ABSTRACT
During learning only the output layer weights have been changed. Pretraining was done by using 1000 random and uniformly distributed input signals yd. After this off-line learning phase, two tests by a previously unseen ramp signal, were carried out. In both simula tions the hidden layer weights were not the subjects o f the learning. In the first simulation (Figure 44 left) the output layer weights were fixed and in the second one, both networks operated in a learning mode by adapting the O L weights (Figure 44 right). The graphs in Figure 44 show that the whole A B C structure can be successfully trained in on line mode as long as the plant surface is a monotonie one. Figure 45 shows that N N is a good model o f this nonlinear plant. There is no big difference between the actual plant surface and the one modeled by N N 2 . Note that all the trajectories o f the controlled plant lie on this surface.
