Exploring Genetic Algorithms Through the Iterative Prisoner's Dilemma

1
Exploring Genetic Algorithms Through the
Iterative Prisoner's Dilemma

• Computer Systems Lab 2007-2008
• Aaron Dufour

2
(No Transcript)
3
(No Transcript)
4
Mutation Rate

• How the mutation rate changes per generation
• Within a generation the mutation rate does not
  change

5
Natural Selection

• How many are removed from the population each
  generation
• Static rate the same number are removed each
  generation
• Fitness-based rate all those below a threshold
  fitness value are removed

6
Recombination

• DoublePoint
• 10010101
• 01000011
• Yields
• 10000001
• 01010111

• SinglePoint
• 10010101
• 01000011
• Yields
• 10000011
• 01010101

7
Initial Population Creation

• Simple
• Random binary
• Flip Half
• Random on first half
• Second half is inverted first half
• Ensures that every bit has 50 1's and 50 0's
• Check for Duplicates
• Same as flip half, except remakes each one that
  has a duplicate
• Ensures that all of the solutions are different

8
Output

• Outputs the average fitness value for each
  generation
• File name is g i p t s m n r f.txt
• g number of generations
• i number of iterations
• p population size
• t number of turns
• s initial population type
• m mutation rate info
• n natural selection info
• r recombination type
• f test number
• Example 10 100 150 s s-0.0050 s-0.5 s t0.txt

9
Data Analysis
10
Data Analysis, contd

• The program analyzes the data to find where the
  fitness stabilizes
• Although we can do this visually, it is difficult
  for the computer
• My algorithm eliminates data from the left side
  until the slope of a fit line gets within a
  certain amount of 0

11
Final Product

• The program automatically loops through each of
  the algorithms for each method that can change,
  as well as certain values of variables that I
  chose based on many trials
• It does 10 runs with each setting, and outputs
  the number of generations each one took in
  addition to an average

12
Analysis Tool

• Graphs the number of iterations that each run
  with a specific attribute took
• First input line is number of graphs
• The following (one for each graph) are a number
  followed by a string
• A second program moved files between two folders
  so I could eliminate those with attributes that
  did not perform well

13
Analysis Tool
14
Analysis

• I could not make any conclusions about the
  mutation rate or initial population creation
• The double-point recombination slightly
  outperformed the single-point recombination
  algorithm
• The fitness-based natural selection was better
  than the static natural selection, and got better
  with a higher percent of the maximum required (up
  to 95)?