Tractable neurocontroller design and application to ship control with actuator limits

This paper extends the popular PID control structure to a nonlinear format by using a building block based neural network. A GA based-optimisation method is used to optimise the neurocontroller. Special training is employed in the design of a feedforward path neurocontroller, in which the network can be trained from a plant model directly. In order to arrive at the simplest structure of a network, the growth training method is developed. Through applications, it is found that if there is a rate limiter in a practical control loop, the automatically designed neurocontroller outperforms an optimised linear controller