Frank Kwasniok and Leonard A. Smith, May 14, 2003
Abstract
An algorithm for optimizing local predictors constructed from data is discussed. The approach considers the real-time selection of a learning data set, extracted from a data stream, i.e., a data source where retaining and processing all observations is impractical. The refined learning set selectively covers those regions of state space which contribute most to the accurate prediction of the underlying dynamical system. The method is illustrated in the context of local linear prediction using the chaotic Ikeda laser map as an example.