ABSTRACT
This structure operates in on-line mode and it trains a neural network to act as a controller in the region o f interest (meaning that it is "goal directed"). In this way this scheme avoids some o f drawbacks o f the previous two structures. Here, in the specialized learning architecture, the controller no longer learns from its input-output relation, but from a direct evaluation o f the system's performance error e3 = yd - y. The network is trained in order to find the best control value u that drives the plant to an output y = yd. This is accomplished by using a steepest decent (EBP) learning procedure. Despite the fact that the specialized architecture operates in on-line mode, in the case o f nonlinear plant, a pretraining or off-line phase is usually both very useful and highly recommended.
