J-038

Adaptive Learning Method of Neural Network Controller using an Immune Feedback Law

Authors: Motohiro KAWAFUKU(*), Minoru SASAKI(*), and Kazuhiko TAKAHASHI(**)
Affiliation: (*) Department of Mechanical and Systems Engineering, Faculty of Engineering, Gifu University
(**) Advanced Telecommunications Research Institute International, Media Integration & Communications Research Laboratories

Abstract
An adaptive learning method for neural network (NN) controllers using an immune feedback law which features rapid response to foreign matter and rapid stabilization of biological immune systems is proposed. Several improvements are made to correct deficiencies in the usual gradient descent NN algorithms. An adaptive learning rate is used in order to make the learning step as large as possible while avoiding oscillations. In the proposed method, the immune feedback law changes the learning rate of the NN individually and adaptively, thus the cost functional is minimized quickly and the training time is shortened.

In the control structure, a reference signal self-organizing control system employing NNs and flexible micro-actuators is used. The micro-actuator is made of a bimorphic piezo-electric high polymer material (Poly Vinylidene Fluoride). The control system consists of both a plant with a feedback loop and a NN with a feedforward loop. In this system, the NN functions as a reference input filter, setting new reference signals in the closed loop system. Numerical and experimental results show that the proposed control system is effective in tracking a reference signal.

Motohiro KAWAFUKU
kawafuku@gifu-nct.ac.jp