Abstract: The concept of leveraging knowledge from previous experience to accelerate learning forms the crux of transfer learning. This notion also applies to reinforcement learning where the agents acquire knowledge via interacting with the environment. In this paper, we investigate the application of transfer learning in the fuzzy reinforcement learning domain, specifically within the context of differential games. We introduce a novel approach for knowledge transfer across analogous tasks, employing fuzzy logic controllers as function approximators, notably within the Fuzzy Actor-Critic Learning (FACL) algorithm. Specifically, we propose a strategy for fuzzy rule transfer (FRT) aimed at mapping fuzzy rules between the source and target tasks. The target task is assumed to be related to the source task yet it contains more complex states. Our approach has been implemented and tested within the domain of differential games in which all state space and action space are continuous. The simulation outcomes demonstrate that the application of knowledge transfer enables RL agents to learn faster and achieve asymptotic performance more rapidly in the target task.
PDF
Keywords: Transfer Learning, Fuzzy Actor-Critic Learning, Differential Games, Reinforcement Learning.