Scribd
Upload
38 views16 pages

ADALINE:Machine Learning Application

This document discusses machine learning and checkers games as an example. It describes training a program to play checkers by: 1) Defining the task as playing checkers and measuring performance as percentage of games won. 2) Providing training experience by having the program play games against itself to learn from different board states and moves. 3) Choosing a target function to predict board value based on features like piece count, in order to learn optimal moves through a least mean squares algorithm.

Uploaded by

kumar shanu rao
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
38 views16 pages

ADALINE:Machine Learning Application

This document discusses machine learning and checkers games as an example. It describes training a program to play checkers by: 1) Defining the task as playing checkers and measuring performance as percentage of games won. 2) Providing training experience by having the program play games against itself to learn from different board states and moves. 3) Choosing a target function to predict board value based on features like piece count, in order to learn optimal moves through a least mean squares algorithm.

Uploaded by

kumar shanu rao
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 16

ADALINE:Machine Learning

Application
Dr.S.G.Sanjeevi
Professor
Department of Computer Science &
Engineering, NIT Warangal
Machine Learning: Definition
• Computer Program learns from
• experience E
• Some Class of Tasks T
• Performance measure P
• If P improves with E
Successful Applications
• Learning to recognize Spoken Words
• Learning to drive Autonomous Vehicle –
ALVINN System
• Classify Astronomical Structures Using
Decision Trees – NASA
• Play Backgammon
• Play Chess
A Checkers learning problem
• Task T: playing Checkers
• P: percent of games won against
opponents
• Training Experience E: playing practice
games against itself
Handwritten recognition
problem
• Task T: recognizing, classifying hand
written words
• P: percent of words correctly classified
• Training experience E: A database of hand
written words with their given
classifications
Designing a learning system
• To make program play checkers in the
world checkers tournament
• Direct training experience – checker board
states and correct moves
• Indirect information – move sequences
and final outcomes
• Credit assignment- degree to which each
move contributed to final outcome
Extent to which learner controls training examples

• Teacher might provide board states with


correct moves
• Learner proposes board states and
teacher provides correct moves
• Learner learns by playing against itself
Distribution of training examples
• Training examples and future test
examples should be from similar
distribution
• If a learner learnt by playing against itself
then it might encounter novel board states
when playing a human champion
Checkers learning problem
• Task T: to play checkers
• Performance measure P:Percent of games
won in the world tournament
• Training experience E: games played
against itself
Choosing the target function
• Program needs to choose the best move
among legal moves
• Function Choose move : B → M where B
is board state, M legal move
• Learn choose move
• Program has indirect training experience
• Target function program learns is
• V : B → R // R is a real value
V(b) for an arbitrary
board state b
• 1. if b is final state with win then V(b) =
100
• 2. if b is final state with loss then V(b) = -
100
• 3. if b is final state with draw then V(b) = 0
• 4. if b is not final state with win then V(b) =
V(b’) where b’ is the best final state that
can be achieved from b playing optimally
Operationally feasible target
function
• V(b) is not efficiently computable above
• So we find operationally feasible target
function Ŷ
• Choosing Target function
• X1 no of black pieces
• X2 no of red pieces
• X3 no of black kings on the board
• X4 no of red kings on the board
• X5 no of black pieces threatened by red
• X6 no of red pieces threatened by black
• Target function Ŷ(b) = wo+ w1x1+w2x2 + w3x3+
w4x4+ w5x5+w6x6 where w1 to w6 are weights,
w0 is an additive constant
• To learn program needs training examples of the
form < b, Vtrain(b)>
• To compute Vtrain(b) if b is final state is easy
• If b is not final state then Vtrain(b) =
Ŷ(successor(b)) where successor(b) is the
next board state for which it is the
program’s turn to move

• Adjusting the weights:


• Error defined by
• LMS Weight Update rule (Least mean
squares rule)
• For each training example <b, Vtrain(b)>
• Use the current weights of Ŷ( b)
• For each weight Wi = Wi + µ(Vtrain(b)-
Ŷ( b))xi where µ is a small constant.

You might also like