Learning algorithm improvements of a neural network based tuning method for robot controller
O. Bossard, Atsuo Kawamura · 2002
A neural network based method has been proposed for control system's gains tuning, but the network training is time consuming. Thus, several improvements of the classical backpropagation algorithm are investigated in order to speed-up the learning process. They are theoretically analyzed, and their effectiveness is confirmed with the application to the control of a nonlinear model of direct-drive two-axis robot manipulator. It is clarified that the convergence efficiency of the optimization process can be drastically increased by those improvements, resulting in a faster training of the network.>