Overview of approaches to our EA

[gianlucatruda]

We could use documentation to choose the most potentially interesting and suitable operators for Crossover, Mutation, Selection, possibly use migration(?). And use lectures and our own ideas too.

Implementation Overview

Representation

• Genotype
  • [x] 265 floats [-1.0, 1.0]
  • [ ] Tree structure ?
• Phenotype
  • [x] (20, 10, 5) neural net
  • [ ] Decision tree ?

Fitness function

• [x] Evoman standard function [TODO formula]

Selection

• Parent selection
  • [ ] Fitness-proportionate selection (FPS)
  • [ ] Rank-based selection
    • [ ] Linear ranking
    • [ ] Exponential ranking
  • [ ] Uniform selection
  • [x] Tournament selection
• Survival selection
  • [ ] "Plus strategy": children + parents
  • [ ] "Comma strategy": children only
  • Fitness-based selection
    • [ ] Elitism
    • [ ] Delete worst
    • [ ] Round-robin tournament
  • Age-based selection
    • ???

Mutation

• Non-adaptive
  • [ ] Uniform mutation
  • [x] Non-uniform mutation (sample noise from normal dist.)
• Self-adaptive
  • [x] Uncorrelated
  • [ ] Correlated

Recombination

• [ ] 1-point crossover
• [ ] n-point crossover
• [x] Uniform crossover
• [x] Blend crossover
• [ ] Multi-parent crossover
  • [ ] Diagonal
  • [ ] Arithmetic

Population control

• [ ] Generational
• [x] Steady state
• Diversity preservation
  • [ ] Fitness sharing
  • [ ] Crowding
  • [ ] Automatic speciation
  • [ ] Geographical (island or cellular EA)

Initialisation

• [x] Uniform distribution from [-1, 1]

  Termination

• [x] Fixed number of generations

Parameter tuning

TODO

Parameter control

TODO

[Serafim179]

Uniform distribution from [-1, 1]

Changing initialization from random.random to random.uniform between -1 and 1 (as it was in _demospecialist.py) provided considerable boost towards fitness diversity in the first generation. The problem could've been that the weights were too skewed before. The changes are in the pull request #10 .