13346_en-US

liliconcom

path of steepest ascent or descent<meta http-equiv="Content-Type" content="text/html; charset=UTF-8" />

Hi,

I have a client who asked that how is ‘path of steepest ascent or descent (POA/POD)’ being performed in JMP? He said that there is an macro function in Minitab for such purpose. I've never used Minitab before and have no idea how to do POA with it. Based on my understanding, the direction is given by the coefficients of the first-order model, and there is a formula to define the distance to move at every step along the path. So would appreciate that if you can share an example or add-in in JMP to address this kind of problem.

Thank you.

Yanfei

Re: path of steepest ascent or descent

Yanfei: Another aspect of the scenario the customer may be pursuing is multiple response co-optimization? If that's the case the logic and concept gets a bit more elegant in JMP because JMP uses a Desirability Function to help insure the solution meets the conditions set forth by the Desirability Function. The JMP Desirability Function algorithm requires a weighting factor (weights I believe are set forth in a geometric average kind of logic?...I have to admit it's been awhile since I last checked?) for each response which in turn influences the path the co-optimizer will take along the surface to a solution. Based on these weighting factors alone, the co-optimizer may find different solutions in the factor space.

Re: path of steepest ascent or descent

Thank you, Lou. It helps a lot.

Now I understand the process and I know it is easy to do that kind of calculation in JMP. And as Martin suggested, I would also recommend more advanced methods like DSD to him.

Re: path of steepest ascent or descent

Hi Martin,

I think the DSD is an interesting alternative to explore. By allowing experimentation over wider ranges, it offers the possibility of:

- identifying a local optima in one experiment (or two experiments if more than three factors turn out to be important)

- a simpler workflow to learn and master

- protects against missing important factors due to confounding with two factor interactions which is a risk in the workflow promoted by Minitab.

Malcolm

Re: path of steepest ascent or descent

Thank you so much, Martin.

I just installed Minitab, downloaded the macro and tried run an example to see what it does. From the output I know it just calculate the coefficients and the increment step of other factors given the step length of the base factor.

While thank you for the detailed calculation process, so that I can show it to client so they can write a formula in JMP when needed. And also DSD is quite a good suggestion for client. They should know they have better options even if they are not ready to change at present. And I think client would be very pleased to know there are advanced techniques to help them address their problems.

Re: path of steepest ascent or descent

Hi Yanfei,

looking at the macro in Minitab it is used like that:

1. Run a Screening Experiment to identify the important main factors

2. Fit a linear model for those

3. Now the macro will be used to identify new runs to prepare a RSM model. Therefore it will choose a base variable and walks along the path. Here is the description you find also in the macro:

4. Now run the new experiments

5. With all the data you now can model a RSM to identify the significant model effects and optimize the Settings.

That said I believe you could offer the customer to rund a DSD which allows for up to 3 variables this procedure in one step. In Addition if you have more variables the DSD will allow the main effects to be orthogonal in regards to the interaction and quadratic effects, which also allows with a good probability to find a good model right away. If you have constraints or many categorical variables you may want to use Custom designer for this purpose, or DSD with confirmation runs.

Anyhow, customers often know a certain way of doing things and rely on it. Therefore they want to have the same functionality again when switching to another tool. Sometimes there is an even better way but it is not what they know. So I believe it will be important to convince him that this Macro is not the only good way of doing what he wants to do.

I'm interested in other opinions to my thoughts.

Martin

Re: path of steepest ascent or descent

Yanfei,

If it's helpful here is the link to the Minitab Macro

http://support.minitab.com/en-us/minitab/17/macro-library/macro-files/quality-control-and-doe-macros/ascent/

Re: path of steepest ascent or descent

Thank you, Martin. Client asked about the functionality of JMP rather than consulting for a specific issue. I need to get more information from him and then let him know what we can do.

Re: path of steepest ascent or descent

Thank you, Don.

Yes, the biggest challenge for me is to decide how far to move at every step. As you said, it depends on a large part of domain knowledge rather than the formula.

While client didn't have any specific needs for this, he just ask "For first order model fitting, how is ‘path of steepest ascent or descent (POA/POD)’ being performed in JMP? Is there an equivalent macro function as in Minitab for such purpose?"  I am just not sure what does he mean about macro function in Minitab.

I’ll check with him later and get more information from him tomorrow.  

Re: path of steepest ascent or descent

Hi Yanfei,

the first thing that came into my mind is what does the customer wnt to achieve. POA/POD usually is used for optimization Problems and there are various of implementations out there. The basic is as Don described, looking at the gradient. However, those methods will likely result into a local Optimum unless you vary the starting Point and also the step width.

May the optimizer in the profiler will already do what he wants to achieve. Maybe not. I guess he has not just a 2D Thing but at least 3D Problem and he wants to find an max/min of a surface?

Martin

Re: path of steepest ascent or descent

POA or POD is the gradient of the prediction equation. The biggest challenge is not in its calculation, but how far to step once your outside of the design space (and you're often starting at the corner of the design space). It almost always requires domain knowledge. I suppose if you were doing one run at a time, you can start with a standard value (maybe a fraction of the design space range), calculate the difference between the expected and actual values at each step, and increase/decrease/stop depending on how large this value is.