Data-Driven Nearly Optimal Control for Constrained Nonlinear Systems

This article develops a novel data-driven policy iteration (PI) to obtain nearly optimal control of nonlinear systems with asymmetric input constraints. The data-driven PI is derived from an early established model-based PI. Owing to the datadriven PI sharing the same solution as the model-based PI, the convergence of the data-driven PI algorithm is guaranteed. The implementation of the newly developed data-driven PI algorithm relies on an actor-critic structure consisting of two kinds of neural networks (NNs). Specifically, the critic NN aims at estimating the value function and the actor NNs aim at approximating the control policies. The weight parameters used in the critic and actor NNs are determined via the least squares method together with the Monte Carlo integration technique. Finally, a nonlinear plant is provided to validate the proposed data-driven PI algorithm.