06 Learning Systems

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Learning Systems : AI Course Lecture 31 34, notes, slides
www.myreaders.info/ , RC Chakraborty, e-mail rcchak@gmail.com , June 01, 2010
www.myreaders.info/html/artificial_intelligence.html

Learning System

Artificial Intelligence

www.myreaders.info
Return to Website
Learning systems, topics : Definition, learning agents, components
of learning system, paradigms of machine learning. Rote Learning :
learning by memorization, learning something by repeating.
Learning from example : Induction, Winston's learning, Version
spaces - learning algorithm (generalization and specialization tree),
Decision trees - ID3 algorithm. Explanation Based Learning (EBL) :
general approach, EBL architecture, EBL system, generalization
problem, explanation structure. Discovery : theory driven AM
system, data driven - BACON system. Clustering : distance
functions, K-mean clustering algorithm. Learning by analogy;
Neural Net Perceptron; Genetic Algorithm. Reinforcement
Learning : RL Problem, agent - environment interaction, RL tasks,
Markov system, Markov decision processes, agents learning task,
policy, reward function, maximize reward, value functions.

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Learning System

Artificial Intelligence

Topics
(Lectures 31, 32, 33, 34 4 hours)

Slides
1. What is Learning
Definition, learning agents, components of learning system; Paradigms
of machine learning.

03-09
2. Rote Learning
Learning by memorization, Learning something by repeating.

10
3. Learning from Example : Induction
Winston's learning, Version spaces - learning algorithm (generalization
and specialization tree), Decision trees - ID3 algorithm.

11-38
4. Explanation Based Learning (EBL)
General approach, EBL architecture, EBL system, Generalization problem,
Explanation structure.

39-43
5. Discovery
Theory driven AM system, Data driven - BACON system

44-52
6. Clustering
Distance functions, K-mean clustering algorithm.

53-62
7. Analogy
63
8. Neural net and Genetic Learning
Neural Net Perceptron; Genetic learning - Genetic Algorithm.

64-67
9. Reinforcement Learning
RL Problem : Agent - environment interaction, key Features; RL tasks,
Markov system, Markov decision processes, Agents learning task, Policy,
Reward function, Maximize reward, Value functions.

68-80
10 References
81
02

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Learning
Machine Learning

What is learning ? Some Quotations

Herbert Simon, 1983
Learning denotes changes in a system that enable a system to do the same
task more efficiently the next time.

Marvin Minsky, 1986
Learning is making useful changes in the workings of our minds.

Ryszard Michalski, 1986
Learning is constructing or modifying representations of what is being
experienced.

Mitchell, 1997
A computer program is said to learn from experience E with respect to
some class of tasks T and performance measure P, if its performance at
tasks in T, as measured by P, improves with experience E.
03

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI - Learning
1. What is Learning
Learning denotes changes in a system that enable the system to do the
same task more efficiently next time.
Learning is an important feature of Intelligence.

1.1 Definition

A computer program is said to learn from experience E with respect to
some class of tasks T and performance measure P, if its performance at
tasks in T, as measured by P, improves with experience E. (Mitchell 1997)
This means :

Given : A task T

A performance measure P

Some experience E with the task

Goal : Generalize the experience in a way that allows to
improve your performance on the task.

Why do you require Machine Learning ?

Understand and improve efficiency of human learning.

Discover new things or structure that is unknown to humans.

Fill in skeletal or incomplete specifications about a domain.
04

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI - Learning

1.2 Learning Agents
An agent is an entity that is capable of perceiving and do action.
An agent can be viewed as perceiving its environment through sensors and
acting upon that environment through actuators.

Environment
Agent
Sensors Actuators
Human agent Eyes, ears, etc Leg, hands, mouth
Robotic agent Cameras, IR range finders motors
Software agent Key stroke, File contents Displays to screen,
write files

In computer science an agent is a software agent that assists users
and acts in performing computer-related tasks.
05
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI - Learning

Intelligent Agent (Learning Agent)
Agent is an entity that is capable of perceiving and do action.
In computer science an agent is a software agent.
In artificial intelligence, the term used for agent is an intelligent agent.
Learning is an important feature of Intelligence.

Percept : agent's perceptual inputs

Percept sequence : history of everything the agent has perceived

Agent function : describes agents behavior

Agent program : Implements agent's function

Learning Agent consist of four main components :

Learning element,

Performance element,

Critic, and

Problem generator.

Components of a Learning System

The components are described in the next slide.
06

E
N
V
R
O
N
M
E
N
T
Problem
Generator
Sensors
Performance
Element
Learning
Element
Critic

learning
goals
knowledge
feed back
Performance
standard
changes
Effectors
Learning Agent
Percepts
Actions
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI - Learning

Components of a Learning System

Performance Element: The Performance Element is the agent itself that
acts in the world. It takes in percepts and decides on external actions.

Learning Element: It responsible for making improvements, takes
knowledge about performance element and some feedback,
determines how to modify performance element.

Critic: Tells the Learning Element how agent is doing (success or failure)
by comparing with a fixed standard of performance.

Problem Generator: Suggests problems or actions that will generate
new examples or experiences that will aid in training the system further.

Example : Automated Taxi on city roads

Performance Element: Consists of knowledge and procedures for
driving actions.
e.g., turning, accelerating, braking are performance element on roads.

Learning Element: Formulates goals.
e. g., learn rules for braking, accelerating, learn geography of the city.

Critic: Observes world and passes information to learning element.
e. g. , quick right turn across three lanes of traffic, observe reaction of
other drivers.

Problem Generator: Try south city road .
07
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI - Learning

1.2 Paradigms of Machine Learning

Rote Learning: Learning by memorization; One-to-one mapping from
inputs to stored representation; Association-based storage and
retrieval.

Induction: Learning from examples; A form of supervised learning,
uses specific examples to reach general conclusions; Concepts are
learned from sets of labeled instances.

Clustering: Discovering similar group; Unsupervised, Inductive
learning in which natural classes are found for data instances, as
well as ways of classifying them.

Analogy: Determine correspondence between two different
representations that come from Inductive learning in which a system
transfers knowledge from one database into another database of a
different domain.
08

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI - Learning

Discovery: Learning without the help from a teacher; Learning is
both inductive and deductive. It is deductive if it proves
theorems and discovers concepts about those theorems. It is
inductive when it raises conjectures (guess). It is unsupervised,
specific goal not given.

Genetic Algorithms: Inspired by natural evolution; In the natural
world, the organisms that are poorly suited for an environment die off,
while those well-suited for it prosper. Genetic algorithms search the
space of individuals for good candidates. The "goodness" of an
individual is measured by some fitness function. Search takes place in
parallel, with many individuals in each generation.

Reinforcement: Learning from feedback (+ve or -ve reward) given
at end of a sequence of steps. Unlike supervised learning, the
reinforcement learning takes place in an environment where the
agent cannot directly compare the results of its action to a
desired result. Instead, it is given some reward or punishment that
relates to its actions. It may win or lose a game, or be told it has made
a good move or a poor one. The job of reinforcement learning is to
find a successful function using these rewards.
09
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Rote Learning
2. Rote Learning
Rote learning technique avoids understanding the inner complexities but
focuses on memorizing the material so that it can be recalled by the
learner exactly the way it was read or heard.

Learning by Memorization which avoids understanding the inner
complexities the subject that is being learned; Rote learning instead
focuses on memorizing the material so that it can be recalled by the learner
exactly the way it was read or heard.

Learning something by Repeating over and over and over again;
saying the same thing and trying to remember how to say it; it
does not help us to understand; it helps us to remember, like we learn a
poem, or a song, or something like that by rote learning.
10

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Learning from example : Induction

3. Learning from Example : Induction
A process of learning by example. The system tries to induce a general rule
from a set of observed instances. The learning methods extract rules and
patterns out of massive data sets.
The learning processes belong to supervised learning, does classification and
constructs class definitions, called induction or concept learning.

The techniques used for constructing class definitions (or concept leaning) are :

Winston's Learning program

Version Spaces

Decision Trees

Note : The Decision tree technique has been explained in all details
while other two techniques are briefly described in the next few slides.
11
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

3.1 Winston's Learning
Winston (1975) described a Blocks World Learning program. This
program operated in a simple blocks domain. The goal is to construct
representation of the definition of concepts in the blocks domain.

Example : Concepts such a "house".

Start with input, a line drawing of a blocks world structure.
It learned Concepts House, Tent, Arch as :
brick (rectangular block) with a wedge (triangular block) suitably
placed on top of it, tent as 2 wedges touching side by side, or an
arch as 2 non-touching bricks supporting a third wedge or brick.

The program for Each concept is learned through near miss. A near miss
is an object that is not an instance of the concept but a very similar to
such instances.

The program uses procedures to analyze the drawing and construct a
semantic net representation.

An example of such an structural for the house is shown below.

Object - house

Semantic net

Node A represents entire structure, which is composed of two parts :
node B, a Wedge, and node C, a Brick.

Links in network include supported-by, has-part, and isa.

12
A
C B
Supported - by
Wedge Brick
has-part has-part
isa
isa
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Winston's Program

Winston's program followed 3 basic steps in concept formulation:

1. Select one known instance of the concept.
Call this the concept definition.

2. Examine definitions of other known instance of the concept.
Generalize the definition to include them.

3. Examine descriptions of near misses.
Restrict the definition to exclude these.

Both steps 2 and 3 of this procedure rely heavily on comparison
process by which similarities and differences between structures
can be detected.

Winston's program can be similarly applied to learn other concepts
such as "ARCH".
13

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Learning from example: Induction

3.2 Version Spaces

A Version Space is a hierarchical representation of knowledge that
keeps track of all the useful information supplied by a sequence of
learning examples without remembering any of the examples. The
version space method is a concept learning process. It is another
approach to learning by Mitchell (1977).

