312158_en-US

liliconcom

DOE: changing Axial Distance for a RSM in Custom design

Hi,

doing a central composite design I can change the axial distance to change the axial points being more within or beyoung the face. Is there something simuilar in Custom Design using a RSM as base? I can do it with augment design too, but not within the custome design.

 

Thanks,

Martin

Re: DOE: changing Axial Distance for a RSM in Custom design

I don't think so, Martin. I can't think of a way to get axial points with Custom Design, other than by augmentation. The reason being that a design with axial points will not be optimal. I guess you could do it with factor constraints? But that would seem like a strange thing to do.

Re: DOE: changing Axial Distance for a RSM in Custom design

I recently saw this interesting paper from Cameron Wilden (a good customer of JMP at WL Gore) about adapting optimal design algorithms to create designs with axial points.

Re: DOE: changing Axial Distance for a RSM in Custom design

I just tested and also believe creating an i-optimal custom design with interaction effects in the model, then augmenting this design with axial points and center points should work. 

Thanks for the reference of the paper, I will check that as well. 

Re: DOE: changing Axial Distance for a RSM in Custom design

What about using a Discrete Numeric factor instead of a Continuous factor?

 

Also, why is it important to have axial points over having an optimal design for the model?

Re: DOE: changing Axial Distance for a RSM in Custom design

I guess you mean discrete numeric with 5 levels then to address the axial points need for 5 levels instead of 3 for a star-point on the face. That would be an option, right. 

I don't have full insight to their problem but I beleive their process can't be operated at one or more of the extremes of the design region. One might argue, using an i-optimal Custom Design with constraints might be the best approach. And due to the fact that given a reasonable magnitude of overall experimental error, say a 5-percent RSD, the reduced prediction error for, and improved estimation of, quadratic (curvature) effects may not outweigh the added complexity of requiring each variable to be run at five levels.