Description
The selection of salient features and an appropriate hidden layer architecture contributes significantly to the performance of a neural network. A number of metrics and methodologies exist for estimating these parameters. This research builds on recent efforts to integrate feature and architecture selection for the multi-layer perceptron. In the first stage of work a current algorithm is developed in a parallel environment, significantly improving its efficiency and utility. In the second stage, improvements to the algorithm are proposed. With regards to feature selection, a common random number (CRN) addition is presented. Two new methods of architecture selection are examined, including an information criterion and a signal-to-noise based procedure. These methodologies are shown to improve algorithm performance.
