On Optimal Projective Fusers For Function Estimators

We propose a fuser that projects diier-ent function estimators in diierent regions of the input space based on the lower envelope of the error curves of the individual estimators. This fuser is shown to be optimal among projective fusers and also to perform at least as well as the best individual estimator. By incorporating an optimal linear fuser as another estimator, this fuser performs at least as well as the optimal linear combination. We illustrate the fuser by combining neural networks trained using diierent parameters for the network and/or for learning algorithms.