This award funds the first year of a three year continuing grant. A number of basic scientific issues remain to be addressed in order to facililtate the wide use of neural networks in control systems. This research focuses on trainable state estimation for neural networks in order to deal with effects of plant and sensor noise and incomplete availability of state measurments. It will also explore neural network implementation of a self-tuning regulator for adapting a controller to track changes in a nonlinear plant; techniques for controller weight initialization that can decrease network training time and also reduce the probability of convergence of the weights to undersirable local minima; and adaptation of networks for navigational obstacle avoidance to robotic manipulator obstacles avoidance.
