The General Utility Problem in Machine Learning

Abstract

Experiments have revealed that uncontrolled application of the analytical learning paradigm results in knowledge having low utility. Because the performance element must consider low utility knowledge along with high utility knowledge, the proliferation of low utility knowledge eventually defeats the goal of improved performance. Experiments in empirical learning have demonstrated a similar phenomenon. Uncontrolled application of an empirical learning paradigm may result in inaccurate knowledge, and a post-processing stage is typically needed to repair the degradation in performance. The results from experimentation in both analytical and empirical learning imply a general utility problem in machine learning. This paper presents evidence for such a perspective and recommends a closer dependence between the learning paradigm and the performance goals for which it is designed. A new approach is presented along with experimentation that illustrates the applicability of the approach to the general utility problem.