부산대학교 도서관

Multi-Agent Reinforcement Learning (MARL) these days are used for modelling numerous real life applications. It is mostly used for modelling such scenarios in which there is very limited information of the environment. Off-policy Reinforcement Learning and Off-policy MARL makes use of buffer to store the experiences or transitions of every agent and later samples it for training purpose. Prioritized Experience Replay (PER) proposed a way to sample transition by the magnitude of the error a transition has during training. The more the error, the more is the probability of that transition of getting sampled for the next training. This method, though effective, encounters a problem. A transition which has low error, might never get sampled, though it is important for the model for that particular transition to be sampled to learn optimal policies. We propose Partitioned PER (ParPER) where we partition the memory into fixed length size and sample equal transitions from them independently. We test our proposed method on Multi-Agent Particle Environment (MPE) and the results were found to be promising.