Average-case improvements when integrating ML and KA

摘要

In this paper we develop a mathematical analysis, based on empirical measurements, of expected average improvements when integrating Machine Learning and Knowledge Acquisition systems in real-life domains. The analysis is based on the characteristics of component systems and combining techniques. Important characteristics include the accuracy of component systems, the degree to which component systems complement each other's weaknesses, and the ability of the combining mechanism to make good choices among competing component systems. Empirical measurements in a real-life application, in the Sendzimir rolling mill, have shown that integrating both approaches enables significant improvements. Improvements when combining systems in two oncological domains were smaller, yet positive again. Analytical average-case integrated models consisting of two systems are introduced. Conditions for improvements over the best, average and the worst system are established and the expected gains are analytically computed based on expected performances. Models strongly suggest that a reasonable integration of two systems offers significant improvements over the best single system in many or even most real-life domains.