MULTI-HEURISTIC THEORY ASSESSMENT WITH ITERATIVE SELECTION by

MULTI-HEURISTIC THEORY ASSESSMENT WITH ITERATIVE SELECTION by Kareem Ammar Chairperson of the Supervisory Committee: Professor Tim Menzies Department of Computer Science and Electrical Engineering Modern day machine learning is not without its shortcomings. To start with, the heuristic accuracy, which is the standard assessment criteria for machine learning, is not always the best heuristic to gauge the performance of machine learners. Also machine learners many times produce theories that are unintelligible by people and must be assessed as automated classifiers through machines. Theses theories are either too large or not properly formatted for human interpretation. Furthermore, our studies have identified that most of the data sets we have encountered are satiated with worthless data that actually leads to the degradation of the accuracy of machine learners. Therefore, simpler learning is more optimal. This necessitates a simpler classifier that is not confused with highly correlated data. Lastly, existing machine learners are not sensitive to domains. That is, they are not tunable to search for theories that are most beneficial to specific domains. Our solution involves multiple heuristics to assess our theories and the production of simplistic intelligible theories. We propose a variable combination of heuristics to gauge theories against one another in order to provide the best theory for a specific class within a specific domain. Our learner, Iterative Quick hull or IQ, provides a set of standard statistical heuristics as well as the option to add any number of domain specific heuristics to assess simple, intelligible theories. .