Notes on Machine Learning - Spring 2003 | ECS 271, Study notes of Computer Science

Download Notes on Machine Learning - Spring 2003 | ECS 271 and more Study notes Computer Science in PDF only on Docsity!

ECS 271 Machine Learning

Rao VemuriSpring 2003

ECS 271: Outline

• Introduction• Concept Learning• Decision Tree Learning• Artificial Neural Nets• Computational Learning Theory• Bayesian Learning• Genetic Algorithms + other topics

Grading

• 45% for a term paper• 25% Mid term (in class, Closed book)• 15% {Home Work Assignments}• 15% Final (Take Home, 24 hr turn-around)
  • Each student reads and critiques two term

papers from other students. These critiqueshave no influence on how I grade the termpapers.

Outside Reading

• The text book is really at an upper division

UG level. You are expected to read outsidepapers to boost the theoretical background.

Term Paper

•^

Choose the topic. Any problem with which youare

very familiar

would be a good topic

•^

Choose a ML Technique discussed in the book

-^

Search the WWW for a good data set

-^

Implement the method, get the results

-^

Write a report (see class web site for more details)

-^

This is a 4-credit course. I expect you to devote aminimum of 20 hours on this project alone andanother 20 hours for the rest of the course.

What is Machine Learning?

•^

Learning is– making

useful

changes in our minds (Minsky)

• constructing or modifying

representations

of what is

being

experienced

(Michalski)

• knowledge acquisition in the absence of explicit

programming (Valiant)

• An adaptive process that enables a system to do the

same type of task more effectively the next time around(Simon)• “type” means “drawn from the same population”

Why ML? Why Now?

•^

Recent progress in algorithms and theory

-^

Growing flood of on-line data

-^

Growing computational power

-^

Easier to build than similar programmed systems– Search engines– Computer games (TD-gammon)– User interfaces/Personal assistants (paper clip asst.)– Robot exploring Mars/ Vacuum Cleaner

-^

Better understanding of human/animal learning

Opportunities for Tomorrow

• Learn across full mixed-media data• Learn across multiple databases
  • Internal databases– Web and Newsfeeds
    • Learn by active experimentation• Learn to make decisions rather than

predictions

Types of Learning

• Rote Learning (memorization)• Learning from examples (

generalize

from

specific instances)

• Learning by being told (

compile

abstract

advise)

• Learning from scalar feedback

( reinforcement

learning)

What is the Learning Problem?

•^

Given– a task T– a performance measure P– some experience E

-^

Learn to do the task

-^

Example: Learn to play T = checkers until P =90% of the campus-wide tournaments are won, byusing E = the opportunity to play against player A.

How Do You Learn Checkers?

•^

What exactly is the meaning of “learning to playcheckers”?– Given a board, what is the next move?– Given a board, what is the best move from among

{legal moves}?

•^

The {legal moves}defines a large search space.From this space we choose a move. That is, we arelearning a target function V, given the board.– Target function, V: Board

Æ

Real Number

• ChooseMove: Board

Æ

Move

Target Functions

• Define a function that assigns a score to the

board

• The better the board, the better the score

assigned.

• Then our search is really to maximize the

score.

V for Checkers

• If

b

is the final board state

–

V(b)

= 100, for a winning state

–

V(b)

= -100, for a losing state

–

V(b)

= 0, for a draw

• If

b

is NOT the final board state

–

V(b) = V(b`),

where

b`

is the best final board

state that can be reached from

b

while playing

optimally (This is non much of a help)

Operational Definition

• What we want is a function that we can

compute.

• What exactly is the value V of a board?
  • We do not know. The best we can do is

approximate with something we can compute

• The most friendly computable functions are

polynomials

0

1

1

"

...

n^

n

V

w

w x

w x

=

+

0 +

1 1

...^

n^ n

V^

w^

w x

w x

=^

+^

0

1 1

...^

n^ n

V^

w^

w x

w x

=^

+^

0

1 1

...^

n^ n

V^

w^

w x

w x

=^

+^

0

1 1

...^

n^ n

V^

w^

w x

w x

=^

+^

Representation of Learned

Function, V”

•^

V” is an approximation to the unknown V

-^

Xi’s are features of the board

-^

X1= # of black pieces

-^

X2 = # of red pieces

-^

X3 = # of black kings

-^

X4 = #of red kings

-^

X5 = # of blacks threatened by reds

-^

X6 = # of reds threatened by blacks

A Concept Learning Problem

• Consider a task• The task is to learn a concept described by

Boolean attributes

• Both positive and negative examples are

provided by a teacher– The concept to be learned is “dog”

is_mammal

has_claws

¬can_fly

has_tail

can_bark

∧

∧

∧

∧