Version Space Characteristics

Represents all the alternative plausible descriptions of a heuristic.
A plausible description is one that is applicable to all known +ve
examples and no known -ve example.

A version space description consists of two complementary trees:

one, contains nodes connected to overly general models, and

other, contains nodes connected to overly specific models.

Node values/attributes are discrete.

Fundamental Assumptions

The data is correct ie no erroneous instances.

A correct description is a conjunction of some attributes with values.
14
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Version Space Methods
A version space is a hierarchical representation of knowledge.
Version space method is a concept learning process accomplished by
managing multiple models within a version space.
It is specialization of general models and generalization of specific models.
Version Space diagram described below consists : a Specialization tree
(colored Blue) and a Generalization tree (colored green) .

Version Space diagram

Top 1st row : the top of the tree, we have the
most general hypotheses.

Top 2nd row : This row is an expanded version of
the first. This row of hypotheses is slightly more
specific than the root node.

Top 3rd row : As training data (positive examples)
is processed, the inconsistent nodes are removed
from the general specification.

Finally, we converge to a solution.

Bottom 3rd row : Any hypothesis that is
inconsistent with the training data (negative
instances) is removed from the tree.

Bottom 2nd row : The specific hypothesis is
expanded to form more nodes that are slightly
more general.

Bottom 1st row : This is the most specific
hypothesis.

Tree nodes :

each node is connected to a model.

nodes in the generalization tree are connected to a model that
matches everything in its sub-tree.

nodes in the specialization tree are connected to a model that
matches only one thing in its sub-tree.

Links between nodes :

denotes generalization relations in a generalization tree, and

denotes specialization relations in a specialization tree.

15

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Version Space Convergence
Generalization and Specialization leads to Version Space convergence.

The key idea in version space learning is that specialization of the general
models and generalization of the specific models may ultimately lead
to just one correct model that matches all observed positive examples
and does not match any negative examples. This means :

Each time a +ve example is used to specialize the general models, then
those specific models that match the -ve example are eliminated.

Each time a -ve example is used to generalize the specific models, those
general models that fail to match the +ve example are eliminated.

Finally, the positive and negative examples may be such that only one
general model and one identical specific model survive.
16

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Version Space Learning Algorithm: Candidate Elimination
The Candidate Elimination algorithm finds all describable hypotheses
that are consistent with the observed training examples.
The version space method handles +ve and -ve examples symmetrically.

Given : A representation language and a set of positive and
negative examples expressed in that language.

Compute : A concept description that is consistent with all the
positive examples and none of the negative examples.

Note : The methodology / algorithm is explained in the next slide.
17
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

[Continued from previous slide - version space]

Version Space Search Algorithm

Let G be the set of maximally general hypotheses.
Initialize G, to contain most general pattern, a null description.

Let S be the set of maximally specific hypotheses.
Initialize S, to contain one element : the first positive example.

Accept a new training example.

For each new positive training example p do :

1.
Delete all members of G that fail to match p ;

2.
For every s in S do

If s does not match p, then replace s with its maximally
specific generalizations that match p;

3.
Delete from S any hypothesis that is subsumed, a more
comprehensive, by other hypothesis in S;

4. Delete from S any hypothesis more general than some hypothesis in
G.

For each new negative training example n do :

1.
Delete all members of S that match n;

2.
For each g in G do

If g match n, then replace g with its maximally
general specializations that do not match n;

3.
Delete from G any hypothesis more specific than some other
hypothesis in G;

4.
Delete from G any hypothesis that is not more general than some
hypothesis in S;

If S and G are both singleton sets, then

1.
If they are identical, output their value and halt.

2.
If they are different, the training cases were inconsistent; output this
result and halt.

3.
Else continue accepting new training examples.

The algorithm stops when it runs out of data;

1.
If 0 , no consistent description for the data in the language.

2.
If 1 , answer (version space converges).

3.
If 2 , all descriptions in the language are implicitly included.
18

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Version Space Problem 1
Learning the concept of "Japanese Economy Car"
[Ref: E. Rich, K. Knight, "Artificial Intelligence", McGraw Hill, Second Edition]

Examples 1 to 5 of features
Country of Origin, Manufacturer, Color, Decade, Type
Example Origin Manufacturer Color Decade Type Example Type
1. Japan Honda Blue 1980 Economy Positive
2. Japan Toyota Green 1970 Sports Negative
3. Japan Toyota Blue 1990 Economy Positive
4. USA Chrysler Red 1980 Economy Negative
5. Japan Honda White 1980 Economy Positive

Step by step solution : Symbol for Prune
1.
Initialize : to maximally general and maximally specific hypotheses

Initialize G to a null description, means all features are variables.
G = { (?, ?, ?, ?, ?) }

Initialize S to a set of maximally specific hypotheses;
Apply a positive example;
Apply Example 1 (Japan, Honda, Blue, 1980, Economy)
S = { (Japan, Honda, Blue, 1980, Economy) }

Resulting Specialization (blue) and Generalization (green)
tree nodes

S = {(Japan, Honda, Blue, 1980, Economy)}

These two models represent the most general and the most specific
heuristics one might learn.

The actual heuristic to be learned, "Japanese economy car",
probably lies between them somewhere within the version space.
19

G = ( ?, ?, ?, ?, ? )
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

2. Apply Example 2 (Japan, Toyota, Green, 1970, Sports)
It is a negative example.

Specialize G, to exclude the negative example;
i.e., G must be specialized in such a way that the negative example is
no longer in the version space.
The available specializations are :
G = { (?, Honda, ?, ?, ?), (?, ?, Blue, ?, ?), (?, ?, ?, 1980, ?),
(?, ?, ?, ?, Economy) }

tree nodes and links

20
(?, ?, ?, ?, Economy)
(?, ?, ?, 1980, ?) (?, ?, Blue, ?, ?) (?, Honda, ?, ?, ?)
G = ( ?, ?, ?, ?, ? )
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

3. Apply Example 3 (Japan, Toyota, Blue, 1990, Economy)
It is a positive example.

Prune G , to exclude descriptions inconsistent with the positive
example; i.e., remove from the G set any descriptions that are
inconsistent with the positive example. The new G set is
G = { (?, ?, Blue, ?, ?), (?, ?, ?, ?, Economy) }

Generalize S, to include the new example. The new S set is
S = { (Japan, ? , Blue, ? , Economy) }

tree nodes and links :

At this point, the new G and S sets specify a version space that
can be translated roughly in to English as : The concept may be as
specific as "Japanese, blue economy car".
21
(?, ?, ?, ?, Economy)
(?, ?, ?, 1980, ?) (?, ?, Blue, ?, ?) (?, Honda, ?, ?, ?)
G = ( ?, ?, ?, ?, ? )
S = {(Japan, ?, Blue, ?, Economy)}
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

4. Apply Example 4 (USA, Chrysler, Red, 1980, Economy)
It is another negative example.

Specialize G to exclude the negative example; (but stay consistent
with S) i.e.; G set must avoid covering this new example.
The new G set is
G = { (?, ?, Blue, ?, ?), (Japan, ?, ?, ?, Economy) }

Prune away all the specific models that match the negative example.
No match found therefore no change in S
S = { (Japan, ?, Blue, ?, Economy) }


It is now clear that the car must be Japanese, because all description in
the version space contain Japan as origin.
22
(?, ?, ?, ?, Economy)
(Japan, ?, Blue, ?, ?)
G = ( ?, ?, ?, ?, ? )
S = {(Japan, ?, ?, ?, Economy)}
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

5. Apply Example 5 (Japan, Honda, White, 1980, Economy)
It is positive example.

Prune G to exclude descriptions inconsistent with positive example.
G = { (Japan, ?, ?, ?, Economy) }

Generalize S to include this positive example.
S = { (Japan, ?, ?, ?, Economy) }


Thus, finally G and S are singleton sets and S = G.
The algorithm has converged. No more examples needed.
Algorithm stops.

23

(?, ?, ?, ?, Economy)
(?, ?, Blue, ?, ?)
G = ( ?, ?, ?, ?, ? )
G = {(?, ?, ?, ?, Economy)}
G = {(Japan, ?, ?, ?, Economy)}
G = ( ?, ?, ?, ?, ? )
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

3.3 Decision Trees
Decision trees are powerful tools for classification and prediction.
Decision trees represent rules. Rules are easily expressed so that humans
can understand them or even directly use in a database access language
like SQL so that records falling into a particular category may be retrieved.

Description

Decision tree is a classifier in the form of a tree structure where each
node is either a leaf or decision node.

leaf node indicates the target attribute (class) values of examples.

decision node specify test to be carried on an attribute-value.

Decision tree is a typical inductive approach to learn knowledge on
classification. The conditions are :

Attribute-value description: Object or case must be expressible as
a fixed collection of properties or attributes having discrete values.

Predefined classes : Categories to which examples are to be
assigned must already be defined (ie supervised data).

Discrete classes: Classes must be sharply delineated; continuous
classes broken up into vague categories as "hard", "flexible", "soft".

Sufficient data: Enough training cases to distinguish valid patterns.

Example : A simple decision tree

24

K=y A
A A
A
K=y
K=y K=y
K=y
C=red
C=true C=false
C=blue
B < 4.5 B < 8.1 B 8.1 B 4.5
A
K=y
Decision node
Leaf node
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Algorithm to generate a Decision Tree
A decision tree program constructs a decision tree T from a set of
training cases. The advantage of a decision tree learning is that a
program, rather than a knowledge engineer, elicits knowledge from an
expert.

ID3 Algorithm (Iterative Dichotomiser 3)
ID3 is based on the Concept Learning System (CLS) algorithm, developed
J. Ross Quinlan, at the University of Sydney in 1975.

