Parent Selection Strategies for Evolutionary Algorithms A Comparison of Parent Selection Strategies Modeled After Human Social Interaction By Michael Ames
Motivation Because existing research is limited, but shows significant improvement in algorithm robustness, additional research is warranted. Evolutionary algorithms modeled after higher level human social interaction could produce significant improvement in convergence times and individual fitness.
Background Assortative mating – mate selection based on the level of genetic similarities between mates. Disassortative mating – mate selection based on the level of genetic dissimilarities between mates.
Background Hamming distance - The number of bits which differ between two binary strings. HD =
Goals Determine difficulty of implementation. Determine how well the strategies handle premature convergence. Measure selective pressure. Rate performance vs. other social strategies Rate performance vs. existing strategies
Test Algorithms N-Queens Binary Knapsack
8-Queens Q Q Q Q Q Q Q Q
Binary Knapsack
Approach Implement the evolutionary algorithms Implement a separate mate selection object.
Results NQueen
Tournament – generations:3925 time: sec AMEA Marriage – generations: 5316 time: sec AMEA Low Cheat – generations: 4876 time: sec AMEA High Cheat – generations: 4215 time: sec
Results Knapsack
Tournament – generations:6975 best: AMEA Marriage – generations: 6350 best: AMEA Low Cheat – generations: 6550 best: AMEA High Cheat – generations: 6825 best:
Conclusions Poor performance Longer time / inferior results
Future Work Develop other evolutionary mating strategies Implement one of the algorithms with two separate populations dynamic in size, one designated male the other female, then apply the same or similar strategies
Any of these?