| dc.description.abstract | Non-regular designs have nice properties regarding run economy. However, standard methods of analysing regular designs are not applicable as a results of possible non-orthogonal contrast columns. We investigated three factor based methods of analysing non-regular factorial fractional designs and also performed follow-up runs in identifying the active factor subspace for an experiment. We studied the six factor 12-run PB and 16-run designs with some simulated models and also on a real data set from the metal cutting experiment by Garson (2000). In our investigation, the 16-run design displayed a relatively significant performance in defining factor activities for models with four active factors over the 12-run PB design. The methods studied in this thesis were found to produce similar results in identifying one, two and three active factors. All the methods performed very well in identifying models with at most three active factors. However, for models with four active factors, the study revealed that the methods have shortcomings in identifying the correct active subspace. The Box-Meyer search estimated variance was lower compared to that of the other two methods. The projection based method is very simple to use, with much less intuition and was robust under various conditions of model's variability. It is not appropriate to use the method alone whenever the results indicate that three active factors are insufficient. However, follow-up experiments help to improve performance of the method. This study recommend the use of the factor based methods in defining factor activities for experiments. | |
