Linguistic Decision Trees for Fusing Tidal Surge Forecasting Models

Abstract

The use of linguistic decision trees as represented within the label semantics framework, is proposed for the fusion of multiple forecasting models. The learning algorithm LID3 is applied to infer a decision tree with branches representing a set of rules each identifying a probability distribution on the available models and where the constraints in each rule are generated from fuzzy labels describing the relevant input attributes. The resulting aggregated forecast for a given vector of input attributes $\vec{x}$, is then taken to be the mean value of the forecasts from each model relative to a probability distribution on models conditional on $\vec{x}$ as determined from the linguistic decision tree. The potential of this approach is then investigated through its application to the fusion of tidal surge forecasting models for the east coast of the UK.