An Approach for Synthesis of Robust Neural Network Controller for Multivariable Non-Linear Systems with Stochastic Parameters

Abstract

In this paper, an approach is proposed to synthesize robust neural network controller for multi-input multi-output non-linear systems with uncertain parameters. The approach is based on the capability of neural networks to approximate arbitrary non-linear mappings. The case of known probability distributions of the uncertain parameters is considered. The control problem is to maintain the system on the required steady state despite of changes of the system parameters. The overall controller structure includes two neural network controllers: a global neural controller which transfers the system near steady state by minimizing the mathematical expectation of the optimality criterion and a local neural controller which generates the correct steady state values of the control variables based on the system response achieved with the global controller.