Abstract
In this paper, we present Committee, a new multi-class learning algorithm related to the Winnow family of algorithms. Committee is an algorithm for combining the predictions of a set of sub-experts in the on-line mistake-bounded model of learning. A sub-expert is a special type of attribute that predicts with a distribution over a finite number of classes. Committee learns a linear function of sub-experts and uses this function to make class predictions. We provide bounds for Committee that show it performs well when the target can be represented by a few relevant sub-experts. We also show how Committee can be used to solve more traditional problems composed of attributes. This leads to a natural extension that learns on multi-class problems that contain both traditional attributes and sub-experts.