relation from B to C
defined by matrix
Cy Cy
b, | 0.2 |-0.7.
b, | 0.3 | 0.8
b3 | 4 0
Find : (1) Max-min composition of R and S.

(2) Max-product composition of R and S.
Soin. :
(1) Max-min composition
T Cy Cy
T=ROS = a, | 0.3 | 0.5
az | 0.3 | 0.8
T (a,c) = max (min (0.4, 0.2), min (0.5, 0.3), min (0, 1) )
max (0.2, 0.3, 0) = 0.3
T (a1, C2) =
max (min (0.4, 0.7), min (0.5, 0.8), min (0, 0) )
max (0.4, 0.5, 0) = 0.5
T (a,c) = max (min (0.2, 0.2), min (0.8, 0.3), min (0.2, 1) )
max (0.2, 0.3, 0.2) = 0.3
T (a2 C2) =
max (min (0.2, 0.7), min (0.8, 0.8), min (0.2, 0) )
max (0.2, 0.8, 0) = 0.8
(2) Max-product composition

T | c C2
T=ROS = ay-| 0.15 | 0.4
a, | 0.24 | 0.64
T (a cy) =
max (0.4 x 0.2, 0.5 x 0.3, 0 x 1)
max (0.08, 0.15, 0) = 0.15

T (ay, C2) =
max (0.4 x 0.7, 0.5 x 0.8, 0 x 0)
max (0.28 , 0.40, 0) = 0.4
max(0.2 X 0.2, 0.8 x 0.3, 0.2x 1)
T (a, cy) =
P_Al&SC (MU sSem.7-Comp)
= max (0.04, 0.24, 0.2)=0.240 © 8.08ys
x 0)
x 0.8, 0.2
T (ay C9) = max (0:2 0.7, 0.8
max (0.14, 0.64, 0) = 0.64
- i,
Ex. 4.9.10: High speed rail monitoring devices sometimes make use of sensitive sensors to measure the deflection
of the
earth when a rail car passes. These deflections are measured with respect to some distance from the rail cay
and, hence are actually very small angles measured in micro-radians. Let a universe of
deflection be
A=[1, 2, 3, 4] where A is the angle in micro-radians, and let a universe of distance be D =[1, 2, 5, 7] where p
is distance in feet, suppose a relation between these two parameters has been determined as follows :
0.3
0.7} 1 0.2
_forlo4a}1
Now let a universe of rail car weights be W = [1, 2], where W is the weight in units of 100,000 pounds.
Suppose the fuzzy relation of W to Ais given by, ,
Using these two relations, find the relation R' oS = T.
(a) Using max-min composition.
(b) Using max-product composition.

Solin. :
First find R’.
R =(D2/0.3/1 |0.7/0.1] and S =]A2/0.5]1
D3/0.1]0.3]1 10.4 A310.3/0.1

D,jO |0.1/0.2]2 Ado |o
o, | 05]
o, |} 03} 03]
D,)02| 04
T(D,. WI) mat (1, 0.2, 0.0)= 1

T(D;. Wy max (0.4, 0.2, 0, 0) = O04
T(D. Wy) max (0.3, 0.5, 0.3, 0) = 0.5

T(D,, W;) max (0.3, 1,0.1,0)= 1
T(D;, W)) max (0.1, 0.3, 0.3, 0) = 0.3
T (Dy, Wy) max (0.1, 0.3, 0.1, 0) = 9.3

T (Dg Wy) max (0, 0.1, 0.2,0) #02
T (Dy, Wy) max (0, 0.1, 0.1.0) = 0.1
position
fa) Using max product com
Tw, [w2
p, {1 | 04
Ts Db, {os |i |
0, |03 | 03)
0, | 0.06 | 0.1 |
03,0%0) = max (1, 0.10 ,= 1
0. Q)
T (Dy Wd max (1 «1,02% 05,0"
O.1,0%0)
max (1 «04, 0.2 * 1.0%
(DI, W2)
= 0.4
max (0.4, 0.2. 0, 0)
O)
mar (0.3 1,1 «x O05, 0.7 O35, O&
(D2, WD
#3
ayar (0 3,0.5.0.21, 9)
7 © O04, 0.1 80)
mar (3 «OA, LX 1.0.
‘7 (D2, We)
t
ryan (0.92, 1. O07, 0) #
bx OS, OA & OF
mar (lx ta x OS,
(p3, WD)
404
ajar (8, 0 13,04, 07
OF EL ROL, Ob & OF
mas (OE id,
y(D3, W2)
Op ad j
evar (UO, O Gb
EL OLAG GPK OR ERO
mas (Ox
Ty (DA, Wt)
OD WOR
pian QUA, 13.00, AN AI,
2e Ot, bk eO)
man (me Od, Oh «ft G@
‘yps, Wa) Ob
" pan Gb, 4nd, Gp
oneet pest SAE OED RHE
a ° Posters —
—
0 fil ete
Let X be a reasonable
age interval of human bei
ng.
X = {0,1,2,3,..., 100}
Then a fuzzy set “Middle age” can be represented using Trapez
oidal MF as follows.
0 ; x $30
(x-30)/10 , 30<x<40
Trapezoid (x; 30, 40, 60, 70) = 1 , 405x560
(70-xV/10 , 60<5x<70
0 » *&>70
BH A
1.04
0.9 +
0.8 + '
O77
0.6 + !
0.5 + '
0.4 +
0.34 !
0.2 +
0.1 +
+ r ey
0 10 20 30 40 50 60 70 @
-a--rb c ohU6Ud
Fig. P. 4.9.11 : Trapezoidal MF
for “Middle age”
Ex. 4.9.12: Represent the set of old people as a
fuzzy set using appropriate membershi
Soln. : p function.
65
Let X = (0,120) set of all possible ages.
“Ages
A
yp
0 20 40 60 80 100 qho x
Fig. P. 4.9.12: Mem
bership function
for “old People”
0, O<x<6@@
Hod (x) = ) (x-60)20, 69 Sx <9
1, x2 80
A | Fuzzy Logic’
Develop graphical representatio 4-53
1993" .
ft “not”. The temperature ran Nn of Membership tuner
linguistic variables “cold”, “warm” and
temperature. 9€S from © to 109°, UNction to describe
and “warm OF hot”
Oc. Also show plot for “cold and warm”
:
ei
_
0 10 29 30 70, 80 90 100 »x (temp)
Fig ba 40 50 60
:
4.9.13 : MF for
cold, 3 warm and hot
- P.
, temp.
rm”
b Plot for “cold and wa
p
rd x Cold aiid warm
—» x (temp)
0 10 20 30 i
40 50 60 70 80 90 100
Fig. P. 4.9.13 (a) : MF for “cold and warm”
), Plot for “warm or hot”

Warm or hot
x (temp)
co. 60. 70 80 2%
or hot”
: MF for “warm
Fig. P. 4 9.13(b)
ee
84.914: Given two fuzzy sets-

~ foi 02 03}
A=\7 +273
0.5
~ 0.6 , 0:5 04 05}
-{98 +s 74
atioions ns on A
r f o r m f o l lowing oper
Pe
(1) Union
(2) Intersection
_ .
(3) Set difference WwW [ukrnomielas E
Re n's law.
Se (4) Verity Demorg®
Al&SC (MU-Sem.
7-Comp)
Soln. ;
1. Union
2. Intersection
=| 0.1 0.2 03
ANB = IT +y+3 I
3. Set difference
~ ~ ~ ze
AIB = AN
~ 0.1 02 03
A = pty +>}
= O04 OS 06 0.5
B= iT +7 +57 +7)
rare = 0.1 0.2 03

ANB = TT +7 +>)
B | A = B nT A
os 0.6 0. O4 0
B = TT +5 +> +23)
7 0.9 O08 0.7

A = +> + z
= 06 05 O04 }
BOA = +> +
4. Verification of Demorgan’s law
To verify Demorgan’s law first normalize both the sets A and B :
" 01 02 03
AS Up y+ +)oO
~ 06 05 04 05
Betpty epee)
LH.S:AUB
~~ 06 05 04
AUB = (pey +e)9,
=——=
AUB =
04apse
05 06 0.
R.H.S:AAB
= 09 08 07 1
e
A = {498,071} |
} ms Ps _———
Fubticatians
-Sem. 7-Comp)
poresRa
= i
B . {%4 05 06° i
SU ee Ob 05 P ; if
AMB = {9405 06 05 (2) ‘
since LH.S. = R.H.S. hence proved, rtp+gty} :
:
<%=oa Aub |i:

LHS. :AMB :
ANB = {2102-030 i
‘tat ah
ANB = eer “afl i
RHS.: AUB ;
=
AUB
= (09 08
= {92,9897 1}
07 1: :|
@)
R.H.S. hence proved.
yu . _ u _
1 ,
0.7
| 1
1 Ay | | 4 mx
J | —x 1 2 3 4 5 6
0 1 2 3 4
Fig. P. 4.9.15
Soln,:
L Weighted average method ON tee

oe wim
individual furzy Y set
method we first find the centre of each individu
Inweighted average
Centre of A, = 2.5 Centre of Az = 4.0

; centre
at the
Next we find the membership value
= 0.7 eve
Membership value of centre of A,=0
Membership value of centre of A2 = 7 0.7*2.5)+(1 x4.0 _L5+4 = 3.38
I yk = 0.7+1 1.7 ‘
Centre of sum method b ene

x urve. es) X (distance between parallel lines
ie Fist find the area of each individual c sum of len hof parallel In -
Mow that, Area of TraP

r ezoid = = =
~ .
Wrsneen
=,
(1+3)*0.7 *
4,
Therefore, Area of A; = 2
~ 2+4)*1 6
Similarly, Area of Ap. =. baat 5.3" 3
Next, find center of each curve.
Centreof Ay = 2.5
Centre of Ay =4
(2.5 x 1.4) + 4x3) 15.5 _
X* = 1443 =44 = 3.52
Ex. 4.9.16: Consider three fuzzy sets C, , C, and C, given below.-Find defuzzified value using :
(1) mean of max (2) centroid
(3) centre of sum and (4) weighted avg. method.
0.75
0.5
0.25
1.0
0.75
05+
~ 0.25 +
Ee ok Fig. P.4.9.16(b) : Fuzzy set C,
(MU-Der._+ -vomp)
Ae
golf *
First find aggregation of all MFs (union)
55 3.6
set of C, ,C, and 6;
Fig. P. 4.9.16(c) : Aggregated fuzzy
4) Using Mean of max
membership value In
we take the mea n (ave rage ) of all the elements having maximum
since C3 is the maximizing MF,
~
Cy
64+7 13
x* = 2 = 2 =6.5
2) Using Centroid method

ae fig (2) 24
z=
bap (z)dz - -
_—
rn) - te ox
, 5.5 [zde+ J @-22dz
{ (05)zdz+ J (2-5) zdz+ Oo ed
— 4 >
1 36
J (0.3z)dz+ iE nde + ROU S5on , +
-| [(.3z)zdz+
1 3.6 4
0 6 © 3 7 Ser 3
Jdz+
_5)dz+ eee J'(8-2)2 dz
zo2 ) da + J osa+ |e
+| 10.32) dz+ 1J (0.3) d2+ 3.6(5 4 ise Pest j
+!
0
= 4.9 é
ae
3) Using centre of sum method yee

i Q
i
fuzzy set -
First find area of each individual
Area of C1 = 12
:
of Cz, = 15.
Area >
.
Area of CG, =2
+ t. . oe
a
individual fuzzy S&
Then find centre of each
of C1 = 2.5
Centre
vive sieve
Centre of C2 5
t OVATE RMS eA
Centre of Cs J BS
— : OS
, 4-58
AN&SC(MU-Sem. 7-Comp)
+(2x 6.9)
(2.5 x 1.2) +. 5x5)
= 12+15+2
4) Using weighted average method
Find centre of each individual fuzzy set.
Centre of C = 2.5
Centre of C 5
Centre of C; 6.5
_ Find membership values of these centres.
Membership value of centre of C =0.3
Membership value of centre of Cy =0.5
Membership value of centre of C; =1

(2.5 x 0.3) + (5X 0:5) +(6.5X1) — 9.75
= 0.3+05+4+1 = 18
x* = 5.146
. .
Ex. 4.9.17: Given fuzzy set..
~ = {9+
f01 03,0811,
Ba Grargt
08
sf
-
Find core and support of fuzzy set A

Soln. :
i : His 3 j
Core ofA = { 4, 5 }—» Membership value equal to 1
:
Support ofA = {1, 2, 3, 4, 5, 6} +> Membership value > 0
Ex. 4.9.18: Consider following two fuzzy sets
0.2 | 98, Oe +4
0.5 }
~ f0.1 02
B= {eS
1
02 a}
A= {+
Find: 1) Algebraic sum . 2) Algebraic product
3) Bounded sum 4) Bounded difference

Soln. :
1). Algebraic sum
Way BO = CHa) + Bp (x) 1 - [ea CO) - pp (x) ]
Seaweed
2) Algebraic product
MA-BO = BZ®- ns @]
— |_|
3) Bounded sum
FA@B® = min[l, pr@

+ ns WO)
{
| = min{ 1,{93493,98,15}} (93,9508,7}

| . 0.3 0. . . 03 405 06 1
4) Bounded difference
PAQB® = max[0,ux@ - HE WO]

max RYO {o,{24,01,02,=051)
UT tats t eg V7
01,91,
tatat4
92,9)
Ex. 4.9.19 : For given.two fuzzy sets
A 03 07 1 ~ 0.4 09
A ={ So art and B = {f2+08 }
XX % My Ya
Perform Cartesian product.

Soln. :
For Cartesian product we consider min operator (denoted by a)
{ 03A04 03A09 O7A04 07009 1A04 1A09 }

+ (x2, ¥1) * (x2, Y2) + (x3, 1) + (X35 Yo)
—<_ {X1, y1) * (x1; Y2)
{ 0.3 0.3 0.4 0.7 0.4 0.9 }

* (x1, ya) * ey) * (a, 92) * sy) * (Ks Ya)
(x1, ¥1)
It can be represented in a matrix form
%41 Yo
AxB=
Ex. 4.9.20: Consider the following fuzzy sets

08 ae O06 04ee ee02
+4327 133 +134* 135* 136
Low temperature= + 737
1 —
4
0. 0.2 04,06 08, 1) — g
High temperature= 4 734 * 135 * 136 * 137 * 138 139
— . WH Techhnewtedys
Publications
:
FUZZY Logi
01 O02 04 06 O08
High pressure= 4 409 + 600* 700* 800* 300 +7006 }
Temperature ranges are 130° F to 140° F and pressure limit is 400 psi to 1000 psi. Find the following
membership functions :
1) Temperature not very low. 2) Temperature not very high
3) Pressure slightly high. 4) ° Pressure very very high.

Soln.: :
1. Temperature not very low
Verylow = low2
131* 132 +733 +134 +135 * 136

Notverylow = 1-very low
0 , 236 | 0.64 0.84 0.96 53}
. = lz131° 132 "133* 134* 135 * 136
2. Temperature not very high
SVery high = high?
«. Temp very high ig4t 135 “136+ 137 138 * 439
- Temp not very high = 1—very high

_1_ 0.96 0.84 0.64 0.36 0
= ) 734* 735 * 136* 137 7138 * 139
3. Pressure slightly high
Slightly high = dilation (high) = J high

. 4/2
Pressure slightly high = (high pressure)
Yo1 0.2 0.4 Nos Vo8 Va
- {, 600 800 * 900 * 1000
700
[93 0.44 0.63 0.77 089 1
400 * 600 *"700 * 00 + 900 * 1000 }
4. Pressure very very high
Very very high thigh)"
Pressure very very high (high pressure)
0.0001 0.0016 0.025 0.12 0.40 1 |
= { 400 600 *"F00* 800 * 900 * i000
dg aigsc (MU-Sem- 7-Comp) Fuzzy Logic.
goln. :
“ ' using max-min composition
T(xl : zl) a max (min (0.6, 1) , min (0.3, 0.8) )

= max (0.6, 0.3)
= 0.6
T(xl, z2) = max (min (0.6, 0.5) , min (0.3, 0.4) )
= max (0.5, 0.3)
= 0.5
T(x2, 23) = max (min (0.6, 0.3) , min (0.3, 0.7) )
= max (0.3, 0.3)
= 03
T(x2, z1), = max (min (0.2, 1), min (0.9, 0.8) )
= max (0.2, 0.8)
= 08
T(x2, 22) = max (min (0.2, 0.5) , min (0.9, 0.4))
= max (0.2, 0.4)
= 04
)
T(x2, z3) = max (min (0.2, 0.3) , min (0.9, 0.7)
= max (0.2, 0.7)
= 07
Xi 0.6 05 0.3
T= ROS xy l 08 04 an |
2, Using max-product
= max (0.6x1,0.3 x 0.8)
T(X1, 21)
= max (0.6, 0.24)
= 0.6 |
x 0.4)
T(X1, Z2) = max (0.6 x 0.5 , 0.3
= max (0.30, 0.12)
= 03
TKI, 73) = max (06X03, 03%07)

= max (0.18, 0.21)
= 0.21
1K, ZI) == max (0.21 ,09%08)

max (0.2, 0.72)
= 0.72
(X22) = max (0.2% qs ‘We TechKnowledge

Publications
AIBSC (MU-Sem. 7-Comp)
= max (0.1, 0.36)
= 0.36
T(X2, Z3) = max (0.2x0.3, 0.90.7)
= max (0.06
, 0.63)
= 0.63
Zl Z2 Z3
X1 fF 06 03 0.21
T= ROS= xo [ 0.72 0.36 0.63 ]
Ex. 4.9.22: Given two fuzzy relations R, and R, defined on X x Y and Y x Z respectively, where
X= {1, 2, 3}, Y= {a, B, Ys 3} andz= {a, b}. Find i
1. Max- min composition.

2. Max - product composition.
01 09
01 02 03 0.5
R 04 03 07 09 | R 0.2 08
i _ - - a2 “| 05 06
ne 0.7 03
Soln.:
1. Max- min composition
_ Here R; is defined on X x Y where X = {1, 2, 3} and Y = (a, B, 7, 5)

a Bp y 8
0.1 02 03 05
WN rR
R,; = 04 03 07 09
0.6 0.1 08 0.2
Z = {a, b}
Similarly Rz is defined on Y x Z where Y= {a B, y, 5}and
So, : a b
0.1 0.9
a
02 03.
R = &
27" YT! 05 06
51 o7 03
and Z = {a, bh.
So composition of Ry and R2 will be defined on X x Z where X = {1, 2, 3}
Here Rio Rzis 3 X 2 matrix
(min (0.1,0.1); min (0.2,0.2
wy nese U-Sem. 7-Comp)
= max (0.1, 0.2, 0.3, 0.5)

= 05
Uri orall, b) = max (min (0.1, 0.9), min (0.2, 0.3), min (0.3, 0.6), min (0.5, 0.3))
max (0.1, 0.2, 0.3, 0.3)
0.3 .
Pri oR2 (2, a) = max (min (0.4, 0.1), min (0.3, 0.2), min (0.7, 0.5), min (0.9, 0.7)
max (0.1, 0.2, 0.5, 0.7)
0.7
0.3))
ih
Mri oR2(2, b) = max (min (0.4, 0.9), min (0.3, 0.3), min (0.7, 0.6), min (0.9,
= max (0.4, 0.3, 0.6, 0.3)
= 06
0.7))
Bri or2 (3, a) = max (min (0.6, 0.1), min (0.1, 0.2), min (0.8, 0.5), min (0.2,
= max (0.1, 0.1, 0.5, 0.2)
= 0.5 .
Uri oR2 (3,b) = max (min (0.6, 0.9), min (0.1, 0.3), min (0.8, 0.6), min (0.2, 0.3))
= max (0.6, 0.1, 0.6, 0.2) .
= 06
2. Max - product composition
a b
0.35 0.18
—
Ryo Ry = 0.63 0.42

WN
0.40 0.54
= max (0.1 x 0.1, 0.2.x 0.2, 0.3 x 0.5, 0.5 x 0.7)
Uri oR? (1, a)
= max (0.01, 0.04, 0.15, 0.35)
= 0.35
= max (0.1 x 0.9, 0.2 x 0.3, 0.3 x 0.6, 0.5 x 0.3)
Uri oR2 (1,b)
= max (0.09, 0.06, 0.18, 0.15)
= 0.18
= max (0.4 x 0.1, 0.3 x 0.2, 0.7 x 0.5, 0.9 x 0.7)
rio R2 (2, a)
= max (0.04, 0.06, 0.35, 0.63)
= 0.63
= max (0.4 x 0.9, 0.3 x 0.3, 0.7 x 0.6, 0.9 x 0.3)
Pri o R2 (2, b)
= max (0.36, 0.09, 0.42, 0.27)
= 0.42
Uri oR2 Os a)
= max (0.6 x 0.1, 0.1 x 0.2, 0.8 x 0.5, 0.2 x 0.7)
max (0.06, 0.02, 0.40, 0.14)
‘wie Toth Keemloan:
WF AlgSc(MU-Sem.7-Comp)
= 0.40 X 0.6 x 0.2
x 0.3)
= max (0.6 x 0.9,
0.1 x 0.3, 0.8
Hei or2 ( 3,b) 20.54
8, 0.0 6
max (0.54, 0.03, 0.4
maturity
the ‘co lor of aa fru it’ and ‘grade of
: wee n fru it’ , wh er e col or, grade ang
Ex. 4.9.23: Let R be the relation that specifies the relationship bet of
and ‘taste
Relation S specifies the relationship between ‘gra de of maturity’
ly as follows.
taste of a fruit are characterized by crisp sets x, y, Z respective
X = {green, yellow, red}
Y = {verdant, half mature, mature}
Z = {sour, tasteless, sweet}

: ' ‘ ’ of a fruit usin g
Consider following relations R and S$ and find the relationship between color and taste’ 0
1. Max- min composition 2. Max - product composition
0.5
0.2
Solin. :
1. Max- min composition
-T (green, sour) = max (min (1, 1), min (0.5, 0.7), min (0, 0))
= max (1, 0.5, 0)
= 1
T (green, tasteless) = max (min (1, 0.2), min (0.5, 1), min (0, 0.7))
= max (0.2, 0.5, 0)
= 0.5
T (green, sweet) = max (min (1, 0), min (0.5, 0.3), min 0,1)
= max (0, 0.3, 0)

= 0.3
T (yellow, sour) = max (min (0.3, 1), min (1, 0.7) min (0.4.0
= max (0.3, 0.7, 0.4) me
‘= 0.7
T (yellow, tasteless) =
mimi ©3,02), min (1), min 0.4, 0.79)
e
¥ Al&SC (MU-Sem. 7-Comp) __ 4-65 : Fu Logic
— =— peer — a
= max (0.2, 1, 0.4)
= 1
T (yellow, sweet) = max (min (0.3, 0), min (1, 0.3), min (0.4, 1))
= max (0, 0.3, 0.4)
= 04
T (red, sour) = max (min (0, 1), min (0.2, 0.7), min (1, 0))
= max (0, 0.2, 0)= 0.2
T (red, tasteless) = max (min (0, 0.2), min (0.2, 1), min (1, 0.7)
= max (0, 0.2, 0.7)
= 07
T (red, sweet) = max (min (0, 0), min (0.2, 0.3), min (1, 1))
= max (0, 0.2, 1)
2. Max- product composition
ur.| Tasteless | Sweet.

0.5 0.15
1 0.4
0.7 1
T (green, sour) = max (1 x 1, 0.5 x 0.7, 0x0)
= max (1, 0.35, 0)

= 1
T (green, tasteless) = max (1 x 0.2, 0.5 x 1,0 0.7)
= max (0.2, 0.5, 0)
= 0.5
T (green, sweet) = max (1 x0, 0.5 x 0.3, 0x 1)
= max (0, 0.15, 0)
= 0.15
T (yellow, sour) = max (0.3 x 1, 1 x 0.7, 0.4 x 0.7)
= max (0.3, 0.7, 0.28)
= 0.7
T (yellow, tasteless) = max (0.3 x 0.2, | x 1, 0.4 x 0.7)
= max (0.06, 1, 0,28)
= 1
T (yellow, sweet) = max (0.3 x0, 1 x 0.3, 0.4 x 1)

= max (0, 0.3, 0.4)
Al&SC (MU-Sem. 7-Comp)_
T (red, sour) = max (0x1, 0.2x0.7, 1x0)

= max (0, 0.14, 0)
= 0.14
T (red, tasteless) = max (0x0.2,0.2x1,1% 0.7)
= max (0, 0.2, 0.7) .

