Convergent Policy Optimization for Safe Reinforcement Learning

Abstract

We study the safe reinforcement learning problem with nonlinear function approximation, where policy optimization is formulated as a constrained optimization problem with both the objective and the constraint being nonconvex functions. For such a problem, we construct a sequence of surrogate convex constrained optimization problems by replacing the nonconvex functions locally with convex quadratic functions obtained from policy gradient estimators. We prove that the solutions to these surrogate problems converge to a stationary point of the original nonconvex problem. Furthermore, to extend our theoretical results, we apply our algorithm to examples of optimal control and multi-agent reinforcement learning with safety constraints.

Cited by in corpus (14)

• Provably Efficient Q-Learning with Low Switching Cost
• Successive Convex Approximation Based Off-Policy Optimization for Constrained Reinforcement Learning
• Cautious Reinforcement Learning via Distributional Risk in the Dual Domain
• Safe Reinforcement Learning of Control-Affine Systems with Vertex Networks
• Policy Mirror Descent for Regularized Reinforcement Learning: A Generalized Framework with Linear Convergence
• Global Convergence of Policy Gradient Primal-dual Methods for Risk-constrained LQRs
• Primal-dual Learning for the Model-free Risk-constrained Linear Quadratic Regulator
• Safe Reinforcement Learning with Linear Function Approximation
• Safe Distributional Reinforcement Learning
• Finite-Time Complexity of Online Primal-Dual Natural Actor-Critic Algorithm for Constrained Markov Decision Processes
• Safe Learning of Uncertain Environments
• FISAR: Forward Invariant Safe Reinforcement Learning with a Deep Neural Network-Based Optimize
• Scheduling and Power Control for Wireless Multicast Systems via Deep Reinforcement Learning
• Risk Conditioned Neural Motion Planning