Aprendizado incremental com memória parcial via grafo de Gabriel

Abstract

This work presents the development of incremental learning algorithms with partial memory, where the partial memory is obtained by the Gabriel graph structure. Four incremental techniques are proposed. In three of them, the partial memory is obtained by Gabriel graph and noise elimination. Besides considering Gabriel graph and noise elimination, the fourth incremental technique selects relevant patterns from a discrepancy measure of the data true distribution. Statistical tests to evaluated the methods were performed. These tests evaluate the equivalence of the incremental approach with the traditional approach (data separated into training and test), and also the comparison of the incremental algorithms with some others in the literature. The comparative results show that the developed techniques are efficient and have the particularity of not requiring experts to determine the parameters (in most technics). A brief analysis of the influence of the size of the data window indicates that the window size might not be a decisive parameter for the success of the algorithms.