Ensemble of perceptrons with confidence measure for piecewise linear decomposition

In this study an ensemble of several perceptrons with a simple competitive learning mechanism is proposed. The objective of this ensemble is to decompose a non-linear classification problem into several more manageable linear problems, thus realizing a piecewise-linear classifier. During the competitive learning process, each member of the ensemble competes to learn from one linear subproblem in a reinforcement learning-like mechanism. The linearity of the ensemble members will simplify the task for interpreting the rule captured by the ensemble. Although the final goal of this study is to generate a “Whitebox” non-linear classifier, this short paper focuses on the explanation of the properties of the proposed model, while leaving the rule extraction part to the existing methods.