= 07
T (red, sweet) = max (0x0, 0.20.3, 1x 1)
= max (0, 0.06, 1)

= 1
4.10 Design of Controllers (Solved Problems) _
1. Domestic Shower Controller
Ex. 4.10.1: Design a fuzzy controller to regulate the temperature of a domestic shower. Assume that:
{a) The temperature is adjusted by single mixer tap.
(b) The flow of water is constant. .
(c) Control variable is the ratio of the hot to the cold water input.
The design should clearly mention the descriptors used for fuzzy sets and control variables, set of rules to
generate control action and defuzzification. The design should be supported by figures where ever possible.
Soln. :
Step 1: Identify input and output variables and decide descriptors for the same.
— Here input is the position of mixer tap. Assume that position of mixer tap is measured in degrees (0° to 180°). It
represents opening of the mixer tap in degrees. 0° indicates tap is closed and 180°indicates tap is fully opened.
— Output is temperature of water according to the position of mixer tap. It is measured in°C. We take five descriptors
for each input and output variables.
— Descriptors for input variable (position of mixer tap) are given below.
EL - Extreme Left
- Left
- Centre
- Right
ER - Extreme Right
{ EL, L, C, R, ER}
— Descriptors for output variable (Temperature of water) are given below :

vcT- Very Cold Temperature
cT - Cold Temperature
Al&SCAME “Sem. /-Comp)
wr - WarmTemperatura Fuzzy Logic
HT - Hot Temperature
VHT- Very Hot Temperature
{VCT, CT, WT, HT, VHT}

: fin j
sep2: Deine memnuership functions for input
We use tria. ngular MFs because of its si sj molicicityand output variables,
, Apli
4, Membership functions for input varia
le- Position of
mixer tap.
My [ELL
c
° ° x(degree)
45 90° 135° 180°
°
Fig.g. P. P. 4.10.1: ; Membership function for position of mixer
tap
45— ,0<x<45
Hey (x) = =
Xx
a5 «OSKS45
My (x)
90-x
45.” 45<x<90
x-45
He (Xx) =
135 —x
45 , 90<x<135
x-90
45 », 908x135
br (x) = iko—x
“45. °? 135 <x < 180
x- 135
lpr (x) = gg > 135 SxS 180
w ater.
put variable - temperature of
2 Membership functions for out
y(’C)temp.
40 80 100
t ions for water temperature
E ° 0 mbership func
ig. P. 4.1 0.1(a) : Me
E
ns _ W=Y¥ osys10
uvcT (y) = ~ 10
:
. + O<y<10
L
pCT (y) = Oy | 10<ys40
Epe
E : TechKnowledyé
Poollcations
LWT (y) =
HT (y)
wt , 80<ys100
LVHT (y) Ss ,80<y<100

Step 3: Forma rule base,
Table P, 4.10.1
EL VCT
L CT
c WT
R HT
ER VHT
We can read the rul |
e base Shown in Tab
le P. 4.10.1 in terms
of If-then rules.
Rule 1: i mixer tap position
is EL (Extreme Left)
then temperature
Rule2: of water is VCT (Ve
If mixer tap position ry Cold Temperatur
is L (Left) then temperat e).
ure of water is CT (Co
Rule 3: ld).
If mixer tap position
is C (Centre) then temp
erature of Water is WT
Rule4: If
(Warm)
mixer tap pasition
is R (Right) then temp
erature of Water is
Rule5: HT (Hot)
If mixer tap position
is ER (Extreme Right)
then temperature of
Thus, we have five rules. Water is VHT (Very
Hot).
Step4: Rule Evaluation,
Assume that mixer tap posi
tion is 75° . This va
lx u
=e75° Maps to
90 -x follow; Ww
Rule2: Uy(x) = > ONOwing two MFs of Rule 2 an
45 d Rule 3 respecct
tively.
Rule3: x-45
p(x) =“s
Now, substitute the value
of x = 75 in above two €quations,
90278 we
Strength of Rule 2 =>, (75) = 5 Bet strength of €ach
=3 rule
Strength of Rule 3 >, (75) 75-

= ea
Step5: Defuzzification
We apply mean of maxi
mum defuzzification tec
hnique,
Ww TechKnowledge
Publications
1g8C (MU-Sem. 7-Comp)
oe
ae find the rule with the
maximum strength
2)
max (Strength of Rule 1, strength of Rule
ii}
Max (Hy (x), He (x))
ul
a2) 2
max (3,3) =3
"
Thus, Rule 3 has the maximum strength,
t MFs of
according to Rule 3, If mixer tap position is c (center) thentwowater temperature is Warm. So, WE use Outpu
warm water temperature for defuzification, We have following equations for warm water temperature.
y-10
Mwr (y)
80-y
Lwr (y) 40
2
Since, the strength of rule 3 is;37 Substitute Ly; (y)=F in the above two equations.
y-10 2
30 3 >y= 30
80-y 2
40 =3>y= 53
Now take the average (mean) of these values.

30 +53
yt = = 415°C
9. Washing Machine Controller
: Design. a controller to determine ihe wash time. of a domestic washing ‘machine ; Assume. that input isis
grease on cloths. Use three ‘descriptors for input variables and five descriptors for output variables. Derive ‘set
of rules for controller action and defuzzification. The design should be supported by figures wherever possibl
- vice- -vers:
Show that Af the: cloths are solled to a.larger, denice) ‘the wash. time :will ‘be. more “and
‘(MUEIEEE ES ees aE PESRE) Ey PETS
Soln.:
decide descriptors for the same.
Step1: Identify input and output variables and
Here inputs are ‘dirt’ and ‘grease’. Assume that they are measured in percentage (%). That is amount of dirt and —
Brease is measured in percentage.
Output is ‘wash time’ measured in minutes.

We use three descriptors for each of the input variables.
Descriptors for dirt are as follows :
SD - Small Dirt
MD - Medium Dirt
1D - Large Dirt
TT
{Sp, MD, LD }
Descriptors for grease are { NG, MG, LG }

TIT
NG . No Grease
‘MG . Medium Grease
SENSE
> Se TechXnowledas
‘| LG -. LargeGrease .
i We use five descriptors for output variable.
H So, descriptors for wash time are {VS, S, M, L, VL}
| VS - Very Short
S - Short
M - Medium ~
L - Large
VL - Very Large
: ariables.
Step2: Define membership functions for each of the input and output v
We use triangular MFs because of their simplicity.
(1) Membership functions for dirt
ae)
| x (dirt in %)
| 0 50 100
; Fig. P. 4.10.2 : Membership functions for dirt
j Usp (x) = Ep»
50—x
OS x S50
a0 (OSKS50
mp (x) = _
a x, 50<x<100
|| bn@® =7=Z- ~50
, 50Sx<100
(2) Membership functions for grease
y(grease in %)
0 50 100
Fig. P. 4.10.2(a) : Membership function

s of grease
50~
Le (y) = a » OSy<50
ei
50 » OSys50
Hua (y) = 100 -
50 , 30<y< 100
y-50 /
Mig (y) = » S0Sy<109
Al 25C (MU-Sem. 7-Comp)
membership functions for Wash time at 71 Fuzzy Logic
0 10 Z wash time (min).

25 40 60
Fig. P, 4, .
10.2(b) : Membership functions for wash time
LVS () = 10~z
“To > OSz<10
BS (z) = |x10 > 0<z<10
25 -7z
15 , 10<z2<25
z-10
15.» 10s2<25
MM (z) =
40-z
15. > 2<z<40
Z-25
LL (z)Z= 60-2
7 +» 40<zs60
LVL (z) = z—40

20° 40<z<60
lep3: Forma Rule base
LD M iL ML
The above matrix represents in all nine rules. For example, first rule can be “If dirt Is small and
No grease then wa
sh
time is very short” similarly all nine rules can be defined using if --- then.
Hep 4: Rule Evaluation
Assume that dirt = 60 % and grease = 70%
ditt = 60.9 maps to the following two Ms of “dirt” varia
“dirt” variable
100-—x _xX=-50
bX) =~ 59 md Peo) ="5Q
Amilanty grease= 70 % maps to the following two MFs of “greas e” variable.
100-y 50
~Y=50
ume(Y) =~ 59. 20d Pua (¥)=""5q
bE Waluate Maio (x) and [yp (x) for x = 60, we get
100-60 _
o
usp (60) = (1)
Ww Tech Knowledge
feptieen ene
SF _Al&SC (MU-Sem, 7-Comp
) : . RwyLoy
ashe
, . ~
Hip (60)
50 = ~ 5
Similarly evaluate Hc (y) mo
and j,¢ (y) for y = 70, we get
. 100-70 3 e
Hac (70) = “Sp 5 Q)
, 70-50 2
Mig (70)
5 = “39 + (4)
The above four Equation leads to
the following four rules that we are
(1) suppose to evaluate
Dirt is medium and grease
is medium.
(2) Dirt is medium and
grease is farge.
(3) Dirtis large and grease
ismedium.
(4) Dirtis large and greas
e is large.
Since the aitedens aoe of each of the
above rule is connected by and oper
strength of each rule. ator we use mini operator to Evalu
ate
Strength of rule 1: S, = min (tian (60), f4y4q (70) ) =
min (4/5, 3/5) = 3/5
Strength of rule 2:S, = min (Ump (60), Hyg (70) )
= min (4/5, 2/5) = 2/5
Strength of rule 3: S3_
= min (yp (60), {tyg (70)
) = min (1/5, 35) = 1/5
Strength of rule 4:5, = min (Mip (60), Hh (70)
fi ) = min (1/5, 2/5) = 1/5
Grase
| Dirt HMa(70) tg (70)
'
x Jos] x Max ME xl xlx
| |
Hyp(60) | x @sy 2/5

| x @) L
Hip( 60 ) [| x 15fh | 1/5 x L VL H
Fig. P.4.10.2(c): Rul

e strength and its
mapping to correspo
Step5: Defuzzification nding output MF
- Since, we use “Mean of max” defuzz
ification technique, we first find
the rule with the m
aximum strength.
Max (S1, $2, $3,
$4)
I
Max (3/5, 2/5, VS,

1/5)
I
3/5
Ul
- This corresponds to rule 1.

~ This rule 1 — “dirt is medium and grease is medium”
has maximum strength 3/5,
~The above rule corresponds to the output MF
- Ly (2). This Mapping is shown in Fig. p
- To find out the final defuzz
ified value, we now take averag -4.10.2(c).
e (mean) Of tty, (2),
z-
Mm (Zz) = “F10
s and MM (2) = “te
WS = TzZ~E-10 3/5 =A
Z = 49
2=3]
oe 19¢ t 31
= 25 min
25 40 60
min(4/5},3/5)
4 L
)
H 4
®
MG
cetlsonaneen2! 1
aan Nae 15
~ ap -b -------- -
: ULLAA.__—-+
50 100 y 0 10 25 40 60 z
min(i/5,3/5)
LG 4
100 y 0 10 25 40 60 z
50
min(1/5,2/5)
j
Aggregation
grease = 70%
dirt = 60%
:
1
| ‘
4 . Max MF
x
Union of all 3/5
3/5 output MF 2/5
2/5 1/5
1/5 2z
0 10 25 40 60
10 25 40 60 Zz
9
Aggregate output
MF
ion
evaluation and defuzzificat
4. 10 .2 (d ) : Process 0 f rule
Fig. P.
4 Water Purifier Gonteole’

r for the water purification plant. Raw. Water Is:
le r {0 co ntro | the feed amount of purifie
‘control ut as am un
su er at ur e an d gr ade of water. Outptro an
cting c spemic al s. As
me put ‘and wale te mp rules to con l acti Sau
purified by inje output variables. Design :
th re e d as criptors for in nev er nec ess ary , Clearly indicate
tha
‘purifier. Use rted- by, figures whe s
hould be suppo ou nt, AUSTELL) ic
ical used is In. large am
5 low then chem
r the same.
nd de ci de descriptors fo
@
ten a, .
ut variables r.
1: Identify inp ut and out P
a nd grade of wate
pera tu re
e water tem
input variables ar
Ww TechKnowledge
Here Publications
_Al&SC (MU-Sem. 7-Comp) _
74 ————=
2 — centage-'
1 in per
— Water temperature is measured in °C. grade of water is. me asured in P
- Descriptors for water temperature are
{C, M, H}
Cc - Cold
M - Medium
Hl
H - High
et — Descriptors for grade are
:
1 {L,M,H} ,
1
L - low
M - Medium
Hl H - High
in — Amount of purifier is measured in grams. Descriptors for amount of purif
ier are
{S, M, L).
S - Small
M - Medium
HH
L - Large
; + Step2: Fuzzification /
Define membership functions for each of the input and outpu
t variables.
We use triangular MFs because of its simplicity.
(1) Membership functions for water temperatur
e
Hy
i 0 20 80-70 409 temp. (°C)

ia
Fig. P. 4.10.3 : Membership function
s for water temp.
SO0-x |
uC (x) = “57 » OSx<50
x
50 > OSx<59
M (x) =
HM ©) 100~x
50.» S0<x<i09
x-50
HH (x) = 50 » S0Sx< 109
(2) Membership functions for grade of wate
r
Hy) L M H
0 50 109 Y(Grade in %)
Fig._ P. 4.10.3(a) : Membership
Qe TechKnowladge P funf, ctio
Publications = ns for Grade of was_
NaSC (MU-Sem. 7-Comp)
4-75 Fuzzy Logic.

HL(y) = S0ny
50’ Osy<so
YY
MM (y) 2 sO » Osysso
sey
50 , 50<y<100
-50 —
BH (y) = 2220
50° 50<y<100 is
Bunetons for amount of purifier

hip functio ha
ga) Membership i
yh
3
wd
4
0 5 10 _2(amount of purifier in grams)
Fig.P. 4.10.3(b) : Membership functions for purifier

5-z
US@ = =, 0<z255
eo O<z<5,
LM (z)Zz) = 10-2
5 : 5<z<10
z-5 5<z<10
pL (z) 5°
Step3: Formarule base

Temp grade L M H
Cc L M/S
M L M M
H MiSs {5
+ ali ni rules. For example, _ — ,

- The above matrix represents In all nine
purifier required is large.
is cold and grade is low then amount of
First rule can be, “If f tempe rature
2. ing if-then rules.
> Similarly all nine rules can be defined using!
‘tp4; - Rule Evaluation 0:
rade =3
"Assume water temperature = Sand 6 Mes of “temperature” variable.
~ | Wata er temperature — 5° maps to the following two x
«) = a- and pm (x) = 50
-—
Uc =
“grade” variable.
wo Mes of "Bt
a . y
Similarly rade = 30 maps to the following t
76 iy) = Aaz~ and bu ) = 50
py ~
: TochKnowledye
Puptications
— Evaluate p(x) and pay (x) for x = 5°
We get,
He (5) = 50-3 5 =0.9

«fae
uM (5) = a-5 =0.1
— — Evaluate p(y) and by (y) for y = 30
13)
yL (30) = = 80 =% =0.4
wo(4)
uM (30) = 3-2 =06
for rules that we need to evaluate.
— The above four equation represents following
ih ,
‘ 1. If temperature is cold and grade is low.
2. {f temperature is cold and grade is medium.
ut 3. If temperature is medium and grade is low.
It 4. iftemperature is medium and grade is medium.

te strength of each
ted by and operator we use min operator to evalua
— Since the antecedent part of each rule is connec
rule
0.4
Strength of rule1 : 51 min (tc (5), LL (30)) = min (0.9, 0.4) =
Strength of rule2 :S2 min (He(5), Ltvi(30)) = min (0.9, 0.6) = 0.6
Strength of rule3 : 53 min (uy(5), HL(30)) = min (0.1, 0.4) = 0.1
Strength of rule4 : Sq min (M(5), (30) = min (0.1, 0.6) = 0.1
Grade
14, (30) Hy (30) Temp Grade
Temp
Hg(8)5 | 0.4 x Mek SHE RO
1yy(5) | 0.1] 0.4] X tlmin
x |x |x misls
(b) Rule base table

(a) Rule strength table
Fig. P. 4.10.3(¢) :Rule strength and its mapping to corresponding output MF
Step 5: Defuzzification
gth
hnique, we first find the rule with ma: ximum streng
a
ion
Since, we use “mean of max” defuzzSx,ifi53,catSa) = tec
max (0.4 , 0.6, 0.1, 0.1) = 0.6
= max (5,
— This corresponds to rule 2.

de Is medium” has maximum strength 0.6
Thus rule 2: “Temperature Is cold and gra
This Is show Is shown in Fig. - P.4.10 ,3(c)
The above rule corresponds to the output MF [(z). P.4,10.3(c).
average (I.e. mean) ofby(z).
— Tofind out final defuzzified value, we now take
4g AlaSC (MU-Sem. 7- Comp)
Fuzzy. Logic
Hy(z)
MIN
Il
“06 =
WIN
N
it
_—
w
NV
° Grade = 30 . Aggregation
:
Temp =5
1 Max MF
Unlon of all 9 ‘
output MF 4
4
A |
|
Aggregate output ; |
MF :
_.. p a4.ald) : Process of rule evaluation and defuzzification |: :
FUZZY Logie,
y Al&SC (MU-Sem. 7-Comp) I ee
4. Train Brake Power Controller
. i st ation.
Ex. 4.10.4 : Design a fuzzy controller for a train approaching or leaving a
se,
and speed of the train. The output is brake power used. U
(i) Triangular membership functions
(ii) Four descriptors for each variables
(iii) Five to six rules.
(iv) Appropriate deffuzication method.
Soin. :
Step1: Identify input and output variables and decide descriptors for the same.
- Here inputs are
Distance of a train from the station, measured in meters and
Speed of train measured in km/hr.
- Output variable is brake power measured in %.
As mentioned, we take four descriptors for each of the input and output variables.
For distance => {VSD, SD, LD, VLD}
VSD Very Short Distance
sD Short Distance
LD Large Distance
VLD Very Large Distance
- Forspeed = {VLS, LS, HS, VHS}

VLS Very Low Speed
Low Speed
HS High Speed
VHS Very High Speed
- For brake power => {VLP, LP, HP, VHP}

- VLP Very Low Power
LP Low Power
HP High Power
VHP Very High Power
Step 2: Define membership functions for each of the input and output variables,
As mentioned, we use triangular membership functions.
1. Membership functions for distance
Hix)
V8D sD
: u t t
0 100 ~~ xX
400 600 (Distance In Meters
: )
Fig. P. 4.10.4 : Member
ship functions for distan
ce (distance In meet
teer
rss)
Bo waked
/alasc (MU-Sem.7-Comp)
BVSD{x) ~ 100~x
¥
i
po
|
|
1
100»
5
ee
I ; i
b 2
BSD (x) =
I
100° O<x<100
i
400 - x .
|
300” 100 <x< 400 :
.
BLD (x): X= 100

300 ' 100 <x < 400
f
Sn0-x
| 100 ' 400<x<500
|
[
L
x~400
BMLD (x) = To9 400 <x <500
f
7, Membership functions for speed
.
>y
0 10 50 60 Speed in km/hr
tions for speed
Fig. P. 4.1 0.4(a) : Membership func
10-y ev et0
= “49 ° O<y
uVLS(y)
o<yy<10
Ls = >
y
S0-40’ 10<y<50
y=10 10Sy<50
pHS(y) = 40 ’
- 60=¥ 50<y $60
10 ’
y= 30 50<y<60
pVHS(y) = 10°
for b rake power

: Membership functions
p(z)
|
Zz ;
(Breakpower In %)
brake power
“Puptications
. | . Fuzzy Log;
WF Al&SC (MU-Sem. 7-Comp) 4-80 —— —e “=
‘PVP (2) = oe, oszs20
WLP (z) = eat o<zs20
os, 20<z<80
MHP (2) = ae 20 <2 $80
a go<2 100
PVHP(z) » = z= < 100

80 <2
Step3: Form arule base.
vsD | HP | HP | VHP | VHP

sD te | LP | HP | VHP
LD vip | vip} LP | HP
VLD vip | vip | LP | LP
The above matrix represents in all 16 rules.
For example, First rule can be “If distance of a train is Very Short (VSD) and speed is Very Low (VLS) then required
brake power is High (HP)”.
Similarly all 16 rules can be defined using If then rules.
_ Step4: Rule Evaluation
Assume distance = 110 meters and speed = 52 km/hr
Distance = 110 maps to the following two MFs of “distance” variable.

400 - x x-
Msp) = 399 and Mip(x) 2S =
Similarly speed = 52 maps to.the following two MFs of “speed” variable.

, ey -
Has(Y) = “yg and biyas(y) = 222
get,
— — Evaluate psp(x) and pyp(x) for x = 110, we
Usp(110) = oa = 0.96 )
| Pyp(110) ia
= = 0.033 (2)
— — Similarly evaluate pys(y) and Lyys(y) for y = 52, we get, |
pys(52) = P= Lo | 8)
Myys(52) = a =0.2 - ip.
The above four equations leads to the following for rules that we needs to evaluate
f nistance is short and speed is high
pistance isIs short
é:- pistance and speed is very high
large and speed is high
pistance is large and speed is very high
operator to evaluate strength of each rule

Strength of rulet: S;
Min (L4s9(120), juys(52)) = min (0.96, 0.8) = 0.8
Strength of rule2 ; s,
min (Hso(110), ys(52))
min (0.96, 0.2) = 0.2
. Strength of rule3 : s, Min (bjp(110), py4s(52))
min (0.033, 0.8) = 0.033

Strength of rule: S, Min (Hyp(110), Wyys(52))
min (0.033, 0.2) = 0.033
Speed
VHS VL§ LS HS VHS
Distance. VLS LS HS
x | xX | xX] xX vsD| —
vsp}|
sel ~=—Ss SD VHP
sp| x | xX
LD Lp | HP
Lo} X | X{0.03}0.03
x| VLD
VLD | x | x | | x | |
(b) Rule base table
table
(a) Rule strength ponding output MF
ta bl e an d it 5 mapping to corres
Rule streng th
Fig. P. 4.10.4 (c ) :
fep5: Defuzzification
echnique-
We use “m ea n of ma x” defuzzification t
~
gth
fin d the rul e wit h ma xi mu m stren
Wefirst max (S152 S3, S4)
M
0.8
2, 0.33, 0.2) =
= max (0.8, 0.
rength 0.8.
This corresponds to rule 1. h” has maximum st
~ Thus rule 1- “If dist is short and speed is a el?) This mappi |s shown in Fig. P. 4.10.4(c).
ng
‘
ds to the output mom f MHP ypl2)-
"The above rule. corresponds OT
100 =z
ke avera be
ified value, We»

c
Too a pap(2) 20
=.Mpute the final defuzz! z _ 100-2
BaP
se 0. = z = 84
:
z = 68
68+84 76%
a* = 2,
zfs pa Ros
Fu: Logic
97 AlRSC (MU-Sem. 7-Comp)
KT SD
0 100 400 500 X 0 10
kf sD
0 100
te LP ©
Zz
0 20 80 100
CS ay
Union of 08, / .
all oulputMF / Yj /
===> / LE: VY tj p
0.03 mt PO»
0 20 80 100
Fig. P. 4.10.4(d) : Process of rule evaluation and defuzzification
5. Water Tank Temperature Controller
Ex. 4.10.5: Design a fuzzy controller for maintaining the temperature of water in the tank at a fixed level. Input variables
are the cold water flow Into the tank and steam flow Into the tank, For cooling, cold water flow is regulated and
for raising the temperature steam flow Is regulated. Define the fuzzification scheme for input variables. Device
a set of rules for control action and defuzzification. Formulate the control problem in terms of fuzzy inference
rules incorporating the degree of relevance for each rule. Design a scheme which shall regulate the water and
steam flows properly,
TechKnowledga
Pubfications eee
=
Here inputs are,
‘ Amount Of valve openin

g for cold wat er,
SUVS
2. Amount of valve Opening for steam
opening is measured in degree

yalvecripto from 0°
to 180°
.
variables.
pes rs for value openin 0°. we take five descriptors for each of the input
g for cold water ¢| OW
are
{ELCV, LCV, CCV, RCV, ERCV}
ELCV Extreme Left Cold Valve

Lev : Left Cold Valve
CCV Centre Cold Valve
RCV Right Cold Valve
ERCV Extreme Right Cold Valve
Descriptors for valve opening of steam flow are {ELSV, LSV, CSV, RSV, ERSV}
ELSV Extreme Left Steam Valve
LSV Left Steam Value
csv Centre Steam Value
RSV Right Steam Value
ERSV Extreme Right Steam Value

measured in °C.
- Output is temperature of wa ter in the tank
VCT, CT, WT, HT, VHT, VVHT}
use seve n desc ript ors for t emperature of water {VVCT,
We
VVCT Very Very Cold Temperature
|
VCT Very Cold Temperature
{
[
}
{
ct Cold Temperature
wt
||
i Warm Temperature
|t
HT Hot Temperature
|?
| VHT Very Hot Temperature
ature
WHT Very Very Hot Temper
| t and outpu t var
iables.
functions for each of the inpu
Step2: Define membership
ini g for cold water flow
sh ip fu | tions for valve open
nc
| 1. Me mb er
|
acy ,cc¥ RCV
3 erev
|
45 90 435180)
0 cold water
cti o! ns for valve opening of TechKnowledyé
fa ; Fig. P. 4.10.5 : Me mbership fun Wey Pustications
Fuzzy Logic: .
WF Alasc (MU-Sem. 7-Comp) 4-84
45-%x <x<s45
MELCV(x) = ~g57 e
uLevix) = geyx 7 <x<45

30-x , 45<x<90
45’
x-45 45 =x <90
pCCV(x) = a5’
= 90 <x< 135
-90 90<xs 135
HRCV(x) = “a5 >
=* 135 <x < 180

x-135
MERCV(x) = —ge >, 135 <x <180
2. Membership functions for valve opening for steam flow
t t T T » y (In degrees)
0 45 90 135 180
i
Fig. P.4,10.5(a) : Membership functions for value opening for steam flow
| ; A5-y
| HELSV(y) 45° O<sys45
ae
MISV(y) = 45, O<ys4s
90-y
45’ —45<ys90
_ y=45 .
HESV(y) = “G5 45<y<90
'135~y
eS 90<y<135
RSV(y) = 7-90
MRSVIy) = “GS .90<ys135
180~y
45 -. : 135 <y<180
LERSV(y) y= 135 . :
45” 135 <y<180
ns membership functions for temperature of a water in tank
3. .
z
t t — T T i re of
0 10 30 50 70 90 100 (Temperature
waterin ©)
. ter in tank
Fig. P.4.10.5(b): Membership functions for temperature of Wa
10-z o0<z<10
pVVCT(z) = to"
z 0<z<10
pVCT(2) = 49°
30-2 10<z<30
20 '
z-10 10 <2 $30
uct(z) = “20 "
50-2 30<2<50
20 ’
2-30 30<z<50
pwi(z) = ~20 ’
TO, 50<z<70
z-50 50< 2570

pHT (2) = ~ 20 ’
90-2 70<2<90
“0?
1-70 70<2£90
wVHT(z) = 20 '
100 - <100
90<2z<1
2-90 90<25100
wHTl2) = “0 ’
Step 3: Forma rule base-
Step 4: Rule evaluation
Assume that x 95° (valve opening of cold water)

y 50° (valve opening for steam flow)
Here x = 95° maps to the following two MFs of variable x :
135-x x-90
Heev(x) = “gg and Urcv(x) ="45
Similarly y = 50° maps to the following two Mfs of variable y :
90 — _y=45
Misv(y) = — and Hcsy(¥)= "45
Evaluate bccy(x) and [pey(x) for x = 95° we get,
Hoov(95) = 135-9
ae 5 = 0.88 sa(l)
rev) = ==? - 0.11 0
— Evaluate p<y(y) and Lesy(y) for y = 50°, we get
Mysy(50) = = = 0.88 (3)

Hcsy(50) = a =0.11 | 4)
Above four equations lead to the following four rules that we
need to evaluate.
Cold water valve is in center and steam flow valve is
in left.
Bown Pp
Cold water valve is in center and steam flow valve is in center.

