A trainable neural network controller architecture is investigated for motion control systems involving significant distributed mechanical flexibility. In general, this neural network based controller can be trained on-fine to learn the behavior of another another controller which serves as the teacher implementing algorithmic or nonalgorithmic control law. To address potential of such a scheme in real time, the weight adjustments of the network connection strengths and biases are based on a nonlinear filtering adaptation rule, extended Kalman filter, to reduce training time and achieve fast convergence rate. Computer simulations are performed to test the performance of this training algorithm. References: 13
展开▼