ニューラルネットワークによるサーボ系のディジタル学習制御 : 高速学習のための制御系の構成とその応用

抄録

Exact inverse systems are not always expected for neural networks in many applications because of the output oscillation of the inverse systems. There are several sets of input signals to the neural networks, which give approximate inverse systems. In this paper, the comparisons of the learning process are first made among them by a computer simulation with a parallelogram-type robot manipulator of two degrees of freedom. It is shown that the choice of the input signals has considerable influence on its learning speed. Secondly, two learning control algorithms are proposed for applying the results of a one-layer linear neural network to the control of a direct-drive (DD) robot. The algorithms are based on preliminary learning by a one-layer linear neural network, and can be used for shortening the time of on-line learning. The effectiveness of these algorithms is demonstrated by the experiment on the control of a DD robot.