Should Every Man be an Island ? Island number and population size for popular benchmark functions

We evaluate 7 optimization algorithms on 21 benchmark functions. A surprisingly robust island-model evolutionary algorithm exhibits excellent performance on all the benchmarks. We then empirically determine optimal population size and island number in island-model genetic algorithms, where optimality is defined by solution quality under the constraint of limited fitness function calls. We conclude that there are specific population sizes and island numbers that are optimal. These values are constant across benchmark functions. Our results suggest that there are non-intuitive rules of thumb for parameterizing island-model GAs and that a priori optimization of island model parameters is possible on new problems. All code, configuration files, and data are made available online along with detailed descriptions of all benchmarks and experiments.