Automated Split Plot Statistical Designer


This project will combine statistical expertise with combinatorial methods to help automate the task of constructing experimental designs. These designs are widely used in industry to determine joint effects among factors affecting the quality of an output commodity. The current design process can be tedious and error-prone, involving the manual comparison of joint-effects graphs, and utilizing incomplete tables of binary codes. We propose to design a suite of algorithms with a view to automating the above process. Our goal is to provide tools which will present the designer with a spread of alternative testing schemes, each scheme coming with a cost-benefit analysis. Freed from tedium and inexactitude, a designer can thus select with confidence suitable and cost-effective testing protocols.The project is a complicated mix of statistics and combinatorics, and will require a focused collaborative effort. Required ingredients include: the classification of small binary codes, an efficient subgraph isomorphism tester, and a specialized traveling salesman calculator. All these problems can be "solved" by modern generation and metaheuristic methods, but they must be carefully implemented with a view to statistical considerations. The end result will be of great utility to industrial designers. There is nothing like this in existence, and we anticipate keen interest from both practitioners and statistical software providers such as SAS.Similar design structures are also used in statistics to form a small number balanced resamples from which estimates of variability can be obtained for large surveys and other large complex data structures.