Dynamic behavioral diversity
Stéphane Doncieux, Jean-Baptiste Mouret · 2012
Behavioral Diversity (BD) aims at improving the evolution of robots by fostering exploration on the basis of their behavior, whereas evolutionary algorithms typically consider the diversity on a genotypic level. Several Behavioral Similarity Measures (BSM), the key component to improve behavioral diversity, have been investigated in the litterature. Current benchmarks show that (1) most tested BSM improve the final performance, (2) they do not lead to the same improvements and, (3) it is hard to predict a priori which BSM will work the best. Instead of trying to find the best BSM, a different approach is proposed here: assuming that several BSM are available, we propose to randomly switch between then each K generations (e.g. K=20). This new approach is tested on a ball collecting task. Results show that better fitness values are obtained with the random switch approach than with any single measure. In effect, the present contribution show that it is possible use behavioral diversity while avoiding to choose between behavioral similarity measures.