On Credit Assignment in Hierarchical Reinforcement Learning

TL;DR

This paper investigates hierarchical credit assignment in reinforcement learning, revealing how multistep backups can be adapted for hierarchy, and introduces a new algorithm HierQ_k(λ) that improves agent performance.

Contribution

It provides a fundamental understanding of hierarchical backups and proposes HierQ_k(λ), a novel algorithm that enhances reinforcement learning through hierarchical credit assignment.

Findings

Hierarchical backups can be viewed as multistep backups with skip connections.

Generalizing to multistep return estimation requires environment trace partitioning.

HierQ_k(λ) demonstrates performance improvements due to hierarchical credit assignment.

Abstract

Hierarchical Reinforcement Learning (HRL) has held longstanding promise to advance reinforcement learning. Yet, it has remained a considerable challenge to develop practical algorithms that exhibit some of these promises. To improve our fundamental understanding of HRL, we investigate hierarchical credit assignment from the perspective of conventional multistep reinforcement learning. We show how e.g., a 1-step `hierarchical backup' can be seen as a conventional multistep backup with $n$ skip connections over time connecting each subsequent state to the first independent of actions inbetween. Furthermore, we find that generalizing hierarchy to multistep return estimation methods requires us to consider how to partition the environment trace, in order to construct backup paths. We leverage these insight to develop a new hierarchical algorithm Hier$Q_k(\lambda)$, for which we demonstrate that hierarchical credit assignment alone can already boost agent performance (i.e., when eliminating generalization or exploration). Altogether, our work yields fundamental insight into the nature of hierarchical backups and distinguishes this as an additional basis for reinforcement learning research.