A methodology to reduce large overshoot in a direct neural control system

Abstract

Direct neural control presented a good relearning performance in tracking a desired trajectory. However, random distribution of the initial weights of the neural network controller results in large initial overshoots. In this paper, a closed-loop training method for a direct neural controller is proposed, aiming to generate ‘good’ initial weights for direct control. The training data are generated on the on-line trajectory tracking using a conventional control guide. Pre-training of the neural network concentrates on a subset that system states mainly fall in. The simulation studies for a single-link manipulator have verified that the trained direct neural control system exhibits a better system response than an untrained neural control system does in trajectory tracking and set-point regulation with significantly reduced initial overshoots.