Robot Controller Evolution with Macroevolutionary Algorithms

Summary

There are certain problems that require using small populations to explore fitness landscapes that are mostly flat, thus offering very little information, where the solutions appear as sparsely distributed narrow peaks. This is the case of the evolution of controllers for many problems in evolutionary robotics. Consequently, for these types of problems it should be useful to consider the use of evolutionary algorithms that cluster the few individuals in the surroundings of the local good solutions permitting an adequate trade-off between exploration and exploitation. Macroevolutionary algorithms cover this need, and through the appropriate selection of the values for its parameters they perform in general better than genetic algorithms for the case of very low population values. In this work we study the influence of the two main parameters governing the search performed by macroevolutionary algorithms as well as the influence of dividing populations into races.