TRL: Discriminative Hints for Scalable Reverse Curriculum Learning
Deep reinforcement learning algorithms have proven successful in a variety of domains. However, tasks with sparse rewards remain challenging when the state space is large. Goal-oriented tasks are among the most typical problems in this domain, where a reward can only be received when the final goal is accomplished. In this work, we propose a potential solution to such problems with the introduction of an experience-based tendency reward mechanism, which provides the agent with additional hints based on a discriminative learning on past experiences during an automated reverse curriculum. This mechanism not only provides dense additional learning signals on what states lead to success, but also allows the agent to retain only this tendency reward instead of the whole histories of experience during multi-phase curriculum learning. We extensively study the advantages of our method on the standard sparse reward domains like Maze and Super Mario Bros and show that our method performs more efficiently and robustly than prior approaches in tasks with long time horizons and large state space. In addition, we demonstrate that using an optional keyframe scheme with very small quantity of key states, our approach can solve difficult robot manipulation challenges directly from perception and sparse rewards.
PDF Abstract
