MAP-Elites for noisy domains by adaptive sampling.

Quality Diversity algorithms (QD) evolve a set of high-performing phenotypes that each behaves as differently as possible. However, current algorithms are all elitist, which make them unable to cope with stochastic fitness functions and behavior evaluations. In fact, many of the promising applications of QD algorithms, for instance, games and robotics, are stochastic. Here we propose two new extensions to the QD-algorithm MAP-Elites --- adaptive sampling and drifting-elites --- and demonstrate empirically that these extensions increase the quality of solutions in a noisy artificial test function and the behavioral diversity in a 2D bipedal walker environment.