Learning Action Translator for Meta Reinforcement Learning on Sparse-Reward Tasks

Meta reinforcement learning (meta-RL) aims to learn a policy solving set of training tasks simultaneously and quickly adapting new tasks. It requires massive amounts data drawn from infer the common structure shared among Without heavy reward engineering, sparse rewards in long-horizon exacerbate problem sample efficiency meta-RL. Another challenge meta-RL is discrepancy difficulty level tasks, which might cause one easy task dominating thus preclude adaptation This work introduces novel objective function an action translator We theoretically verify that value transferred with can be close source our (approximately) upper bounds difference. propose combine context-based algorithms for better collection moreefficient exploration during meta-training. Our approach em-pirically improves performance ofmeta-RL on sparse-reward