Local and hybrid learning models in forecasting natural phenomena

Modular models (committee machines, mixtures of experts) are comprised of a set of specialized (local) models each of which are responsible for a particular region of input space, and trained on a subset of the training set. The known algorithms for allocating such regions typically do this in automatic fashion. In forecasting natural processes domain experts, however, want to see more domain knowledge behind such allocation, and to have certain control over such allocation and the choice of models, making thus the overall model hybrid. The paper presents some of the approaches to building modular and hybrid models, new algorithms and reports case studies in the area of river flow forecasting.