Auto-adaptive neural network tree structure based on complexity estimator

Abstract

We present in this paper an auto-adaptive neural network tree structure that is used to solve classification problems. Two main ideas are behind this neural network tree structure. The first one is that it is easier to solve simple problems than difficult ones, so we divide an initial hard classification problem into several simpler/easier ones in a recursive way: "Divide To Simplify". In order to perform the classification problem decomposition, we need to answer two following questions: is the problem difficult? And then, how to simplify it? The second idea is to use some classification complexity estimation methods to evaluate the problem difficulty. We propose an algorithm which estimates the problem complexity, divides it into several easier ones and builds a self organized neural tree structure that solves this problem efficiently. © Springer-Verlag Berlin Heidelberg 2003.