Seminar on Computational Learning and Adaptation



 
Average-Case Analyses of Induction
 
Pat Langley  and  Wayne Iba
Institute for the Study of Learning and Expertise
Palo Alto, California
{langley, iba}@isle.org
 
 
In recent years, researchers have made considerable progress on the worst-case analysis of inductive learning tasks, but for theoretical results to have impact on practice, they must deal with the average case. In this talk, we review progress on the average-case analyses of a variety of simple induction algorithms for which experimental results exist. The analyses to date assume two classes, independent Boolean attributes that follow a single distribution, and a certain class of target concepts. Given knowledge about the number of training cases, the number of irrelevant attributes, the amount of class/attribute noise,and the class/attribute distributions, one can derive the expected classification accuracy over different training sets. One can then use the derivations to explore the behavioral implications of the analysis by comparing predicted learning curves on artificial domains.  After giving examples of this reasoning process, we discuss some unresolved questions raised by the approach and outline directions for future research.

This talk describes joint work with Stephanie Sage and Kevin Thompson. For a recent paper on the average-case analysis of induction, see <http://www.isle.org/~langley/papers/abayes.ps>.


Date: Thurs., March 11 
Time: 4:30PM
Place: Cordura 100

Return to the seminar schedule