Constructing and analyzing supersaturated designs for experiments involving a large numbers of factors

Experimental designs are required to develop an industrial process, or evaluate a product or a system performance. In most of the cases in these experiments, practitioners may deal with a large number of factors that they wish to screen, to minimize testing or computational resources required. Supersaturated designs (SSDs) are helpful in this situation. SSDs are cost-effective testing strategies at the preliminary stages of scientific investigation, because the number of experiments n is less than the number of factors m to be investigated.; In this research a new approach to construct and analyze the SSDs is proposed. For construction, the proposed method utilizes the most recent works in the development of SSDs, and introduces a genetic algorithm approach for optimizing the criterion E(s2) that used in comparing SSDs. A comprehensive comparison was performed between the results of the proposed method and the previous published results. Most of the results obtained from the proposed method were better than the previously published results. New improved SSDs based on E( s2) were achieved, and are presented in this research.; Analyzing data from experiments using SSDs is challenging for two reasons. The first reason is that because n < m, the model with all factors will be over-defined. The second concern is the existence of the multicollinearity in this type of design. Methods for analyzing SSDs are reviewed and summarized. The most widely used analysis method for the data in SSDs is stepwise regression, which can often fail to identify the true significant factors in the model. A new proposed biased estimation technique for analyzing the data of SSDs is proposed. The new proposed technique combines ridge regression, and a modified best subset variable selection technique, to select the significant factors in the model.; In this research Monte Carlo simulation is utilized for developing and evaluating eleven case studies, with different configurations. The performance measures are the observed average errors, both in terms of failing to identify significant factors (type II) and selecting insignificant factors (type I). For the proposed method the errors were low for most cases and acceptable for the more difficult cases. A detailed design of experiments study was conducted on the different factors affecting type I, and type II errors. The results show that the proposed method performs better than stepwise regression in all the cases. Some opportunities are presented for furthering and improving this research as well.