A machine-discovery approach to the evaluation of hashing techniques

摘要

This paper, describes an inference technique based on machine discovery for drawing conclusions from experimental results. Given access to the results of a full-factorial experiment, the inference technique finds three types of empirical generalizations. First, the best and worst values for each independent attribute, in terms of their effect on the dependent attribute, are identified. Second, direct and inverse relationships are found by applying regression to rank frequencies. Finally, cases where restricting a variable to a single value yields different behaviour from usual are identified. These three types of generalizations are produced in the form of English sentences. Experimental results using a Prolog implementation indicate that the inference technique finds many of the same generalizations as human researchers did in a fundamental study of the performance of hashing techniques.