Statistical Guarantees for Lifelong Reinforcement Learning using PAC-Bayes Theory

TL;DR

This paper introduces EPIC, a PAC-Bayes-based algorithm for lifelong reinforcement learning that learns a shared policy to adapt quickly to new tasks while retaining past knowledge, with proven theoretical guarantees and strong empirical results.

Contribution

The paper presents EPIC, a novel lifelong RL algorithm using PAC-Bayes theory, providing theoretical analysis and demonstrating superior performance over existing methods.

Findings

EPIC outperforms existing lifelong RL methods in various environments.

Theoretical analysis links generalization performance to the number of stored tasks.

EPIC offers both theoretical guarantees and practical efficiency.

Abstract

Lifelong reinforcement learning (RL) has been developed as a paradigm for extending single-task RL to more realistic, dynamic settings. In lifelong RL, the "life" of an RL agent is modeled as a stream of tasks drawn from a task distribution. We propose EPIC (Empirical PAC-Bayes that Improves Continuously), a novel algorithm designed for lifelong RL using PAC-Bayes theory. EPIC learns a shared policy distribution, referred to as the world policy, which enables rapid adaptation to new tasks while retaining valuable knowledge from previous experiences. Our theoretical analysis establishes a relationship between the algorithm's generalization performance and the number of prior tasks preserved in memory. We also derive the sample complexity of EPIC in terms of RL regret. Extensive experiments on a variety of environments demonstrate that EPIC significantly outperforms existing methods in lifelong RL, offering both theoretical guarantees and practical efficacy through the use of the world policy.