Genetic Algorithms and Artificial Life

1
Genetic Algorithms and Artificial Life

• Ilan Bronkesh
• Noga Amit

2
What did we talk about?

• Introduction
• Overview of genetic algorithms
• Interaction between learning and evolution
• Summary

3
Genetic Algorithms

• Evolution by natural selection is a central idea
  in biology
• Evolution of artificial systems as an important
  component of artificial life
• 2 function
• Tools of solving practical problems
• Scientific models of evolutionary processes

4
Genetic Algorithms structure

• Representation of the problem
• The fitness function and scaling
• Heredity to the next generation

5
Social systems

• GAs play a major role in understanding and
  modeling social systems.
• One central field in this research is modeling
  the evolution of cooperation.
• The subject will be viewed in the context of the
  Prisoners Dilemma.

6
The Prisoner Dilemma (PD)

• A simple 2-person game
• Was studied extensively in game theory, economics
  political science
• Considered to be an idealized model for real
  world phenomena such as arm races.

7
The Prisoner Dilemma (cont)
Player B
Player A

• On a given turn, each player independently
  decides whether to cooperate or defect

8
The Prisoner Dilemma (cont)
Player B
Player A

• If both players cooperate they each get 3 points

9
The Prisoner Dilemma (cont)
Player B
Player A

• If player A defects player B cooperate player
  A gets 5 points and player B gets none, and vice
  versa.

10
The Prisoner Dilemma (cont)
Player B
Player A

• If both players defect, they each get 1 point.

11
The Prisoner Dilemma (cont)

• What is the best strategy to take?
• For one turn defect!
• The problem is finding an optimum strategy for an
  iterated game
• This question takes on special significance when
  cooperating defecting correspond actions of
  the real world.

12
TIT FOR TAT

• Early work of Axelrod, suggested that the
  strategy is simple
• TIT FOR TAT cooperate in the first move, after
  that always do what the other player did in his
  last move.
• Axelrod performed experiments to see if a GA
  could evolve strategies to play the game
  successfully

13
TIT FOR TAT (cont)

• The fitness function the average score of the
  player
• Strategies were encoded in a lookup tables
• Most of the strategies that evolves were similar
  to TIT FOR TAT.
• But strikingly, the GA occasionally found
  strategies that scored much higher that TIT FOR
  TAT

14
TIT FOR TAT (cont)

• Did a GA really evolved better strategies that
  human designed ones?
• The performance of a strategies depends on the
  environment the other strategy which is playing
• TIT FOR TAT is generalist, whereas highest score
  evolved strategies were specializes to given
  environment

15
TIT FOR TAT (cont)

• GA is good at doing what evolution often does
• Developing highly specialized adaptation to
  specific environment

16
Effect of dynamic environment

• The experiment allowed strategies to play with
  each other.
• The environment is changing from generation to
  generation due to the evolved strategies
• At each generation every strategy plays with all
  the other, its fitness is the average score of
  all the games

17
Effect of dynamic environment

• Axelrod observed an interesting phenomenon
• The GA initially evolves uncooperative strategies
• After about 10-20 generations the trend start
  to reverse!

18
Effect of dynamic environment

• Further experiments by Lindgren includes the
  possibility of noise
• More open-ended kind of evolution, more free
  gene duplication
• Interesting evolutionary dynamics, including
  periods of relative stasis

19
Effect of dynamic environment

• Other work discover PD strategies with imperfect
  information about the past .
• Both works making PD a more realistic model of
  social political interaction.
• Immune system not direct mapping from genotype
  to phenotype
• But connection simple enough to be studied
  analytically

20
Summary

• Strategies are represented as look-up tables
• Fitness function the average score in the
  prisoner dilemma game
• Heredity by mutations in the look-up tables

21
Open discussion

• Try to think of other implementations of genetic
  algorithms
• In social systems
• In other fields

(now is the time)