Expert systems for guiding backpropagation training of layered perceptrons

摘要

The backward error propagation (or backpropagation) model for training neural networks is time consuming. Some theorems and “rules of thumb” are available to guide the process, but the former tend to be nonconstructive and the latter are not universally applicable. Neural networks are appealing for representing nonlinear, stochastic, or ill-conditioned transformations by searching “transformation space” for a good approximation. However, very little is known about navigating through transformation space. In this study, we describe the use of expert systems to learn about and supervise that navigation process.