1997 年 33 巻 9 号 p. 882-889
This paper proposes a generic design method to develop an optimal state feed-back control system using multilayered neural networks. Because Back-Propagation can not be applied to this case, two new methods are proposed for training neural networks: one is based on gradient method and the other on Powell's conjugate direction method. These two methods can be applied to various problems such as ones for non-linear systems where the feed-back controller can not be obtained by conventional methods. Especially, the method based on the direction method can deal with criterion functions whose derivatives can not be obtained analytically. Illustrative examples of optimal regulator and minimal time control problems show the effectiveness of the proposed methods.