Author(s)

J. Voracek

Abstract

This lecture explains the design method of an adaptive algorithm, which is able to solve input-output formulated problems. Desired output classes, for example possible declarations, solutions or control actions, in which a input vector is placed, are made by the terminal nodes of a binary tree. From a theoretical point of view, the described task consists of two basic sections. The first is to find a way of automatic tree creation during the learning phase The second is to formulate a uniform decisive rule, applied to nonterminal nodes. An experimental basis was a node element modified perceptrone (neurone), adapted by a back propagation method. The tree branches are derived from normalised distances of single images in feature space to guarantee their linear separability. The chosen way of solving guarantees full adaptation, even for complicated tasks. Algorithm

Keywords