TRAINABLE NEURAL-NETWORK FOR MECHANICALLY FLEXIBLE SYSTEMS BASED ON NONLINEAR FILTERING

A trainable neural network controller architecture is investigated for motion control systems involving significant distributed mechanical f lexibility. In general, this neural network based controller can be tr ained on-fine to learn the behavior of another another controller whic h serves as the teacher implementing algorithmic or nonalgorithmic con trol law. To address potential of such a scheme in real time, the weig ht adjustments of the network connection strengths and biases are base d on a nonlinear filtering adaptation rule, extended Kalman filter, to reduce training time and achieve fast convergence rate. Computer simu lations are performed to test the performance of this training algorit hm.