Pavel F. Chernavin, Nikolai P. Chernavin, Fedor P. Chernavin

Ural Federal University

When solving practical problems, quite often there is a need for the simultaneous application of machine learning and operations research methods. Since many methods for solving problems in both areas can be based on fundamentally different mathematical tools, it will be impossible to combine their results into a single model. This article presents an interconnected set of machine learning and operations research models designed to select the parameters of the technological processes. All models have a common mathematical formulation based on the mathematical programming problems with the partial-integer variables. The models have been tested on a real problem of selecting the composition of the charge and technological parameters of sinter production. It is presented in the sequence of their occurrence in the process of solving the problems set by the customers of the research. The first stage is based on solution of the regression problems with a selection of the most informative features and the degree of their influence on the output features is carried out. Then, based on the classification problems, the recommended areas of controlled input features are determined to obtain the high-quality products. These areas can have a rather complex geometric configuration in a feature space. Further, within the framework of the operations research problems, the reference states of a process are determined, to which it is necessary to strive. At the final stage, the results of all previous studies are combined into a single optimization model, which can be supplemented with the results of the researches obtained from other sources of information, if these results can be represented as the linear constraints. The proposed approach to the parameter optimization can be used in the various subject areas.

mathematical programming, machine learning, operations research, regression, classification, strength of sinter