Cold water valve is in right and steam flow valve is in left.
Cold water valve is in right and steam flow valve is in center.
Table P.4.10.5 shows rule strength table.
Table P. 4.10.5: Rule strength table
Hetsv Hisv csv Hrsy HEpsy

HELsv] xX X x x x
MHsv | xX x x x x
| — Hesv | x o11| x x
é
Hasv | x 0.11 | O41-] x 4
HEnsy| xX X x | x x
Step5: Defuzzification
— We find the rule with maximu
m strength
max (0.88, 0.11, 0.11, 0.11)

u
0.88
UW
Which corresponds to rule

1,
The rule 1 corresponds to the Output MF Hex(z). To comp
ute final defuzzified value we take average of Herz):
vale TarhWamul.i.>
AI&SC (MU-Sem. 7-Comp) _ Fuzzy Logic
z-10
WCT@) = “3-= 0.88
‘z-10
= 99 2 2=27.7
50 - 0.88
MCT) = “=*—
50 -z
= Tp 25323
2k =
27.7 + 32.3
2
= 30°C
Q.1 Explain support and core of a fuzzy set with examples.
Q.2 Model the following as fuzzy set using trapezoidal membership function : “Numbers close to 10”.
Q.3 Using Mamdani fuzzy model, Design a fuzzy logic controller to determine the wash time of a domestic washing
machine. Assume that the inputs are dirt and grease on cloths. Use three descriptors for each input variables and five
descriptor for the output variable. Derive a set of rules for control action and defuzzitication. The design should be
supported by figures wherever possible.
G4 LetA= {a,,a9}, B= (bj, bg, bs), C = (c}, C2)

Let R be a relation from A to B defined by matrix :
Let S be a relation from B to C defined by matrix :
nd$
Find (i) Max-min composition of Ra
R and S.
(ii) Max-products composition of
Qs Define Supports, Core, Normality, Crossover paints and o.— cut for fuzzy set.
QQ.6 High speed rail monitoring devices sometimes make use of sensitive sensors to measure the deflection of the earth
when a rail car passes. These deflections are measured with respect to some distance from the rail car and, hence are
actually very small angles measured in micro-radians. Let a universe of deflection be A = [1, 2, 3, 4) where A is the
angle in micro-radians, and let a universe of distance be D =[1, 2, 5, 7] where D is distance in feet, suppose a relation
between these two parameters has been determined as follows : ;
W is the weight in units of 100, 000 pounds. Suppose the
Now let a universe of rail car weights be W = [1, 2], where
fuzzy relation of W to A is given by,
Using these two relations, find the relation R'oS=T.
(a) Using max-min composition. (b) Using max-praduct composition.
Design a fuzzy logic controller for a train approaching or leaving a station. The inputs are the distance from the station
and speed of the train. The output is the amount of brake power used. Use four descriptors for each variable use
Mamdani Fuzzy model.
Explain the Fuzzy Inference System in detail.
Explain the different Fuzzy membership function.
Explain any four defuzzification methods with suitable Example.
State the different properties of Fuzzy set.
Determine all « - level sets and strong a - level sets for the following fuzzy set.
A= {(1, 0.2) (2, 0.5), (3, 0.8), (4, 1) (6, 0.7), (6, 0.3)}.
Design a fuzzy controller to determine the : wash time of a domestic washing machine - Assume that the inputs
tu are dit
and
al action
each /P variable. Device a estates % a
and grease on clothes. Use three descriptors for lor contro
defuzzification. The design should be supported by. figures wherever possible. » C Clearly indi s are
, . y indicate that iif the clothe
soiled to a larger degree the wash time required will be more.
Q.14 Explain different methods of defuzzification.
Q. 15 Explain cylindrical extension and projection operations on fuzzy relation with example
Model the following as fuzzy set using trapezoidal membership function “Middle age”
Fuzzy Logic
Comp)
ig nigsC (MU-Sem.7-
OO . by any 2 methods.
; he given memb ershipi functio .
Q. 17, det erm ine the defuzzified output value
q.17 Fort n as shown in Fig.
Fig. Q. 17
fuzzy models
Q.18 Compare Mamdani and Sugeno ter as the
ature of wa
purif icati on plant. Assu me the grade of w ater and temper
Design fuzzy logic controller for water va riables. Derive
g.19
: as the output. Use three descriptors for input and output
inp uts and t amount of purifier should be supported by figures.
Clearly indicate that
inputs and the required on and deff uzzi fica tion . The desi gn
set of rules for control the acti
fier required is large.
temp erat ure is low and grad e of water is low, then amount of puri
if water
mple.
techniques with suitable exa
Q.20 Discuss fuzzy composition
given by
Q.21 Two fuzzy relations are y2
yl
x1 lo 0.3 |
R =
x2 LO2 09
zl z2 23
| 1 05 0.3 |
yl
S = 0.4 0.7
y2 L08
the fuzzy relations.
ma x- pr od uc t co mposition between
osition and
Tas amax-min comp
Obtain fuzzy relation zzy inference sy
Fu
stem.
s in a Mamdani
the f ‘o rm at io n of inference rule
Q.22 Describe in detail
ooo
wW TechKnowl edge
Puotications:
f
‘
i
‘
Network
fr
Artificial Neural
i
t
———SS=
:. Important Terminolog
inologies of
Introduction : Fundamental concept : Basic Models of Artificial Neural Networks
ANNs : McCulloch-Pitts Neuron
Single layer Feed Forward ANN, Multilayer Feed Forward ANN,
5.2 Neural Network Architecture: Perceptron,
ation algorithm.
Activation functions, Supervised Learning : Delta learning rule, Back Propag
5.3 Un-Supervised Learning algorithm : Self Organizing Maps
5.1. Fundamental Concept - Artificial Neural Networks _

5.1.1. Artificial Neural Networks
They basically mimic the
Artificial Neural Networks (ANNs) are simplified models of the biological nervous system.
working of a human brain.
.
An ANN, in general, is a highly interconnected network of a large number of processing elements called neurons.
An ANN can be considered as a highly parallel network. Distributed processing is a typical feature of a neural network.
Neural networks work on the principle of learning by examples. They are presented with known examples of a
the network can be effectively
problem called ‘training set’, to acquire knowledge about the problem. After training,
employed in solving instances of the problem previously unknown to the network.
Architecture of a simple artificial neural net
Fig: 5.1.1 Architecture of a simple artificial neural net
As shown in Fig 5.1.1, There are two input neurons and one output neuron.
Each neuron has an internal state of its own called ‘activation of a neuron’,
The activation signal of one neuron is transmitted to another neuron.
— This Activation of a neuron can be considered as a function of the inputs a neuron receives.
— Consider a set of neurons; say X, and Xz, transmitting signals to another neuron, Y,
They transmit signals to output neuron Y
- As shown in Fig 5.1.1 there are two input neurons X; and X,.
li . swan
— Input neurons X; and X2 are connected to the output neuron Y, over a weighted interconnection
ction links (W; an
the net input is calculated as :
For the above simple neural net architecture,
Yin = Ww, Xy+ Wy X2
Comp) __
(MU: -Sem. 7-
|
of the input neurons x;_ and Or espective
A there x32and xjare activations : Le.
over the net input,
the output neuron Yca
nb by applying activations
. sre output y OF . obtained
ey 2 Fin)
tion function
y ov er the net input is called activa
ap pl
thee function that we
Networks s.in a
f giological Neural a neuron. Th er e are ap pr ox imately 40neuron
el em en t in the human brain is called
l computin g ere are
+ The fu ndamenta d sy rapses- Th
ns an
‘human brain. ug h a co nn ec ti on network of axo
ens thro
communica tion betwe
en neurons happ
neuron.
10° synapses per
approxim ately al impulses.
a t e wi th ea ch ot he r by means of electric
nic
_ Neurons commu s the network
of neurons.
ironmen t surround
chemical env complicated
stem and has a very
rm at io n fl ow in nervous sy em el y co mp lex interconne
ctions
in fo made up of ex
tr
ne tw i the brain is
ork in
ur al
- The huge ne
. within the
structure.
e hu ma n nervous system li ve r st im ul j both from
th de
shows inform
atio n flow in ork. Receptors
- Fig. 5.1.2 s pr ov id e in put to the netw
nsor y recept
or d.
e ext ernal worl
- Asshownin
Fig. 5.1.2, se
th e st im ul i originate in th twork o f neur
ons.
se ns or y or gans, when in fo rm at io n into the ne
body and from e s that convey
e fo rm of el ectrical impuls
are in th
-- The stimuli
Internal
; .
External
;
feedback
system
rm at io n in the nervous ous System
Fl ow of in fo ss ing in the Central Nerv
Fig. 5- 1. 2: io n pr oc e
accoun nt of informat
olled on ou s actions.
e s in t he form of vari
espon s
otor organs.
ctors then pr ‘tted to the m
" (CNS), These effe feedback.
command an id external
When necessary, te rn al
both in stem.
ni to re d in the CNS by cs of a closed loop sy
mo racteris ti
Motor organs are has the ¢ ha
ucture
~
ous system str a typical
The overall nerv te s. A schematic of
th e de nd ri
B on ~ | the axon and
ological Neur e cel l bo dy , ca lled the soma,
th
3 major regions: .
Atypical cell has 7 body.
1.3- neuron’s
og ic al ne ur on is shown in Fig.5- bush of th in fi br es around the
_ bi ol is @ very fine TechKnowledy s
at, which
Dénendridrtecites fo£2 rm a dendrite tree, mete Pudticacipns
| Al&SC (MU-Sem. 7-Comp) Artificial Neural Network
Dendrites receive information from neurons through axons.

An
axon isj a long cylindrical
. . 7
fibre 7
that carries impulses from the neuron.
Towards the end, an axon splits into fine branches. Each branch of
an axon terminates in a small end bulb, almost
touching the dendrites of neighbouring neurons
.
The contact between an axon and a dendrite is called a synapse. Via the synapse, a neuron introduces its signal to a
neighbouring neuron. The receiving neuron either generates an impulse to its axon oF produces no responses, The
neuron is able to respond to the total of its inputs aggregated within a short time interval called the period of latent
summation.
The neuron generates a response and sends it to its axon only if the conditions necessary for firing are fulfilled,
———— Incoming axons

soy From other neurons
Dendrites
Cell body
Axon
\ Impulse
Axon
Synapse |
=3 =
Terminal receiving neuron
Fig. 5.1.3 : Neuron-schematic diagram
Incoming impulses can be excitatory if they cause the firing, or inhibitory if they hinder the firing of the response
the inhibition by the amount called the threshold of the
The condition for firing is that the excitation should exceed
40 mV.
neuron, typically a value of about
by neighbouring neurons and by thesneuron itself
The incoming impulse to a neuron can only be generated oon 7 at
are more likely to cause
spaced in and arrive synchronously
Hence, impulses that are closely
The characteristic feature of the ; biological neuron Is that the signals generated do not differ.
e significantly
ie "
either absent or has the maximum value
magnitude; the signal in the nerve fibre is
The information is transmitted between the nerve cells by means of binary signals
a state of total non-excitability ' in cciee of
After an axon fibre finishes generating a pulse, it goes Into or a certai over| if the
conduct any signals during the refractory period
time, called the refractory period. The nerve does not yp 1 OE Tae
excitation intensity is very high.
milliseconds. However the ref act get not
The time for modelling biological neurons Is of the order of owever, ractory period.”
uniform over the cells. “ mes NSB atet
“ nan prain an
s). They do
Low compapared
re to ANN. The speed is in milliseconds. Human (in nanosecond
eed 7 Speed is » High30 teatigue’
may y fati fatigue ifi too much informatio : Fatigue -
afr Brain n and stress is | not experience
presented. i
on the
inci
The no. ; of neuronsi in Brains is 10"
1
i
and total | The size of ANN depends
——————
7€ are 10°> Thos the sami (usually hundreds = OF
no. of connections | application
; us the complexity of human and the developed
who-
brain is very high thousands) chosen
—
develops the application.
ae
. also perform Massive parallel
i simultaneously. | Can
i
complexity | © an perform Massive parallel computation but much
computation simultaneously
brain.
more faster than Human
=
=
=
the information in continuous
- Stores
a in the synapse.
Storage Stores the information memory locations.
are not modelled
capacity
their topology are fault- | Artificial neural networks
due to
Biological neuron networks tolerance or self-regeneration.
Tolerance so minor failures | for fault
stored redundantly
tolerant. Information is
loss.
will not result in memory unit for controlling
There is a control
external to the computing
No specific control mechanism computing activities
Control
Mechanism | task. Nvidia GPU runs on 250.
body's | A single
The brain consumes about 20% of all the human and requires a power
Power watts alone,
on about 20 watts. supply. Machines are way
less efficient
An adult brain operates
consumption | energy.
than biological systems.
other
to connect Artificial neural networks ‘in the
grow and reach out where no
learning in human brain, Brain fibres to | hand, have'a predefined model,
allows new connections further neurons or connections
can b e
to other neurons; ‘ neuroplasticity or weaken based
.
and synapses may strengthen iaadded
or removed.
be created ynap
al
etic | on their importance
NN
saci
jbetween Biological Neuron and A . Cell Body
Comparison
: Dendrites
v -
Threshold
: Vv
Summation
icial neuron
uron and a rtif
b jological ne
logy petween ; ,,
Fig. 5.1.4: Ana Neuron.

d artificial
Biolog ical an
TechKnowledgs
ow s ana log y be tw ee n tions
Efe. 5.1.4 sh
_Y Publica
Artificial Neural Netwoik
Al&SC (MU-Sem. 7-Comp) a 5-5
: inputs based on their SyNaptic
synapy
inpu
— The dendrites in the Biological Neural Network are analogous to the weighted
interconnection in the Artificial Neural Network. :
; ‘ summation and t hresholq
in the ANN which also comprises of
— The cell body is comparable to the artificial neuron unit
unit.
is modeled using
of Artificial Neural Netw ork. So, ANN
- Axon carries output that is analogous to the output unit in case
the working of basic biological neurons.
5.1.4 Features of ANN
from the existing environment and.also adapt to the new

1. Adaptive learning : An ANN possess the ability to learn
environment.
representation of the information it receives during
2. Self-organization : An ANN can create its own organization or
learning time. |
in parallel, they can be used for real time
3. Real-time operation : Since ANN computations may be carried out
and manufactured to take advantage of this capability of
applications. Special hardware devices are being designed
ANNs and to reduce the response time.
if some portions of the neural net is removed
4. Fault tolerance : Neural networks are fault tolerant in the sense even
there will be only a small degradation in the neural: network
(For example some connections are removed),
performance.
: After learning from the available inputs and their relationships, ANN has capability to infer unseen
5. Generalization
relationships on unseen data, thus making the model generalize.
In most of the real-life
6. _Non-linearity : ANNs have the ability to learn and model non-linear and complex relationships.
problems, many of the relationships between inputs and outputs are non-linear as well as complex.
very efficient.
7. Parallel distributed processing : ANNs have massive parallelism which makes them
5.1.5 Disadvantages of ANN
1. The best-known disadvantage of Neural Networks is their “black box” nature. This means that you don’t know how
and why your NN came up with a certain output. For example, when you put in an image of a cat into a neufal
network and it predicts it to be a car, it is very hard to understand what caused it to came up with this prediction
than traditional Machine Learning algorithms, as in at least
2. Neural Networks usually require much more data ; 7
thousands if not millions of labeled samples.
computationally expensive than en
Expensive : Usually, Neural Networks are also more
3, Computationally
tch. Most
algorithms. State of the art deep learning algorithms can take several weeks to train completely f rom scratch.
traditional Machine Learning Algorithms take much less time to train.
5.1.6 Applications of ANN

ata-intensive applicati
Neural networks have been successfully applied to a broad spectrum of data-j , are
- Pplications. Few of them 2” .
listed below.
1. Forecasting
Neural network can be used very effectively in forecasting exchange rates

proved th
ie
Predicting stock values, inflation and cash.
forecasting, forecasting weather conditions etc. Researchers have
model.
tend to excel over that of the linear regression
. As it, th transmis.
) ae .
re qu ir e a larggee amount of memory for storage. . a wres
d ult
t h , the ‘
ages d band i
pigigital im h e r ca n b e v
e n s i v e in te rm s of ti me an
re qu ir ed . ves
ano t ery e x p
andwidth at remo
ommpputer to
co . a t e c h nique th
is e
osion of In imeageswi. th. | Imaggee comcomprsre
ssion reducing th
ith the explre
+ th e dunddant inf ternet, timo
on
:
re pres est arine
siten thuseingimag g its pe rc ep ti bi lity, thus,
wi un or ma ithout affectin
5 ome 0 e the image inB
quired to stor f-organiz
storage size re es ch as K o h o n en’s sel
chniqu su
ss the image. Several NN te
an be effectively ly us used to compre ompressi
on.
NN ¢ ; network etc. can be used for imag e c
s, Ba ck proopagatioi n algorithm, Cellular neural
pr
ma p’
l
rocess contro
4, industrial p
s-
l of dyn amic system
sfu lly in ind ust rial process contro modelling
or ks ha ve been applied suc ces
e be st choice for
ne tw th
Neural oved to be eir univer
sal
lt i- la ye r pe rceptrons) hav e been pr, ro ll er s, due to th
mu nt
ks (especiad lly lementing general — purpos e non-linear ure machine
co
Neural networ sy st em s an imp
non-n-li
near st em
ae ple contro land management
of agri cult
. atio”n capabilities. For exam
approxim
are
Recognition ) tools that
Ch ar ac te r ct er Re co gnition (OCR
Optica l l Char a
4, on is the Optica
ic ati on us in g image recognititi
pl me computer.
Well known ap ftware for the ho recognizing
both
e st an da rd sc an ni ng so
d sy st em for correctly
th e ba se
available with NN with a rul
su cc es s in combining
d grea t
Scansoft has ha el of accuracy.
an d wo rd s, to g¢ t a high lev
characters
nagement
5. Customer Rel
ationship Ma ip Ma na
gement (CRM). ch
me r Re la ti on sh collected for ea
io n fo r NN is Custo
be de ri ve d from raw da ta
pular applicat y informatio
n to
- Another po requires ke data infor mation
.
ip Ma ne ge me nt
ls us ing historical
Custome ; Rela
tionsh eved b y bu
ilding mo de
si nes s processes. Th
ey are doing
- an be ac hi da y to ele v. bu .
. This c
stomer in their
help ty.”
individual cu nology to sed productivi
g neural tech lo pme' nt and increa ©
Many compan ies are NO
W us in
ht, fa st er de ve
e different types
-~
e, gr eater i nsig er ns ca n be identifie d for th
e better pe rformanc s, pa tt
this to achiev ning i n
the database :
fo r da ta mi he company. mand and
etwork s mation tot centre loading, de
> By using Neura | N le custom
er in for
ch as fo recastin g call
to CRM, su
d en hancing databa
ses, clustering and
customers, th
us givi ks related l e t i n g an
, comp
for validating
d be useful em arket, .
~ Also, NN coul a" analysing th ; .
ination,
sales levels,
monitoring ;
sa le s of ti ck et s in relation to dest
base etc. predict
profiling client io n gy ‘ em AM
st T, which could
rese rv at
e is| the airline
- One exampl d ticket price: itechs
time of year an dad techno; logies
.
an advanc e
e la te st °
6 nce ken benefits
i from th science.
Be Medical scie ys ta in te re st In medical .
d that has alwa
of ms in
omising area omedical proble
B:e > > Medicine is the fiel
: .
the next pr e application to bi
‘
cu rr en tl y ha ve ex te ns iv
is NN) s) will work
al Ne ur al Networ k Ss (A Iso DeeP ne
t
bio chemical
~ Artifi ci : systems,
that neural such as diagnostic
Itis believed ars: P lications
\> w ye in me dical ap
the next fe lly applied elopment.
ready be h Knowledge
ANN has al
; Tec
n, Pudlications
> ase detectio
analysis, dise
‘ eee Anificial Neural Network
y _AI&SCiC (MU- -Sem. 7-Comp) _
5.2 Basic Models of Artificial Neural Networks
The basic model of ANN is specified by three basic entities.

1. The Model’s synaptic connections.
2. The training or learning rules for updating weights.
3. The activation functions.
5.2.1. Connections
There
defines the n eural network architecture.
The arrangement of neurons in layers and connection between them
are four basic types of neuron connection architectures. They are:
1. Single Layer feed forward network

2. Multi-layer feed forward network
3. Single layer Recurrent network
4. Multi-layer Recurrent network
(We will discuss above points in Section 5.5]
5.2.2 Learning
behaviour brought about by experience.

