Multi-objectivization and surrogate modeling for neural network hyper-parameters tuning

Abstract

We present a multi-objectivization approach to the parameter tuning of RBF networks and multilayer perceptrons. The approach works by adding two new objectives - maximization of kappa statistic and minimization of root mean square error - to the originally single-objective problem of minimizing the classification error of the model. We show the performance of the multi-objectivization approach on five data sets and compare it to a surrogate based single-objective algorithm for the same problem. Moreover, we compare the multi-objectivization approach to two surrogate based approaches - a single-objective one and a multi-objective one. © Springer-Verlag Berlin Heidelberg 2013.