Initial Population Diversity as a Wild Goose Chase: Phenotypic Effects Sometimes Dominate

In evolutionary computation (EC), genetic diversity (or the lack thereof) gets the credit (or the blame) for a multitude of effects — and so mutation operators, population initialisation, and even pseudo-random number generators, all get probed and prodded to improve genetic diversity. This paper demonstrates how extra initial diversity can appear to cause improvements in EC performance, but the true cause is in unforeseen effects of the problem-specific representation. The problem considered in this paper is to learn to play the Iterated Prisoner’s Dilemma (IPD) game, where players have a fine-grained range of choices between cooperation and defection.