— Learning is a relatively permanent change in
nse to a specific input.
— Learning in a network is a process that forces the network to yield a particular respo
mental learning.
— There are two types of learning used to train the neural network: batch learning and incre
— Batch learning takes place when the network weights are adjusted in a single training step.
information produced
-— In this mode, the complete set of input training data is needed to determine weights. Feedback
by the network is not involved in developing the network. This learning technique is called recording.
= Incremental learning is most commonly used and can be broadly classified into three basic types.
5.2.2(A) Supervised Learning
: ‘diagram supervised leaming.

u yervi sed Jeaming ie
ls the rarget
sponding output pattern associated with it, which
— In this type of learning,eeach input pattern has a corre
or the desired pattern.
_ ° Here, a comparison is made between the actual output of the network and the desired output to find out the ‘errors
— The computed error can be used to adjust network parameters (like connection weights, threshold etc.). As the
network parameters are modified, the performance of the network improves,
— The training could continue until the network Is able to produce the €xpected or desired response
ro
5-8 Artificial Neural Network,
S e ired
As shown in Fig. 5.2.1, the distance
function fd, oft kes two values as an input. Actual network output 0 and desi
output d for an input X and
then computes error inca
sure,
hat is to
since we have assumed adjustable weights network performance t
reduce the error.
the weights can be adjusted to improve
Input Ai Network
|
xX = —————> output
. ial
0 ery
desired
<— response
pfd,0]
error/distance
function
Fig. 5.2.1 : Block diagram of supervised learning
if needed by
- This is analogous to classroom learning with the teacher's questions answered by students and corrected,
the teacher.
5.2.2(B) Unsupervised Learning

Dec. 12, May 13, Dec. 15
if there is no teacher to present

In this learning method, the desired output is not presented to the network. It is as
adapting to structural features
the desired or expected output. Hence the system learns on its own by discovering and
present in the input pattern.
Network output
> oO
Fig. 5.2.2 : Block diagram of unsupervised learning

learning must somehow be accomplished based on the
- Since no information is available to correct the network,
inputs.
observations of responses to
learning ¢ an be used for finding the boundary between classes of input patterns
~ For example, unsupervised
distributed as shown in Fig. 5.2.3.
in Fi g. 5.2.3 (a), cluster boundaries can be found based on the large and representative
- In favourable case, as shown
sample of inputs. XQ
x o| ‘ Xy
(a) Clustered (b) Unclustered

Fig. 5.2.3 : Two dimensional patterns
Artificial Neural Network
AI&SC (MU-Sem.7-Comp) 5-9
S| uitable weights self-adaption
mechanism |'
— Since no external instructions regarding potential clusters is availabl e,
pta tion rule that can

be applied is “A pattern
needs to be embedded in the trained network. One possible network ada
than to the other ¢ Justers’ centres”.
added to a cluster has to be closer to the cluster centre
patte rns, regularities or
separating
— Thus in unsupervised learning, the network must itself discov er possibly existing
properties. ;
s to learning the subjec but not
— This type of learning is analogou t from a videotape lecture covering the material
e for clarification of
lecture bu t is not availabl
including any other teacher’s involvement. Here the teacher delivers the
unclear questions or to check answers.
ning
supervised and unsupervised lear
Table 5.2.1 : Difference between
a type of learning where the desired

1. |Supervised Learning is that type of learning where|Unsupervised Learning is
the desired output is known output is unknown
uses procedures that attempt to find

2. JA supervised learning algorithm analyses the Unsupervised le arning
of patterns given to the neural network.:
training data and produces an inferred function, natural partitions
which can be used for mapping new examples.
ring
3. |Supervised learning is achieved by means of|Unsupervised learning is achieved by means of cluste
classification or regression
a Unsupervised learning is known as learning without
a teacher's
.4. [Supervised learning is known as learning with
presence as desired output is unknown
teacher’s presence as the desired output is known
beforehand
small and With unsupervised learning it is possible to learn larger and
more
5. |Supervised learning is limited to learning
simple models complex models than with supervised learning
learning, Winner-take-all learning are examples of

6. |Perceptron learning, Delta learning, Widrow-Hoff|Hebbian
learning, are examples of supervised learning]unsupervised learning algorithms
algorithms
images and there is no labelling on them,
‘7. |Example: Giv en several images of faces and not-/Example: Given several
wilijan unsupervised algorithm clusters the data into different
faces (i.e. labelled data), a supervised algorithm
eventually learn and be abl
le to predict whether or/groups, e.g. it can distinguish that faces are very different from
; landscapes, which are very different from horses.
not an unseen ima ge is a face. - _—
ning
5.2.2(C) Reinforced Lear thé staan
s not pre sen t the exp ect ed answer but aiy'tnaidSted iF
ugh available, doe
In this method, a teacher tho
orrect.
output is correct or inc
rning process.
helps the network in its lea
— The information provided
is a ve'Y
A reward is given for a correct answer computed and a penalty for a wrong answer. Reinforced learning
d learning is not
general approach to learning that can be applied when the knowledge required to apply supervise
available.
sed learning, because they are
— However, it is usually better to use other methods such as supervised or unsupervi
8,
more direct.
this case may not be indicated as the desired
Reinforcement training is related to supervised learning. The output in
output, but t he condition whether it is ‘success’(
. ndicated, Based on this, error may Pe
1) or ‘failure’ (0) may be indi mo
ed her e is bin ary .

calculated.The error signa | produc
5-10
z “0 (MU-Detl sowvuip)
Network
Artificial Neural
this learning attempts to lear
the system kn
sically, . N the input-ou ows
Correct or not, but does not kno tput mapping through trial and error. Here
whether the output is . W the correct out put.
table 5.2.1 shows the difference between Supervised and Unsupe
rvised learning
‘23 activation Functions
t response of a neuron i '
_ The outpu P is calculated using an activation function (also called transfer function).
ighted in i :
_ Thesum of weighted inputs (net input to the neuron) is applied with an activation to obtain the neuron’s response-
Neurons placed in the same layer use same activation,
_ There are basically two types of activation functions: linear and non-linear
| The non-linear activation functions are used in a multi-layer net
The most commonly used activation functions are :
1 Unipolar Binary
- {t's unipolar in nature meaning it generates only two values 0 or 1.
- {tis used only at output layer.
- Itis computed as,
1 net >O
f(net) = 0 net < 0
4f(net)
el —___—_—_—> net
0
function
Fig. 5.2.3 : Unipolar binary activation
2 Bipolar Binary
s two values +1 or -1.
- It’s bipolar in nature meaning it generate
It is used only at output layer.
Its computed as,
+1 net>9
f(net) = sgn (net) Pe net <0
4 {(net)
+1
ee net
——
-1
activation function
Fig. 5.2.4: Bipolar binary
ting activation functlons.
polar binary are called hard limi
The two functions bipolar binary and uni
Tech Knowledge
Pubtications
Artificial Neural Network:
3... Sigmoidal Function / Unipolar Sigmoidal

— Also known as unipolar Continuous or (Binary sigmoidal )
— It’s range is 0 to 1.
- It’s non-linear in nature. 0 or 1, as value for sigmoid function
— Usually used in output layer of a binary classification, where-result is eithe
. . i r
lies between 0 and 1 only.

— It is computed as,
n
1
f(net) = sere Where, net = D w,x;
i=l
f(net)
+1
t » net
30-20-14 0 1 2 3
Fig. 5.2.5 : Sigmoidal or Unipolar continuous activation function
4. Tanh Function / Bipolar Sigmoidal
- Also called Bipolar Continuous (or Bipolar Sigmoidal).
— Therange is between +1 and—-1.
-— It is computed as,
n
2 .
f(net) = Teel Where, net = » w,X;
i=l
- This function is related to the hyperbolic tangent function. is called the steepness parameter.
“Usually used in hidden layers of a neural network as it’s values lies between -1 to 1.
— {t's output is zero centered because its range in between -1 to 1. Hence in Practice it is always preferred over
Sigmoid function .
f(net)
Flg. 5.2.6 : Tanh or bipolar continuous

activation functio
Se 5-12
-, Wecan easily observe that at 1-0 in bipolar co i ones oT en nction.

' idal)
m nuous (sigmo function it becomes the sgn {net) fu
4 The two activation functions bi
Polar continuous an iy eristics and
hence called soft limiting activat ion functions, d unipolar continuous have sigmoidal charact
peLu- Rectified Linear units

It is mathematically represented as
R(x) = max(0,x)
ie ifx <0, R(x) =O and ifx>=0, R(x) =x,

_ Itis simple and efficient.
from Tanh function.
- _ Itwas recently proved that it had 6 times improvement in convergence
- it avoids and rectifies vanishing gradient problem . Almost all deep learning Models use ReLu nowadays. But its
limitation is that it should only be used within Hidden layers of a Neural Network Model.
‘w=
ReLU
Fig. 5.2.7 : ReLu
§.. Softmax Function

of sigmoid function.
~ The softmax function is also a type
classes. What if there are multiple classes.
- The sigmoid function can handle only two
~ Softmax is commonly used for multi - classification problems.
It is non-linear in nature.
the outputs for each class between 0 and 1 and would also divide it by the
~ The softmax function would squeeze bability of the input being in a particular class.
sum of the outputs. This gives the pro
ed in the output layer of the classifier where we actually try to attain the
7 The softmax function is ideally us ee
probabilities to define the cl ass
of each input.
Other activation functions:
Linear
>
Al&SC (MU-Sem, 7-Comp) Artificial Neural Network
Piecewise linear
, .
1 if mi>l
0
O=‘jgl if |mi<1| +1
-1 if ml>-1
g=tang _ I
-1
Step function (unipolar binary) _ { 1 if 1>=6 1

“lO. if |<0
Sigmoidal functions. “(unipolar continuous SoD emo) aare” ‘used dine multiiay
back-propagation networks, Radial Basis Function Networks (RBFN) etc :
The reason is, in multilayer network, the actual Inputs are effectively masked off from the outp
the intermediate layer. e .
The hard- -limiting | threshold functions8 (unipolar binary and bipolar binary) remove the informa on
needed if the network i is to successfully learn.
Hence the network. iss unable to determine which of the input weights should be increased ‘and whic
5.3 Important Terminologies of ANN

5.3.1 Weights
In the architecture of neural networks, each neuron is connected to other neurons by means of directed
communication links. Each of these links are associated with weights. These weights Carry the information about the input
signals. The weights can be presented in terms of a matrix called the weight matrix. The weight matrix is also called
connection matrix, and is defined as,
Wir Wig ee Wim
t
W, Wo, Waar +++ Wam
t
Ww =| W, |= .
(
W,, War Wa oe Mam
ft o.04 te
Where W, = [We Woe e+ Wim) for i= 1, 2,3, «+ mis the Weight vector for processing element
W, is the individual element of weight vector that represents the weight on the communication link from i” neuron to
h
j neuron. _ ee
TachKnewledge
Publicakions
2 Bias
abias isa weight

weig on aco Nnection from an additi 0 Nalin .
Put unit whose activation is always 1.
,
ng the bias in
;
__iereasin€ creases the net input to the unit,
4
b
x4 w
be? S Output
Xe
- Fig.9. 5.3.1 : A single neuron with a bias
ifthe bias is present, the net value is calculated as

n
net = b+ DY x,we i=l,2,...,0

i=1
- forthe above example, net is computed as,

net = D+W, X,+W2%2
specific value.
bias should also be initialized to some
= Similar to initialization of weights,
Why do we need to use bias?

s to improve performance of the net.
= Using bias in a neural network help function to the left or right which
may be critical for
us to shift the activation
- For example, the bias value allows
’
riate
successful learning.
th ro ug h the ori gin . This may not be approp
(or plane) passes
i a s i s not use d, 4 the separating line us to shift the separating line (or
plane) either to the
= ustaabl
: an adjjust
e bias help s
Usually, when a bias r's Pproblem. Adding
for solving a par tic ula
the origin.
right or to the left of 7 5.3.2.
eke work shown in Fi w
~ Consider the following net
Output 0
ba bias
weight
k with a bias
Fig. ;5:3-2: Networ
Iculated as,
Be. utput 0 15
then P
d | activation function,
tw e use sigmoida
gd xXx + b)
= sigmoid ;
oO
sigmoid &* P) . 5.3.4.

lef t or to the right as shown in the Fig
the
__ Where| b is the bias.
bias,
Using different values of
Sigmoid (1 x x +- 5)
1.0 f
0.8 + Sigmoid (1 xx +0)
os + ——> Sigmoid (1 xx +5)
047
027
—>x
-10 -5 0 5 10
Fig. 5.3.4: Sigmoid functions with different bias values
5.3.3 Threshold
Threshold is a set value based upon which the final output of the network is calculated.
Usually a net input to the neuron is calculated and then compared with the threshold value.
If the net value is greater than the threshold, then the neuron fires, otherwise, it does not fire.
5.3.4 Learning Constant
The learning constant is used to control the amount of weight adjustment at each step of training. It controls the rate
of learning.
The effectiveness and convergence of learning algorithm-depends significantly on the value of the learning constant.
However, the optimum value of learning rate depends on the problem being solved and therefore there is no single
learning constant that can be used for different training cases. ‘ ce
For example, to solve a problem with broad minima, a large value of learning constant will result in a more rapid
convergence. However, for problems with steep and narrow minima, a small value of learning
constant must be
chosen to avoid overshooting the solution. But choosing a small value of learning constant also
increases the total
number of steps in training.
5.3.5 Momentum Factor
Convergence is made faster if a momentum factor is added to the.weight update process. This is generally done in
B y
back propagation algorithm.
The method involves supplementing the current weight adjustment with a fracti
Q , ; raction Plof th Uy mos’ recent weight
weigh
adjustment.
AW(t) = -nV E(t) + ocAW(t- 1)
where, argument t and t — 1 indicate the current and the most recent training step respectivel
: ; ively.
1 is a learning constant and a is the user defined selected positive momentum constant
In the above equation, 2 the term : adW/(t — I)indicates
‘ a scaled most
; re cent ad : ust . : the
momentum term. Justment of weights and is called
Typically a is chosen between 0.1 and 0.8.
N etwork
Al&SC (MU-Sem. 7-Comp
) Artificial Neural
ter
3.6 Vigilance Parame
quired
the vigilance parame ter denot ed by i the degree of | similari, ty re
pis control
: in ART network s. It is used to
to the same clus er uied
unit. The range of vigilanc e parameter is usually 0.7 to 1.
for pattern to be assigned
uron Model
54_McCulloch Pitts Ne
*:
8 ae
biological
simplified considerations of the
5
tion of | neuro n model based on highly ;

q
synthetic
.
The first formal defini

a
model was formulated by McCulloch and Pitts in 1943
The model is shown in the Fig. 5.4.1. impulse at

presence of the input
i
| 3, ..., n are either O or 1, depending : on the absence or
_ The inputs x, i for i = 1, 2,3
instant k. w;
: x
Pitts Neuron model

Fig. 5.4.1: McCulloch
nal is denoted by o.
- The neuron’s output sig
to neuron.
con nec ted by direct weighted paths inhibitory (weight is
- The inputs are
lin ks ma y be exc i tatory(weight is positive)or
unication
ated with the comm
- The weights associ
the threshold, then the
negative). to the neuron is greater than
a
ne uron , and if the net input

uron does not fir e i.e., It goes into an
- - There is a fixe d th
reshold for each s th an th e threshold, then the ne
er
i s les
input to the neuron
e net
neuron fires. If th
inhibition state. as follows,
fo r th e ne ur on is defined
rule
~ Accordingly, the firing n
k
1 if x wx,2T
i=l
ke1 n
oO = k
0 if wx, <T
Fhidt 1
Sr 1 <!
k = 1, 2, 3, ue
no te s di sc re te tim e instances,
t k de signal will have
on, superscrip ctiv e synapse which: transm
its. a stronger
In the above equa ti
oo
, an ef fe
neuron aller weights.
biological nap se will have sm
By analogy with the ile a we ak sy
weights, wh
correspondingly larger :
model,
~~ In McCulloch Pitts’ tatory synapses, -
ci
= +1 for ex napses- ;
: 5
—1 for inhibitory
Artificial Neural Netwoik
Al&SC (MU-Sem. 7-Comp) 517 —areew
i
by the weighted s um of input
1p ignals f forthe
signals
And T is the neuron’s threshold value, which needs to be exceeded
neuron to fire.
}
~— Although, this neuron model is very simplistic, it has subs tantial ‘computing potent
ial. It ¢ an perform’ basic logi,
lected . =
threshold are app ropriately se
Operations such as AND, OR, NOT etc. provided its weights and performed to
- The McCulloch Pitts neuron does nat have any particular training algorithm. An analysis has to be
determine the values of weights and the threshold.
Problems based on McCulloch Pitts Model :
Ex. 5.4.1: Design two input AND logic using McCulloch Pitts Neuron model.
Soln.: Consider the truth table for.the AND logic function shown in Table P. 5.4.1.
Here we need to find the appropriate values of w,, w, and threshold T such that, they satisfy the truth table of AND
logic.
Here, the firing rule is,
W,X,+w,xX, 2T ~(l)
And inhibition rule is,
W, Xt W2X) <T ---(2)
From the truth table of AND logic, we get following four inequalities.
0 <T
w, <T
w, <T
Ww, tw, 2T
Table P.5.4.1 : Truth table of AND
0] 0 0
x
o}1 0 '
1/0 0
Xo
1] 1 1
| Fig. P.5.4.1: MCP Neuron Model for Logical AND

Se e _ above four inequalities, use Equation (1) when neuron must fire and use Equation (2) when neuron
For x, = 0, x, = O=> y = 0 (Le. inhibitory)

So substitute x, = 0, x, = 0 In Equation’(2), we get,
0 <T
For x, = 0, x, = 1=> y =0 (Le. inhibitory)
ee a
AlgSC (MU-Sem. 7-Comp) twork _.

Artificial Neural Ne
0 substitute x = 0, x, = 1 in Equation (2),we get,
w, <T
For Xy = 2, X2 = 0=> y=0 (ie. inhibitory)
so, substitute x, = 1 and x, = 0 in Equation (2), we get,

: Wi <T
for X, = 1,%,=1> y= 1 (ie. Firing)

so, substitute x, = 1 and x, = 1 in Equation (1), we get,
w,+Ww, 2T
Here W, and w, can only take either + 1 or —1values.
So, for Wy = 1, W, = 1 and T = 2, all of the four inequalities are satisfied. Hence, solution is,
w, =1
i
w, =1
T =2
for AND logic

Fig. P.5.4.1(a) : Final weights and threshold
Pitts Neuron model.

§x.5.4.2: Design two inputs OR logic using McCulloch
Soln.:
OR logic function
j shown iin the Table P.5.4.2.
Consider the truth table for
Table of OR
Table P.5.4.2 : Truth
PIiPlolol
PilPilPilo
rlojrRio
Output y
X2@ We
Fig. P.5.4.2 :MCP Neuron Model for Logical OR
ew opriate values of Wy w, and T such that they satisfy the truth table of OR logic.
€ need to find the appr
Here, the firing rule is, Ql)
WX, + WX = T
and the inhibitory rule is, .++(2)
W,X,+W xX, <T inequalities.
get the following four
we
From the truth table of OR logic,
0 <T
> T Vale Teoh rents
WF _Alasc (MU-Sem. 7-Comp) 5-19
w, 2T
Equation (2) when th e@

(1) when the neuron must fireire an and use
ri e the above fi four inequalities, use Equation
T © deriv
neuron must be inhibitory. ;
. j t,
For x, = 0, x, = O= y =0 (i.e. inhibitory), so substitute x, = 0, xX, =Oin Equation Ewe Ee
Q0<T
x, = 0, x, = 1 in Equation (1),we get,
For x, = 0, x, =1=> y = 1 (i.e. firing), so substitute
w, 2T
= 1, x, = 0 in Equation (1) we get,
For x, = 1, x, = O=>y = 1 (i.e. firing) so substitute x,
w, 2T
For x, =1,x,=1=>y=1 (i.e. firing).
So, substitute x, = 1, x, = 1 in Equation (1) we get,
Ww, +w, 2T
Hence, the solution is,
the above inequalities are satisfied.
For w, =1, w, = 1andT=1, all of
w, =!
w, =1
T =1
Output y
xg ewe”
OR logic
Fig. P.5.4.2(a) : Final weights and threshold for
P.5.4.3 using McCulloch-Pitts neuron.

Ex. 5.4.3: Implement the logic function shown in Table
Soln. :
Table P.5.4.3 : Truth table
Fig. P.5.4.3: MCP Neuron Model for given Logical function
that they satisfy the truth table.

We need to find the appropriate values of W,,W, and T such
The firing rule Is,

pill)
W, X, + Wo Xo 2 T
And the inhibitory rule Is,

Wy X + Wa Xo <T
go poo L— — m.
(mu-Se
From the truth table , we get the following four inequalities,
0<T
w, <T
w, 2T
wytw, <T
is,
go for Wy = 1, W2 =~ 1 and T = 1, all of the above inequalities are satisfied. So, the solution
wi =]
w, =-1l
T =]
XQ We* .
given logic
Fig. P.5.4.3(a) : Final weights and threshold for
neuron model.
1.5.4.4: Design NOT logic using McCulloch Pitts
Soln.:
function.
Consider the truth table of NOT logic
NOT
Table P. 5.4.4: Truth table of
T=?
t w=? © Output y
of * .
1/0
l for Logical NOT
Fig. P.5.4.4 : MCP Neuron Mode
y the truth table of NOT logic.
app rop ria t e values of w and T, such that, they satisf
out the
Here, we need to find
The firing rule can be,
w, 27
And inhibitory rule can be,

-
w, <T .
lowing two inequalities
NOT logic, We get fol
From the truth table of
0 2T
i qualitit ies are sa tisfied.

e mentioned two ine
ye Sr
0, forT =O and w=w
-1t, he thea abov
So, Solution is,
Tech|Knowled
Publteatin
° : , spear
W_AlaSc (MU-Sem. 7-Comp) 5-21 “eee itificial Neural Network
T=0
xe—W==1 ()
Output y
Fig. P.5.4.4(a) : Final weights and threshold for NOT

5.5 Neural Network Architecture
— As we know, an artificial neuron is defined as a data processing system consisting of a large number of simple highly
interconnected processing elements inspired by the biological neuron.
— Generally, an ANN structure is represented using a directed graph, where neuron outputs are connected, through
weights, to all other neurons including themselves; Both lag-free and delay connections are allowed.
— Based on these connection types, there are four fundamental architectures of neural networks.
5.5.1 Single Layer Feed Forward Networks
— Fig 5.5.1 shows the architecture of a‘single layer feed forward network.
-— Asingle layer feed forward network consists of neurons arranged in two layers.
— The first layer is called the input layer and the second layer is called the output layer.
— The input signals are fed to the neurons of the input layer and neurons of the output layer produce output signals.
— Every input neuron is connected to every output neuron via synaptic links or weights.
Output
layer
Input layer
x; : Input neurons
y; : Output neurons
_ Wy: weight
Fig. 5.5.1 : Single Layer feed forward network
— The output vector is given as,

O = [04 02 O3eer-.-Oml’
And input vector is,
— Weight Wy connects the /” output neuron with the/" input neuron.

