Neural network control of nonstrict feedback and nonaffine nonlinear discrete-time systems with application to engine control

In this dissertation, neural networks (NN) approximate unknown nonlinear functions in the system equations, unknown control inputs, and cost functions for two different classes of nonlinear discrete-time systems. Employing NN in closed-loop feedback systems requires that weight update algorithms be stable. This dissertation is comprised of five refereed journal-quality papers that have been published or are under review. Controllers are developed and applied to a nonlinear, discrete-time system of equations for a spark ignition engine model to reduce the cyclic dispersion of heat release. In some of the papers, the controller is also tested on a different nonlinear system using simulation.; An adaptive neural network-based output feedback controller is proposed to deliver a desired tracking performance for a class of discrete-time nonlinear systems, which are represented in non-strict feedback form. A spark ignition engine can be viewed as a nonstrict-feedback nonlinear discrete-time system. An NN controller employing output feedback is designed to reduce cyclic dispersion of heat release in a spark ignition engine that uses three NNs to estimate the unknown states, generate the virtual control input, and to generate the actual control input. Another NN controller uses state feedback to minimize cyclic dispersion caused by high levels of exhaust gas recirculation (EGR). Adding another state for EGR to the engine model, an adaptive NN controller is designed with a separate control loop for maintaining an EGR level where output feedback of heat release is used. The system becomes nonaffine with spark timing as the control input, and a novel controller based on reinforcement learning is proposed for the affine-like nonlinear error dynamic system.