Analysis results on the wrapper model due to Ng [14] provided a general explanation of why
heuristic guided wrapper approaches is effective. Since the ANNIGMA heuristic can be handily
calculated from trained weights after a training cycle, it is the choice to guide a wrapper model based feature selector for neural nets. The ANNIGMA heuristic is also reliably accurate. However, it will be interesting to nd a bound on the degree to which a neural net needs to be trained before the ANNIGMA heuristic can be reliably applied and thereby minimize the training eort.