Reinforcement Learning for Robust Neuro-Control of Constrained Nonlinear Systems

This article considers the robust neuro-control problem of unknown nonlinear systems subject to asymmetric input constraints. Initially, with a discounted value function being introduced for the nominal systems associated with the studied nonlinear systems, the robust constrained control problem is converted into a nonlinear-constrained optimal control problem. Then, in the reinforcement learning framework, an actor-critic architecture is employed to solve the nonlinear-constrained optimal control problem. Two neural networks (NNs) are utilized to implement such an architecture. Specifically, the actor and critic NNs are, respectively, constructed to approximate the control policy and the value function simultaneously. Meanwhile, the actor and critic NNs’ weights are determined via the least square method and the Monte-Carlo integration technique. Finally, simulations of a nonlinear plant are provided to validate the theoretical results.