120 likes | 240 Views
Comparison of Evolutionary Techniques. Motivation. EAs are general problem solvers EAs are themselves problems EAs should be simpler to apply to new problems. Local Mating. Pick a location, L Take a random walk from L, for each parent Generate a new child at L. Local Survival.
E N D
Motivation • EAs are general problem solvers • EAs are themselves problems • EAs should be simpler to apply to new problems
Local Mating • Pick a location, L • Take a random walk from L, for each parent • Generate a new child at L
Local Survival • All sites, including L, are occupied • Each offspring must displace some individual, normally at L, or die
Complexity Analysis • Rank-based: O((m + l) * log(m + l)) – sorting dominates • Fitness-proportional: O(m + l) – linear, but generally poor results • Grid: O(l) – linear, and independent of m! 1-2 orders of magnitude faster than rank-based!
Parallelization • Divide the population along each dimension of the grid • Exchange information only upon need • Asynchronous execution • No compromises necessary
Delta-Entropy Fitness Holes • Define an entropy function • Rank the population normally • Calculate incremental entropy • Randomly use in place of fitness
Floating-point Entropy • Given alleles {x0…xn-1} • DH(i) = nearest neighbor to xi in {x0…xi-1} • Disfavors clumps of alleles • Inhibits rampant breeding
Complexity Analysis • O((m + l)* log(m + l) * dim) for sorting each set of alleles • ~1 order of magnitude slower