Inductive Modeling of Objects and Phenomena by the GroupMethod of Data Handling: the Shortcomings and Ways of TheirElimination

Original results of a research of an efficient computing method - group method of data han-dling are presented. Key shortcomings on each significant procedure of a classical algorithm arerevealed and systematized, and also ways of their elimination, including author’s modificationsare presented. In particular, the use of dispersion and an assessment of dispersion (Fischer’scriterion) is proposed as an assessment of accuracy of the received result, additional “internal”criterion for evaluation of adequacy of model in various tests during the fixing of input dataand changing of characteristics of an algorithm, and determining the optimal complexity of themodel. To solve the convergence problem of the classical algorithm, it was proposed to usethe methods of dispersion, factor and correlation analysis to eliminate non-informative features,modify the criterion for stopping the algorithm. The use of regularizing functionals is suggestedto solve the problem of multicollinearity of input characteristics and increase the stability of theobtained model, etc. A complex of computer modeling programs was developed, realizing an ef-ficient modified algorithm of GMDH with the considered modifications and also methods of adispersion analysis, correlation analysis, component analysis, elements of the regression analy-sis and others. The conducted researches and the received practical results can become a basisfor development with use of Machine Learning and Data Science technologies of the automaticsystem of computer modeling, the intellectual analysis and the data processing.