An Evolutionary Pattern Recognition Network

A pattern recognition network with two types of adaptation has been investigated. The network output is a weighted sum of the outputs of elements which compute real functions of the discrete network inputs. The first type of adaptation involves the adjustment of the weights while the second type involves the periodic replacement of the least valuable network elements with new ones. The expected error of the network in realizing arbitrary input-output functions has been found by Monte-Carlo simulation for simple weight adaptation and for the case where the population of network elements is allowed to evolve. Three heuristics for determining which elements are to be replaced in each generation have been evaluated and compared. These were based on the size of the weight associated with each element after training, a normalized weight size, and the cross correlation between the elemental function and the desired network function. All three selection criteria resulted in improvements of the network performance over the nonevolutionary case. The normalized weight size criterion was most effective while the cross-correlation criterion was least effective.