Evolving Artificial Neural Networks for Nonlinear Feature Construction


We use neuroevolution to construct nonlinear transformation functions for feature construction that map points in the original feature space to augmented pattern vectors and improve the performance of generic classifiers. Our research demonstrates that we can apply evolutionary algorithms to both adapt the weights of a fully connected standard multi-layer perceptron (MLP), and optimize the topology of a generalized multi-layer perceptron (GMLP). The evaluation of the MLPs on four commonly used data sets shows an improvement in classification accuracy ranging from 4 to 13 percentage points over the performance on the original pattern set. The GMLPs obtain a slightly better accuracy and conserve 14% to 54% of all neurons and between 40% and 89% of all connections compared to the standard MLP.
Helmut A. Mayer

Last modified: Sep 29 2014