A new system for learning by induction, called LAB, is presented. LAB (Learning for ABduction) learns abductive rules based on a set of training examples. Our goal is to find a small knowledge base which, when used abductively, diagnoses the training examples correctly, in addition to generalizing well to unseen examples. This is in contrast to past systems, which inductively learn rules which are used deductively. Abduction is particularly well suited to diagnosis, in which we are given a set of symptoms (manifestations) and we want our output to be a set of disorders which explain why the manifestations are present. Each training example is associated with potentially multiple categories, instead of one, which is the case with typical learning systems. Building the knowledge base requires a choice between multiple possibilities, and the number of possibilities grows exponentially with the number of training examples. One method of choosing the best knowledge base is described and implemented. The final system is experimentally evaluated, using data from the domain of diagnosing brain damage due to stroke. It is compared to other learning systems and a knowledge base produced by an expert. The results are promising: the rule base learned is simpler than the expert knowledge base and rules learned by one of the other systems, and the accuracy of the learned rule base in predicting which areas are damaged is better than all the other systems as well as the expert knowledge base.

PDF, PS

Masters Thesis, Department of Computer Sciences, The University of Texas at Austin. 53 pages.

@MASTERSTHESIS{thompson:masters93, title={Inductive Learning For Abductive Diagnosis}, author={Cynthia A. Thompson}, month={August}, school={Department of Computer Sciences, The University of Texas at Austin}, address={Austin, TX}, pages={53 pages}, url="http://www.cs.utexas.edu/users/ai-lab?thompson:masters93", year={1993} }