Augmented Policy Optimization for Safe Reinforcement Learning

Luoyang Sun, Ning Yang, Jiwen Jiang, Rasul Tutunov, Haifeng Zhang, Jun Wang

Keywords: Safe reinforcement learning, Constrained policy optimization, Constrained Markov decision processes

Abstract:

Safe reinforcement learning (RL) holds a critical role in acquiring policies that conform to explicit constraints, ensuring their suitability for safety-critical applications. However, methods rooted in the primal-dual concept demonstrate inherent instability.
Meanwhile, owing to policy initialization and algorithmic approximation errors, prior methods relying on trust region invariably produced infeasible policies during training, rendering the constructed local optimization problem insoluble.
In this paper, we present the Augmented Constraint Policy Optimization (ACPO) algorithm, which encompasses a novel approach to constructing local policy search problems and an optimization problem decomposition method. Specifically, this method introduces an approach for constructing local policy search optimization that guarantees a solution without relying on hypothetical premises. Utilizing the Alternating Direction Method of Multipliers (ADMM) algorithm as a foundation, we partition the original optimization problem into simpler subproblems that can be efficiently and robustly solved using first-order methods.
Comprehensive experimental evaluations have conclusively demonstrated that the proposed method consistently outperforms the baselines in terms of both performance and constraint satisfaction.

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7097