Multitask learning applied to evolving fuzzy-rule-based predictors

Abstract

This paper converts the online learning framework denoted FBeM (Fuzzy set Based evolving Modeling) to an online multitask learning evolving system. FBeM is data flow driven and recursively updates information granules that can be interpreted as the antecedent parts of functional IF-THEN fuzzy rules. The intersection of those information granules is directly interpreted here as defining a sparse graph of structural relationships among the IF-THEN fuzzy rules, so that the consequent terms of the rules, corresponding to linear regression models, can be produced by the solution of a regularized multitask learning problem. Being an evolving system, every time that the information granules are updated, including possibly merging and splitting of existing information granules, the multitask learning step should be re-executed, thus motivating our investment on scalability issues such as sparsity of the graph and convexity of the regularized formulation. Weather temperature and wind speed in eolian farms are taken as the two case studies devoted to online time series prediction. When compared to the original FBeM framework, which treats the learning of the regression models as independent tasks, and also to several other state-of-the-art evolving systems in the literature, our approach guides to an expressive gain in performance in most cases, even consistently resorting to a reduced set of IF-THEN fuzzy rules to synthesize the online predictors.