Specialization with NeuroEvolution in a collective behaviour task

In Nature, behavioral specialization is ubiquitous. Groups benefit from complementary and specialized behaviors in individuals, especially in tasks requiring collective behavior. We apply four multiagent NeuroEvolution approaches to such a task: Enforced SubPopulations [5], Parallel and Coevolutionary Enforced SubPopulations [16] and Collective NeuroEvolution [11]. Rather than just single controllers we evolve teams of simulated robots to search an unexplored area and gather certain object types for collective construction of a specific sequence. Teams are composed of agents that may evolve from initially homogeneous behavior into specialists that effectively complement each other. Results show that CONE outperforms in the collective behavior task when assisted with target behavior heuristics for lifetime learning to speed up the search. Some evolved specialists however become what we call all-rounders, taking on some more tasks to compensate for their lack in number.