Neural-network based online policy iteration for continuous-time infinite-horizon optimal control of nonlinear systems

Abstract

A new policy-iteration algorithm based on neural networks (NNs) is proposed in this paper to synthesize optimal control laws online for continuous-time nonlinear systems. Latest advances in this field have enabled synchronous policy iteration but require an additional tuning loop or a logic switch mechanism to maintain system stability. A new algorithm is thus derived in this paper to address this limitation. The optimal control law is found by solving the Hamilton-Jacobi- Bellman (HJB) equation for the associated value function via synchronous policy iteration in a critic-actor configuration. As a major contribution, a new form of NN approximation for the value function is proposed, offering the closed-loop system asymptotic stability without additional tuning scheme or logic switch mechanism. As a second contribution, an extended Kalman filter is introduced to estimate the critic NN parameters for fast convergence. The efficacy of the new algorithm is verified by simulations.