MAJORITY AND LOCATION-BASED FUSERS FOR SYSTEMS OF PAC CONCEPT LEARNERS

A system of probably and approximately correct learners of Valiant typ e that infer concepts from a sample is considered. Each learner had be en trained by a sample using the methods of minimizing the empirical e rror, and no examples are available to the fuser. A majority fuser is known to make the composite system better than the best of the learner s in terms of normalized confidence (that corresponds to the same prec ision value). An analysis of general majority fusers is carried out to obtain bounds on actual and expected errors. Conditions under which t he r-of-N fuser performs better, in terms of normalized confidence or precision, than best of the individual learners are obtained. For a sp ecial class of statistically independent learners, slightly weaker con ditions are obtained. Two fusers that use the location information of a test point are proposed, and are shown to be better than a learner w ith least empirical error.