Reliable nearest neighbors for lazy learning

A key problem of memory based learning methods is the selection of a good smoothing or bandwidth parameter that defines the region over which generalization is performed. In this article we present a novel algorithm to answer this question by utilizing the information from confidence intervals, to compute a bandwidth. The basic idea is the usage of confidence intervals to get a statistical statement about the quality of fit between estimated model and process. As long as the prediction intervals of a certain model include the neighboring data points of an incremental growing validity region, it is considered to be a good fit.