Andrzej Skowronek

The article presents a method used in the process of generating smart designs of experiment in a dedicated computer program with the application of pseudo-random numbers, the use of which may have a considerable impact on the quality of generated designs. In order to increase the probability of generating the optimal design for the defined parameters of generating, a special method was applied, in which several designs were generated and the best of them selected, based on the equipartitional analysis parameters. The results confirm a significant positive influence of the analyzed method on the smart design’s quality.

This paper presents an analysis of comparing the effects of the application of the smart and traditional designs of experiment. The smart designs of experiment generated according to two various methods were compared to the central composite designs for two and three inputs. To check how the use of selected designs will affect the results of the experiment a computer simulation was performed, where a real research object was replaced by two special testing functions whose values were compared to the values predicted by the neural networks trained with the use of data sets based on compared smart and traditional designs of the experiment.

This paper presents results of computer simulation made to evaluate how the quality of the smart designs of experiment depends on the methods of generating. The quality of smart designs was evaluated by comparing the known values of special testing functions simulating the real research object and approximated values predicted by neural networks trained with the sets based on smart designs of experiment. The results suggest the possibility of significant reducing the number of experiments runs using studied smart designs.

This paper presents the results of a computer simulation, performed to check how the number of smart designs of experiment units affects the quality of information to be obtained in experimental research conducted on the basis of the analyzed designs. In the simulation, a real research object was replaced by a special testing function whose values were compared to the values predicted by the neural networks trained with the use of data sets based on smart designs of experiments containing various numbers of units.