Data-based decentralized control of nonlinear-constrained interconnected systems using reinforcement learning.

Abstract

This article designs a data-based decentralized controller for mismatched interconnected nonlinear systems having asymmetric input constraints. Initially, it is proved that the decentralized controller for original interconnected systems is composed by the solutions of a set of unconstrained-input optimal control problems of auxiliary subsystems with the preassigned cost functions. Then, in order to solve the Hamilton-Jacobi-Bellman equations arising from these optimal control problems, a data-based policy iteration (PI) algorithm in the reinforcement learning framework is introduced. The implementation of such a PI algorithm relies on the actor-critic structure, which consists of actor and critic neural networks (NNs). After that, the weighted residuals' method and the Monte Carlo integration technique are combined to determine the actor and critic NNs' weight parameters simultaneously. Finally, an interconnected power system is provided to validate the present decentralized controller.