A toolkit for the search of the most general interpretable hypotheses

We apply first order logic (FOL) to formalize the problem of "meaningful generalization", finding the most general and easily interpretable hypotheses. Our software toolkit, LogicMill, is designed to solve this meaningful generalization problem as well as two other related problems: search for a maximal subsystem of mutually independent attributes, and aggregation of the hypotheses in the concise rules. We describe all three algorithms used for these purposes. Application of the toolkit to the data from various public domains demonstrates that LogicMill not only produces concise interpretable hypotheses and decision rules, but also it can compete in the prognostic power with traditional predictive learning algorithms.