Sunspot number prediction by a conditional distribution discrimination tree

This paper describes a constructive learning system for conditional probability distribution estimations. First, the system carries out an unsupervised partitioning of the input space into small regions containing input vectors of the training set. It then computes coarse estimates of the output value conditional probability distribution, knowing the region of the corresponding inputs. Finally, a supervised region-gathering occurs when the output probability distribution, which are associated to neighboring regions, are not significantly different. This is done by using statistical tests. The region gathering stage allows the refinement of the estimates while reducing the complexity of the system structure. A neural discrimination tree implementation is presented and applied to sunspot number prediction. It is fast, parameter free, easy to use and provides excellent results.