Seminar on Computational Learning and Adaptation

  Finding Plausible Explanations of Anomalies

Will Bridewell

University of Pittsburgh

I will present a perspective on theory revision that characterizes the resulting revisions as "explanations" of anomalous data (i.e., data that contradict a given model). The plausibility of these explanations, as judged by a domain expert, is emphasized as opposed to their performance within a revised theory. An explanation generator implementing (part of) John Stuart Mill¿s Method of Induction was constructed that divides the available data into meaningful subsets to better resolve the anomalies. Experimental results showed that using relevant subsets of data can provide plausible explanations not generated when using all the data. Additionally, this work suggests that plausible explanations can be used to identify learning opportunities with new, unclassified data.

Return to the seminar schedule