On predicting the optimal number of hidden nodes

Abstract- Determining the optimal number of hidden nodes isthe most challenging aspect of Artificial Neural Network(ANN) design. To date, there are still no reliable methods ofdetermining this a priori, as it depends on so many domainspecificfactors. Current methods which take these intoaccount, such as exhaustive search, growing and pruning andevolutionary algorithms are not only inexact, but alsoextremely time consuming - in some cases prohibitively so. Anovel approach embodied in a system called Heurix isintroduced. This rapidly predicts the optimal number ofhidden nodes from a small number of sample topologies. It canbe configured to favour speed (low complexity), accuracy, or abalance between the two. Single hidden layer feedforwardnetworks (SLFNs) can be built twenty times faster, and with ageneralisation error of as little as 0.4% greater than thosefound by exhaustive search