— Then the activation value of the i” neuron is given as,
n
net, = >, wx, Fori=1,2,3,....,m

j=o
— The transformation performed by each of the m neurons in the network, is non-linear mapping and expressed a5,
O, = £(W,X), fori=1,2,3,..m
- teers, : ee
TechKncwladge ‘ at
ny Funlscations
am. 7-Comp)
;contai ns weights leading

W,conta ; =
where wel ight vector toward the i” output n node and is defined as ie
Wi = lw, wy Wig
se Wi, ]'
only the outP ut
'
ite of having two layers, the network is still named as ’single layer’ because out of the two layers,
InsP!
- layer perfo
P rms j
computations. The e input signal is simply transmitted to the output layer by the input layer.
Activation function
x() =>] five) syn
Fig. 5.5.2 : Feed forward network block diagram

nis not
Thus, feed iamete networks are characterised by the lack of feedback. That is, the output of any notgerne
influence future
or through other neurons. Thus, present output does
fed back to itself directly or indirectly
output.
orks
55.2 Multilayer Feed Forward Netw
tput layer, this
d networ k, there are multipl e layers. Thus, besides having input layer a nd ou
Ina multilayer feed forwar
called hidden layers.
network has one or more intermediary layers
Output
Hidden
Input
layer layer
layer
feed forward network
Fig. 5.5.3 : Multilayer
za aries
ee
layer Output layer
input layer Hidden
ee
Piel 23,0409
% Input neurons
|_
jul, 2,3, en
Hidden neu rons
¥ wt He ——— e
ke hLadiuoe
qh Output neurons TT eaeeeateeaneanmemammmamecenimm
mmaniiamll
—
|
weight ee e
y4 Input hidden layer oO nemas
_
-output layer veight |
wi Hidden
EO
a ; k
forward nelwior’
cae
Fig. 5.5.3 shows a multilayer feed ron s OF hidden units. Before

directing the Input
hid den nuu
are known as
units 5 of the hiddan layer useful intenmedlary computations.
;
~The computu ationarl the hid den layer per for ms

tothe aipuriave weights on these connections are
aye
Y ected to the neurons in hidden layer and the
Neurons in input layer are conn weights”.
Cferrag to as “Input-hidden layer
_Al&SC (MU-Sem. 7-Comp) ’ Artificial Neural
Se
. .
Similarly, the hidden layer neurons are connected to the output layer neurons ard.th e corresponding
P 8 weighWeights are
referred to as “Hidden-output layer weights”.
.
A multi-layer feed forward network with m input neurons, n; neurons ,in the first t hi hidden layer,
yer, n, nz neurons in the
second hidden layer and k output neurons is written as m—,-n,—k network.
5.5.3 Single Layer Feedback Network

i ast one feedback lo Op.
— The feedback networks differ from the feed forward network in the sense that there exists at le
i
- When the output of the output neurons is fed back as an input to the same or preceding layer n odes, then thist ype of
network is called feedback network.
—. Fig. 5.5.4 shows one such type of feedback network. ~
Ease
= { ie |
x, (0) @ Wi4 >p 0,(t+A)

Wo4 Wi2
0,(t)
Xp(0) eo W22 nee

ee Z b—» 0,(t+A)
2
Oo(t)
o Won ®
Lag-free
® neurons
Wi
° e
Wm2
Ww. .
X,(0) eo mn p—> 0,(t+A)
o,(t)
See
Delay elements
Fig. 5.5.4 : Single layer discrete time feedback network
—— O(t+A)
Fig. 5.5.5 : Block diagram of feedback network
oft + A).
The present output, say oft), controls the output at the following instant,
logy to the refractory period of
: a
t + 4. Here the time delay 4 is in an ana

.
A indicate s the time ela pse d bet wee n t and i

3 .
:
2.
' basic biological neuron model.
into o(t+ 4) can be written
, as,
3 : Thus, 1 the mapping of oft)
o(t+A) =F [Wot]
tg Al&SC (MU-Sem. 7-Comp)
ce Artificial Neural Network
" 5-24
, 5.5.4 Multilayer Feedback Networke
Fig. 5.5.6 : Multiplayer recurrent network (Redraw Figure)
- Fig 5.5.6 shows architecture of multilayer feedback network.

a preceding layer, forming a
it can be noted that a processing element output can be directed back to the nodes in
multilayer recurrent network.
back to the processing element itself and to
- Also, in these networks, a processing element output can be directed
other processing elements in the same layer.
5.6 Supervised Learning

(or
data meaning, for every input, the corresponding output
As discussed earlier, supervised learning works on labeled
’
class) is known.
e main supervised algorithms.
Following section discusses thre
1. Perceptron learning rule

2. Delta learning rule
thm
3. Back propagation algori
Rule .
5.6.1 Perceptron Learning MU.— Dec. 11, May'12
5.6.1(A) | Rosenblatt’s Perceptron

Association unit A and Response unit R.
, the net wo rk consists of 3 units, the Sensory unit S,
As shown in Fig.5-6-2 t .It contains 400 photo-detectors, -
nd generates 0 or 1 electrical signal as outpu
The S unit: receives
‘ves input
inpU images 4pared with threshold. If the input signals exceed a threshold, then the photo-detector
then com
These input signals are
9. comprises of feature
outputs 1 otherwise y connected to the Association
unit A. The association ‘ unit A /
s ar e randdoml
e
The photo-detector
es-
demons or predicat ;
TechKnowledys,
Pudlicatio
Se
Al&SC (MU-Sem. 7-Comp) 5-25 Artificial Neural Network
Sensory unit Association unit © | Response unit
(S) (A) (R)
ME
> Output
weights O Adjustable
weights
Photodetectors
Fig. 5.6.1 : Rosenblatt’s original perceptron model

— The 3 unit R contains pattern recognizers (also called perceptrons) which receive the results of the predicate.
— The weights of S and A units are fixed while those of R are adjustable.
— The output of R is 1 if the weighted sum of its inputs is greater than 0, otherwise it is 0.
5.6.1(B) The Perceptron Model

— The perceptron model is based on the perceptron learning rule in which the learning signal is the difference between
the desired output and the neuron’s actual output.
— Here learning is obtained in supervised environment and the learning signal is equal to:
r =d-o,
t
Where 0, = sgn(W,X)
and d; = desired (or target) output
— Here the weights are adjusted using following weight update formula.
Aw, = ¢ [d.-sen ( W, X)] x
AW; y c [4,-sen (w, x)| X;; forj=1,2,.
c(learning constant)
Fig. 5.6.2 : The perceptron model
neuron response only.

This rule is applicable for binary
ghts are adjusted if and only if O, isi inert (ie. if #0)
From the above formula of AW, it is clear that ,wei
adjustment reduces to
Since the desired output/response is either 1 or — 1, the weight
AW; =+2cX
ee
4 Alb gc (MU-Sem.7-Comp) 28 artificial Neural Network

(W!
: where AW + 2cX when d, =1land Sen X)=o=~1
i
a nd AW,= ~2cX when d= 1 and Sen(Wx)

* =o=1
:
The weight adjustment is inherently zer 0, W hen d==pli
ofi.e. d= sen(w,x)
‘ ). Thus, when the desired output Is the same
as the actual output, there will not be any change in weigh .
weights
6.1(C) Algorithm for Training a Single Perceptron Network
single perceptron can so i jong

s and two output classes (0/1).
. iven° slow is the al a a classification problem with ‘n’ input feature
algorithm for training such a network i i
Discret Percep tron ' Trainin
e Percep t in g
as and iis called SDPTA (Single
Algorithm)
spPTA (Single Discrete Perceptron Training Algorithm)

_° Given are P training pairs
{Ky dy Xqr dy» X3y ds, .., Xp, dy} where X/ is (n x 1), djis (1x 1),i=1,2,...P
_ Here augmented input vectors are used.

X.
Y, = [*| fori=1,2,...,P
and it is 1.
i.e. there is an additional input for the bias
counter within a cycle.
ing steps and p denotes the step
In the following, k denotes the train
Step1: c>Ois chosen
random values.
Step2: Weights are initialized as W at small
Wis(n+1)x1
Counters are initialized
K-1,pe-1, EO
and output Is computed
here. Input is presented
Step3: The training cycle begins
YCY,, d ed,
Otsgn (W'Y)
Step4: Weights are updated
1
Wew+s c(d-9o)Y
Step5: Cycle error is computed

1
Fes (d—o)'+ E
Step6: fp <p then

Pepe+l kekt4
and go to step 3; othe rwise go to steP 7
leted.
Step? the training cycle is comP lghts and k.
ssion. Display we
ining s¢
oc For E = 0, terminate the tra
: IFE>0, then £ <-0,P<* wap
by going to step 3
and enter the new training cycle Tech Knowledge,
,Publicatioa
—,
°
¥ Al&SC (MU-Sem. 7-Comp) : Artificial Neural Network
————— SS
5.6.1(D) | Model for Multilayer Perceptron
In multilayer perceptron networks, the perceptrons are arranged

in layers.
This model has three (or more) layers:
a. An input layer
b. An output layer and
¢. One (or more) layers between input layer and output layer. These layers are called hidden layers.
ion and for the perceptr ons in the hidden and output
For perceptrons in the input layer, we use linear transfer function p
layers, we use sigmoid (continuous) function.
;
Here, since the single layer perceptron’s model is modified : changing the transfer
by adding a hidden nl “ vie ee So'tiw
(activation) function from linear function to a nonlinear function, we need to alter the learning :
the multilayer perceptron network should be able to recognize more complex things.
The input-output mapping of multilayer perceptron is shown in Fig. 5.6.3 and is represented by,
O = f,[W [f,[XV]] Where, f, and f, represent non-linear mapping.
Input layer Hidden layer

Output layer
Fig. 5.6.3 : Multilayer perceptron
5.6.1(E) |The EX-OR Problem and Need for Multi-layer Perceptron

— The simplest and most well-known pattern recognition problem in neural network literature is the Exclusive-OR
problem,
- The truth table of Ex-OR function is given below,
Tablo 5.6.1: EX-OR Truth Table
My | 2 | Output (y)
1 1 -1
1 |-1 1
mt 1 1
-1}~-1 ~4
- The task of the network Is to classify a binary Input vector to output el
the inputs +1 Or both the Inputs -2) or assign It to output class = 4 othe S=— 1 ’ ifthe v ector has similar inp’inputs (Both
wise,
— Let us try to solve this problem using a single-layer perceptron Model, Consider the follo eptron
model shown in Fig. 5.6.4, wing g single-layer
single-layer p perceP ae
EF Techinowledgi Se , _—
AlaSC (MU-Sem. 7-Comp) Artificial Neural Network
a:
— =
xy
7
a j
0.45 Sin
Fig. 5.6.4: gle layer recent . a
dary is preceptron model
ror the above model, the decision boun ~
_
b+ Ww, X,+ WX, =0
ie.net =0
1
if net > 0, y (the output)= 1, else y=-
following four inequalities.
- 5Sofrom the truth-table and above conditions, we have the
b+w,+w, <0
b+tw,-Ww, >0
b —w,+w, >0
b-w,-w, <0
which must be satisfied.

when
- However, the set of inequalities is clearly self-contradictory
considered as a whole.
ly separable. This can
- This is because the EX-OR problem is not linear
Fig. 5.6.5. Fig. 5.6.5 : EX-OR function plot
easily be observed from the plot given in
the 2D input space into
ot use a singl e layer perce ptron to construct a straight line to partition
- Inother words, we cann
two regions (class 1 and class 2).
perceptron
Solving EX-OR problem using Multilayer
XOR gate
:
OR gate
AND gate
Output
NOT AND gate
multi - layer perceptron

Fig. 5.6.6 : EX-OR using
arable Patterns
ly Sep
Linearly and Non-Linear
i
5.6.1(F) MU — Dec. 10, May 12, Dec. 13, Dec. 15)
able Patterns desired

4. Linearly Separ assification
problem, w here
if the input pattern is a member of its class, then the
s
~ Considera particular acl otherwise it
is “no”.
:
( tput) IS aby
response (ou yes an of 1, a “no” by an output signal of— 1 (considering
bipolar signals).
is represente
output signal Tech Know!
~~ A “yes”
Pudtica
Al&SC (MU-Sem. 7-Comp) 5-29 Artificial Neural Network
— To achieve this, a step p function can b e ivation fu
used as an activation function. That is, the output is + 1, if the net inputt>
the net input to the neuron ig
the neuron is positive, and -1 if the net input to the neuron is negative. Since
computed as i
n
net =b+ > Xi,
i=l
i is determined by the i
— Itis: clear that the boundary between the region where net > 0 and the region where net < Oi
relation. {
n
b+ > XW, = 0
i=l
— This boundary is also called decision boundary.

— Depending on the number of input units in the network, this equation represents a line, a plane or a hyperplane.
if — If there exist weights (and a bias) such that all of the training input vectors for which correct response !s + ; - on
correct response is — 1 lie on
one side of the decision boundary and all of the training input vectors for which he
the other side of the boundary then we say that the problem is “linearly separable”.
Class 0
f4—— Decision boundary
(linear line)
é:) Class 1
Fig. 5.6.7 (a) : Linearly separable pattern

the network shown in
The problem shown in Fig. 5.6.7 (a) is a linearly separable problem and can be solved using
Fig. 5.6.7 (b).
xy
Wy
! X2 we O
p
;
Fig. 5.6.7 (b) : Network to solve linearly separable problem In Fig. 5.6.7 (a) ;
iy — The region where the output is positive is separated from the region where it is negative by the line,
- b
hy
24,
= W, xy
omW2
~ Here,
net is computed as
n
net = b+ WX
jel
- Here,
net = b+ WyX, + W)%,
~ So, the decision boundary Is
b+ WyxX,+W2x, = 0
. ‘nto two classes-
;
in the above example, there are man diff ;
erent lines that will serve to sepa rate the inpu t vect ors In
ces ofy
there could also be many choi i
Wy, W, and b that gigive exactly the same line.
i .
_ Toun derstand the conc ept of linear separability further, consider simple logic gates AND and OR.
The AND gate can be represented by the truth-table shown in Table 5.6.2.
1 1 1
1 )]-1 -1
-1/ 1 -1
-1|-1 -1
nted as shown in Fig. 5.6.8(a).
_ Then, the desired response for the AND function can be represe
Fig. 5.6.8(a) : AND function plot

shown in Fig. 5.6.8 (b).
ible deci sion bou nda ry for AND function could be as
- The poss
Decision
boundary
decision boundary
Fig. 5.6.8 (b) : AND function
separable.
~ Hence it is linearly is clear from the Fig. 5.6.9.
is als o lin ear! y separable which
Similarly, the O R function Xo
x4
Decision
boundary
ution
Fig. 5.6.9: OR Function sol
f Tech Knowledge,
Peoiter ere
a Artificial Neural Network
— The examples that we have seen are in 2D space. mo an that thay ;

. le, we me 1 can-be
When we say that a set of points in two dimensional space are linearly separable,
:
—
separated by a straight line. :
. , le if they can be separated b y
- More generally, a set of points in n-dimensional space are said to be linearly separab
_ ahyper plane.
- Alinearly separable problem can be solved by a perceptron network.
:
layered ed P perceptron network.
— However, problems which are non-linearly separable cannot be solved by a single
b. Non-Linearly Separable Patterns

- Tounderstand non-linearly separable patterns, consider the Exclusive-OR problem.
= — 1),
1), iiff t the vector has ev en
- Here, the task of the network is to classify a binary input vector to class 1 (output
number of 1s, or assign it to class 2 (output = 1) otherwise.
Fig. 5.6.10 : EX-OR function plot
— Fig. 5.6.10 shows EX-OR function plot. The EX-OR is nonlinearly separable since there is no line that can partition
: the 2D input space into two regions (class1 and class2). This can easily be observed from the plot given in
Fig. 5.6.10.
— Fig. 5.6.11 shows examples of linearly separable and non-linearly separable patterns.
Linearly Separable/Example Nonlinearly Separable Example
Fig. 5.6.11 : Linearly separable and non-linearly separable Patterns
5.6.2 Delta Learning Rule

— The problem with perceptron learning is that many times it is unable to solve non-linearly separable problems. The
initial solution to such problems was to use multiple layer perceptron. P
— In such a network, for the perceptron In the first layer, the input comes fram the actual
inputs of the problem, while
for the perceptron in the second layer, the inputs are outputs of the first layer,
— So, the perceptrons of the second layer do not know which of the real inputs of the first layer were on or off. The
actual inputs are effectively masked off from the output units by the intermediate | ayer.
- The problem here is, the hard limiting thresholdin 8 activation function removes the information that is needed.to-
train the network.
Up Techknawledye
Se
0" ea
582 Artificial Noural Network

e sion 0 this problem i is to use non-jilin
to :
ear (continuous ) activation function such as sigmoidal or '
ae adial basis
eo
drow-Hoff introduced the Delt a learning rul oh are tion for
yenice withe multilayer perceptron networks.
Jing
| 8 *
rule that uses non-linear . :
differentiable ac tivation func
mn Delta Learning Model
-pelta learning is a supervised form of learning which uses continuous activation functions.
Fe jearning signal for this rule
is called delta and is defined
as
= [d,-f0W; 014" (WX)
t
The term f (wx) is the derivative of the activation function f(net) computed for,
Set
net = WX
xy
0, = fw)
r (d;-0})
rning model
Fig. 5.6.12: Delta lea
ween 0, and d,.
of least squared error bet
e d from the condition
- The learning rule can be readily deriv
uared error defined as
with respect to W, of the sq
Calculating the gradient vector -.(5.6.1)
B =4(d-o)
_ Which is equivalent to 4 [a- (WX) ] . ...(5.6.2)

E
ul
ch is obtained by taking partial derivative of E with respect to each of

e which!
ctor valu
> We obtain the error gf adient ve
the weights. gp = - (dr) f’ (wx) x (5.6.3)
N
I jent vec tor are ;

The components of the gradien OE --( d= o) f (Wx) X; 3j= 1,2,...,n ... (5.6.4)
aw; .
take
the negative gradient direction, we
~ Since the minimization of the error red vires the weight changes to bein
aw _ VE .. (5.6.5)
; t
. ~
Where Nis a positive constant aw =n de o)f (WX) x : »+.(5.6.6)

caiman i Wey. TechKnowledaa
AI&SC (MU-Sem. 7-Com P)
For a single weight, the adjustment becomes.
Aw, = 1(d,-0) f’ (net) x35 =1,2,3, 2G

Thus,
AW, = c(d-o) f (net)X

where c = 7 (learning constant)
— The initial weights can be any random values.
— This rule is sometimes also called continuous perceptron learning rule.
5.6.2(B) Proofs
ous or unipolar continuous, and further assuming that

— Depending on whether the activation function is bipolar continu
of the output f(net) or o.
2 = 1, we can express f’ (net) in terms
on, we have,
-— For bipolar continuous activation functi
f(net) = 3 A -0%) (5.6.8)

|
tion function.
— Similarly, for unipolar continuous activa
inet) = o(t-o) 16.6.9)

(5.6.9)
— Now, we will prove results (5.6.8) and
Proofs
(I) f’ (net) (1- 0’) ...[Bipolar continuous]
Proof:
2
(net) 1 + exp(—Anet) — .
Assume A = 1
3 _ 2
0 = f(net) = TT exp(cnet) -1
LHS = f (net)
- d { 2 d 2°
d(net) (1 + exp(— net) — 1 } = d(net) iz
exp(— = me :
_ d 1
= 2 d(net) iz exp(— =a |
‘S
ele Tachiinewledy’
a
5-34 Artificial Neural Network ~
d
2 ; d(nety (1 + exp(- =|
=|)an.
(1 + exp(— net))
— —2exp(— net
(1 +exp(- net))
LHS = —2exp( net ..(5.6.10)

(1 + exp(— net))
1
R.H.S = 2 c= 0’)
1 2
al
= Trace
“ (- net)— i} ]
1 ;
=3 Stree
+ aor net) — 1|
-14
272
241]
¢ + =e ned)? 2) “+ exp net)) ~
2
"(+ nt net) 1 +exp(- net)
[- 14d tener)
= 2
(1 + exp(- net))
2 exp(— net ...(5.6.11)
~ (1 +exp(- net)
(5.6.11)
From Equations (5.6.10) and
LH.S = R.HS.
i _ )
Hence proved f (net) => (1-0

vf
(I) f’(net) = o(1 - 0) --- (Unipolar continuous)

;
Proof f(net)
WW
1
es
1 + exp(- Anet)
u
= 1
Assume ©
0 f(net)
W
1 + exp(- net)
At}
f’ (net)
L.H.S
u
” ae
d ——
+ exp(- =|
(- net))
ii
! Gnet) adi
* a+ exp(-net))
(—exp\= net
wat
—
= a + expe net))
° rtificial NeuralN rk
GF _AisSc (MU-Sem.7-Comp) 535 Eee
_ . LHS... = exp(— »(— ne net (5.6.
(5.6.12) — - ae
(1 + exp(- net))
R.H.S. = o(1-0)
__1_f,_-—_ 1
|
net) [ 1—T+ exp(- net)
= 1 + exp(-
1 + exp(— net) -1
(1 + exp(— net))
_ — expe net -» (5.6.13)
(1 + exp(— net))
From Equations (5.6.12) and (5.6.13), -
LH.S. = R.H.S.
Hence proved, f (net) = 0(1 - 0)
5.6.2(C) Algorithm-Delta Learning
SCPTA-Single Continuous Perception Training Algorithm

Given are P training pairs {X,, di, Xz,d,....X,, do}
Where Xs (n x 1) and djis (1 x 1)
i = 1,2,3...,P
Here, augmented input is used,

.
X;
,P
eet fori=1,2,...
Xi = [
- In the following, k denotes the no. of training steps and p denotes the no. of steps within a training cycle
Enax> 0 chosen .
Step1: n>1, A=1,
Step 2: Weights W are initialized at some small random values.

Wis(n+1)x1
4
Counters an initialized
i
pel, kei, EO
Step 3: The training cycle beings here.
Input is presented and output is computed.
Y<Cy, pe! ded p
oe
f (W'Y)
2
With f(net) = Tre-anet~ 4
* afer Tach Wemwladaa
”¢ AlxSC (MU-Sem. 7-Com _ artificial Neural NeMworS
step 4: Weights are updated
wwe 1/2n (d-o) (1~-0°)y

step 5: Cycle error is computed
Ee£+1/2(d-o)
step 6: lf p< P then
kek +1 |
p <p +1 and go to Step 3
Otherwise, go to Step 7.
hts, kand E.
step 7: The training cycl e i is completed. For E<E,,.,terminate the training session. Output weig
ycle
If E2 Ea, then,
E<0
to Step 3.
p <1 and enter the new training cycle by going
| -5.6.3 Back Propagation Algorithm
|
oe
the 5 help of a fi wehart.
jon taining galgothn wt ‘
‘ (Dec 11, ‘May 12,
ts netwo rks. This learning

in neural
important developmen
n algorithm is one of the most cessing
The back propagatio forward network s consist of pro
- net wor ks. Multi layer feed
orward
t o multilayer feed-f
algorithm is applied on functions.
us di ff er en tiable activati hts
elements with cont
in uo to the process by which the weig
erent from the oth er networks in respect
pr op ag at io n algorithm is diff
- The ba ck
ess.
learning proc la yer in an efficient way that wou ld result in a very
are calculat ed during we ig ht s of the hidden
at in g th e
rception calcul
- Forthe multilayer Pe
ng task.
error isa challengi put layer. But at the hidden lay
er
small or zer o out put
act ual and desired output at the out
rror between the er perception is
e a s u r e the @ Th er ef or e, ad ju sting th e weights of hidden lay
Wecanea sily m lable.
-
io 1 ° f the error aval
n
ct in fo rm at
there is no dire ights at hidden
on for updating the we
providing the equati
difficult.
this problem by
algorithm so
Ives tput err or is minimized.
op ag at io n rs su ch that the overall ou
~The back pr ula at output
la ye
ht upda te for m ases:
i layer along with weig tains two ph
jthm con
algor
-> The back-propagation
d ror.
d recall an tion of the er
1. Feedforwar th e b a ck-P ropaga
that c alculates
; 2. Feedback P' hase 3
TechKne
‘
a . PRaeer:
e
sejatiyee e
———
'_Al&SC (MU-Sem. 7-Comp) 5-97 Artificial Neural Network
5.6.3(A) Architecture of Back Propagation Network
Consider the layered feed forward neural network with two continuous perceptron layers shown in Fig: 5.6.13,
Layer| of neurons Layer k of neurons
manna
OK
\ 7-1 SJ
°K
zT
yy=-1
(Fixed input Dummy _ (Fixed input)
z;=-1) neuron
ith th
th j column k” column
i column
of input nodes of hidden nodes of output nodes
Fig. 5.6.13 : Back propagation network architecture
In the above network, | input neurons are connected to each of the J” hidden layer neuron.
— Weights connecting the input layer to the hidden layer are represented by matrix V. Similarly, ’
outputs of J hidden
neurons are connected to each of the K output neurons, with weight matrix W.
— EBPTA uses supervised learning mode, therefore the training patterns Z should be arranged in pairs with desired
response d provide by the teacher.
Feed forward recall phase

— " Inthis phase, the input pattern vectors are presented and mapped to the output vector O as follows :
oO =T (WI [VZ]]
where, I [VZ] Is the internal mapping and relates to the hidden layer mapping 2 > Y
— The operator T is a non-linear diagonal operator.
= Vand Ware adjusted so that error value proportional to | |d—-o| lis minimized,
Al&SC (MU-Sem. 7- Comp)
igo! back propagation training

: g
on
a rt Error Back propagati
Fig. 5.6.15 : Flowch
Tech Knowledge
WF Teck Knomioays
__Artificial Neural Network:
6-39 oes
- Consider the flowchart for EBPTA shown in Fig. 5.6.15.
ing begin s with the feed forwa rd recall phase (Step 2). After a single pattern vector Z is submitted at the
— The train
input, the layers response Y and O are computed.
— Then the error signal computation phase (Step 4) follows. The error signal vector must be datertnined! in the output
input nodes.
layer first and then it is propagated back towards the network
— Next, KxJ weights are subsequently adjusted within the matrix w (Step 5).
— Finally J x1 weights are adjusted within the matrix V (Step 6).

ed (in Step 3)
— The cumulative cycle error of input to output mapping is comput
calculated after each completed pass through the training
— The final error value for the entire training cycle is
set {2,, Z,, Z3, .---Zp}
than E,pax is obtained (Step 8)
— The learning stops when the final error value less
5.6.3(B) Algorithm (Error Back Propagation Training)
— Given are P training pairs

{Z,, dy, Zy, dy... Zp» dy}
(— 1) since input vector have
..., P. The i" component of each Z, is of value
Where Z, is (Ix 1), d; is (1x 1) and i= 1, 2, 3,
been augmented.
of value (- 1),
output Y. Note that the "component of Y is also
There are total J — Jneurons in hidden layer having
also been augmented.
since hidden layer outputs have
1)
— — Thus Yis (J x 1) and O is (K x
Step1: n>OandE,,.. chosen.
at small random values.
Weights W and V are initialized
Wis (K XJ), Vis UX)
qe1pe1LEO
Step 2: Training step starts here

outputs computed.
Input is presented and layer’s
Z <Z, d <d,
y, f(V,Z), forj= 1,2,
it row of vector V and
where V,, a column vector, is the
o, MWY), fork=1,2,-4K
Is the K” row of W.
Where W, is a column vector, and
Step 3: Error value is computed
E <4 (,-0,) +E, fork=1,2,3,...K

layers are computed.
Step 4: Error signal 5, and 6, of both the
a (MU-Sem. 7-¢1) _ Artificial Neural Network
ee ') .
Vector 6,, is (K x 1), dis (1x 1)
1 .
6, ok =7>2 (4k - 0)
(1~0F) fork=1,2
, F1,2,0,K
I| and 6,yi => (1
| | 7y, 2j x 8 okWy, forjj =1,2,..J
Step 5: Output layer weights are adjusted

