The Limits of Temporal Abstractions for Reinforcement Learning with Sparse Rewards

Abstract

Skills are temporal abstractions that are intended to improve reinforcement learning (RL) performance through hierarchical RL. Despite our intuition about the properties of an environment that make skills useful, there has been little theoretical work aimed to characterize these properties precisely. This work studies the utility of skills in sparse-reward environments with a discrete state space and finite action space. We show, both theoretically and empirically, that RL performance gains from skills are worse in environments where successful trajectories are less compressible. In environments with a highly incompressible distribution of successful trajectories, using unexpressive skills such as macroactions will provably worsen RL performance. We hope our findings can guide research on automatic skill discovery and help RL practitioners better decide when and how to use skills.