Valery Frolov
The algorithm Fitness function Crossover Mutation Elite individuals Reverse mutations Some statistics Run examples
The most standard evolution algorithm While (condition isn’t met) Select fittest individuals Crossover Mutation Update population The condition Simulation time
Ideal crossover function for our problem Learns the relation between different parameters Performs the crossover accordingly My crossover Interleaves two parents with equal probability Creates one sibling
Each new population consists of Elite individuals from previous population Individuals created by crossover Several random individuals
Each population consists of K individuals Best M individuals are copied to the next population. M is 1/3 of population size Crossovers are done with at least one elite parent
Problem : Mutations can make things worse Elite individual that survived many generation can be removed because of one bad mutation Our solution : Reverse previous mutation if condition is satisfied
Size of population is 16 Time per each individual is 4450 ticks Individuals get bonus points for surviving Individuals get penalty for non-completed laps
In most cases starting population consists of pair of “good” and many “bad” individuals At least half of final generation are “very good” individuals