Improving Experience Replay with Successor Representation
This work addresses a bottleneck in reinforcement learning for agents by enhancing experience replay efficiency, though it is incremental as it builds on existing prioritized replay methods.
The paper tackled the problem of prioritizing experiences in reinforcement learning by incorporating both gain and need, leading to significant performance improvements in benchmarks like the Dyna-Q maze and Atari games.
Prioritized experience replay is a reinforcement learning technique whereby agents speed up learning by replaying useful past experiences. This usefulness is quantified as the expected gain from replaying the experience, a quantity often approximated as the prediction error (TD-error). However, recent work in neuroscience suggests that, in biological organisms, replay is prioritized not only by gain, but also by "need" -- a quantity measuring the expected relevance of each experience with respect to the current situation. Importantly, this term is not currently considered in algorithms such as prioritized experience replay. In this paper we present a new approach for prioritizing experiences for replay that considers both gain and need. Our proposed algorithms show a significant increase in performance in benchmarks including the Dyna-Q maze and a selection of Atari games.