Improving ANN generalization using a priori knowledge to pre-structure ANNs

This is a continuation of work reported by Lendaris at el. (1994) whose objective has been to develop a method that uses certain a priori information about a problem domain to pre-structure artificial neural networks (ANNs) into modules before training. The method is based on a general systems theory methodology, based on information-theoretic ideas, that generates structural information of the problem domain by analyzing I/O pairs from that domain. The notion of performance subset of an ANN structure is described. Extensive experiments on 5-input/1-output and 7-input/1-output Boolean mappings show that significantly improved generalization follows from successful pre-structuring. As the previous work already showed, such pre-structuring also yields improved training speed