Binary versus Real-valued Reward Functions under Coevolutionary Reinforcement Learning

Models of coevolution supporting competitive and cooperative behaviors can be used to decompose the problem while scaling to large environmental state spaces. This work examines the significance of various design decisions that impact the deployment of a distinctionbased formulation of competitive coevolution. Specifically, competitive coevolutionary formulations with and without point population speciation are compared to stochastic sampling of the environment under both binary and real-valued rewards. The additional structure implicit in the competitive coevolutionary models is shown to be of significant benefit under binary rewards, however, stochastic sampling results in more dependable performance under real-valued feedback. It is also observed that cooperation between multiple solutions is much more prevalent under real-valued rewards than under binary rewards.