Wi:
Kj Wij + N5,, Yj
fork = 1, 2,3,...K
updated j=1,2,3,..,J
Step 6: Hidden layer weights are
i Veji Vv.ji +76..
n yj
z.
Zi, for j =1, 2, 3,...,J
i=1, 2,3,...,1
Step7: | Ifp<Pth
p ' “a P <p +1, q —q +1 and go to step 2, “other
otherwise go to steps.
Step8: The training cycle is completed.
For E<E,,,, terminate the training and
Output weights W, V, q and E.
step 2.
IfE>E max?
then E — 0, p< 1, and initiate the new training cycle by going to
Organizing Maps
5.7 Un-Supervise d Learning Algorithm: Self
Forse Oey eon PEW a ee
Self-Organizing Maps: d a two-dimensional map

of a problem space.
sed lea rni ng to buil
(SOM) uses unsupervi nt
- ASelf-Organizing Map bac kpropagation with gradie
learning (which is used in
tit ive lea rni ng as opp osed to error-correction
a It uses CO mpe
.
. aona hexagonal or rectangular grid.
descent)
: enerate a visual representation of dat
- Aself-organizing maP cané ouping similar dat

a together.
co ncepts by gr and gas exploration.
esents clustering ph y, pr oj ect prioritization, and oil
~ SOM also repr cean og ra
e meteorology; © ur' ¢ Map (SOFM)
or a Kohonen map.
“7 Applications incl ud a Se lf -O rganizing Feat
as
known
~ As el f- or ga ni zi ng ma P i s also |
: g maps, are used for
organizing maps
Kohonen’s Self
s fea tur e map s or topology preservin
honen
Iso called Ko
ing networks, 4
erty’. .
_~ Kohonens’s self orga? th at a:certain topological prop
on the output unit such
constrai
ishbourhood
nt
°
»Sata clustering:
~ Networks of this type '™
iS reflec
in the innut.data
541 autficial Neural Networks
¥ Al&SC (MU-Sem. 7-Comp)
- Fig. 5.7.1 presents relatively simple Kohonen’s self organizing network with two inputs and 49 outputs.
\SSS
ESS
Input —»X, Xp
units
Fig. 5.7.1 : Kohonen’sself organizing network with two Inputs and 49 outputs
:
foo 0 60 0 0 0}
Pio
o pasencnnanonawnnnnmennacn
fo0 0 0 GO O;0 i
' NO eee me ee 1 1
‘olotfe o ofoio}
poy ro off
fo }oj;o % FO} 0} 0}
'
! . 1 1
Lotoio o ofoio!}
Pop
t '
Tmrenencegeces "4
'
a Re
!o 0
10 OO 1 8o}]o
8880 of i
NB, (t= 2)
NB, (t = 1)
NB, (t= 0)
Fig. 5.7.2 : Nelghbourhood of unit C
eas
The learning process here is similar to asthat of : competitive learning networks. 5 That at (ori is, Scere
imil ar;
nea
oe to be the one with the largest
measure is selected and the winning unit is considered activation function.
wei
not only the winning unit’s weigh ' :
However, for Kohonen’s feature ‘ maps, ‘ we update ght but also all of the weights in a
;
neighbourhood around the winning unit.
in Fig. 5.7.2
The neighbourhood size generally decreases slowly with each iteration, as shown —
neuron, we can also use a nei
— Instead of defining a neighbourhood of a winning neighbourhood function Qc (i) around a.
winning unit c.
Where,
2
Qe) = en Fe )
20
— Here, o reflects the scope of the neighbourhood and P, and P, are the positions of the output ul nitit// and 7 c respectively:
tively.
Algorithm (Kohonen’s self organization map)

: aj
Step1: Initialize the weights at some random values — set topological neiighborhood rate ees CE f
decreases. Initializ e the learning
progresses, the radius of the neighborhood
©, It should be slowly
decreasing function of time. _—————
ay TechK avwledss
publications
sc (MU-Sem. 7-Comp)
Perform step 3
5 Artificial Neural Network
“Still Stoppni Condition
2: Pping =
. S true,
3; Perform steps 4-6 for
teP:
et each input vector Xx
Compute the sq
uare of the Eucl
id eand istance j ef .
n mm. - for each j = 1 to m (there are m no. of output neurons)
DG) = 2 2 (x; - w,)?

i=l j=1 4
L:
rep 6: i
For all units jj withwithin : :
step a specific neighborhood ofJ and for all i, calculate new weights
Wi = wj(old) +a [x, —w;;(old)]
step 7: Update the learning rate o using formula o (t + 1) = 0.5 a: (t)
step 8: Reduce radius of topological neighborhood at specified time intervals.
step9: Test for stopping condition.
5.8 Solved Problems
Soin. :
n network.
ecture of the give
Fig.P. 5.8.1 shows t he archit
5 Mo
4
6 X20 2 output 0
3
7. %3° 4
x49
of given network
‘to p. 5.8.1 = Architecture
net = WX
5
6
= [1234] 7
= 5412421432
70
"
)=2* net = 2% 70= 140

= f (net
N
o
. . . ni ors are as follows.

Ex. 5.8.2: Use perception learning rule to train the network. The set of input training veo!
1 0 4
-2 15 1
X=] o]| %=]_95| Xs=lo5
-4 -1 -1
The learning, constant c = 0.1. The teacher’s desired response for X,X_ and Xz are qq
=-1, d2=-1 and
d3 = 1 respectively. Calculate the weights after one complete cycle.
Soln.:
For perceptron learning rule,
net; i WwW xX
oO, = sign (net;)
AW, = c(d-o0)X
Step 1: Take first training pair X = X,,d=d,
1
-2
X = o|> d=-l
-1
1
t -2
Compute : net=W, X = [1 -1 0 0.5] 0 =2.5
—o-4
0,= sign (net,) = sign (2.5) = 1
AW, =c(d-0,)X
1
= 01C¢1-1) “3
~1J
= 0.2
_| 04
a 0
0.2
W, = W, +AW,
1] f-0.2 0.8
_|-1 0.4 - 0.6
= 0 + o| = 0
0.5 0. 0.7
Step 2 : Take second training pair,
SetX - =X, d= d,
_Artificia
| Neural Network
Ee W_Al&SC (MU-Sem.7-Comp)
1.5
X =]_o5| 457!
-l
0
— net, = Ww; X=[0.8 -0.60 0.7] | - fe
Compute =-1.6
Let
O,= Sign (net,) = Sign (— 1.6)=- 1
Here d=- 1 ando, =-1

So, (d— 02) = 0
Hence, AW, = 0. Thus there is no change in weight
, 0.8
-0.6
= Ww, = 0
0.7
Step 3: Take third training pair,
, Set X =X,,d=d,
-1
J
X =los{> ¢=1
-1
Compute— net,= w,; x -1

1
= [0.8 -0.6 0 0.7] 95 =-2.1
-l
0,= Sign (net,) = sign (-2.1)=-1

d = land 0o,=-1
Here
AW, = ¢ (d-03)X
-1 - 0,2
1 0.2
= (0.1) -C1) Josl=] on
-l -0.
W, = W,; +A W;
0.8 -0.2 0.6

_|- 0.6 0.2] _|-04
=} oft} otf =] ot
0.7 - 0.2 0.5
neuron network.
training for perceptron learning rule of a’single cea
Dalar nine the weights ‘itor -four steps of eee ah UPS OIGae ae
@
acid with initlal wolghts+
=(0 of, inputs a X= (2 2),
ite Mae sree bsg : Pn eS
si 2), X= 22)
ag eis ac ta ERE nie : .10 M-
13,
U- Dec
4,20, dy2vandonts.
Al&SC (MU-Sem. 7-Comp) ~ Artificial Neural Network
Soln.:
Here, we use unipolar binary activation function because desired values given are 0’s and 1's for perceptron learning,
net, = Ww; xX
0, = f (net)
AW, =c(d-0)X
1 , net20
Here 0, =
O , net;<0
Step1: Take first training pair,
2
set X = X, and d=d> x=[ 3 | andd=0
net; = W, xX |
Here W, is the initial weight, W, = [0 0]
net, = [0 OQ] [3]
=0
Oo; = fO)=+1
WwW, _[ 0
=- wi+aw, =[5]+[ -2 ]=[72]
75
-2
Step 2: Take second training pair,
sett X = XK, and d=d,

1 :
xX = [3] and d=1
‘ 1
net, =W,X=62-21[ 1]
=-2+4=2
o = £(2)=+1
AW, = c-(d-o0,)-X
- ma-v[
_
1,1 Jeo
Hence there is no change in weights.
, -2:
Ws = w.=[ 73]
Step3: Take third training pair
set X = X, and d=d,

-2
net, = WwW; x
Se
SSE SS Ea aan
Al&SC (MU-Sem. 7- twork

5-46 Artificial Neural Ne
= [2-2] LY]
= 4-4=0
03 = f(0)=+1
AW, = c (d-0)-X
wool 3142 ~2
-2 0
2
2
“= Ws + aw} | (3 l4 ‘]
Step4: Take fourth training pair
X =X, and d=d,

:
-1
x =| | and d=1
t -1 .
net, = W,:X=[0 -4| 1 |0-4=-4
“04 = f(-4)=0
AW, =c(d-0,)X
=Ma-9} ATT]
|) |,
W; = W4+ AW,
[E455]
iH
)}—> output 0
Fig. P. 5.8.4
p
Here f (net) = | [Binary
i sigmoid
i i means the unipolar continuous]
+e
Assume 4 =
0.8
- 0.2] a6 == 0,08 + 0.18 — 0.08 = 0.18

Computenet = W' X= [0.1 0.3
.18
Compute o f (net)
1 __1 __. 9.5448
(| =- T+ 0.8352
ite
Soln. :
The T ruth table for a 2 iinput OR function is (assuming

bipolar inputs
| target).
-1/}-1]-1
-1) 41/41
+1/-1)+1
+1) +1) +1
Where x,, x, are the 2 inputs and ‘d’ is the desired output (target)
Perceptran N/W for OR
Fig. P. 5.8.5
Shown in Fig. P. 5.8.5 is a perceptron to represent OR function.
The output ‘O’ of the network is given by

O = sign (x,w, + X,W, + bw,)
Now, let us train the n/w to obtain the OR function.

Assume that the initial weights are w, = 0.5 and w, =—0.5
Assume bias weight w, = 0.5
Further assume bias input, b =1
. Our initial weight vector

0.5
w, = -0.5
0.5
Efile)
And, our training pairs would be:
(ons elifo
Assume c = 1 (learning constant)
TathKnowtedge
Publications
:
Mi&SC (MU-Sem. 7-Comp) Artificial Neural
Network
: steP 1 ‘
Compute net, = w, x;
t . -1
WX, = [0.5 -0.5 0.5]] -1
; +1
= -0.54+05+05 =0.5
0, = sgn (net,) = sgn (0.5)=+1
Now,d, =-1
“AW, = —-2c¢ x,
-1 2
=-2(1)} -1] =| 2
~ L4l -2,
Aw, =w,tAw
0.5 2 2.5
=|-05/4+] 2 |=] 15
0.5 -2. -15
Step 2:
net, = W,%
-1
= [2.5 1.5-1.5]| +1
+1
= -254+15-15
= -25
0, = sgn (net) = —1
d z+1
1 |=| 2
Awa
= 2cx
?
= 2(1)
il l2
WwW; = w, +A W2
2.5 -2 0.5
15 {=| 2] 3.5
=|
1,5 2 0.5
t
net, = W,%3
1
= [0.5 3.5 0.5) 7
2 0.5-3.5 +05 2-25
0, = fnet)= -1
d, = +l
Ee a
WF _AI&SC (MU-Sem. 7-Comp) 46 Artificial Neural Network
1 [=
“AW; 2¢x;=2(1)|-1]=) —2
” of L2
Wg w,+Aw;
0.5 2 2.5
3.5 |=|-2] =| 1.5
0.5 2 2.5
Step 4:
net, W, Xs
+1
(2.5 1.5-2.5] 3
+1
25415425 =6.5
sgn(net,) = sgn (6.5)=+1
+1
No change in weights
ie. Aw,
Ws
Step 5:
net,
-1
[2.5 1.5 28-1]
+1
22.5-154+2.5=-1.5
sgn (net 5) = sgn (— 1.5) = -1
-1
. No change in weights
2.
Aw, We = Ws= 3
0
2.5
Now, after Step 5, we have,

We Ws= W4
and this means that the training
i.e. Weight vector is not changing after 2 consecutive steps. Hence, we stop training
Is
in complete final weight vector after tralning
2.5
WwW 1.5
2.5,
ror ne cree
k
. Artificial Neural Networ
meat Ne
Fig. P. 5.8.5 (a)

Where O = sgn (2.5x, + 1.5x, + 2.5)
——
eee
Ex.5.8.6:. Consider the following set of input training vectors.
_| 15 _| 1
X=!
=| o|
-1
%2=|-05|%=|05- -1
d,=—-1, ce for X,, Xz and X, respectively. Initial weight vector

— 1 and dg= + 1 are desired responses
-1 , aD oe, tome aug RTED

Wi=l oO
0.5.
the final weights. Use bipolar
ning cons tant c = 0.1 and 4 = 1.Use delta learning rule to calculate
The lear
ction.
continuous activation fun
Soin.:
activation function.
:
with bipolar |‘ continuous
’ For delta learning
I ..(1)
net, = W, %
(2)
o, = f(net)= Ty exp (-Anet) ~
i 2
| (3)
: eet) = 2 (2-9) (4)
AW, = ° (d- o)f’ (net,) X
- i
ly andd=d
: 1
Step1:; Take 1° training pair, X =x, and ¢
-2
= . 0 , d = -1
x
-1
t _1 0 0.5) a)
= W, * =(1
Compute— net, -1 =f Rig
2 mate ~1= 9-848
o, = = Tr
flmets)exp (= 2.5) -1=T7 0.0820
1: 2) 2 (4 0,848") = 0.140
f’(net,) = 5 (17%) *2 O co pits 2 RRM
4 *
‘Artificial Neural Network
“Al&SC (MU-Sem. 7-Comp) 551 ;
AD re , is _2
ne aS Fat
— 0.848) (0.14) | 9
c
AW, =c(d-o,) f (net) X=O.1) C1 -1
1 — 0.026
: -2} :| 0.052
= -— 0.026 o|= 0
1 0.026,
WwW, = W, +AW,
1 - 0.026 0.974
1 - 0.052 —0.948
= 0 + 0 = 0
0.5 0.026, 0.526.
Step 2: Take 2™ training pair, X =X, andd= d,

0
1.5 :
X =} gg [dat
-1
0
t 1.5
Compute —net, =W, X=([0.974 -0.948 0 0.526] 05
-1
= -— 1.422 -—0.526
= — 1.948
2
-1=s
O,= finet,) = tap a = 8.0146 ~1 =~ 9.75
f'(net,) = $ (i -0,) =3 (1 -(-0.75)’) = 0.2187
0
AW, = ¢(4~0,)£" (net) X = (0.1) (1 -(0.75)) x (0.2187) | _'¢
-1
0 0
= (- 0.00546)| _5nee ae ‘hee
a 0.00546
W,#AW,= -|-| 094974
8) , ~f008t907 _| -0.£0974
Ws; 956
.00273 |=] 0.0027
L:0.52 0.00546] 0.5315
Step3: Take 3 training pair, X = X, and d=d,
' oy
Compute net, =W, X =[0.974-0.956 0.0027 0.531] 05
x
Artificial Neural Network -

'-0.974 4 - -0,956
+ 0.00135 — 0.
5315 =~ 2.462
2
o,=f(nets)
= f(net.
== Trop (2.462) -1pt
. ~1=-0 843
#(nety)d= = 315 (2-0;3)=5=4 (1-(-0.843)%)Ls
= 0,145
AW; = (0.1) (d-o,) 3) f’f (nets)

X = (0.1) (1-(-0.843)) (0.145)
3:
0.5
-1
-1] [-0.0267
AW, = (0.0267) re =
Coss
-1) L-0.0267
0.974] [-0.0267 0.947
- 0.956 0.0267 -0.929
a = W3+4W3=)
WwW =
9.0027 | +] 0.0133
=
0.016
0.5315 L-0.0267 0.504.
for 4 = 1, ¢ = 0.25. Train the network using

Ex.5.8.7: Perform two training steps of the network using delta learning rule
the following data pairs. -
1 ,
2 =[1 0 4]
(x -| 0| d,=- |(x -|- 2| do=+ ) _ The initial weights are W,
-1 -1
. Soln.:
For delta learning rule : t
net; = W, X
2
95 f(net;)= T+ exp (— net, ) vA
ul
f’ (net) 31-4) |
AW, «a-oi[z(1-°)]
jes
X= d=4d,.
Step1; Take 1° training pa!" X=X,andd=C
x
i}
Jy ot
e enh et ey
tye a) } Of=2-4
0 4) E
|
-eS Compute —» net, = W, X= / 0.463
__ 2 _ 1=
7 1+ exp 1)
il
= 0.393
f (net)
e ; a- (0.463))
2
,
463) (0.393) E
AW,
‘= (0.25)C 1- 0.
samo
2 Artificial Neural Network
|
}
}
| = (-0.1437)} 0] = ao
-l
1 ons)
. W, = WtAW =. |0- ale=)

; re tid £01437 1137
Step2: Take 2™ training pair, X =X, and d =d,
X =s/-2]|,d=1
-1
Bea boty. 1
Compute TP netgn Wa X= (OTS: 1.1437] |-2
= [0. 713-1. 1437) == -0.4307

0, : hans
(net) = 4 (=©0.2119)") Ne— 0.0449)
= 0.477
pen RW,” =! 0.25) (1 - € 0.2119) x (0.477) X,
1 0.14
= (0.145)|-2]=] -0.29
voye Lend L+o.14s ; J
0.713 0.145] 70.858 ,
| W,=W,tAW, = O}4+} 0.29) =! -0.29 . osigk
1.143] L0.145] Lo.998
Soln. : feleQait 2 + ys eyed

(a), Backtracking
3
Step3: W, = EOF (le CIEOb
ES. f
Aw
=
Wy i Wat AW, 3; coe

W = c:(d,—0,)-X,
“a
hh
-AN&SC (MU-Sem.:7-Comp)
work
esvney) Aftificlal Neural Net
We know that the correction
ha S been P perf tformed at every step this means, actual output and desired outputs were
not same.
.
This means they we exactly opposite of each
i ¥ oth er,
Since d, ° =-1>0,=41
d, =+1>0,=-1
d, =-1>0,=+1 sts ?
-2 44
0 0
“AW, a, = =()¢6
(Q)EG1- 1) 3 = 6
é | -4 2
. W, = W,+AW;
» W3 = W,-AW,
3 4 -1
2 0 2
“!6 || 6 |] 0
1 2 -1
Step2: W; = W,+AW, .
AW, = 0(d,-9)Xo
0 0
Voy 2
= (1)(1-C 1) 2 \F 4
-1 -2
i Now W, = W,-AW,
t “ -1 0 l “1
|| =| 2of) |e 21)4} | -44

-1 -2 L 1
A W,=¢ (di 02) % aes 24,

Step 1: W, = w,tAwi and , V7 ' ' 2
-2 #2 [. 4
Pe) Q-C)) 3 =(-2) 3 || -6
: =ah = - :
Wi = W2-AW1
4
: 177277 |
4 4 a
=| .4|7| -6 |] 2
{ ait -!
Puolicatigns
A
‘ ' Artificial Neural Network
oe eee eee 5-55 a

Thus,
1 -1 -1
4 wile 2
0 0
W.= _4 3
W, = 2/
1 -1
-1
(b) Find W,, W, and W,

Step 4: Take first training pair
set X = X,and d=d,
1
-2
x = 3 |. and d=-1-
-1
net, = w,X
1
-2
= B26 1] 3
-1
= 3=44+18-1=16
0, = sgn(net,)=+1
AW, = c(d=0,):X
1 2
1
; 2 4
=
=Me1-D} 4 |-MC2D) |=] _g
4 >
-1
3 -2 1
2 4 6
Ws = Wat AWel 6 |*) _6 IFl

1 2 3
Step5: Take second training pair,

set X =X, and d=d,
0
-1 :
X = 2 =1
-1
0
-1
net, © = W, t X=[1
6 0 3]
-1
0-6+0-3=-9
as
Os = sgn(-9)=-1
AWs =c(d-o,):X
-1 -2
2 { -1 4 4
=(d-C))| 2 =2|
-1 -2
-1
-
oe Ws ‘= W,+AW,
0 1
1
6 -2 4
=|o [7] 4°] | 4
3 -2 1
step6: Take third training pair,.

