Minimize Surprise MAP-Elites: A Task-Independent MAP-Elites Variant for Swarms

Swarm robotics controllers are often automatically generated using methods of evolutionary computation with a task-specific fitness function to guide the optimization process. By contrast, our minimize surprise approach uses a task-independent fitness function to generate diverse behaviors over several independent evolutionary runs. Alternatives are divergent search algorithms rewarding behavioral novelty, such as novelty search, and quality-diversity algorithms generating diverse high-quality solutions, such as MAP-Elites. These approaches usually rely on task-dependent measures. We propose Minimize Surprise MAP-Elites, a task-independent MAP-Elites variant that combines MAP-Elites with our minimize surprise approach. Our first experiments result in high-quality solutions that lead to behavioral diversity across tasks and within tasks.