Gofman Y. Methods of decision trees construction in intelligent systems of diagnosis

The thesis is devoted to development of methods for inductive learning of decision trees to improve the levels of interpretabylity and generalization, as well as the speed of synthesis of recognizing models in intelligent systems. Object of research - the process of synthesis of decision trees. The subject of research - the methods of inductive learning of decision trees in intelligent systems of diagnostics. The purpose is the creation of methods of decision trees identification by using a stochastic approach and additional information about the investigated objects, which will increase the accuracy, speed of construction and operation, the levels of generalization and interpretability of the synthesized decision trees. Methods of research are: evolutionary and multi-agent search, the theory of decision trees, the theory of belief functions, mathematical statistics. In the thesis the process of construction and methods for the synthesis of decision trees in intelligent systems are analyzed. The evolutionary method for the synthesis of decision tree is developed. It is based on a stochastic approach and does not use a greedy search strategy that allows to construct decision trees with sufficient generalization and approximation properties with the small amount of nodes. The modification of the decision tree synthesis method ID3 is proposed. It is calculated the pignistic probabilities of instances referring to classes based of the theory of belief functions that allows to classify the instances under the uncertainty or incompleteness of the data. The method of decision trees constructing is created. It allows to induct the linguistic rules and to develop the expert systems based on a more interpret linguistic rules databases. A method for the synthesis of neuro-fuzzy networks based on decision trees is proposed. It does not require solving optimization to adjust the values of model parameters. The automated system for the synthesis of decision trees is created. It allows to construct simple and convenient model for further analysis in the form of decision trees. The problem of technical diagnostics of vehicle bodies is solved using the proposed methods and software. The experiments of comparison of the proposed methods with known analogues are carried out.