A Reinforcement Learning Technique with an Adaptive Action Generator for a Multi-robot System

We have developed a new reinforcement learning (RL) technique called Bayesian-discrimination-function-based reinforcement learning (BRL). BRL is unique, in that it does not have state and action spaces designed by a human designer, but adaptively segments them through the learning process. Compared to other standard RL algorithms, BRL has been proven to be more effective in handling problems encountered by multi-robot systems (MRS), which operate in a learning environment that is naturally dynamic. Furthermore, we have developed an extended form of BRL in order to improve the learning efficiency. Instead of generating a random action when a robot functioning within the framework of the standard BRL encounters an unknown situation, the extended BRL generates an action determined by linear interpolation among the rules that have high similarity to the current sensory input. In this study, we investigate the robustness of the extended BRL through further experiments. In both physical experiments and computer simulations, the extended BRL shows higher robustness and relearning ability against an environmental change as compared to the standard BRL.