Creating decision criteria from examples: the criteria learning system (CRLS)

This thesis describes research on a machine learning approach to automated knowledge acquisition. It focuses on the needs and expectations of problem-solvers in the domain of medicine. It outlines an approach to learning criteria-based knowledge from examples and describes the implementation of a program called the CRiteria Learning System (CRLS) which learns rules in the form of criteria tables. The program learns with a bias for unate (monotone) boolean functions which display non-equivalence symmetry. These biases are described along with their applicability to the problem of learning decision criteria. The thesis details the results of the application of CRLS to ten different biomedical problem domains, and shows that the unate bias results in more comprehensible decision rules which have better diagnostic performance than rules induced with a conjunctive bias. The system produces rules that are simple, understandable, and have appropriately tuned diagnostic performance. Comparison with other machine learning programs shows that CRLS also requires less processing time. The explanation for these very favorable results is related to the appropriateness and strength of the two components of learning bias applied here: unateness and non-equivalence symmetry. It is concluded that criteria tables can be learned efficiently by machine learning techniques; they appear to be especially appropriate in biomedical domains.