An ILC-based Minimum Entropy PI Controller for Unknown and Non-Gaussian Stochastic Systems

Abstract: In this paper a new adaptive control algorithm is presented for unknown nonlinear and non-Gaussian stochastic systems. The method combines the minimum entropy control with an Iterative Learning Control (ILC) framework, where the control horizon is divided into a number of time-domain intervals called Batches. Within each batch a PI controller is used to control the plant so as to achieve the required tracking performance, where a neural network is used to learn the dynamics of the unknown plant. Between any two adjacent batches, a D-type ILC law is applied to tune the PI control coefficients so that the tracking error entropy for the closed loop system is reduced batch by batch. The analysis on the ILC convergence is made and a set of demonstrable experiment results on a test rig are also provided to show the effectiveness of the obtained adaptive control algorithm.