Optimal control with a recurrent neural network and a priori knowledge of the system

The authors present a method of making a control network efficiently with the K-identifier using a priori knowledge for the object being controlled. This method is useful for the rapid and efficient construction of the controller and adaptation to time changes of the object using a neural network identifier. Specifically, the authors show how to construct a controller of a two-link manipulator and apply it to the trajectory formation with a recurrent network. Although the cascade network introduced by M. Kawato et al. (1990) is also applied to the trajectory formation, the recurrent network is different from that in the sense of interpolation of the target position and boundary conditions at terminating time. The authors also show a way to speed up the construction of the controller with an identifier which uses a priori knowledge of the dynamics of the manipulator, and they discuss its validity with numerical simulations