Bootstrapping aggregate fitness selection with evolutionary multi-objective optimization

Aggregate fitness selection is known to suffer from the bootstrap problem, which is often viewed as the main inhibitor of the widespread application of aggregate fitness selection in evolutionary robotics. There remains a need to identify methods that overcome it, while requiring the minimum amount of a priori task knowledge from the designer. We suggest a novel two-phase method. In the first phase, it exploits multi objective optimization to develop a population of controllers that exhibit several desirable behaviors. In the second phase, it applies aggregate selection using the previously obtained population as the seed. The method is assessed by two non-traditional comparison procedures. The proposed approach is demonstrated using simulated coevolution of two robotic soccer players. The multi objective phase is based on adaptation of the well-known NSGA-II algorithm for coevolution. The results demonstrate the potential advantage of the suggested two-phase approach over the conventional one.