The creation of human-friendly rules for long-range weather prediction using the LAPART neural architecture

Often the user of a model, such as one represented by a neural network, desires greater insight into the process which leads to the classification of a pattern or prediction of a future state in time series. The laterally primed adaptive resonance theory neural network (LAPART) has been applied to a weather index time series in order to assess the predictability of future states of the atmosphere given certain antecedent conditions. LAPART creates classes based on weights learned during training which are interpreted as “rules” that are easily analyzed and understood by the human analyst, and which may be integrated into expert systems or processed by the human with other modeled or observed data. Many problems require the system to process vast amounts of diverse data, as in weather prediction, which eventually form the basis for a human decision by the weather forecaster. LAPART provides a logical approach for compressing vast amounts of data into a few “human-friendly” rules