An Evolutionary Bootstarp Approach to Neural Network Pruning and Generalization.

This paper combines techniques drawn from the literature on evolutionary optimization algorithms along with bootstrap based statistical tests. Bootstrapping is used as a general framework for estimating objectives out of sample by redrawing subsets from a training sample. Evolution is used to search the large number of potential network architectures. The combination of these two methods creates a network estimation and selection procedure which finds parsimonious network structures which generalize well. The bootstrap methodology also allows for objective functions other than usual least squares, since it can estimate the in sample bias for any function. Examples are given for forecasting chaotic time series contaminated with noise.