& Faculdade de Ciências,

Cooperative coevolution algorithms (CCEAs) facilitate the evolution of heterogeneous, cooperating multiagent systems. Such algorithms are, however, subject to inherent scalabil-ity issues, since the number of required evaluations increases with the number of agents. A possible solution is to use par-tially heterogeneous (hybrid) teams: behaviourally heteroge-neous teams composed of homogeneous sub-teams. By hav-ing different agents share controllers, the number of coevolv-ing populations in the system is reduced. We propose Hyb-CCEA, an extension of cooperative coevolution to partially heterogeneous multiagent systems. In Hyb-CCEA, both the agent controllers and the team composition are under evo-lutionary control. During the evolutionary process, we rely on measures of behaviour similarity for the formation of ho-mogeneous sub-teams (merging), and propose a stochastic mechanism to increase heterogeneity (splitting). We evalu-ate Hyb-CCEA in multiple variants of a simulated herding task, and compare it with a fully heterogeneous CCEA. Our results show that Hyb-CCEA can achieve solutions of simi-lar quality using significantly fewer evaluations, and in most setups, Hyb-CCEA even achieves significantly higher fitness scores than the CCEA. Overall, we show that merging and splitting populations are viable mechanisms for the cooper-ative coevolution of hybrid teams