N-learners problem: system of PAC learners
Nageswara S. V. Rao, E.M. Oblow · University of North Texas Digital Library (University of North Texas) · 1997
A system of Probably Approximately Correct (PAC) learners, where each learner had produced a hypothesis by employing empirical risk minimization methods, is considered. When no access to additional examples is available, our objective is to make the system at least as efficient -- in terms of normalized precision or confidence -- as best of the learners. Two separate cases are studied. In the first case, the training samples used by the individual learners are known; a method that approaches (in a weak convergence sense) the optimal Bayesian fuser is proposed. In the second: Case, the training samples are not known; majority and location-based fusers are shown to achieve the objective.