Description

ID3 builds a decision tree from a set of "examples".
The "examples" have several attributes.
They belong to a class (yes or no).

Leaf nodes contain the class name.

Decision node is an attribute test .
25
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Attribute Selection
How does ID3 decide which attribute is the best?
There are different criteria to select which attribute will become a test
attribute in a given branch of a tree. A well known criteria is
information gain. It uses a log function with a base of 2, to determine
the number of bits necessary to represent a piece of information.

Entropy measures information content in an attribute.

Shannon defines information entropy in terms of a discrete random
variable X , with possible states (or outcomes) x
1
. . . x
n
as:
H(X) = p(xi) log
2
( 1/p(xi) ) = - p(xi) log
2
p(xi)
where p(xi) = P (X = xi) is the probability of the i
th
outcome of X.

Information gain of an attribute X with respect to class attribute Y.

Let Y and X are discrete variables that take values in
{y
1
. . . y
i
. . . y
k
} and {x
1
. . . x
j
. . . x
l
} .
The information gain of a given attribute X with respect to the class
attribute Y is the reduction in uncertainty about the value of Y
when we know the value of X, is called I(Y ; X).

Uncertainty
Uncertainty about the value of Y is measured by its entropy, H(Y).

uncertainty about the value of Y when we know the value of X is
given by the conditional entropy of Y given X, i.e., H(Y |X).

Information Gain of a given attribute X with respect to the
class attribute Y is given by : I(Y ; X) = H(Y ) - H(Y |X). where

H(Y) = - p(yi) log
2
p(yi) is the initial entropy in Y,

H(Y |X) = p(xj) H(Y |xj) is conditional entropy of Y
given X where H(Y |xj) = - p (yi | xj) log
2
p(yi | xj) is
the entropy in Y given a particular answer x
j
.
26

i=1
n
i=1
n

yi=1
k
xj=1
l
xj=1
l
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Example
Decide if the weather is amenable to play Golf or Baseball.

Training Data Collected are Over 2 weeks

Day Outlook Temp Humidity wind Play
1 sunny 85 85 week no
2 sunny 80 90 strong no
3 cloudy 83 78 week yes
4 rainy 70 96 week yes
6 rainy 65 70 strong no
7 cloudy 64 65 strong yes
8 sunny 72 95 week no
9 sunny 69 70 week yes
11 sunny 75 70 strong yes
12 cloudy 72 90 strong yes
13 cloudy 81 75 week yes
14 rainy 71 85 strong no

Learning set
In the above example, two attributes, the Temperature and
Humidity have continuous ranges. ID3 requires them to be
discrete like hot, medium, cold, high, normal. Table below
indicates the acceptable values.

Attribute Possible Values
Outlook Sunny Cloudy Rainy
Temperature Hot Medium Cold
Humidity High Normal
Wind Strong Week
Class play no play
Decision n (negative) p (positive)

27

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Assign discrete values to the Attributes
Partition the continuous attribute values to make them discrete,
following the key mentioned below.

Temperature : Hot (H) 80 to 85 Medium (M) 70 to 75 Cold (C) 64 to 69
Humidity :
High (H) 81 to 96 Normal (N) 65 to 80
Class :
Yes (Y) play No (N) no play

Day Outlook Temp Humidity Wind Class (play)
1 Sunny 85 Hot 85 High week no
2 Sunny 80 Hot 90 High strong no
3 Cloudy 83 Hot 78 High week yes
4 Rainy 70 Medium 96 High week yes
5 Rainy 68 Cold 80 Normal week yes
6 Rainy 65 Cold 70 Normal strong no
7 Cloudy 64 Cold 65 Normal strong yes
8 Sunny 72 Medium 95 High week no
9 Sunny 69 Cold 70 Normal week yes
10 Rainy 75 Medium 80 Normal week yes
11 Sunny 75 Medium 70 Normal strong yes
12 Cloudy 72 Medium 90 High strong yes
13 Cloudy 81 Hot 75 Normal week yes
14 Rainy 71 Medium 85 High strong no

28
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Attribute Selection
By applying definitions of Entropy and Information gain

Entropy : From the definition, When Y is a discrete variables
that take values in {y
1
. . . y
i
. . . y
k
} then the entropy of
Y is given by H(Y)

H(Y) = p(yi) log
2
( 1/p(yi) ) = - p(yi) log
2
p(yi)
where p(yi) is the probability of the i
th
outcome of Y.
Apply his definition to the example stated before :
Given, S is a collection of 14 examples with 9 YES and 5 NO
then Entropy(S) = - (9/14) Log
2
(9/14) - (5/14) Log
2
(5/14)
= 0.940
Note that S is not an attribute but the entire sample set.
Entropy range from 0 ("perfectly classified") to 1 ("totally random").

[Continued in next slide]
29

i=1
k

i=1
k
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

[Continued from previous slide Attribute election]

Information Gain : From the definition, a given attribute X
with respect to the class attribute Y is given by :
I(Y ; X) = H(Y ) - H(Y |X) where

H(Y) = - p(y
i
) log
2
p(y
i
) is the initial entropy in Y,

H(Y |X) = p(x
j
) H(Y |x
j
) is conditional entropy of Y
given X where H(Y |x
j
) = - p (y
i
| x
j
) log
2
p(y
i
| x
j
) is
entropy in Y given a particular answer x
j
.
Apply this definition to the example stated above :
Information Gain (S, A) of the example set S on attribute A is
Gain(S, A) = Entropy(S) - ((|Sv| / |S|) * Entropy(Sv)) where
sum over each value v of all possible values of attribute A
Sv is subset of S for which attribute A has value v
|Sv| is number of elements in Sv
|S| is number of elements in S

Given, S is a set of 14 examples, in which one of the attributes is wind
speed. The values of Wind can be Weak or Strong. The classification of
these 14 examples are 9 YES and 5 NO. Given, for attribute wind :
8 occurrences of wind = weak and
6 occurrences of wind = strong.
For wind = weak, 6 of the examples are YES and 2 are NO.
For wind = strong, 3 of the examples are YES and 3 are NO.
Gain(S, wind) = Entropy(S) - (8/14) Entropy(Sweak)
(6/14) Entropy(Sstrong)
= 0.940 - (8/14) 0.811 - (6/14) 1.00 = 0.048

Entropy(Sweak) = - (6/8) log
2
(6/8) - (2/8) log
2
(2/8) = 0.811
Entropy(Sstrong) = - (3/6) log
2
(3/6) - (3/6) log
2
(3/6) = 1.00
For each attribute, the gain is calculated and the highest gain is
used in the decision node.
30

yi=1
k
xj=1
l
xj=1
l
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Step-by-Step Calculations : Ref. Training data stated before

Step 01 : "example" set S
The set S of 14 examples is with 9 yes and 5 no then
Entropy(S) = (9/14) Log2 (9/14) (5/14) Log2 (5/14) = 0.940

Step 02 : Attribute Outlook
Outlook value can be sunny, cloudy or rainy.
Outlook = sunny is of occurrences 5
Outlook = cloudy is of occurrences 4
Outlook = rainy is of occurrences 5
Outlook = sunny, 2 of the examples are yes and 3 are no
Outlook = cloudy, 4 of the examples are yes and 0 are no
Outlook = rainy, 3 of the examples are yes and 2 are no

Entropy(Ssunny)
Entropy(Scloudy)
Entropy(Srainy)
= (2/5) x log2(2/5) (3/5) x log2(3/5) = 0.970950
= (4/4) x log2(4/4) (0/4) x log2(0/4) = 0
= (3/5) x log2(3/5) (2/5) x log2(2/5) = 0.970950

Gain(S, Outlook) = Entropy(S) (5/14) x Entropy(Sunny)
(4/14) x Entropy(Scloud)
(5/14) x Entropy(Srainy)
= 0.940 - (5/14) x 0.97095059 (4/14) x 0
- (5/14) x 0.97095059
= 0.940 - 0.34676 - 0 - 0.34676
= 0.246

31

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Step 03 : Attribute Temperature
Temp value can be hot, medium or cold.
Temp = hot is of occurrences 4
Temp = medium is of occurrences 6
Temp = cold is of occurrences 4
Temp = hot, 2 of the examples are yes and 2 are no
Temp = medium, 4 of the examples are yes and 2 are no
Temp = cold, 3 of the examples are yes and 1 are no

Entropy(Shot)
Entropy(Smedium)
Entropy(Scold)
= (2/4)x log2(2/4) (2/4) x log2(2/4) = -0.99999999
= (4/6)x log2(4/6) (2/6) x log2(2/6) = -0.91829583
= (3/4)x log2(3/4) (1/4) x log2(1/4) = -0.81127812

Gain(S, Temp) = Entropy(S) (4/14) x Entropy(Shot)
(6/14) x Entropy(Smedium)
(4/14) x Entropy(Scold)
= 0.940 - (4/14) x 0.99999 (6/14) x 0.91829583
- (4/14) x 0.81127812
= 0.940 - 0.2857142 -0.393555 - 0.2317937
= 0.0289366072

32

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Step 04 : Attribute Humidity
Humidity value can be high, normal.
Humidity = high is of occurrences 7
Humidity = normal is of occurrences 7
Humidity = high, 3 of the examples are yes and 4 are no
Humidity = normal, 6 of the examples are yes and 1 are no

Entropy(Shigh)
Entropy(Snormal)

= (3/7) x log2(3/7) (4/7) x log2(4/7) = -0.9852281
= (6/7) x log2(6/7) (1/7) x log2(1/7) = -0.5916727

Gain(S, Humidity) = Entropy(S) (7/14) x Entropy(Shigh)
(7/14) x Entropy(Snomal)
= 0.940 - (7/14) x 0.9852281 (7/14) x 0.5916727
= 0.940 - 0.49261405 - 0.29583635
= 0.1515496

33
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Step 05 : Attribute wind
Wind value can be weak or strong.
Wind = week is of occurrences 8
Wind = strong is of occurrences 6
Wind = weak, 6 of the examples are yes and 2 are no
Wind = strong, 3 of the examples are yes and 3 are no

Entropy(Sweak)
Entropy(Sstrong)
= (6/8) x log2(6/8) (2/8) x log2(2/8) = 0.811
= (3/6) x log2(3/6) (3/6) x log2(3/6) = 1.00

Gain(S, Wind) = Entropy(S) (8/14) x Entropy(Sweak)
(6/14) x Entropy(Sstrong)
= 0.940 (8/14) x 0.811 (6/14)*1.00
= 0.048

34
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Step 06 : Summary of results are

Entropy(S) = 0.940

Gain(S, Outlook) = 0.246

Gain(S, Temp) = 0.0289366072

Gain(S, Humidity) = 0.1515496

Gain(S, Wind) = 0.048

Step 07 : Find which attribute is the root node.

