Choosing GA parameter settings
Remarkably little is known about the best choices for GA parameters:
- population size
- type and strength of selection
- mutation and crossover probabilities
- degree of elitism.
Settings are proposed in A parameter-less genetic algorithm by G. Harik and F. Lobo, Proceedings of GECCO-99:
- tournament selection with a selection rate of 4
- crossover probability 0.5 and mutation probability 0 (we use 0.01)
- nested runs of populations of size 256, 512, 1024, etc. in parallel
We investigate restart and local search policies:
- under what circumstances to restart and to do local search
- how to initialize a restart (cf. demes with limited inter-deme transfer)
- when to use caching to reduce duplicate function evaluations