X =X; and d=d,
set
-2°
d==-1 :
X =|o-, B
-1
net, = WX
-2
1=-14
=-2+0~-12-
=f144 1]
-1
o, = sgn(-14)=-1 -2
1-1) 23 =0
A We = ¢(d-06) X=)
-1
1
4
= Wo= 4
OW
:
_1
1
1
1
4 Ww. 4
6 w =|.
4
'
7=
4
=_
=| 0 6T
Hence Ws 1
3 1
MATS
oxy Publications
““Al&SC.(MU-Sem. 7-Comp) : Artificial Neural Netwat
Soln. :
Desired output is 1.0
Fig. P. 5.8.9(a)
(I) Forward Pass:
Computing forward activations.
Z,)
(i) Computing hidden layer activations (Z, and
netZ) = WipXyt Vy. X2 + Voy
+ 1(1)
= 2( 0) +01
= 041401 —
= 1,1
Z, = f (netz,)
we have
Using bipolar continuous activation function,
m 1 A =e1]
f(net) ©-—
14+” [Assum
ee:
"The™ 7!
Z,= {(netz,)
flac)» eee
Teg Tyg I
2
z, = 0.5005
_
Net =
72 = Va X1+ V2 Xo+ Von
= 10)'+2(1) +03
= 2.3
2
= f(netz)= os
2
(net 2) = Tygt!
; Z, = 0.8178
(i) Computing output layer activations (y)
net, = wW,Z,+ W22)+ Woi

= 04(0.5005) + 0.2 (0.8178) + (-0.4)
- net, = — 0.03624
2
=f y)
y=f(net == [oma - 1
y =-0.0181
(l) Backward Pass :

layers.
Computing error signal vectors for both
for output layer
() Computing error signal terms 5,,
Sy = ; (4,-0)) (1-0;), fork=1,2...-K
‘y’
Since, there is only one neuron
8.1 = 5 (d-y)Q -y)
6, = aC - (-0.0181)) (1-G 0.0181))
5, = 0.5089
;
or sig .
nal ter ms of the hi idden layer (zi)
(i) Computing err 1 2 >
.
Sq. Wid = 1,2, wd
8; 5( = -2,)
:
: k=1
1 2
5 = 3(! -2,) 5, 4
4)
= 4 (1 0.50057) (0.5089) (0.
8, = 0.0763
bn = AC -1,) 5, Wa
= 3 (1 0.8178") (0.5089) (0.2)
by = 0.0169
= Wy ndy Zz (Assume n= »
(iii) Adjusting output lay
Wie
= Wiold + 5,21
|
Cee
AI&SC (MU-Sem. 7-Comp) ____ Artificial Neural Network
= 0.4 + (0.5089) (0.5005)
Ww, | = 0.6547
W. = Wot 5,2,
=. 0.2 + (0.5089) (0.8178)
W, = 0.6162
Bias Wo = woi.8y(1)
= -0.4 +0.5089
Wo, = 0.1089
(iv) Adjusting hidden layer weights (Vas Vage Voge Yoo)
vi = vgtn8,xiforj=1,2 Tandi=1,2.. wl
Vip = Vu + 8y1%,= 2+ 0.0763 (0)= 2

Vig 2 °Vyq+ 8, Xp= 1 + 0.0763(1) = 1.0763
Vo, = Voy + 8,2 X= 1 + 00169 (0) =1
Vag = Vogt ene 2.0169
Yo = 1 (1)= 0.1 + 0.0763(1)= 0. 1763
Vor = Vont+ 5,2 (1) = 0.3 + 0.0169(1) = 0.3169

1 = 0.5183
CycleerrorE = %(d-y)*+E= 5 (I —(-0.0181))’
Final weights: wi = 0.6547, w,= 0.6162, wol= 0.1089 , v,, = 2, Vj.= 1.0763
Voy =. 1, ¥q2= 2.0169, Vo; = 0.1763, Vo = 0.3169
with examples.
Q.1 Explain linearly separable and non-linearly separable patterns
the help of flowchart.
Q.2 Explain Error back Propagation Algorithm with
network.
Q.3 Write a short note on Kohonen self organizing
of an example.
Q.4 Explain McCulloch Pitts Neuron Model with help
on function is linear, that is, the activation
Q.5 Aneuron with 4 inputs has the weight vector w = [1, 2, 3, ay. The activati
7, a} then find the output of the neuron.
function is given byf (net) = 2% net. If the input vector is X = [5, 6,
help of a flowchart.
Q.6 Explain Error back propagation training algorithm with the
trained using the pairs of (X,, d) as given below :
Q.7 A single neuron network using f (net) = sgn (net) has been
X,=[1-23-1]},d)=-1
X,=[0-12-1],d,=1
=[-20-3-1],d,=-1
rule are ;
The final weights’ obtained using the perceptron
i

5-60 Artificial Neural Network
W,=[326 1]
Knowing that correctj on has been . .
;
Perform edi
c = 1, dete rmin e the foll owin g weights :
the traini in each step for
(a) Wa Wa, W, by backtracking
(b) Ws, We, W; obtained for st
eps 4; a in
(Xq, dy), (X2,°d,) (X3; d3) Segre ee
g.8 Explain with example perceptron learning rule
with initial weights
Q.9 Determin iterati for hebbian learning of a single neuron | network starting
ett, - e ty theinp weights aft €F one iteration
; puts as X, =[1, - 2], X, = [2, 3], X=[1, - 1]ande=1.
Use (i) Bipolar binary activation function

(ii) Bipolar continuous activation function.
Q. 10° Explain Single Continuous Perceptron Training Algorithm (SCPTA).

learning.
Q.11 Explain with neat diagram supervised and unsupervised
is binary sigmoidal activation
has the weight vector w = [0.1 0.3- 0.2]. The activation function
Q. 12 A neuron with 3 inputs
0.6 0.4] then find the output of neuron.
function. If input vector is [0.8
sed learning.
supervised and unsupervi
Q.13 Differentiate between
network.
self organizing neural
Q. 14. Explain Kohonen’s le neuron network starting
with
tra ini n g for per cen ptron learning rule of a sing
hts after four st eps of
Q.15 Determine the weig
initial weights:
=[0 oj, inputs as X1== [22],
X21 1
=[1 -2], Xs
=[-22);,
o =1.
, d 3 = 0, dds ,== 1and
no d,=1
ee
ties :
llowing identi
Q.16 Prove the fo tivation fu
n t i n u o u s ac
r co
(ii) For unipola
;
:
function |
nuo us 2°tivation
laroc onti
|
(jiii) Foare bii poo

n vise d and unsupervi
sed learning.
iate between super
_O0
) k s? Different
{' (net) a=2 l
née 0
l e a r n i n g in neural noo
What is
OF Tetk newietel
eo eae . Pubtications
Expert System
6.1 Hybrid Approach - Fuzzy Neural

Systems
. ;
ert s' . : ch
6.2 E
xp ystem : Introduction, Characteristics, Architecture, Stages in the development of expert system,
6.1 Hybrid Approach

6.1.1 Introduction to Hybrid Systems
- . Hybrid systems are those for which more than one soft computing technique is integrated to solve a real-world
problem.
Neural networks, fuzzy logic and genetic algorithms are soft computing techniques which have been inspired by
biological computational processes.
Each of these technologies has its own advantages and limitations. For example, while neural networks are good at
recognizing patterns, they are not good at explaining how they reach their decisions.
Fuzzy logic systems, which can reason with imprecise information, are good at explaining their decisions but they
cannot automatically acquire the rules they use to make those decisions.
These limitations have been a central driving force behind the creation of intelligent hybrid systems where.two or
more techniques are combined in a manner that overcomes the limitations of individual techniques.
6.1.2 Types of Hybrid Systems
Hybrid systems can be classified as:

4. Sequential hybrids
2. Auxiliary hybrids
3. Embedded hybrids
1. Sequential hybrid systems

ologies are used in a pipe-line fashion. The output of one technology
In sequential hybrid systems, al | the techn
gy (Fig. 6.1.1). .
becomes the input to another technolo
rate different technologies into a single
one of its weakes t, because It does not Integ
This kind of hybridization form Is
unit.
dg Al&SC (MU-Sem, 7-Comp)
Outputs
Fig.6.1.1: A’sequentlal hybrid system
2. Auxiliary hybrid systems
. . whatever
In this type, a technology treats another technology as a “subroutine” and calls it to process OF generate
information is needed by it. Fig: 6.1.2 illustrates the auxiliary hybrid system.
Inputs
; cr
!
|' a 1
i) !'
rn ee
.
. 3!
In 1
\
\
ye
<— “Technology
| 13
i It} !
\ Technology
B |
1 J
' !
1. '
7 ‘
\
!
aeee nae 1
See!! ae aeeon eee
leee-eee
Outputs
Fig. 6.1.2 : An auxiliary system: .
ro = genetic system In whichecturan'NN employs a GA to optimize its structural parameters, ie.

~ An example is a neu Network
s arc hit t ure.’
Neural
de fine
i whic h
P ‘amet ers
®
3. Embedded hybrid systom
be intertwined,
5 ys
In embedded hybrie id system? the technlaet negates 7 such a manner that they appear to
~
lete that it would appear that no technologyem.can be used without the other for solving the
~ The fusion is so comP embedded hybrid
syst
hema for an
Fig . 6.1 - 3 depicts the sc
problem.
N-FL (Neural Network Fuzzy Logic) hybrid system that has NN which receives information
of this syste isa N outputs as well.
Example
fuzzy : iginahl
processes it a nd generates
WH team nleitgs
AI&SC (MU-Sem. 7-Comp) 6-3 Expert'System
1
Technology A
‘
!
1
!
!
'
le--—
V Technology B
Outputs
Fig. 6.1.3 : An embedded system —

6.1.3 Fuzzy Neural System
Fuzzy neural system is a system with seamless integration of fuzzy logic and neural'networks
While fuzzy logic provides an inference mechanism under cognitive uncertainty, Neural networks offer advantages,
such as learning, adaptation, fault-tolerance, parallelism and generalization.
To enable a system to deal with cognitive uncertainties like humans, we need to incorporate the concept of fuzzy logic
into the neural networks.
The computational process for fuzzy neural systems is as follows.
The first step is to develop a “ fuzzy neuron” based on the.understanding of biological neuronal morphologies
This leads to the following three steps in a fuzzy neural computational process
1. Development of fuzzy neural models.
2. Models of synaptic connections which incorporates fuzziness into neural network
3. Development of learning algorithms, which helps in adjusting the synaptic weights
There are two possible models of fuzzy neural systems:
(a) Model 1:
In this model the output of fuzzy system is fed as an input to the neural networks.
— . The input to the system is linguistic statements. The fuzzy interface block converts these linguistic statements into
an input vector which is then fed to a multilayer neural network. The neural network can be adapted (trained) to
yield desired command outputs or decisions.
Decisions
———>
Perception as
neural inputs a .
one {Neural
outputs)
Linguistic
’ statements
Fig 6.1.4 Fuzzy Neural System : Model 1
war TechKnowledge
EE
‘Expert System
Neural outputs Decisions |
Fig 6.1.5 Fuzzy Neural System : Model 2

empl jloyed as decision-
- Here, neural net works are used to tune membership functions of fuzzy systems that are
process and
s for
ing systemred
mak tantially i ning techniques can automate this
ork lear
subs controlling equipment. Neural netw
g performance.
y reduce development time and cost while improvin
6.1.4 Adaptive Neuro - Fuzzy Inference System (ANFIS)

“VUE cian oe Lk)
ained
Neuro —fuzzy Inference system) expl
pra cti c ally used Neur o-Fuzzy system is ANFIS (Adaptive
- Awell —known and rence system developed in
1990s.
Takagi-Sugeno fuzzy infe |
below. It is based on len t to a fuzzy inference model.
ona lly equ iva
or k which is functi
- ANFIS is a neural netw ential to capture the benefits of both in a sin
gle
networks and fuzzy logic, it has t he pot
es both neura |
- Since it integrat
framework.
: n in Fig. 6.1.7.
ANFIS Architecture iva len t ANF IS architecture is show
and its equ
Sugeno model
shows the first-order
- 6.1.6.6 sno
Fig.Fig. 6.1
f, = pxt nyt 4
h
. wy fit 2 =w, f+ Wo
=> f -” Wi +W
sah 2 oops noes

f, = poxt nyt” .
:
— hs
MVALLLA------- PRs rir
BA. "1
\e
x - —> y
pe Loi
“PPT RET We
<a
et 3 y
; 16:4 two-Iinput fl rst-order Sugeno model
Ag.
1.6 1 A ay panes
Layer 1
Layer 4
- Fig. 6.1.7 : Equivalent ANFIS architecture

a
Assume that the fuzzy inference system under consideration :
has two inputs x and y and or ne output.
p
For a first-order Sugeno fuzzy model, a common rule set
with two fuzzy rules is,
Rule 1: If x is A, and y is B, then f, = p.x+q.y
+r,
Rule 2: If xis A, and y is B, then f, = p,x+a,y +f,
The architecture contains five layers.
Layer1 : Fuzzification layer

Neurons in this layer represent fuzzy sets used in the anteced
ents of fuzzy rules. A neuron in the fuzzification layer
receives a crisp input and determines the degree to which this
input belongs to the neuron’s fuzzy set.
Every node in this layer is an adaptive node and a node
function can be represented as,
QO, = Hy; (x), fori=1,2or
O1j; = Upi-2(y), fori= 3,4

Where x (or y) are the input to the node
i and A; (or B;.2) isa linguistic label associated
with this node. °
Any appropriate membership function can be used
such as generalized bell function
1
Ba(x) = X=¢, 2b
1+
where, {a, b, c i} is the parameter set. Parameters

in this layer are called Premise Paramete
rs.
Layer 2 : Fuzzy rule layer
Every node in this layer is the fixed node.
Each node (neuron) in this layer corresponds

to a single fuzzy rule, A
fuzzy rule neuron receives inputs
fuzzification layer neurons that repres from the
ent fuzzy sets in the rule antecedents
The output of i” node is the product of all
incoming signals.
07 = Wi= by (x) Mai (y) i 1,2

Each node output represents the firin
g strength (W,) of a rule,
The number of neurons (or node
s) In this layer will be equal to
the number of rules
Tech Knowledge
Publications
: Al&SC (MU-Sem. 7-Comp) . Expert System
. ‘Layer 3: Normalization | ayer 6-6 eS
Every node in this layer j th ’s firing strength to

'S a fixed ith io of the? rule's"
the sum of all rules’ firing senate The i node of the layer calculates the rat
: s.
The output of this la

yer js the normalized fir
ing strength.
Layer 4: Output
membership layer
si
Nodes in this layer represent fuzzy sets used in the consequent of fuzzy rules.
represented as,
a « Everyjnode iq] this layer is an adaptive node with a node function
O,
= Ww, f=W(pxtaytn>
ai i=l,2
where w, is the normalized firing strength from layer 3 and {p, q, 1} is the parameter set of this node.
s this layer are referred to as consequent parameters.

Parameterin
Layer 5 : Defuzzification layer
- Node in this layer is a fixed node.
ppped by
o-fuzzy system. It takes the output fuzzy sets clip
- Every node in this layer represents a single output of the neur
a single fuzzy set.
the respective integrated firing strengths and combines them into
- Output of this node can be represented as,
O,, =
= Lwf= Sit
yw, i=1,220
cture is shown Fig. 6.1.8.

ork is not unique. The alternative stru
7 The e structure of this adaptive netw
Fig. 21.0 © alternative

6.1.8: ANFIS architecture for Sugeno fuzzy model
layer.
n jis performed in the last
normalizatio
‘ .
Here weight
=
6.2 Expert System

6.2.1 Introduction follows:
aefin et Ex pert Systems as
has
~ Jackson (1999) ter program that represents and reasons with knowledge of some specialist subject with
~ “nn expert syste™
proble
a view to ee eee—™” : Wp TeaKneateas
Publications
Al&SC (MU-Sem. 7-Comp) 6-7 ae System
Artificial intelligent aims to implement intelligence in machines by developing computer programs that exhibit
intelligent behavior. To develop systems which are capable of solving complex problems, it needs efficient access to
substantial domain knowledge, a good reasoning mechanism and an effective and efficient way of representing
, knowledge and inferences; to apply the knowledge to the problems they are supposed to solve. Expert systems also
need to explain to the users how they have reached their decisions.
Expert systems are generally based on the way of knowledge representation, production rule formation, searching
mechanism and so on. Usually, to build an expert system, system’s shell is used, which is an existing knowledge
independent framework, into which domain knowledge can be added to produce a working expert system. This avoids
programming for each new system from scratch.
Often, the two terms, Expert Systems (ES) and Intelligent Knowledge Based Systems (IKBS), are used synonymously,
Expert systems are programs whose knowledge base contains the knowledge used by human experts, in place of
knowledge gathered from textbooks or non-experts. Expert systems have the most widespread areas of artificial -
intelligence application.
6.2.2 Characteristics of Expert Systems
Following characteristics distinguish expert systems from conventional computer applications.
Simulation of human reasoning
Expert systems simulate human reasoning process in the given problem domain, whereas computer applications try to
simulate the domain itself.
Representation of human knowledge
Expert systems use various methods to represent the domain knowledge gathered from human expert. It performs
reasoning over representations of human knowledge, in addition to numerical calculations or data retrieval. In order
to do this, expert systems have corresponding distinct modules
’
referred to as the inference engine and the
knowledge base. Whereas in case of computer applications it might be just the calculations performed on available
data, without inference knowledge.
Use approximations .
Expert systems tend to solve the problems using heuristics or approximate methods or probabilistic methods which
are very much like how human do in general. While in case of computer applications strict algorithms are follows to
produce solutions, most of the times which do not guarantee the result to be correct or optimal.
Provide explanations
Expert systems usually need to provide explanations and justifications of their solutions or recommendations in order
to make user understand their reasoning process to produce a particular solution. This type of behavior is hardy
observed in case of computer applications.
Some typical tasks performed by existing expert system are as follows:
Data interpretation : There are different types of data to be interpreted by expert system, which have various formats
and features. Example: sonar data, geophysical measurements.
Diagnosis of malfunctions: While collecting data from machines or from | experts there can be shortfall of accuracy or
mistakes in readings. Example equipment faults or human diseases.
p_DSE (MU-Sem. 7-Compy tem
Expert Syste a
: 6-8
_- Structural analysis . If syst =<
mM is buj : unds, computer
systems are used; conf
| uration
Uild for a domain where complex objects like, chemical compo
of
- Planning : Expert these complex objects must be studied by the expert system.
Syste . .
task. Example: actions
Quired to plan sequences of actions in order to perform some
. ms are requi
that might be perform €d by robots,

~ Prediction : Ey, :
* S*Pert syster
to predict the future basis on past knowledge and current information available.
need
Example: weather fo recast,ms exchange
rates, share prices, etc
High Performanc © : Expert syste :
process huge amount of data toMs are 8 generallally preferred because of theirir hihigh perform ance. In the sense, they can
p le amount of
s, in in acceptab time. The
; is very less dering several detailils,
Produce results consiideri
response time
:
Highly responsive - can ask any query
© + Expert system are required to be highly responsive and user friendly. User
.
system shoul
¥ d be able to produce the appropriate reply to it. Even if the query asked by user is not answerable wih
the existing k
8 knowledgebase, expert system should give some informative reply about the question.
Reliable : Expert systems are highly reliable as they process huge amount of database. Hence the results produce by
the system are always close to exact.
6.2.3 Real Time Expert Systems

There are no fundamental limits on what problem domains an expert system can be built to deal with. Expert systems
ideally be one in
can be developed almost for every domain for which requires human expert. However, the domain should
of the expert systems which have
which an expert requires a few hours to accomplish the task. This section explains some
been created for such domains.
Dendral identifies the molecular
(i) Dendral : This system is considered to be the first expert system in existence.
by Stanford University.
structure of unknown compounds. It was developed
contributions to the medical field;
(ii) Mycin : This is a milestone in expert system development which, made significant
to physicians in the diagnosis and treatment of meningitis
but was not used in actual practice. It provides assistance
d by Stanford University.
and bacterial infections. It was develope
service department
of certain models of Toyota cars, used in a central
(iii) Altrex : It helps to diagnose engine troubles d by their
actually servicing the cars for assistance, if required. It was develope
hich can be called up by those
:
whic! | :
.
.
research lab. of several minerals, includi ;
t
locate deposits
system
f
iis success ul to |
i exp ert
: This
riT
(i iv) ) Pro specto
rnational.
d by SRI Inte
was develope
ngs. It was
time re quired to construct high rise buildi
(v) Predicate : This is expert systent provides esfor
em Corporation the Lease,
te byof Land
timause Australia.
develope
d by Digital Equipm | |
n g B l o c k s of Ex pert Systems
6.3 Build i MU -Dec. 15, May 16
eee (May 16,5 Marks)

=
d the componen ts of expert systems.
~ Let's first understar
WF ennmealedgs
¥ Al&SC (MU-Sem. 7-Comp) 6-9 Expert System
- Following are the two principal components of every expert system.
1. Knowledge base
The reasoning or inference engine
Every expert system is developed for a specific task domain.

It is the area where human intelligence is required. Task
refers to some goal oriented problem solving activity and Domain refers to the area within
which the task is being
performed.
— Consider the expert system architecture in Fig. 6.3.1.
Knowledge ;
Acquisition Knowledge Engineer | Domain Expert
“- Subsyste’ 3
rane >| User Elendy ___,
Explanation : intewace :
- Subsystem ‘ User
4 4 y
Interface Engine Workplace

y
Knowledge Base Rule > Problem... -

Interpreter Description
Problem
Inference : Status .- -
-;Rules. . » -
Fig. 6.3.1 : Expert System Architecture
- _ If we consider inference engine as the brain of the expert systems, then knowledg
e base is the heart. As the heart is
more powerful, the brain can function faster and efficient way. Hence the success
of any expert system is more or less
depends on the quality of knowledgebase it works on.
1. Knowledge base
There are two types of knowledge expert systems can have about the task domain.
| (i) Factual knowledge

(ii) Heuristic knowledge
(i) Factual knowledge :It is the knowledge which is accepted widely as a standard
knowledge. It is available in text
books, research journals and internet. It is generally
accepted and verified by domain experts or resear
chers of
that particular field.
(ii) Heuristic knowledge : It is experiential, judgmental
and may not be approved or acknowledged publica
lly. This
type of knowledge is rarely discussed and is largely
individualistic. It doesn’t have standards for evalua
tion of its
correctness. It is the knowledge of good practice,
good judgment and Probable reasoning in the domai
n. It is the
knowledge that is based on the “art of good guessi
ng.” It is very subjective to the Practitioner’s knowl
edge, and
experience in the respective problem domain.
Expert System
; i nce, the
t uses iis based on his learning from various sources over a time period. Hence,
an expert ; ‘ : : im to
formation j increases with number of years of experience in the given field, which allows hi
in hi s :
2. Inference engine N his databases to advantage in diagnosis, analysis and design.
The infere
nce engine
ha S$ a problem: Solving module which organizes ; he
and controls the steps required to solve t
‘
but powerful rea soning.8
rorwan eng ine inv olv es cha ini ng of IF...THEN rules to form a line of
inf
.. ng Pra . erence
oO f chaini
ctices to solve a proble
m.
concluard Chaining: NE: Thistr; type of reasoning strategy starts from a set: of conditions and moves toward some
n, also called as dat
a driven approach.
Backward Chaini ini
ng : Backward chaining is a goal driven approach. In this type of reason
ing, the conclusion is
nown an ‘ soot
that d the path to the conclusion needs to be found out. For example, a goal state is " given, but the path tomt
at state from start state :is not known, then
backward reasoning is used.
Inf erence engine
ine jis nothing
. but these methods implemented as program modules. Inference engine
i i
manipulates
and uses knowledge ;in the knowledge base to generate .
a line of reasoning.
es Data. > Rules»

x21 ifx=1 and y=2thenz=3,
y=2 ifz=3thena=4
——
Fig. 6.3.2 : Forward Chaining
Data i
x=1 ifx=1andy=2thenz=3,
y=2 ifz=3thena=4
—
Fig. 6.3.3 : Backward Chaining
uncertainty in
Knowledge is most of the times incomplete and uncertain. There are various ways to deal with
knowledge. One of the simplest methods is to associate a weight or a confidence factor with each rule. The set of
uncertain knowledge in combination with uncertain data in the reasoning process is called
m ethods used to represent
intell
artificialsystem
area of “fuzzy
reason ing with uncertainty. “Fuzzy Logic’ is an important called igence which provides methods for
s use it are as s”
i and the system that
i
reasoning i
with unce rtainty
:Most of the times, human experts can guess or use their gut feeling to approximate or
on subsystem
3. Explanation Its. As an expert system try to mimic human thinking, it uses uncertain or heuristic
knowledge as we
estimate the rem cult its credibility of the system is often in question, as is the case with humans. As an answer toa ©
humans do. Asa r =afiondble, one urges to know the rationale or the reasoning behind the answer. If the rationale
problem ee eeptable generally the answer is considered to be correct. So is the case with expert systems!!
seems to be a ’ ;
rt systems have the ability to answer questions of the form: "Why is the answer : X?" Infe rence engine
i
Most of the expeplanations by tracing the line of reasoning used by it.
x
can generate & ;
: Building an expert system is also known as “knowledge engineering” and its practitioners are
4. Knowledge & ngineer
ge engine's The primary job of knowledge engineer is to make sure that the expert system has all the
called knowle ‘ d to solve a problem. He does a vital task of gathering knowledge from domain experts.
knowledg' e neede
WF TechKnowledys
Pooticarions
ett Expert System -
mp)__
¥. Al&SC (MU-Sem. 7-Co
them. Then
Knowledge engineer needs to learn how the domain expert reasons with their knowledg
e by interviewing
the interface engine. There might Be some
he translates his knowledge into programs using which he designs
how to integrate this with
uncertain knowledge involved in the knowledgebase; knowledge engineer needs to decide
be required by the user,
the available knowledgebase. Lastly, he needs to decide upon what kind of explanations will
and according to that he designs the inference levels.
6.4 Development Phases of Expert Systems

Steps in developing an
To develop an expert system following are the general steps followed. It is an iterative process.
expert system include:
1. Identify problem domain

— The problem must be suitable for an expert system to solve it.
— Find the experts in task domain for the ES project.
— _ Establish cost-effectiveness of the system.
2. Design the system

— Identify the ES Technology
Know and establish the degree of integration with the other systems and databases.
Realize how the concepts can represent the domain knowledge best.
3. Develop the prototype

From Knowledge Base: The knowledge engineer works to —
- Acquire domain knowledge from the expert.
Represent it in the form of If-THEN-ELSE rules.
4. Test and refine the prototype
The knowledge engineer uses sample cases to test the prototype for any deficiencies in performance.
—- End users test the prototypes of the ES.
5. Develop and complete the ES

Test and ensure the interaction of the ES with all elements of its environment, including end users, databases,
and other information systems.
- Document the ES project well.
- Train the user to use ES.
6. Maintain the ES
Keep the knowledge base up-to-date by regular review and update.
Cater for new Interfaces with other information systems, as those systems evolve,
6.5 Representing and Using Domain Knowledge
Knowledge affects the development, efficiency, speed, and maintenance of the system. Knowledge representation is a
way to transform human knowledge to machine understandable format. It Is a very challenging task in expert systems,
as the knowledge is very vast, unformatted and most of the times It Is uncertain. Knowledge representation formalizes
and organizes the knowledge required to bulld an expert system.
TechKnowtedgé
Publications
ee
|
. e
s
Expert Sy—
BASS (MU-Sem.7-Comp)
,
4
The knowl iC t in
- tiyfy = ledge should be pake
engineer must
idenntif o
ne or mo re fo rm s in which the required know
the syste m.edgeH © must also ens re oning
as
ith the selecte
:
Y of knowl
ed Be matters, the representation used also matters 4
programm er to write
i the Code for the .
system.
— Anumber of k nowledge-r sent the know ledg e effic ientl y,
have been devised till date to repre
. :
“repres
poir eiae l techn iques on tech niqu es are
Few of the knowledge representati
.
7
but finally it depends on the
mentions below. pplication, the design of system.
Production rules
Decision trees
000000000
Semantic nets
Factor tables
Attribute-value pairs
Frames
Scripts
Logic
Conceptual graphs antic |
are Production Rules, Sem
used met hod s for repr esenting domain knowledge
monly
— Out of these the most com |
}
l study of these methods.

