Fast bounded smooth regression with lazy neural trees

Abstract

Proposes the lazy neural tree (LNT) as the appropriate architecture for the realization of smooth regression systems. The LNT is a hybrid of a decision tree and a neural network. From the neural network it inherits smoothness of the generated function, incremental adaptability, and conceptual simplicity. From the decision tree it inherits the topology and initial parameter setting as well as a very efficient sequential implementation that out-performs traditional neural network simulations by the order of magnitudes. The enormous speed is achieved by lazy evaluation. A further speed-up can be obtained by the application of a windowing scheme if the region of interesting results is restricted. >