Neural controller based on back-propagation algorithm

Investigates the possibility of using a simple approximation for evaluating the error which must be back-propagated to allow a neural net to learn to control a plant in an adaptive way. The algorithm is based on an approximation of the Jacobian of the plant. The method is applied to five simulations. The first two simulations allow a comparison between the proposed algorithm and the standard back-propagation, for which the error to be back-propagated is precisely known. The results for the two methods show equivalent performances, and equivalent convergence time, for the test problems. This shows that the rate of convergence of the neural net does not seem to depend crucially on the values of the Jacobian. The last three simulations investigate the possibility of online adaptive learning. The results show that control based on some approximation of the Jacobian is possible for a neural network