Gain(S, Outlook) = 0.246 i s highest.
Therefore "Outlook" attribute is the decision attribute in the root node.

"Outlook" as root node has three possible values - sunny, cloudy, rain.

S.No Day
Outlook Temp Humidity Wind Play
1 11 sunny 75 Medium 70 Normal strong yes
2 2 sunny 80 Hot 90 High strong no
3 1 sunny 85 Hot 85 High week no
4 8 sunny 72 Medium 95 High week no
5 9 sunny 69 Cold 70 Normal week yes
6 12 cloudy 72 Medium 90 High strong yes
7 3 cloudy 83 Hot 78 High week yes
8 7 cloudy 64 Cold 65 Normal strong yes
9 13 cloudy 81 Hot 75 Normal week yes
10 14 rainy 71 Medium 85 High strong no
11 6 rainy 65 Cold 70 Normal strong no
12 10 rainy 75 Medium 80 Normal week yes
13 5 rainy 68 Cold 80 Normal week yes
14 4 rainy 70 Medium 96 High week yes

35
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Step 08 : Find which attribute is next decision node.

Outlook has three possible values,
so root node has three branches (sunny, cloudy, rain).

Ssunny = {D1, D2, D8, D9, D11} = 5 "examples". "D" represent "Days".

With outlook = sunny

Gain(Ssunny, Humidity) = 0.970

Gain(Ssunny, Temp) = 0.570

Gain(Ssunny, Wind) = 0.019

Humidity has the highest gain 0.970 is next decision node.

Step 09 : This process goes on until all days data are classified
perfectly or run out of attributes.
36
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Thus, the classification tree built using ID3 algorithm is shown
below. It tells if the weather was amenable to play ?

Decision tree
37
Outlook
Humidity
sunny
cloudy
rainy
High Normal Strong Weak
No play
D1 D2 D8
Yes play
D9 D11
No play
D6 D14
Yes play
D4 D5
Yes play
D3 D7 D12 D13
Wind
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Applications

ID3 is easy to use.

Its primary use is replacing the expert who would normally
build a classification tree by hand.

ID3 has been incorporated in a number of commercial rule-
induction packages.

Some specific applications include

medical diagnosis,

credit risk assessment of loan applications,

equipment malfunctions by their cause,

classification of plant diseases, and

web search classification.
38

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Explanation Based Learning (EBL)

4. Explanation Based Learning (EBL)
Humans appear to learn quite a lot from one example.
Human learning is accomplished by examining particular situations and relating
them to the background knowledge in the form of known general principles.
This kind of learning is called "Explanation Based Learning (EBL)".

4.1 General Approach
EBL is abstracting a general concept from a particular training example.
EBL is a technique to formulate general concepts on the basis of a specific
training example. EBL analyses the specific training example in terms of
domain knowledge and the goal concept. The result of EBL is an explanation
structure, that explains why the training example is an instance of the goal
concept. The explanation-structure is then used as the basis for formulating
the general concept.
Thus, EBL provides a way of generalizing a machine-generated explanation
of a situation into rules that apply not only to the current situation but to
similar ones as well.

4.2 EBL Architecture
The overall architecture of the EBL learning method.

39

Specific
goal /
problem
Generalizer
Partial
external
solution
Problem solver
(under-stander)
Knowledge
base
Explanation
New general
concept
Inputs
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

4.3 EBL System Schematic
The schematic below shows explanation based learning.

Explanation Based Learning

Input to EBL :
The blocks color yellow are external to EBL.
They are the 4 different kinds of input that EBL algorithm accepts.

1. Training example - Situation description, facts .
An example is a set of facts describing an instance of the goal concept.

2. Goal concept - Some predicate calculus statements .
A high level description of concept that the program is supposed to
learn; e.g., the definition of an object "Cup" in function-structure
involves functional properties (lift-able, stable, open-vessel . . .) rather
than structural features (light, part of, pointing . . .).

3. Domain theory - Inference rules .
It represents facts and rules that constitute what the learner knows.
The facts describe an instance of the goal concept and the rules
describe relationships between objects and actions in a domain; e.g.,
the cup domain includes facts: concavities, bases, and lugs, as well as
rules: about lift ability, stability and what makes an open vessel.

4. Operational criterion - description of concepts .
A description in an appropriate form of the final concept.
A predicate from domain theory over concept definition, specifying the
form in which the learned concept definition must be expressed.
40
Domain
theory
Operationality
criterion
Training
Example
Operationality
pruner
Generalizer
Problem
solver
Goal
Concept
Results
general
trace
trace
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Description of EBL algorithm
Given the four inputs just described, the task is to construct an
explanation in terms of the domain theory and then determine a set of
sufficient conditions under which the explanation holds.
Thus, EBL algorithm consisting of two stages:
1. Explain: Construct an explanation in terms of the domain theory that
shows how the training example satisfies the goal concept definition. This
explanation must be constructed so that each branch of the explanation
structure terminates in an expression that satisfies the operationality
criterion.
2. Generalize: Determine a set of sufficient conditions under which the
explanation holds, stated in terms that satisfy the operationality criterion.
This is accomplished by regressing (back propagating) the goal concept
through the explanation structure. The conjunction of the resulting
expressions constitutes the desired concept definition.

Algorithm : The step involved are

1. Given an example, an "explanation" is first constructed by applying a
problem solver to relate the example to some general domain theory.
The result of this operation is a trace or "proof" of the example with
respect to the theory.

2. Next, generalizes the explanation using the method of goal regression.
This involves traversing the tree from top to bottom, replacing constants
by variables, but just those constants that are not embedded in the rules
or facts used to create the proof. The result of this operation is the
generalized proof tree.

3. The tree is then pruned by removing irrelevant leaf nodes until no
operational predicates appear at internal nodes. The result of this
operation is that explanation structure satisfies the operationality criterion
that applies to other examples too.

4. Finally, the operational rules are extracted from the general explanation.
The definition of the target concept is conjunction of the remaining leaf
nodes.
41

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

4.4 Example : "CUP" Generalization problem

The EBL methodology described before is applied to concept "CUP".
The training example, domain theory, goal concept, operationality
criterion, and a proof tree generalization using a goal regression
technique are presented. The logical operators used are
^ and; implies; if .. then; if and only if; iff

Goal Concept
lift-able(x) ^ stable(x) ^ open-vessel(x) cup(x)

Training example
colour(Obj23, Blue) ^ has-part(Obj23, Handle16) ^ has-part(Obj23, Bottom19) ^
owner(Obj23, Ralph) ^ has-part(Obj23, Concavity12) ^ is(Obj23, Light) ^
is(Ralph, Male) ^ isa(Handle16,Handle) ^ isa(Bottom19, Bottom) ^
is(Bottom19, Flat) ^ isa(Concavity12, Concavity) ^
is(Concavity12, Upward-Pointing)

Domain theory
has-part(x,y) ^ isa(y,Concavity) ^ is(y, Upward-Pointing) open-vessel(x)
is(x, Light) ^ has-part(x,y) ^ isa(y,Handle) liftable(x)
has-part(x,y) ^ isa(y, Bottom) ^ is(y,Flat) stable(x)

Operationality Criterion

The concept definition is expressed in terms of structural features.

The first stage of the EBL process is to show why the training example is an
example of a cup. This represents a proof.

This proof is expressed only in terms of the operationality criterion and
irrelevant details discarded relating to the Owner and Color.

Explanation structure of the cup.
42

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Explanation Structure of the cup
The structure below represents the first stage of the EBL process.
It proofs why the training example is an example of a cup.

Proof tree generalization using a goal regression technique.
The above proof is now generalized by using a goal regression technique. In
this example replacing the constants with variables gives the required
generalization of the cup.
has-part(x, y) U isa(y, Concavity) U is(y, Upward-Pointing) U
has-part(x, z) U isa(z, Bottom) U is(z, Flat) U has-part(x,w) U
isa(w, Handle) U is(x, Light)

43

Cup(Obj1)
Open-vessel(Obj1)
Stable(Obj1)
Lift-able(Obj1)
i sa(Concavity1, Concavity)
has-part(obj1, Concavity)
is(Concavity1, Upward-Pointing)
Is(Obj1,light)
isa(Handle1,Handle)
has-part(Obj1,Handle1)
i sa(Concavity1, Fl at)
i sa(Concavity1, Bottom)
has-part(Obj1,Concavity1)
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Learning : Discovery
5. Discovery

Simon (1966) first proposed the idea that we might explain scientific discovery
in computational terms and automate the processes involved on a computer.
Project DENDRAL (Feigenbaum 1971) demonstrated this by inferring structures
of organic molecules from mass spectra, a problem previously solved only by
experienced chemists.

