Provable Hierarchy-Based Meta-Reinforcement Learning

Abstract

Hierarchical reinforcement learning (HRL) has seen widespread interest as an approach to tractable learning of complex modular behaviors. However, existing work either assume access to expert-constructed hierarchies, or use hierarchy-learning heuristics with no provable guarantees. To address this gap, we analyze HRL in the meta-RL setting, where a learner learns latent hierarchical structure during meta-training for use in a downstream task. We consider a tabular setting where natural hierarchical structure is embedded in the transition dynamics. Analogous to supervised meta-learning theory, we provide "diversity conditions" which, together with a tractable optimism-based algorithm, guarantee sample-efficient recovery of this natural hierarchy. Furthermore, we provide regret bounds on a learner using the recovered hierarchy to solve a meta-test task. Our bounds incorporate common notions in HRL literature such as temporal and state/action abstractions, suggesting that our setting and analysis capture important features of HRL in practice.

References in corpus (8)

• Trust Region Policy Optimization
• FeUdal Networks for Hierarchical Reinforcement Learning
• Data-Efficient Hierarchical Reinforcement Learning
• A Laplacian Framework for Option Discovery in Reinforcement Learning
• Learning Multi-Level Hierarchies with Hindsight
• Few-Shot Learning via Learning the Representation, Provably
• Why Does Hierarchy (Sometimes) Work So Well in Reinforcement Learning?
• Exploration--Exploitation in MDPs with Options