Nets and Frames. Let’s have detai
a. Production Rules tree. A rule has a

is the set of pro duc tion rules, or simply rule
entation of knowledge a set of
One widely used repres by keyword “IF”. It lists
—
wit h it. The con dit ion part is i identified
ion associated
condition and an act in “THEN” part.
s are specified
ical combination. Action en. The piece of knowle
dge
conditions in some log THE N part actions can be tak
; con seq uen tly , the wle dge is
of th e rule is satisf
ied whose kno
As the IF part to pro duc e the line of reasoning. Expert systems
—
ted by the
production rule is used We have studied rule-b
ased agents named as sim
ple
represen ru le-based systems.
d in rul e form are called
represen te
del is
in chapter 1. on basis, . this mo
reflex agents co nc lu si on or condition—acti
uation
evolved on th e basis of sit
n g is
As Ss human thinki esentin g kn ow le dg e in ES.
y used repr resent expert’s knowledge
. It takes following form.
predominantl d effi cien t way to rep
th e simplest an
IF-THEN rules are -
bn
Ag e? a, THEN by, by Dy +
IF ay, Ag r an
conditions
ar e th e situation oF
4 ay
Where ere ay ay ns or action
s.
th e co nclusioio
are
by, By +
Consider the foll
evel count OF ©
NTIs
lity , nt jek
por resetabl pouiads
&
ENT.
stability of ELEM EMENT greate
r than2
jal level count oF FL
TechKnewledge
Puatications
¥ Al&SC (MU-Sem. 7-Comp) 6-13 Expert System
b. Semantic Nets
- Asemantic net or semantic network is a knowledge representation technique used for propositional information,
so it is also called a propositional net. In semantic networks the knowledge is represented as objects and
relationships between objects. They are two dimensional representations of knowledge. It conveys meaning.
Relationships provide the basic structure for organizing knowledge. It uses graphical notations to draw the
networks, Mathematically a semantic net can be defined as a labelled directed graph. As nodes are associated
with other nodes semantic nets are also referred to as associative nets.
—- Semantic nets consist of nodes, links and link labels. Objects are denoted by nodes of the graph while the links
indicate relations among the objects. Nodes can appear as circles or ellipses or rectangles to represent objects
such as physical objects, concepts or situations. Links are drawn as arrows to express the relationships between
objects, and link labels specify specifications of relationships.
— The two nodes connected to each other via a link are related to each other. The relationships can be of two types:
“IS-A” relationship or “HAS” relationship. IS-A relationship stands for one object being “part of” the other related
object. And “HAS” relationship indicates one object “consists of” the other related object. These relationships are
nothing but the super class subclass relationships. It is assumed that all members of a subclass will inherit all the
properties of their super classes. That’s how semantic network allows efficient representation of inheritance
reasoning.
‘Mammals
SubsetOf
SubsetOf
MemberOf
SisterOf Legs
Fig. 6.5.1:Semantic net example
— For example Fig..6.5.1 is showing an instance of a semantic net. In the Fig. 6.5.1 all the objects are within
rectangles and connected using labeled arcs. The links are given labels according to the relationships. This makes
the network more readable and conveys more information about all-the related objects. For example, the
“Member Of” link between Jill and Female Persons indicates that Jill belongs to the category of Female Persons.
— It does also indicate the inheritance among the related objects. Like, Jill inherits the property of having two legs
as she belongs to the category of Female Persons which in turn belongs to the category of Persons which has a
boxed Legs link with value 2.
re than one
Seje
ob mact
ntsntii csn net S als
oO can re presen oug h whi ch, an obj ect can belong to mo
} a inheritance thr mmon form of
d an object can bea ste obj ect s. It doe s also allows a co
much easier to answer
e 2 more than one anothe r
inferen ce
[i the job of inference algorithms
rse
knieows n as inve se links. The inverse lin ks ma ke
rev erse quer
a is a query “such as
For exam ple, in Fi pia s inv ers e of Sister Of link. If there
there ine IS- A HAS Sister link which is the of, to make the
t h e s i t ig.
Jac6.5
k .1?” The Has Sist er is the inverse of Sister
who he sister of cover that
rence system will dis the query.
inference algorithm f from Jack to Jill and answer
ollow the link HAS Sister
A dvantages of Semantic Nets
it can
using semantic nets. For example,
1. Semanti Ic nets are very exp res si ‘ ute detail. s can be represented
min by explicitly mentionin
g as
ues for siv e and be rep res ent ed
val ‘ 6.5.1, Jack has one leg can
represent default c i . In Fig.
ategories
person object.
shown. . Thi This specified v: a
the default value 2 for number of legs for
lue replaces
an nets canre present semantics
Semantic manti of relationships in a transparent manner.
pwn
m nets are simple and easy to understand.
Semanti
ic net s are eas y to im plement using PROLOG.
ant
involved
Disadvantage of Semantic Nets o objects are
re la ti on s wh er e more than tw Ba ng alore,
ent only binary relation s.
In i,
obj ect s rep res ai Express, Chenna
1. The links betw ee n the tence Run (Ch enn
For example, t he sen
be rep res ent ed using semantic nets.
cann ot
resented directly.
Today) cannot be rep
link names.
rd definition of
There is no standa
situations. For
are typ ica l to stereotypical
cts th at decomposed
Frames uc tu re for re presenting obje thi s te ch ni qu e knowledge is
nvenient $ tr physical objects
etc. In
Frames provide a co e of comp lex
— structur language
sual scen es, sion and natural
example, vi at. ns in cl ud in g vi
odular form igence applicatio
into highly m in ma n Y Artificial Intell kn ow le dge.
of sc he ma used pr es en ti ng commonsens e
tween : concepts
, and also
ty pe re be
is regarded as 3-D
is 2 eful fo r ts, rela ti on sh ip s
Frame are also us ae utes of concep he nc e
am es at tr ib ur es an d
des to have struct
. Fr
proces sing ts, situations
,
resent coi ncep janshipsp . It allows no :
repi™ n their relations ; ject and
_ As frames can ex pl ai t of a lis: t of properties of the ob lers. The
ro cedures . nsists slot fil
ow le dg e Ty pi ca ll y, 4 frame co es; also called as slots and
of kn as unit schema, OF lis t. lu So
epresentatio es ; si mi la r to the fields
and va ;
procedures, etc.
e proper ti e
kn ;
pers, functions, le frame, fram
_ A frame !5 also st in g, ob je ct . Ra th er than a sing an ot he r
ea pical be
associate yalues
ll er s th at de fi nes @ stereoty of th e at tr ib ute values can
and fi One
te to ea’ ch other.
contents no am es connected .
of ecec ekti book obje ct
A frame is 3B a ow s a frame fora
= F 6.5.2
suall .
systems
frame: F xampler
e
. opup s
iie arione
t
WF _AlaSc (MU-Sem. 7-Comp)_ 6-15 XPM Syston
- This is one of the simplest frames, but frames can have more complex structure in real world. A powerful
knowledge system can be built with filler slots and inheritance provided by frames. Following Fig. 6.5.3 is the
example for generic frame.
Slot Fillers
Name Laptop
specialization_of a_kind_of machine
Types (desktop, laptop, mainframe, super)

if-added : Procedure ADD_LAPTOP
Speed default: faster

if-needed : Procedure FIND_SPEED
Location (home, office, mobile)
under_warranty (yes, no) j
Fig.6.5.3 : Generic Frame Example

— From Fig.6.5.3 we can conclude that fillers are of various types and it may also include procedural attachments.
Let's see what those types are.
1. Value: “laptop” in the “name” slot.
2. Range : “types” slot has (desktop, laptop, mainframe, super) as fillers.
3. If-needed : It’s a procedural attachment. It will be executed when a filler value is needed.
4 Default :“Default” value is taken if no other value exists. It represents common sense knowledge, when no
specific value is available.
5. If-added :It’s a procedural attachment. It is required if any value is to be added to a slot. In the
above
example, on arrival of a new type of laptop, ADD_LAPTOP procedure should be executed
to add that
information.
6. . If-removed :It’s a procedural attachment. It is used to remove a value from the slot.
— - Finally, the domain knowledge in an expert system will reside in a Knowledge Represent
ation Language (KRL),
such as an expert system shell. Let’s understand what it is.
5-6 Expert System-Shell

In early days each of the expert system was built from scratch. As numbers
of expert systems are developed,
researchers found that there are few parts of the systems which were common.
As in each of those systems, there was an interpreter, which was develop
ed separately for each system to interpret
the rules. It was vividly noticed that, there can be a common generic interpre
ter developed independent of the
domain.
é'Also, there are only a handful of ways to represent knowled
ge, or ta make inferences, o r to generate explan
ations; it
is better to have a generic system without any domain specific knowle
dge. .
Such systems are known as skeletal systems, shells, or simply
Al tools. A new ES ca n be developed by addin
g domain
knowledge to the shell.
: ae : = Expert System
= : 6-16 ae
Fig. - 6 6.6.1. depj acquisition syst em
epicts generic , includes Knowledge
Knowl wledgeba compon ents o f expert system shell. It
oO e, Inference anelherseec and iinference
shais interface. _ K Knowledgebase
led
rethec re cdiipon esol, planation sub system, and user
m m
Knowledge
engineer
Fig.g 6.6.1 : Components of Expert System Shell

Let us u nderstand iins hort what each component is, and what it is used fo r.
i
st
mental al step i n build
building
dal
ilding Pp rt
an expe
1. Knowledge
ge acquisition s ystem : K Know ledge isition is the fir:first and the f fundamen
acquisitio
U
led
problems and build the k nowledgebase.
system. It hel ps to collect the e xperts knowledge required to solve the
But this iis also the biggest
i bottleneck of building ES
| and heuri stic
tl l fo| ge
k nowle!
tt euris
th e h eart of f e xpiert 1
syste m wit
t
store Sa Il f. actua
pone nt is
T com
the
s s fo for all the data, as
2 . Know! ledg e bas e ‘ his ‘Oo
i ; res ent ati on tec hni que

que
about the app
ppl tioon”
icaati
lic n domain. |
with the varing s rep
iouly
are req ui to pro t pro
Igra m vid thees system accord
programmers equired
b
Inf rence engini n e e i
is the braini of expert syste m T his comp I onenentt is mainly responsible 10
3. eemechanism :
Inference ism : Inte
| line ofa
and produce uce Meal ing g : ini turn
reasone
given knowledge from the knowledgebase ;
generating inference rom the the give
query.
the result of the user’s a rmedi fin inte ate
or justifying the inal or
i explaining
:: Thi s part
Oo
f shell is responsible for
subsystem Thi
Explanation wleddge.
itional knowle
ponsible to justify need of add
4 .
of use r que ry. It is also res

result it wasnot use to be a part of expert
h the user . Earlier
It Is the mea ns of communication wit
5. User interface : .
associated wi ith user iinterface But
later on it was also recognized as an
si gn if i can ce
as there wa s no
, systems,
as it decides the utility of expert system
co mp on en t of the system,
important . ll velop ex pert
ng sh ign ifi
el shel s has significant adva
l nt ag es . It is always advisablesysttoemuse sh sheshell to de
expert system using ter
usi
€% pert systems b y using
scr atc h. To bui ld an cre she ll, one needs to en
Buil di ng he sy stem from rt can himsel f
by
a shell. The expe
system as it avoids building t
e abou ta task
do ma in int o
ate the knowledgebase
ry knowledg. that applies the knowled ge to the given task
all the necessa
0" how ow to to u use |
the shell. The inference engine
me training
undergoing 5° hell.
the 5 ved from MYCIN ; ; “I “Inter modeller” used to develop
is built into cial shells available today, “EMYCIN” deri ells
any commer ells on PC
, s, to sh
re are variety of si zes for shells , starstarting from sh
ti
There are m to name a few. T he
systems, d, sesed d systems.
pert ty from simple, forward-chaine , rul e-baba
educational ex uters. They range in complexi f thousands of do llars. Applicati ion
frame comP ce range from
hundreds to tens of *
on large main n d co mp lexity, shell pri al--time

i ncial planning or reeal
he size a ls t om iz ed o nes , su ch as fi na process
t pose shel st
accordingly t° e from 6 ene ral-pur
o cu
they rang
wise,
TechKnowledge
control. Puolications
¥ Al&SC (MU-Sem. 7-Comp) <a
6-17
— So
Expert System
SS
6.7 Explanations
— Expert systems are developed with the aim of efficient and maximum utilization of technology in place of human
expertise. To achieve this aim, along with the accuracy in the working, also the user interface must be good.
— User must be able to interact with system easily. To facilitate user interactions, system must possess following two
properties:
1, Expert systems must be able to explain its reasoning. In many cases user are interested in not only knowing the
answers to their queries but also how the system has generated that answer. That will ensure the accuracy of the
reasoning process that has produced those answers. Such kind of reasoning will be required typically in medical
applications, where doctor needs to know why a particular medicine is advised for a particular patient, as he
owns the ultimate responsibility for the medicines he subscribe. Hence, it is important that the system must store
enough meta-knowledge about the reasoning process and should be able to explain it to the user in an
understandable way.
2. The system should be able update its old knowledge by acquiring new knowledge. As the knowledgebase is
where the system’s power resides, expert system should be able to maintain the complete, accurate and up to
date knowledge about the domain. It is easy said than done!! As the system is programmed based on the
available knowledgebase, it is very difficult to adapt to the changes in knowledgebase. It must have some
mechanism through which the programs will learn its expert behavior from raw data. Another comparatively
simple way is to keep on interacting with human experts and update the system.
— TEIRESIES was the first expert system with both these properties implemented in it. MYCIN used TEIRESIES as its user
interface.
— As the TEIRESIAS-MYCIN system answers the user questions, he might be satisfied or might want to know the
reasoning behind the answers. User can very well find it out by asking “HOW” question.
— The system will interpret it as “How do you know that?” and answers it by using backward chining starting from the
answer to one of the given fact or rule. TEIRESIAS-MYCIN can do fairly goad job in satisfying the user’s query and
providing proper reasoning for it.
Types of Explanations
There are four types of explanations generally the expert system is asked for there are:
Report on rule trace on progress of consultation.
Pen PP
Explanation on how the system has reached to a particular conclusion.

Explanation of why the system is asking question.
Explanation of why the system did not give any conclusion.
6.8 Knowledge Acquisition—

Knowledge is the most important ingredient in any expert system. The power of expert systems resides in the specific,
high quality knowledge they contain about task domains. The more quality knowledge a system is given, the more
competent it becomes.
Even if expert systems shells simplify programming by providing a general skeletan; knowledge acquisition has to be
done separately for each system. The choice of reasoning method, or a shell, is important,
but it’s equally important
to accumulate high quality knowledge for developing any expert system. Knowledge engineers
perform the task of
knowledge acquisition.
° : : stem
SF _Al&SC (MU-sem, 7-Comp) Expert System
—==
6-18
The knowledge acquis
== =>
ise into the expert system.
ition co mponent allows the expert to enter their knowledge
It can be refined as or expertise r nteract aceusiy with
and whe n required. Nowadays automated systems that allow the ore acs enepacr
the system are becomi
ng inc reasingly common, thereby reducing the work pressure of knowlecg ,
The knowledge acquisition process has three principal steps. Fig. 6.8.1 depicts them in. a diagrammatic
. i
for
m. They
Y
are
as follow :
1. Knowledge elicitation : Knowledge engineer needs to interact with the domain expert and get all the knoeene:
He also needs to format it in a systematic way so that it can be used while developing the expert system shell.
.
2. Intermediate knowledge representation : The knowledge obtained from the domain expert needs to be store in
some intermediate representation, such that, it can be worked upon to produce the final refined version.
3. Knowledgebase representation : The intermediate representation of the knowledge needs to be complied and
transformed into an executable format. This version of knowledge is ready to get uploaded to system shell as it
is.
©.B. production rules, that the inference engine
can process.
In the process of expert system development, numbers of iterations through these three stages are required in.
order to equip the system with good quality knowledge.
The iterative nature of the knowledge acquisition process can be
represented in Fig. 6.8.1.
Reformulations
Redesigns
a *
c Refinements
5 2 2
oO
y 5os y g ¥ — go - : y — 8
identi a Find 5 Designa» | 2 | Fomulate | 3
eeecen ify | ,| conceptsto.| O | structureto | © | tulesto =| &
“Problem represent. organize: “} embody. -
characteristics _ knowledge.” ~ knowledge| knowledge©
Identification Conceptualization Formalization Implementation Testing

Fig. 6.8.1 : Stages of knowledge acquisition
As quality of knowledge base determines the success of an expert system, Al researchers are continually exploring and
adding new methods of knowledge representation and reasoning. Future of expert systems depends on breaking the
knowledge acquisition bottleneck and codifying and representing a large knowledge infrastructure.
6.8.1 Knowledge Elicitation
The knowledge elicitation is the first step of knowledge acquisition. In this process itself there are several stages
Generally knowledge engineer performs these steps.
these steps need to be carried out before meeting the
domain expert to collect the quality knowledge. They
are as
follows.
mum possible data about
. the problem
: domain from books, manuals, internet, etc.,
1. Gather maxi i
become familiar w! h specialist terminology and :jargons of the problem domain.
rt it!
“In arder to
Identify the types of reasoning and problem solving tasks that the system will be required
e i to perform .
d domain expert or team of experts who are willing to work on the Project. So metimes exp
Find do I erts are frightened
of being replaced by a computer system!
jaw the domain experts multiple times during the course of buildin g the system,
4. Elemis that, the system is expected to solve. Have them check Find out how they sol
a nd refine the int ve
the pro ’ ermediate knowledge
representation.
Ww TechKnowledga
Publications
Expert
knowle
proce sg. There
exists automated
ic it at io n is a time consuming co mmon modern al
ternatives.
ow le dg e el in g us ed as
Kn easingly be
es which are incr
learning techni qu
Expert Syste m vs.

Traditional Sy stem
6.9
structures
na l sy st em s are Algorithms + d ata
+ inference Traditio
Expert system Is knowledge base
they do not
Engine
ot do pr ed ic ti on ta sks so efficiently as
ed on TS cann
t future events bas
Expert syste ms can predic have a strong infere
nce engine.
their
the current data input patterns using
inference process. . inference sys tem to deduce
does no t have a strong
TS
system to deduce
ES have very strong infere nce
knowledge.
s
knowledge from given fact the results.
not have mechanism to justify
any
TS do
tem which can explain
ES have explanation subsys ed to be done
edi ate stag e. Man ual debugging is requir
interm
and justify the results at any intervention.
g, | TS can not do exp ert tasks without human
planning, schedulin
ES can do tasks like :
require to deal with
prediction, diagnosis; which
wledge from past
current data input and kno
handled by human
experiences, which generally
experts.
data based on available data,
to match human expertise | TS systems can only provide
Expert systems are able the domain.
ple te | It cann ot prov ide user with knowledge about
ided with a com
in a particular domain prov to analyse the data
ul inf ere nce eng ine. Hence human expertise are required
knowled gebase and a pow erf
it. Hence TS c annot
further to deduce knowledge from
t them.
eliminate Human experts it can only assis
Systems
6.10 The Applications 0 f Expert
rt systems are used in the
t syst ems find their way into most areas of knowledge work. Expe
The applications of exper problems.
to in dustrial and commercial
place of human interactions
big in number that, they
wide s pectr um of appli catio ns for expert systems. These applications are so
There exists a seven major classes. -
All these applications are clustered into
cannot be categorized easily.
Diagnosis and troubleshooting of
devices and systems of all kinds
1.
est corrective actions fora
of applications comprises of expert syste ms that detect faults and sugg
This category which expert system
diagnosis was one of the first knowledge areas to
malfunctioning device or process. Medical
technology was applied.
Planning and scheduling

ve the
provides with the plan of action in order to achie
These expert systems analyze set of goals and scheduling
commercial potential. For example, airline flight
predetermined goal. These applications has great
acturing process planning.
manufacturing job-shop scheduling; and manuf
Expert System
3. Confi eurati i on
Configure Os =
of manufactured objects from
subassemblies
ation « sole
applications is histori aly One of the most important of expert system applications. In this category of
by a set of constraints.
These applicati ution to a problem is synthesized from a given set of elements related
include modular home building,
manufacturing con used in many different industries. Examples
? m i j '
4. plex engineering design and manufacturing.
Financial decision
making
; .
Expert systems ;ar i. rams
ces, typically in foreign exchange trading. Advisory prog
Ms .
are widely used te wee vigorously in financial servi s expert systems to
used
sis © assist bankers to take decisions about loans. Insurance companie
assess the risk associated with the customer in order to determine price for the insurance. Atypical application in~
i i
& tinancial market
s is
5. nae
Knowledge Publishing
3
. .
Usage of ex further exploration. There are
appli pert
many Y applic systems in this area is relatively new, but has good potential for
ations based on user information access preferences.
6. Process monitoring and control
In thi . .
from
h :al eee system s perfor ming analysi s of real time data are designed. These systems obtain data
Physic devices and produce results specifying anomalies, predicting trends, etc. We have existing expert
systems to monitor the manufacturing processes in the steel making and oil refining industries.
7. Design and manufacturing
These expert systems assist in the design of physical devices and processes. It is ranging from high level
- conceptual design of abstract entities all the way to factory floor configuration of manufacturing processes.
jew Guestions
Q.1 Explain the need of hybrid approach.
Q.2 Explain Fuzzy neural system.
0.3 Explain the architecture of ANFIS with the help of block diagram.
examples of re al time expert systems.
Q.4 What is expert system? Give
of expert system?
Q.5 Whatare the characteristics |
blocks of expert system.

Q.6 — Explain in building
in case of expert system?
to represent domain knowledge
Q.7 Whatare the various methods
em shell.
te on expert syst :
Q.8 Writeasho rt no
systems?
perfor med in ca se of expert
Q.9 Howreas oning is
isiti on? What are the
steps of knowledge acquisition? ql
le dg e ac qu
Q.10 Whati 's know
dge elicitation.
short note on kn owle
Q.11 Writea
—
UF Tecthnamieags
ae

AISC - Techmax (Searchable) With Index - Compressed

Uploaded by

Copyright:

Available Formats

AISC - Techmax (Searchable) With Index - Compressed

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

AISC - Techmax (Searchable) With Index - Compressed

Uploaded by

Copyright:

Available Formats

Scanned by CamScanner

14.1. Introduction and Definition of Artificial Intelligence.

1.1. Introduction to Artificial Intelligence

1.2 Foundations and Mathematical Treatments

1.2.1. Acting Humanly : The Turing Test Approach

Fig.1.2.1 : Turing Test Environment

Capabilities a Computer needs to Possess

_ Natural _ Knowledge Automated. Q ‘Machine —

1.2.2 Thinking Humanly : The Cognitive Modelling Approach

2. Psychological experiments can be carried out by observing a person in action.

3. Brain imaging can be done by observing the brain in action.

1.2.3 Thinking Rationally : The “Laws of Thought” Approach

_ There are two problems in this approach :

resources of any computer.

advantages over the other approaches

In our syllabus we are going to study acting rationally approach.

1. Artificial Narrow Intelligence/ Weak Al

2. Artificial General Intelligence / Strong Al

3. Artificial Super Intelligence

Natural language processing (Language-understanding)

Fig. 1.4.1: Components of Al

1.4.1. Computational Intelligence vs. Artificial Intelligence

1.5 History of Artificial Intelligence

4. Education : Language : Speaking English, etc,

4. Linguistic : Theories of the structure and meaning of language have emerged.

1.6 Applications of Artificial Intelligence

Fig. 1.6.1 : Fields of Al Application

5. Business and Manufacturing

6. Automated planning and scheduling

1.7 Sub Areas/ Domains of Artificial Intelligence

Al Applications can be roughly classified based on the type of tools/a

Fig. 1.7.1 : Sub-areas in Intelligent

Current Trends in Artificial Intelligence _

4. Internet of Things (loT)

6. Alin shopping and customer service

1.9 Intelligent Agents .

1.9.1 What is an Agent?

Fig.1.9.2 : Generic robotic agent architecture

[A, Dirty], [B, Clean], [A, absorb],[B, Nop], etc.

1.9.2 Definitions of Agent

Agent will perform

1.9.3 Intelligent Agent

1.9.3(A) Structure of Intelligent Agents

ucture of intel ligent agent

should be taken when you go near hot elements?”

1.10 Rational Agent

- For problem solving, if an agent makes

hammer (b) Agent Is using nail and hammer efficiently

‘Input | —> Goal:

127 Alternate possible actions

Ez Selected actions which leads to

Fig.1.10.2 ; Rational Agent

1. Fully observable vs. Partially observable

Known vs. unknown

Fig.1.11.1; Environment types

2. Single agent vs. Multl-agent

Intro. to Artificial Intelligence( 1) and t Computing

3. Deterministic vs. Stochastic

4. Episodic vs. Sequential