Later, a knowledge based program called AM the Automated Mathematician
(Lenat 1977) discovered many mathematical concepts.

After this, an equation discovery systems called BACON (Langley, 1981)
discovered a wide variety of empirical laws such as the ideal gas law. The
research continued during the 1980s and 1990s but reduced because the
computational biology, bioinformatics and scientific data mining have convinced
many researchers to focus on domain-specific methods. But need for research
on general principles for scientific reasoning and discovery very much exists.

Discovery system AM relied strongly on theory-driven methods of discovery.
BACON employed data-driven heuristics to direct its search for empirical laws.

These two discovery programs are illustrated in the next few slides.

44
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

5.1 Theory Driven Discovery

The Simon's theory driven science, means AI-modeling for theory building.
It starts with an existing theory represented in some or all aspects in form
of a symbolic model and one tries to transform the theory to a runable
program. One important reason for modeling a theory is scientific discovery
in the theory driven approach, this means the discovery of new theoretical
conclusions, gaps, or inconsistencies.
Many computational systems have been developed for modeling different
types of discoveries. The Logic Theorist (1956) was designed to prove
theorems in logic when AI did not exist. Among the more recent
systems, the Automated Mathematician AM (Lenat, 1979) is a good
example in modeling mathematical discovery.

AM (Automated Mathematician)
AM is a heuristic driven program that discovers concepts in elementary
mathematics and set theory. AM has 2 inputs:
(a) description of some concepts of set theory: e.g. union, intersection;
(b) information on how to perform mathematics. e.g. functions.
AM have successively rediscovered concepts such as :
(a) Integers , Natural numbers, Prime Numbers;
(b) Addition, Multiplication, Factorization theorem ;
(c) Maximally divisible numbers, e.g. 12 has six divisors 1, 2, 3, 4, 6, 12.
[AM is described in the next slide.]
45
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

How does AM work ?
AM employs many general-purpose AI techniques.
The system has around 115 basic elements such as sets, lists, elementary
relations. The mathematical concepts are represented as frames. Around
250 heuristic rules are attached to slots in the concepts. The rules present
hints as how to employ functions, create new concepts, generalization etc.
about activities that might lead to interesting discoveries.

The system operates from an agenda of tasks. It selects the most
interesting task as determined by a set of over 50 heuristics. It then
performs all heuristics it can find which should help in executing it. The
heuristics represented as operators are used to generalize, to specialize or
to combine the elementary concepts or relations to make more complex
ones. Heuristics can fill in concept slots, check the content of the slots,
create new concepts, modify task agenda, interestingness levels, etc.
Because it selects the most interesting task to perform at all times, AM is
performing the best-first search in a space of mathematical concepts.
However, its numerous heuristics (over 200) guide its search very
effectively, limiting the number of concepts it creates and improving their
mathematical quality.
46

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

5.2 Data Driven Discovery
Data driven science, in contrast to theory driven, starts with empirical data
or the input-output behavior of the real system without an explicitly given
theory. The modeler tries to write a computer program which generates the
empirical data or input-output behavior of the system. Typically, models are
produced in a generate-and-test-procedure. Generate-and-test means
writing program code which tries to model the i-o-behavior of the real
system first approximately and then improve as long as the i-o-behavior
does not correspond to the real system. A family of such discovery models
are known as BACON programs.

BACON System
Equation discovery is the area of machine learning that develops methods
for automated discovery of quantitative laws, expressed in the form of
equations, in collections of measured data.
BACON is pioneer among equation discovery systems. BACON is a family of
algorithms for discovering scientific laws from data.
BACON.1 discovers simple numeric laws.
BACON.3 is a knowledge based system, has discovered simple empirical
laws like physicists and shown its generality by rediscovering the
Ideal gas law, Kepler's third law, Ohm's law and more.

The next few slides shows how BACON1 rediscovers Keplers third Law and
BACON3 rediscovers Ideal Gas Law.
47
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

BACON.1 : Discovers simple numeric laws.
Given a set of observed values about two variables X and Y, BACON.1 finds
a function Y = f(X) using four heuristics:

Heuristic 1 : If Y has value V in several observed cases, make the
hypothesis that Y = V in all cases.

Heuristic 2 : If X and Y are linearly related with slope S and intercept I
in several observed cases, then make the hypothesis that Y = S X + I
is in all cases.

Heuristic 3 : If X increases as Y decreases, and X and Y are not
linearly related define a new term T as the product of X and Y
ie., T = X Y

Heuristic 4 : If X increases as Y increases, and X and Y are not
linearly related, define a new term T as the division of X by Y
ie., T = X / Y

Note : BACON1 iteratively applies these 4 heuristics until a scientific law is
discovered. Heuristics 1 and 2 detect linear relationships. Heuristics 3 and 4
detect simple non-linear relationships. Heuristics 1 and 2 produce scientific laws.
Heuristics 3 and 4 are intermediate steps.

Example : Rediscovering Keplers third Law
Keplers third Law is stated below. Assume the law is not discovered or known.
"The square of the orbital period T is proportional to the cube of the mean distance
a from the Sun." ie., T
2
= k a
3
, k is constant number, is same for all planets.
If we measure T in years and all distances in "astronomical units AUs" with 1 AU
the mean distance between the Earth and the Sun, then if a = 1 AU, T is one
year, and k with these units just equals 1, i.e. T
2
= a
3
.

Input : Planets, Distance from Sun ( D ) , orbit time Period ( P )
Planet D P
Mercury 0.382 0.241
Venus 0.724 0.616
Earth 1.0 1.0
Mars 1.524 1.881
Jupiter 5.199 11.855
Saturn 9.539 29.459

[continued in next slide]
48
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Apply heuristics 1 to 4 : (Iteration 1)
Try heuristic 1: not applicable, neither D nor P is constant
Try heuristic 2: not applicable, no linear relationship
Try heuristic 3: not applicable, D increasing P not decreasing
Try heuristic 4: applicable, D increases as P increase,
so add new variable D/P to the data set.

Adding new variable D/P to the data set:
Planet D P D/P
Mercury 0.382 0.241 1.607
Venus 0.724 0.616 1.175
Earth 1.0 1.0 1.0
Mars 1.524 1.881 0.810
Jupiter 5.199 11.855 0.439
Saturn 9.539 29.459 0.324

Try heuristic 1: not applicable, D/P is not constant
Try heuristic 2: not applicable, no linear relationship between D/P and D or P
Try heuristic 3: applicable, D/P decreases as D increases and as P
increases, so the system could add two new
variables: D (D/P) = D
2
/P or P (D/P) = D
but D already exists, so add new variable D
2
/P

Adding new variable D
2
/P to the data set:
Planet D P D/P D
2
/P
Mercury 0.382 0.241 1.607 0.622
Venus 0.724 0.616 1.175 0.851
Earth 1.0 1.0 1.0 1.0
Mars 1.524 1.881 0.810 1.234
Jupiter 5.199 11.855 0.439 2.280
Saturn 9.539 29.459 0.324 3.088

49
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Try heuristic 1: not applicable, D
2
/P is not constant
Try heuristic 2: not applicable, D
2
/P is not linearly related with any other
variable
Try heuristic 3: applicable, D
2
/P decreases as D/P increases, so add the
new variable: (D
2
/P) (D/P) = D
3
/P
2

Adding new variable D
3
/P
2
to the data set:
Planet D P D/P D
2
/P D
3
/P
2

Mercury 0.382 0.241 1.607 0.622 1.0
Venus 0.724 0.616 1.175 0.851 1.0
Earth 1.0 1.0 1.0 1.0 1.0
Mars 1.524 1.881 0.810 1.234 1.0
Jupiter 5.199 11.855 0.439 2.280 1.0
Saturn 9.539 29.459 0.324 3.088 1.0

Try heuristic 1: applicable,, D
3
/P
2
is constant

Conclusion : D
3
/P
2

This is Keplers third law. (took about 20 years to discover it !)

A limitation of BACON.1
It works only for target equation relating at most two variable.

50
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

BACON.3 :
BACON.3 is a knowledge based system production system that discovers
empirical laws. The main heuristics detect constancies and trends in data,
and lead to the formulation of hypotheses and the definition of theoretical
terms. The program represents information at varying levels of description.
The lowest levels correspond to direct observations, while the highest
correspond to hypotheses that explain everything so far observed. BACON.3
is built on top of BACON.1.

It starts with a set of variables for a problem. For example, to derive
the ideal gas law, it started with four variables, p, V, n, T.
p - gas pressure,

V - gas volume,

T - gas temperature,

n - is the number of moles.

Values from experimental data are inputted.

BACON holds some constant and try to notice trends in the data.

Finally draws inferences. Recall pV/nT = k where k is a constant.

BACON has also been applied to Kepler's 3rd law, Ohm's law,
conservation of momentum and Joule's law.
51

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Example :
Rediscovering the ideal gas law pV/nT = 8.32, where p is the pressure on a
gas, n is the number of moles, T is the temperature and V the volume of
the gas. [The step-by-step complete algorithm is not given like previous
example, but the procedure is explained below]

At the first level of description we hold n = 1 and T = 300 and vary p
and V. Choose V to be the dependent variable.

At this level, BACON discovers the law pV = 2496.0.

Now the program examines this phenomenon :
when n = 1 and T = 310 then pV = 2579.2. Similarly,
when n = 1 and T = 320 then pV = 2662.4.

At this point, BACON has enough information to relate the values of pV
and the temperature T. These terms are linearly related with an
intercept of 0, making the ratio pV/T equal to 8.32.

