Accelerated Primal-Dual Policy Optimization for Safe Reinforcement Learning
This work addresses sample inefficiency for researchers and practitioners in safe reinforcement learning, but it is incremental as it builds on existing primal-dual methods.
The paper tackled the problem of sample inefficiency and slow convergence in safe reinforcement learning by proposing Accelerated Primal-Dual Optimization (APDO), which incorporates off-policy data for dual updates, resulting in better sample efficiency and faster convergence than state-of-the-art methods in simulated robot locomotion tasks.
Constrained Markov Decision Process (CMDP) is a natural framework for reinforcement learning tasks with safety constraints, where agents learn a policy that maximizes the long-term reward while satisfying the constraints on the long-term cost. A canonical approach for solving CMDPs is the primal-dual method which updates parameters in primal and dual spaces in turn. Existing methods for CMDPs only use on-policy data for dual updates, which results in sample inefficiency and slow convergence. In this paper, we propose a policy search method for CMDPs called Accelerated Primal-Dual Optimization (APDO), which incorporates an off-policy trained dual variable in the dual update procedure while updating the policy in primal space with on-policy likelihood ratio gradient. Experimental results on a simulated robot locomotion task show that APDO achieves better sample efficiency and faster convergence than state-of-the-art approaches for CMDPs.