Work

Complexe reguliere fractionele factoriële proefopzetten: enumeratie, evaluatie en applicatie

Year: 2023
Type: dissertation
Abstract: Whether it is to minimize the operating costs of a production line, or to optimize the taste of a new dish in a kitchen, experiments are present all around us in many shapes and sizes. Any experiment can be boiled down to the study of a system, where input factors are changed to see how they affect the output. In most cases, there is some kind of restriction on the available resources. There are many different types of restrictions. For example, the restrictions can be staff-related (e.g., due to the fact that specific tasks need to be performed by a skilled technician), time-related (e.g., due to the fact that experimental tests are time-consuming), or logistic (e.g., due to the fact that certain experiments require the shutdown of a production line in order to conduct an experiment). Statistical design of experiments allows experimenters to maximize the information obtained from an experiment while taking into account all constraints present and the limited resources available. However, as the number of restrictions grows, a feasible highly informative experimental design can become quite hard to find. Therefore, there is a need for catalog containing a broad range of experimental designs. In the first phases of any investigation, it is usually unclear which input factors from a list of potentially influential ones really affect the studied outcomes. In these cases, experimenters employ so-called regular screening designs to quickly identify the influential factors. In such designs, the factors are often tested at two levels so as to capture linear trends (when the factors are quantitative) or differences in output between the two levels (when the factors are categorical). However, some specific cases require more than two levels for one or more factors, whether it is to test for non-linear trends, or to test for differences in output between more than two categories. One kind of designs that can accommodate such cases has factors at both four levels and two levels. Even though they are quite useful, these designs have not received much attention in the literature. Therefore, they are the main focus of this thesis. As suggested by the title, this thesis is a triptych, similarly to the Adoration of the Mystic Lamb. In the first part, we perform an enumeration of four-and-two-level designs. While some catalogs of such designs exist, they do not cover large run sizes or large numbers of factors. We compare three enumeration algorithms, adapted from the literature on two-level designs, in terms of their efficiency to enumerate four-and-two-level designs. With the most efficient algorithm, we generate a catalog of four-and-two-level designs covering experiments with up to 128 tests, up to three four-level factors and up to 20 two-level factors. We also present a web application built to access the catalog in a user-friendly way. In the second part of this thesis, we discuss three ways in which four-level factors are most often used in the literature: as blocking factors, as continuous factors, and as categorical factors. For each of these scenarios, we develop a criterion that allows us to select the best four-and-two-level design, for a specific run size, and a specific number of factors. We then use these criteria to characterize our catalog of designs presented in the first part. To facilitate the use of these designs by practitioners, we present tables with the best designs, for all run sizes and numbers of factors in the catalog. We also adapt the web application to permit the ranking of the designs according to the criteria we developed. In the last part of the thesis, we present an application of a regular screening design to study the effect of several factors on the variability in the measurements in a cell-based bio-assay. As the experimental tests involved many logistic restrictions, it was not straightforward to obtain a feasible high-quality experimental design. We present a step-by-step approach to design the experiment, taking into account all of the restrictions. Finally, we further consolidate our approach by constructing designs for alternative sets of logistic restrictions.
Source: Lirias
Author: Bohyn, Alexandre
Institution: KU Leuven
Language: English
Open Access status: green
OpenAlex ID: W7045949248
Cites: 60
Cited by:
Related to: 10