Comparative Evaluation of Convergence's Speed of Learning Algorithms for Linear Classifiers by Statistical Experiments Method
Leonid S. Fainzilberg, Nataliia A. Matushevych · Kibernetika i vyčislitelʹnaâ tehnika · 2018
Introduction.One of the main tasks of artificial intelligence is pattern recognition, which is often reduced to determining the discriminant function parameters in the multidimensional feature space.When recognizable objects can be completely separated by a linear discriminant function, the task is reduced to the linear classifier learning.There are many algorithms for linear classifiers learning, two of which are the Rosenblatt learning algorithm and the Kozinets algorithm. The purpose of the article is to investigate the properties of the Rosenblatt and Kozinets learning algorithms on the basis of statistical experiment by the Monte Carlo method.Methods.Two algorithms for linear classifiers learning have been studied: Rosenblatt and Kozinets.A number of researches have been performed to compare the convergence rate of algorithms for a different number of points and for their different location.Variation of the iterations number of algorithms spent on samples of different sizes was analyzed.Results.Statistical experiments have shown that for a small sample size in approximately 20% of cases the convergence rates of the Rosenblatt and Kozinets algorithms are the same, but with the increase of observations number, the Kozinets learning algorithm proved to be the absolute leader.Also, the convergence rate of the Kozinets learning algorithm is less sensitive to the location of points in the learning sample.