Now the discovery system can vary its third independent term.
while n = 2, the pV/T is found to be 16.64,
while n =3, the pV/T is found to be 24.96.

When it compares the values of n and pV/T, BACON finds another linear
relation with a zero intercept. The resulting equation, pV/nT = 8.32, is
equivalent to the ideal gas law.
52
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Learning : Clustering
6. Clustering
Clustering is a way to form natural groupings or clusters of patterns.
Clustering is often called an unsupervised learning.

Example : Three natural groups of data points, i.e., 3 natural clusters.
Y

X

Clustering is one of the most utilized data mining techniques.

The data have no target attribute.
The data set is explored to find some intrinsic structures in them.

Unlike Classification where the task is to learn to assign instances to
predefined classes, in Clustering no predefined classification is required.
The task is to learn a classification from the data.
53

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

6.1 Distance Functions

Euclidean geometry is study of relationships between angles and
distances in space. Mathematical spaces may be extended to any
dimension, called an n-dimensional Euclidean space or an n-space.

Let R denote the field of real numbers. The space of all n-tuples of real
numbers forms an n-dimensional vector space over R, denoted by R
n
.
An element of R
n
is written as X = (x
1
, x
2
, x
i
., x
n
), where x
i
is a
real number. Similarly the other element Y = (y
1
, y
2
, y
i
., y
n
) .

The vector space operations on R
n
are defined by
X + Y = (x
1
+ y
1
, X
2
+ y
2
, . . , x
n
+ y
n
) and
aX = (ax
1
, ax
2
, . . , ax
n
)

The standard inner product (ie dot product) on R
n
, given by
X Y = (x
1
y
1
+ x
2
y
2
+ . . . . + x
n
y
n
) is a real number.

This product defines a distance function (or metric) on R
n
by

d(X , Y) = ||X Y|| = (x
i
y
i
)
2

The (interior) angle between x and y is then given by

= cos
-1
( )

The inner product of X with itself is always nonnegative.

||X || = (x
i
- y
i
)
2

54
i=1
n
X Y
||X|| ||Y||
i=1
n
i=1
n
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Euclidean distance or Euclidean metric is the "ordinary" distance
between two points that one would measure with a ruler. The Euclidean
distance between two points P = (p
1
, p
2
, . . p
i
. . , xn) and Q = (q
1
, q
2
,
. . q
i
. . , q
n
) in Euclidean n-space, is defined as :

(p
1
q
1
)
2
+ (p
2
q
2
)
2
+ . . + (p
n
q
n
)
2
= (p
i
- q
i
)
2

Example : 3-dimensional distance
For two 3D points, P = (p
x
, p
y
, . . p
z
) and Q = (q
x
, q
y
, . . q
z
)
The Euclidean 3-space, is computed as :

(p
x
q
x
)
2
+ (p
y
q
y
)
2
+ (p
z
q
z
)
2

Manhattan distance also known as city block distance between two
points in an Euclidean space with fixed cartesian coordinate system is
the sum of the lengths of the projections of the line segment between
the points onto the coordinate axes.

Example : In a plane, the Manhattan distance between the point P
1
with coordinates (x
1
, y
1
) and the point P
2
with coordinates (x
2
, y
2
) is
|x
1
x
2
| + |y
1
y
2
|

Manhattan versus Euclidean distance:
The red, blue, and yellow lines represent
Manhattan distance. They all are of same
length as 12.

The green line represent Euclidian distance of
length 62 8.48.
55

i=1
n
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Minkowski Metric

Let Xi and Xj are data points (vectors) then Minkowski distance
between these two data points is :
dist(Xi, Xj) = ( (x
i1
x
j1
)
h
+ (x
i2
x
j2
)
h
+ . . + (x
ir
x
jr
)
h
)
1/h

where h is a positive integer.

Euclidean distance IF h = 2

dist(Xi, Xj) = ( (x
i1
x
j1
)
2
+ (x
i2
x
j2
)
2
+ . . + (x
ir
x
jr
)
2
)
1/2

= ( (x
i1
x
j1
)
2
+ (x
i2
x
j2
)
2
+ . . + (x
ir
x
jr
)
2
)

Weighted Euclidean distance

dist(Xi, Xj) = (w
1
(x
i1
x
j1
)
2
+ w
2
(x
i2
x
j2
)
2
+ . . + w
r
(x
ir
x
jr
)
2
)

Manhattan distance IF h = 1

dist (Xi, Xj) = |x
i1
x
j1
| + |x
i2
x
j2
| + . . + |x
ir
x
jr
|

Squared Euclidean distance: to place progressively greater weight
on data points that are further apart.
dist(Xi, Xj) = (x
i1
x
j1
)
2
+ (x
i2
x
j2
)
2
+ . . + (x
ir
x
jr
)
2

Chebychev distance: to define two data points as "different" if they
are different on any one of the attributes.
dist(Xi, Xj) = max (|x
i1
x
j1
| , |x
i2
x
j2
| , . . , |x
ir
x
jr
|)

56
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

6.2 K-Mean Clustering
k-means clustering is an algorithm to classify or to group objects based on
attributes/features into K number of group.
K is positive integer number.
The grouping is done by minimizing the sum of squares of distances
between data and the corresponding cluster centroid.

K-Mean Clustering algorithm

The algorithm consists of 3
steps. First take any random
object as initial centroid, then
the three steps mentioned
below are iterated until
converge.

Step-1 Determine centroid
coordinate
Step-2 Determine distance of each
object to the centroids

Step-3 Group the object based on
minimum distance

The next slide explains step by step the complete algorithm
57

Start
Objects distance
from Centroid
Grouping based on
min distance
Number of
cluster P
No Object
Move to group

Centroid
END
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Example : K-Mean Clustering

Objects : 4 medicines as A, B, C, D.

Attributes : 2 as X is weight & Y is PH

Objects Attributes
X Y
A 1 1
B 2 1
C 4 3
D 5 4

Initial value of centroids:
Suppose medicine A and medicine B
be first centroids. If P
1
and P
2
denote
coordinates of the centroids,
then P
1
= (1, 1) and P
2
= (2, 1).

Let centroid P1 be for cluster group-1
Let centroid P2 be for cluster group-2.

58

4.5
4.0 D
3.5
3.0 C
2.5
2.0
1.5
1.0 A B
0.5

Y
0
0 1 2 3 4 5

X

4.5
4.0 D
3.5
3.0 C
2.5
2.0
1.5
1.0 A B
0.5

Y
0
0 1 2 3 4 5

X

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

1. Iteraion 0

(a)
Objects Clusters centers stated before :
Group-1 has center P
1
= (1, 1) ;
2
= (2, 1) ;
Objects : A , B, C, D
Attributes : X and Y
A B C D
X 1 2 4 5
Y 1 1 3 4

(b)
Calculate distances between cluster center to each object
1st, calculate the Euclidean distances from cetroid P
1
to each point A, B,
C, D. It is the 1st row of the distance matrix.
2nd, calculate the Euclidean distances from cetroid P
2
to each point A, B,
C, D. It is the 2nd row of the distance matrix.
The ways to calculate just two distance matrix elements D
13
and D
23
are :

D
13
= (Cx P1x)
2
+ (Cy P1y)
2
= (4 1)
2
+ (3 1)
2
= 3.61

D
23
= (C
x
P
2x
)
2
+ (C
y
P
2y
)
2
= (4 2)
2
+ (3 1)
2
= 2.83

Similarly calculate other elements D
11
, D
12 ,
D
14 ,
D
21
, D
22 ,
D
24

(c)
Distance matrix becomes
0 1 3.61 5
D
0
=
1 0 2.83 4
A B C D

1st row indicates group-1 cluster
2nd row indicates group-2 cluster

(d) Objects clustering into groups:
Assign group to each object based on the minimum distance. Thus,
medicine A is assigned to group 1;
medicine B is assigned to group 2,
medicine C is assigned to group 2 , and
medicine D is assigned to group 2.
Group Matrix : matrix element is 1 if the object is assigned to that group

1 0 0 0
G
0
=
0 1 1 1
A B C D

1st row as group 1

2nd row as group 2

59

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

2. Iteration 1 :
The cluster groups have new members. Compute the new centroids of
each group. Repeat the process of iteration indicated below.

Group-1 has one member A, the
centroid remains as P
1
= (1, 1).
Group-2 now has 3 members B,
C, D, so centroid is the average
of their coordinates:
P
2
= ( , )
= (11/3 , 9/3)
= (3.66, 3)

(a)
Objects Clusters centers stated above :
1
= (1, 1) ;
2
= (3.66, 3) ;
A B C D
X 1 2 4 5
Y 1 1 3 4

(b)
[The method is same as in Iteration 0 (b)]
1
to each point A, B,
C, D You get D
11
, D
12 ,
D
13
, D
14
as the 1st row of the distance matrix.
2
to each point A, B,
C, D. You get D
21
, D
22 ,
D
23
, D
24
as the 2nd row of the distance matrix.

(c)
0 1 3.61 5
D
1
=
3.14 2.36 0.47 1.89


Assign group to each object based on the minimum distance.
medicine A no change, remains in group 1;
medicine B moves to group 1,
medicine C no change, remains in group 2 ,
medicine D no change, remains in group 2.

1 1 0 0
G
1
=
0 0 1 1
A B C D


60

2+4+5
3
1+3+5
3

4.5
4.0 D
3.5
3.0 C
2.5
2.0
1.5
1.0 A B
0.5

Y
0
0 1 2 3 4 5

X

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

3. Iteration 2 :
The cluster groups have new members. Compute the new centroids of
each group. Repeat the process of iteration indicated below.

