Statistical Guarantees for Lifelong Reinforcement Learning using PAC-Bayes Theory
Zhi Zhang, Chris Chow, Yasi Zhang, Yanchao Sun, Haochen Zhang, Eric Hanchen Jiang, Han Liu, Furong Huang, Yuchen Cui, Oscar Hernan Madrid Padilla
TL;DR
This work tackles lifelong reinforcement learning by modeling a non-stationary stream of tasks and mitigating catastrophic forgetting through a world-policy distribution learned with PAC-Bayes theory. The proposed method EPIC maintains a shared policy distribution and samples policies from it to rapidly adapt to new tasks, guided by a generalization bound that couples empirical performance with a KL penalty. The authors establish a bound that ties long-term performance to the number of past tasks retained and provide a RL-regret-based sample complexity, while empirically achieving state-of-the-art results on diverse lifelong benchmarks. Overall, EPIC offers a theoretically grounded, practically effective framework for continual adaptation in non-stationary RL environments.
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.