Feature selection for time-series prediction in case of nondetermined estimation

Abstract

The issues of factors selection are discussed in the article for the case when estimation of a set of factors is not stochastic. Here the quality comparison of two sets of factors is only possible with some probability, and modification of existing methods is required for their correct operation. For this purpose there is a proposal of CGA Compact Genetic Algorithms utilization the scheme of factor selection being indicated. For stochastic estimation of a set of factors the step of training is updated for genetic algorithms. Results are obtained for the standard benchmarks and Internet - traffic forecasting task.