Directed product term selection in Sigma-Pi networks

An earlier paper presented a framework for training Sigma-Pi networks without incurring a combinatorial increase in the number of product terms employed, or knowledge regarding terms required. This paper summarises refinements to the basic framework in order that the search for polynomials be guided. Consequently, product terms added fit the mapping under construction at the local neuron. Furthermore, an overlearning test determines whether the increase in complexity attributed to the new product term is warranted, given the accompanying reduction in error provided. Finally, the original magnitude based weight significance measure is replaced by the more rigorous OBS technique, for both dynamic and static pruning stages within the product term context. Simulations indicate significant performance improvements when applied to constrained product term count situations