Group-1 has 2 members, so the
new centroid is the average
coordinate among A and B :
P
1
= ( , ) = (1.5, 1)
Group-2 has 2 members, so the
new centroid is the average
coordinate among C and D :
P
2
= ( , ) = (4.5, 3.5)

(a)
Objects Clusters centers stated above :
1
= (1.5, 1) ;
2
= (4.5, 3.5) ;
A B C D
X 1 2 4 5
Y 1 1 3 4

(b)
[The method is same as in Iteration 0 (b) or Iteration 1 (b)]
1
to each point A, B,
C, D You get D
11
, D
12 ,
D
13
, D
14
as the 1st row of the distance matrix.
2
to each point A, B,
C, D. You get D
21
, D
22 ,
D
23
, D
24
as the 2nd row of the distance matrix.

(c)
0.5 0.5 3.2 4.61
D
2
=
4.3 3.54 0.71 0.71


Assign group to each object based on the minimum distance.
medicine A no change, remains in group 1;
medicine B no change, remains in group 1;
medicine C no change, remains in group 2 ,
medicine D no change, remains in group 2.

1 1 0 0
G
2
=
0 0 1 1
A B C D


61

4.5
4.0 D
3.5
3.0 C
2.5
2.0
1.5
1.0 A B
0.5

Y
0
0 1 2 3 4 5

X

1+2
2
1+1
2
4+5
2
3+4
2
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

4. Finally, the result state that Group Matrix G
2
= G
1
.
This means the grouping of objects in this last iteration and the one
before does not change anymore. Thus, the computation of the k-mean
clustering has reached its stability and no more iteration is needed.
Results of final grouping are :

Objects Feature 1 (X)
Weight index

Feature 2 (Y)
pH
Cluster Group
Medicine A 1 1 1
Medicine B 2 1 1
Medicine C 4 3 2
Medicine A 3 4 2

62
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Learning by analogy
7. Analogy

Learning by analogy means acquiring new knowledge about an input entity by
transferring it from a known similar entity.
This technique transforms the solutions of problems in one domain to the
solutions of the problems in another domain by discovering analogous states and
operators in the two domains.

Example: Infer by analogy the hydraulics laws that are similar to Kirchoff's laws.

Qa = 3 Qb = 9

Qc = ?

Hydraulic Problem

I1 I2

I3 = I1 + I2

Kirchoff's First Law

The other similar examples are :

Pressure Drop is like Voltage Drop

Hydrogen Atom is like our Solar System :

The Sun has a greater mass than the Earth and attracts it, causing the Earth to
revolve around the Sun. The nucleus also has a greater mass then the electron
and attracts it. Therefore it is plausible that the electron also revolves around the
nucleus.

63
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Learning - Neural net and Genetic learning

8. Neural net and Genetic Learning
The Neural net, the Genetic learning and the Reinforcement learning are the
Biology-inspired AI techniques. In this section the Neural net and Genetic
learning are briefly described.

8.1 Neural Net (NN)
A neural net is an artificial representation of the human brain that tries to
simulate its learning process. An artificial neural network (ANN) is often
just called a "neural network" (NN).

Neural Networks model a brain learning by example.

Neural networks are structures "trained" to recognize input patterns.

Neural networks typically take a vector of input values and produce a
vector of output values; inside, they train weights of "neurons".

A Perceptron is a model of a single `trainable' neuron.

64
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Learning - Neural net

Perceptron is a model of a single `trainable' neuron.

Inputs are x1, x2, . . . , xn as real numbers or
boolean values depending on the problem.

w1, w2, . . . , wn are weights of the edges and are real valued.

T is the threshold and is real valued.

Y is the output and is boolean.

If the net input which is (w1 x1 + w2 x2 + ... + wn xn) is greater than
the threshold T then output Y is 1 else 0.

Neural network uses supervised learning, in which inputs and outputs
are known and the goal is to build a representation of a function that
will approximate the input to output mapping.
65
X
1

X
2

X
3

X
n

T Y
w
1
w
2
w
3
w
n
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Learning - Genetic learning

8.2 Genetic Learning
Genetic algorithms (GAs) are part of evolutionary computing.
GA is a rapidly growing area of AI.

Genetic algorithms are implemented as a computer simulation,
where techniques are inspired by evolutionary biology.

Mechanics of biological evolution:

Every organism has a set of rules, describing how that organism is
built, and encoded in the genes of an organism.

The genes are connected together into long strings called
chromosomes.

Each gene represents a specific trait (feature) of the organism and
has several different settings, e.g. setting for a hair color gene may
be black or brown.

The genes and their settings are referred as an organism's
genotype.

When two organisms mate they share their genes. The resultant
offspring may end up having half the genes from one parent and
half from the other. This process is called cross over.

A gene may be mutated and expressed in the organism as a
completely new trait.

Thus, Genetic Algorithms are a way of solving problems by mimicking
processes the nature uses ie Selection, Crosses over, Mutation and
Accepting to evolve a solution to a problem.
66

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Learning - Genetic learning

Genetic Algorithm steps

(1)
[Start] Generate random population of n chromosomes (Encode
suitable solutions for the problem)

(2)
[Fitness] Evaluate the fitness f(x) of each chromosome x in the
population.

(3)
[New population] Create a new population by repeating following
steps until the new population is complete.

(a) [Selection] Select two parent chromosomes from a population
according to their fitness.

(b) [Crossover] With a crossover probability, cross over the parents
to form new offspring (children). If no crossover is performed,
the offspring would be the exact copy of parents.

(c) [Mutation] With a mutation probability, mutate the new offspring
at each locus (position in chromosome).

(d) [Accepting] Place new offspring in the new population.

(4)
[Replace] Use new generated population for a further run of the
algorithm.

(5)
[Test] If the end condition is satisfied, stop, and return the best
solution in the current population.

(6)
[Loop] Go to step 2.

Genetic Algorithms does unsupervised learning - the right answer is not
known beforehand.
67

R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Learning - Reinforcement learning

9. Reinforcement Learning
Reinforcement learning refers to a class of problems in machine learning which
postulate an agent exploring an environment.
The agent perceives its current state and takes actions.
The environment, in return, provides a reward positive or negative.
The algorithms attempt to find a policy for maximizing cumulative reward for
the agent over the course of the problem.

In other words, the definition of Reinforcement learning is :
" A computational approach to learning whereby an agent tries to maximize the
total amount of reward it receives when interacting with a complex, uncertain
environment."

Reinforcement learning is a way of programming agents by reward and
punishment without needing to specify how the task is to be achieved.
[Kaelbling, Littman, & Moore, 96]

The RL Problem and the tasks are illustrated in the next few slides.

68
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

9.1 Reinforcement Learning (RL) Problem
Reinforcement Learning is a computational approach to understanding
and automating goal-directed learning and decision-making.

Agent - Environment interaction

RL emphasizes learning from interaction with its environment.
RL uses a formal framework defining the interaction between a learning agent
and its environment in terms of states, actions, and rewards.
The state-action pairs form value functions that does efficient search in the
space of policies guided by evaluations of all policies.

Agent - Environment interaction in Reinforcement learning

The agent and environment interact in a sequence of discrete time
steps, t = 0, 1, 2, 3, 4, - - -

At each discrete time t, the agent (learning system) observes state
s
t
S and chooses action a
t
A

Then agent receives an immediate numerical reward r
t+1

and
the state changes to s
t+1

At each time step, the agent implements a mapping from states to
probabilities of selecting each possible action.

This mapping is called the agent's policy
t
,
where
t
(s,a) is the probability at a
t
= a if s
t
S.

RL methods specify how the agent changes its policy as a result of
its experience.

The agent's goal, is to maximize the total amount of reward it
receives over the long run.
69
Agent
Environment
state s
t

reward r
t
action a
t

s
t + 1
r
t + 1
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Key Features of RL

The learner is not told what actions to take, instead it find finds out
what to do by trial-and-error search.

The environment is stochastic; ie., the behavior is non-deterministic
means a "state" does not fully determine its next "state".

The reward may be delayed, so the learner may need to sacrifice short-
term gains for greater long-term gains.

The learner has to balance between the need to explore its environment
and the need to exploit its current knowledge.
70
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

9.2 Reinforcement Learning tasks

Most RL research is conducted within the mathematical framework of
Markov decision processes (MDPs).
The MDPs involve a decision-making agent interacting with its
environment so as to maximize the cumulative reward it
receives over time.
The agent perceives aspects of the environment's state and
selects actions.
The agent may estimate a value function and use it to construct
better and better decision-making policies over time.

Thus the elements of Reinforcement Learning are

Markov Decision Process

Policy

Agents Learning Task

Reward

Value Functions

Each of these elements are explained in the next few slides. However, first
the Stochastic process, Markov chain, and few notations are explained.
71
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Recall : Stochastic processes, Markov chain , and the notations.

Stochastic processes
A physical stochastic processes is a sequences of events or path
governed by probabilistic laws.
A stochastic process X = { X(t), t T } is a collection of random
variables. For each t in the index set T, the X(t) is a random
variable. The t is often interpreted as time and X(t) the state of
the process at time t.
If the index set T is a countable set,
then we have a discrete-time stochastic process;
Else if T is a non-countable continuous set,
then we have a continuous-time stochastic process.
Any realization of X is named a sample path, which can be
discrete or continuous.

Markov chain
A Markov chain, is a stochastic process with the Markov property.
Markov property means, given the present state, the future states
are independent of the past states; means the future states depends
only on its present state, and not on any past states.

In other words, the description of the present state fully captures all
the information that could influence the future evolution of the process.
Future states will be reached through a probabilistic process instead
of a deterministic one.

At each step the system may change its state from the current
state to another state, or remain in the same state, according to a
certain probability distribution. The changes in state are called
transitions, and the probabilities associated with various state-changes
are called transition probabilities.

