DOE - "Reverse" model

[#1]

I am looking in to making a DOE design, that should result in "reverse" model. Description of the inputs and outputs:

- 6 inputs that I can control during the experiments

- 6 outputs that I can measure during the experiments (but not control directly)

But the resulting model should make me able to predict what the input was, by measuring the outputs, thus the "reverse" model.

Is there anything that I need to take in to account when designing the experiment, or will a given design work as well one way as the other?

[#2]

Hi @Knekse,

I get your idea in theory, but in practice, this may be very difficult to do, as you have no evidence or certainty that there would be a unique combination of factor levels giving a unique specific output (or have you ?).
Just to show an example with the sample data Coffee Data.jmp, you can reach an optimum coffee strenght (here at 1,30) in a lot of different ways with only two factors (Time and Charge) : depending if charge is low, medium or high, you just have to adjust time to get the optimal coffee strenght.

So if your only information is the output (coffee strenght here), you may have no idea how you did get this specific value (with which combination of factors levels).
And since you have 6 inputs and 6 outputs, the complexity may increase.

Hope that helps you understand my point of view,

Optimum 1

Optimum 2

Optimum 3

 

[#3]

Hi @Victor_G 

Thanks for the input. In the example, you are linking two inputs to one output (Time and Charge to Strength). What I would be doing, is linking the 6 outputs, to each one of the inputs. In order to figure out, what the individual inputs might have been, from the 6 outputs. That would a somewhat different situation than the coffee example you are describing?

BR Knekse

[#4]

Hi @Knekse,

My point was to mention that even if you think factor 1 is linked to output 1, factor 2 to output 2, and so on, you might have interactions between your factors, or just several factors contributing to different responses (so you might get a solution with several possible factor levels changes), and also perhaps correlations between your outputs (so a change in one output is not independant from a change in another output). 

If your point is to link precisely one output to one factor, I would first look at data or create a DoE to check if these hypothesis of "independance" and absence of interactions look acceptable and reasonable based on data or not.

I hope this is clearer,

[#5]

Hi @Victor_G 

I don't think factor 1 is linked to output 1, quite the contrary. I have reason to believe that factor 1 is linked to all the outputs, and the same for the other factors.

I also expect to have interactions between the factors, so that is a very good point to keep in mind.

Thanks for the input