Computer aided neural network engineering

Abstract

The lack of rigorous techniques to initiate and complete in detail a neural network design renders neural network engineering to be more of an art than science. Dimensioning of neural network layers is an example of an important practical design detail typically overlooked in the neural network theoretical investigations and missed by the practitioners. We develop a computer aided neural network engineering tool based on a hybrid expert system architecture merging both knowledge-based and neurode-based components. We demonstrate our approach by mechanizing the design of counter-propagation neural network. Our automatic Kohonen layer configurator combines A* and simulated annealing search techniques to achieve both automated dimensioning and simultaneous selection of synaptic weights.