Abstract
In this paper, firstly we propose a neural network with a mixed structure, which consists of multilayer and recurrent structure, for learning the dynamics of a nonlinear plant. A neural network with a mixed structure can learn time series, therefore, it can learn the plant dynamics without knowning the plant order.
Next, we consider the optimal control synthesis problem using the neural network with a mixed structure, which has learned the plant dynamics completely, as a plant model. Procedures are as follows: (1) the neural network is expanded into an equivalent feedforward multilayer network, (2) it is shown that the gradient of criterion functional to be optimized can be easily obtained from this multilayer network, and then (3) the optimal control is generated by applying any of the existing nonlinear programming algorithms based on this gradient information.
The proposed method is sucessfully applied to the optimal control synthesis problem of a nonlinear coupled vibratory plant with a linear quadratic criterion functional.
