THE MLP-METAMODELS APPLICATION IN THE SURROGATE OPTIMIZATION TASKS

A computational technology for constructing metamodels was developed using modern achievements in the field of the experimental planning theory, intellectual data analysis and artificial intelligence, and experimentally determined patterns that allow metamodels efficient construction. The metamodels construction is performed on numerous examples with a goal function, which depends on two variables. As an experiment plan, a point generator is used, which fills the search space and during the implementation of which the Sobol LPτ-sequences are used. It has been established that by increasing the points number in the experiment plan and the hidden neurons number to a certain level, the metamodel parameters are improving; the use of various weakly correlated LPτ-sequences from their complete set does not significantly affect the quality indicators of the final metamodel. The results of numerous experiments testify to the possibility of using the proposed computational technology for constructing MLP-metamodels for approximating the primary goal functions with a rather complex response surface.