A neural network controller by adaptive interaction

In this paper, we propose an approach to neural network controllers by using a new adaptation algorithm. The algorithm is derived from the theory of adaptive interaction. The principle behind the adaptation algorithm is a simple but efficient methodology to perform gradient descent optimization in the parametric space. Unlike the approach based on the back-propagation algorithm, this approach will not require the plant to be converted to its neural network equivalent, a major obstacle in early approaches. By applying this adaptive algorithm, the same adaptation as the back-propagation algorithm is achieved without the need of backward propagating the error throughout a feedback network. This important property makes it possible to adapt the neural network controller directly. Control of second and third order plants are simulated to demonstrate the effectiveness of the algorithm.