Sharing best practice in design of experiments in UK
Sep 27, 2013 9:57 AM
JMP was delighted to welcome Peter Goos and Bradley Jones, two thought leaders in the area of design of experiments (DOE), to our Marlow office in the UK last week to talk about best practice in DOE.
Box, Hunter and Hunter explained in their seminal text, Statistics for Experimenters, that there is a gap in understanding between our model of a system and reality. When there is a problem meeting a customer’s requirements, then we need to bring our model closer to reality in order to find a solution. We can update our model by analysing historical observational data and designing experiments to capture new data in the most efficient way. JMP Pro allows us to perform a better analysis of our historic data through techniques such as generalised regression to robustly identify which variables are important. This holistic analysis approach means that we can get to the knowledge that we require in fewer iterations, thereby saving time and cost in helping our organisation make the right decision.
The latest technique for collecting more information in fewer runs is definitive screening designs. This award-winning design was developed by Bradley Jones and Chris Nachtsheim and allows you to understand curvature and interactions at the same time as identifying the main effects, reducing the need for follow-up experiments. Definitive screening designs are causing quite a stir in the industry and are regarded as so important that the paper that Jones and Chris Nachtsheim wrote about them won the prestigious ASQ Brumbaugh Award.
Jones and Peter also talked about two key concepts covered in their book Optimal Design of Experiments:
They introduced the concept of blocking and explained its importance in managing variation, using the vitamin stability example found in their book.
They explained the importance of managing hard to change factors by using split plot designs, using an automotive wind tunnel example from their book. Two key designs that they covered are D-optimal designs, which focus on the precise estimate of factor effects and which are ideal for screening and identifying which factors are active, and I-optimal designs, which focus on precise predictions with your model and are ideal for process optimisation.
It was a privilege to listen to two of the leading thinkers in DOE, and we look forward to welcoming them to office again in the near future.