Ch9' Genetic Algorithms

1
Ch9. Genetic Algorithms

• 2000. 10. 18
• ? ? ?

2
??

• 1. Motivation
• 2. Genetic Algorithms
• 2.1 Representing Hypotheses
• 2.2 Genetic Operators
• 2.3 Fitness Function and Selection
• 3. An Illustrative Example
• 1. Extensions
• 4.Hypothesis Space Search
• Schema Theorem
• 5. Genetic Programming
• 6. Models Of Evolution And Learning
• 7. Summary

3
1. Motivation

• GA method is motivated by an analogy to
  biological evolution.
• Evolution is known to be a successful, robust
  method for adaptation within biological systems.
• GA can search spaces of hypotheses containing
  complex interacting parts.
• GA is easily parallelized and can take advantage
  of the computer hardware.

4
2. Genetic Algorithms

• Problem addressed by GAs
• Search a hypotheses space to identify the best
  hypothesis.
• Fitness
• The accuracy of the hypotheses
• Problem of approximating an unknown function
• Number of games won by the individual
• Playing against other individuals in the current
  population
• Characteristics
• Iteration of updating a pool of hypotheses
• Population evaluation by fitness function
• Genetic Operators Selection, Crossover, Mutation
• Randomized, parallel beam search

5
2. Genetic Algorithms(cont.)
Table 9.1
6
2.1 Representing Hypotheses

• Bit String representation

Precondition
Postcondition
Bit strings must be syntactically legal! 111 10
11 gt cannot constrain the attribute PlayTennis
7
2.2 Genetic Operators
Table 9.2
8
2.3 Fitness Function and Selection
9
3. Illustrative Example

• GABIL system
• GAs parameter r(0.6), m(0.001), p(1001000)
• Using of GAs for concept learning of
  classification

10
3. Illustrative Example
Crossover with Variable-Length Bit strings
D1(d2) distance from the leftmost(rightmost) of
two crossover points to the rule boundary to its
left.


If crossover points of h1 is lt1,8gt then d11,
d23. Allowed points for h2 is lt1,3gt,
lt1,8gt,lt6,8gt, lt1,3gt.
11
3.1 Extensions
These two new bits are themselves altered and
evolved using the same operators that operate on
other bits in the string.
12
4. Hypothesis Space Search

• Comparison with Neural network
• Gradient descent search in Backpropagation moves
  smoothly from one hypothesis to a new hypothesis
  that is very similar.
• GA search can move much more abruptly.
• Less likely to fall into the same kind of local
  minima.
• Crowding
• More highly fit individual is quickly reproduces.
• It reduces the diversity of the population.
• Alter selection function
• Fitness sharing
• Restrict the kinds of individuals allowed to
  recombine to form offspring

13
4.1 Schema Theorem
Schema is any string composed of 0s, 1s, and s.
14
4.1 Schema Theorem(cont.)
15
5. Genetic Programming
Figure 9.1
16
5.1 Representing Programs (cross over)
17
6. Models of Evolution And Learning

• Lamarckian Evolution
• He believed individual genetic makeup was altered
  by lifetime experience.
• But current evidence contradicts this view.
• Baldwin Effect
• Assume
• Individual learning has no direct influence on
  individual DNA
• But ability to learn reduces need to hard Wire
  traits in DNA
• Then
• Ability of individuals to learn will support more
  diverse gene pool.
• More diverse gene pool will support faster
  evolution of gene pool.
• Individual learning increases rate of evolution.

18
6. Models of Evolution And Learning(cont.)

• Boldwin Effect(example)
• New predator appears in environment
• Individuals who can learn (to avoid it) will be
  selected.
• Increase in learning individuals will support
  more diverse gene pool.
• Resulting in faster evolution
• Possibly resulting in new non-learned traits such
  as instinctive fear of predator

19
7. Summary

• Conduct randomized, parallel, hill-climbing
  search through H
• Approach learning as optimization
  problem(optimize fitness)
• Nice feature evaluation of Fitness can be very
  indirect
• Consider learning rule set for multistep decision
  making
• No issue of assigning credit/blame to indiv. steps

20
(No Transcript)