Designing experiments for single output multilayer perceptrons

Abstract

Where should a researcher conduct experiments to provide training data for a multilayer perceptron This question is investigated and a statistical method for selecting optimal experimental design points for multilayer perceptrons is introduced. Two-class discrimination problems are examined in which the multilayer perceptron is viewed as a univariate nonlinear regression model. A selection criterion is developed based on the volume of the joint confidence ellipsoid for the weights in a multilayer perceptron. The criterion is minimized to find optimal design points. Minimization is accomplished using Powell's algorithm when the feature space is continuous and a discrete exchange algorithm when the feature space is discrete. An example is used to demonstrate the superiority of optimally selected design points over randomly chosen points and points chosen in a grid pattern. In addition, two measures are developed to rank the design points in terms of their relative importance.

Keywords:

• Neural networks
• experimental design
• nonlinear regression
• multilayer perceptron