Most artificial neural networks are trained with supervised learning methods. A simple model for supervised learning is given in Fig. 7.1. The outside world is measured and the measurement vector, x, is given to a knowledge expert who outputs a desired response, f(x). The learning system is exposed to the same measured variable and also computes a result, fÌ(x) . The error between the output of the learning system and the desired response from the knowledge expert is measured. The error signal is then used to modify the response of the learning system, adapting weights for neural networks, so that its response more closely matches that of the knowledge expert. The knowledge expert can be a human expert, a function, a set of rules, a set of measured system outputs, and so forth. The learning system can be trained by using any number of adaptation methods such as backpropagation, fuzzy logic, expert-system rules, evolutionary computation, statistical methods, or an ad hoc method. The key principle is that a set of input data and the desired system responses are used to adapt the learning system.
Online access to SPIE eBooks is limited to subscribing institutions.