Reinforcement learning neural network used in a tracking system controller

Abstract

This paper presents a method of designing a controller for nonlinear systems based on a recurrent neural network which is trained in real time using the reinforcement learning (RL) procedure. The advantage of this method is to overcome the difficulties implied by the direct solving method of the differential models which are necessary in a classical approach. Moreover, this new technique using a real-time training is better then the MLP network controller as well as the RBF network implementation which needs both of them in a preliminary training process, based on a set of input-output data that has to be a priory experimentally determined.