Screening designs for large numbers of variables

In any system or process, the first step in achieving great improvements in product quality and process efficiency is to identify those factors that may affect the performance the most. Design of experiments is the tool of choice when it comes to the screening of significant variables or factors. In any design of experiments, the goals are to find the critical variables, screen out the irrelevant variables and establish the cause and effect relationships.;The screening method used depends on the number of factors that need to be screened. Fractional factorial designs are amongst the most widely used types of design in industry (Myers and Montgomery, 1995). However, those are not practical when the number of factors exceeds 20. Group screening may be used in those cases, however, it is not always obvious how the groups should be formed. Other screening methods for large numbers of variables include sequential bifurcation and iterative fractional factorial designs (IFFD). Sequential bifurcation is limited to quantitative variables, same-sign effects and monotone response functions and IFFD requires that only very few factors dominate.;The Trocine Screening Procedure (TSP) is an iterative method that uses a genetic algorithm and other heuristics to generate the experiments. At each iteration it uses information gathered from previous experiments to generate the next set. TSP worked reliably when the number of factors was less than 50 with three to four important factors.;This research determined ways of improving TSP's performance and determined its limitations in terms of the maximum number of factors that could be screened with acceptable Type I and Type II errors.;A new screening method, the B&C Screening Procedure, based on a one-at-a-time design is presented. The number of experiments required to screen n number of factors is always n + 1. As long as the system is under control B&C works well and there are no errors. Main effects and two-way interactions can be identified. Cases ranging from 20 to 100 factors were investigated. B&C's performance does not depend on the number of factors, the number of significant variables or the signs of the effects.