Using neural networks as a tool for constructing rule based systems

摘要

The paper presents a method for deriving rules from a neural network and abstracting meaning from them. This produces a connectionist rule induction system that combines some of the useful attributes of connectionist and traditional symbolic systems. A tutorial example of the noise immunity enjoyed by neural networks is presented together with examples drawn from the field of intelligent control. Finally, the limits of the technique are explored by applying the technique to a large scale neural network drawn from visual character recognition.