Boundary Points Do Not Improve the Accuracy of Neural Net Classifiers1

It seems reasonable to expect that the inclusion of cases on the decision boundaries in a training set would assist in deriving a more accurate neural network classiier, indeed there are reports in the literature of success with this approach. This paper describes a number of experiments using a genetic algorithm to generate boundary points and two strategies for including them in training. While the genetic algorithm was very successful in nding points uniformly distributed on the decision boundaries, extensive experimentation on three classiication problems failed to nd any situations in which using boundary points was a signiicant improvement over adding the same number of random points. In some cases the boundaries were pushed in the wrong directions.