Rule extraction from a trained artificial neural network

This paper describes a system for rule extraction from a trained artificial neural network. Using a hierarchical Kohonen feature map and a novel `rule transfer agent´, procedural rules in the input data that are classified by the network can be extracted from the neural network and migrated to an attached knowledge based system. The system has been tested on a wide range of fractal patterns and is shown to `recover´ the rules needed to generate the original image. A wide range of potential practical applications is investigated for this dynamic rule migration system