72
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Notations followed

t discrete time step
T final time step of an episode
s
t
state perceived by agent at time t
a
t
action selected by agent at time t
r
t+1
reward received by agent at time t
s
t +1
resulting next state perceived by agent at time t

R
t
return (cumulative discounted reward) following t
R
t
(n)
n - step return
R
t
- return
policy, decision-making rule
(s) action taken in state s under deterministic policy
(s,a) probability of taking action a in state s under stochastic policy

S set of all non-terminal states
S
+
set of all states, including the terminal state
A(s) set of actions possible in state s
p
a
ss
probability of transition from state s to state s under action a

r
a
ss
expected immediate reward on transition from state s to state
s under action a

73

r
t+1
r
t+2
r
t+3

a
t
a
t+1 a
t+2
a
t+3
s
t
s
t+1
s
t+2
s
t+3
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Markov System

A system that can be in one of several specified states with a specified
time step, and at each step the system will randomly change states
or remain where it is according to some fix probability. The probability
of going from State i to State j is a fix number p
ij
, called the
transition probability.

Example : Laundry Detergent Switching
A market analyst for Detergents is interested to know whether consumers
prefer powdered detergents or liquid detergents.
Two market surveys taken 1 year apart revealed that :
20% of all powdered detergent users had switched to liquid 1 year
later, while the rest were using powder.
10% of liquid detergent users had switched to powder 1 year later,
with the rest using liquid.

Analysis of the survey

1. Every year a consumer may be in one of two possible states:
State 1, he is a powdered detergent user;
State 2, he is a liquid detergent user.

2. There is a basic time step of 1 year;

3. If a consumer during a given year is in State 1, a powdered detergent
user, then the probability that he will be in State 2, a liquid detergent
user, in the next year is 20% = 0.2 denoted by p
12
= 0.2 .
Therefore probability for him to remain in State 1 is thus (1- 0.2 = 0.8),
denoted by p
11
= 0.8.

4. If a consumer during a given year is in State 2, a liquid detergent user,
then the probability that he will be in State 1, a powered detergent
user, in the next year is 10% = 0.1 denoted by p
21
= 0.1.
Therefore probability for him to remain in State 2 is thus (1 - 0.1 = 0.9),
denoted by p
22
= 0.9.

The state transition diagram is shown in the next slide.
74
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

[Continued from previous slide - Example : Laundry Detergent Switching]

We can draw a state transition diagram or use a matrix to represent these
probabilities called transition probabilities.

State transition diagram State transition matrix

The Figure 1 and 2 : the numbers are transition probabilities, from
State i to State j , a number p
ij
, and the arrows show the switching
directions.

75
2 1
0.2
0.1
0.8 0.9
0.8 0.2
P =
0.9 0.1
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Markov State Transition
State transition probabilities mentioned before are explained here.

Transition probability :
If a Markov system is in state i, there is a fixed probability p
ij
of it
going into state j in the next time step. The probability p
ij
is called
a transition probability.

Transition diagram :
A Markov system is illustrated by a state transition diagram, that
shows all the states and transition probabilities.

Example : Transition Diagram (Missing arrows indicate zero probability.)

Transition matrix :
The matrix P whose ij
th
entry is p
ij
is called transition matrix
associated with the system. The entries in each row add up to 1.
Example :
A 2 x 2 transition matrix P
To

P11 P12 Arrows originating in state 1

P21 P22 Arrows originating in state 2

The 3x3 Matrix corresponding to Transition Diagram

To

0.2 0.8 0
0.4 0 0.6

0.5 0.35 0.15

76

1 2
From
1
2
1 2
From
1
2
3
3
2
3
1
0.8
0.4
0.6
0.3
0.5
0.2
0.1
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Markov Decision Processes (MDPs)
Markov decision processes (MDPs), provide a mathematical framework
for modeling decision-making in situations where outcomes are partly
random and partly under the control of the decision maker.

Description :
MDP is a discrete time stochastic control process characterized by a
set of states;
In each state there are several actions from which the decision maker
must choose. For a state s and an action a, a state transition function
Pa(s) determines the transition probabilities to next state.
The decision maker earns a reward r for each state transition. The
state transitions possess the Markov property, ie., given the state of
the MDP at time t, the transition probabilities to the state at time t+1
are independent of all previous states or actions.
MDPs are an extension of Markov chains. The differences are the
addition of actions (allowing choice) and the rewards (giving motivation).
If there are only one action, or if the action to take were fixed for each
state, a Markov decision process would reduce to a Markov chain.

State transitions, Actions and Rewards

Set of states S

Set of actions A(s) available in each state S

Markov assumption: s
t+1
and r
t+1
depend only on s
t
, at and not
on anything that happened before t

Rewards: on transition from state S to state S under action a

r
a
ss
= E r
t
+1 | s
t
= s, a
t
= a, s
t+1
= s

Transition probabilities : from state S to state S under action a

p
a
ss
= P ( s
t+1
= s |s
t
= s, a
t
= a )

77

r
t+1
r
t+2
r
t+3

a
t
a
t+1 a
t+2
a
t+3
s
t
s
t+1
s
t+2
s
t+3
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Agents Learning Task
Agent execute actions in the environment.

Observe results, and learn policy which is a strategy (a way of behaving).
: S A [0, 1] with (s, a) = P(a
t
= a |s
t
= s)

If the policy is deterministic, then write more simply as
: S A with (s) = a
giving the action chosen in the state S.

Policy
It defines the learning agent's way of behaving at a given time.
A mapping from states to actions (deterministic policy), or the distributions
over actions (stochastic policy).

It is a mapping from perceived states of the environment to actions
to be taken in those states.

It in some cases may be a simple function or lookup table;
it may involve extensive computation like a search process.

It is the core of a reinforcement learning agent;
it alone is sufficient to determine behavior.

In general it may be stochastic (random).
78
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Reward Function
The function defines the goal in a reinforcement learning problem.

Maps each perceived state (or state-action pair) of the environment
to a single number, indicating intrinsic desirability of that state.

Indicates what is good in an immediate sense.

Tells what are the good and bad events for the agent;
Example : a biological system identify rewards high with pleasure
and low with pain

RL agent's sole objective is to maximize the total reward it receives
in the long run.

Rewards servers a basis for altering the policy;
Example : an action selected by the policy if followed by low
reward, then the policy may be changed to select some other action
in that situation.

In general, the reward function may be stochastic.
79
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o

Maximize Reward
Agent's goal is to maximize reward it receives in long run.
"AgentEnvironment" interaction is generally a sequence of episodes.
If the interaction result into a sequence of separate episodes, it is known
as episodic tasks. If the interaction does not break into identifiable
episodes, but continue without limit is called continuing tasks.

If Episodic task then the simplest return R
t
is sum of rewards
Rt = r
t+1
+ r
t+2
+ + r
T
,
where T is the time when a terminal state is reached.

If Continuing tasks then the discounted return R
t

Rt = r
t+1
+ r
t+2
+ 2 r
t+3
+ = t+k-1 r
t+k
,
where parameter is discount factor, 0 1.

If Average-reward tasks then return R
t

Rt = limt (1/T) (r
t+1
+ r
t+2
+ + r
T
)
T

Value Functions

The value of state s under policy is the expected return when starting from
s and choosing actions according to :

(s) = { R
t
| s
t
= s } = {
t+k-1
r
t+k
| s
t
= s }

80

V

K=1

E

E

K=1
K=1
R
C

C
h
a
k
r
a
b
o
r
t
y
,

w
w
w
.
m
y
r
e
a
d
e
r
s
.
i
n
f
o
AI Learning - References
10. References : Textbooks

1.
"Artificial Intelligence", by Elaine Rich and Kevin Knight, (2006), McGraw Hill
companies Inc., Chapter 17, page 447-484.

2.
"Artificial Intelligence: A Modern Approach" by Stuart Russell and Peter Norvig,
(2002), Prentice Hall, Chapter 18-21, page 649-788.

3.
"Computational Intelligence: A Logical Approach", by David Poole, Alan Mackworth,
and Randy Goebel, (1998), Oxford University Press, Chapter 11, page 397-438.

4.
"Artificial Intelligence: Structures and Strategies for Complex Problem Solving", by
George F. Luger, (2002), Addison-Wesley, Chapter 10- 13, page 385-570.

5.
"AI: A New Synthesis", by Nils J. Nilsson, (1998), Morgan Kaufmann Inc., Chapter
10, Page 163-178.

6.
"Artificial Intelligence: Theory and Practice", by Thomas Dean, (1994),
Addison-Wesley, Chapter 5, Page 179-254

7.
Related documents from open source, mainly internet. An exhaustive list is
being prepared for inclusion at a later date.

81

06 Learning Systems

Uploaded by

Copyright:

Available Formats

06 Learning Systems

Uploaded by

Document Information

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

06 Learning Systems

Uploaded by

Copyright:

Available Formats

R

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example: Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Learning from example : Induction

AI Explanation Based Learning (EBL)

AI Explanation Based Learning (EBL)

AI Explanation Based Learning (EBL)

AI Explanation Based Learning (EBL)

AI Explanation Based Learning (EBL)

AI Learning - Neural net and Genetic learning

AI Learning - Neural net

AI Learning - Genetic learning

AI Learning - Genetic learning

AI Learning - Reinforcement learning

AI Learning - Reinforcement learning

AI Learning - Reinforcement learning

AI Learning - Reinforcement learning

AI Learning - Reinforcement learning

AI Learning - Reinforcement learning

AI Learning - Reinforcement learning

AI Learning - Reinforcement learning

AI Learning - Reinforcement learning

AI Learning - Reinforcement learning

AI Learning - Reinforcement learning

AI Learning - Reinforcement learning

AI Learning - Reinforcement learning

You might also like

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.