Finding Optimum Design Parameters using Taguchi method?

[DavidSt]

Hi all,

I have a design with 6 factors (parameters) with levels varying from 8 to 4.
The goal is to find the combination of factors that gives the largest output value in the least number of experimental trials.

Most of the factors affect the output non-linearly, and many factors interact with one another non-linearly.

I have some prior distant experience with Taguchi methods, but not other DOE methods.
From what I recall, Taguchi methods work best to predict the optimal output value when the factors are linear and independent of one another (little to no interaction).

I was wondering whether people have had success implementing Taguchi methods for the design described above (i.e. multiple, multi-level, non-linear, interacting factors)? I would like to stick to Taguchi methods if possible if it will do the job.

To emphasize, although it would be nice to analyze the interaction between the factors, it is not of prime importance.
Also, accurately predicting the optimal output value is not of prime importance.
However, accurately predicting the factor levels that result in the optimal output value (or within 5% of the optimal for instance) in the least number of trials is of prime importance.

Thanks,
David

 

[Miner]

Taguchi had many important contributions to the concept of robust design, but there are more efficient approaches to optimization than those he proposed. In your situation, I recommend a sequential experimentation approach.

1. If not much is known, start with 2^k screening designs with center points.
2. If a lot is known, use response surface designs to optimize
3. If starting with a 2^k screening design:
   1. If there are significant main effects, but no interactions or curvature, explore the path of steepest ascent(descent)
   2. If there is an interaction, move the next experiment's center point to the "best" corner of the current design
   3. If there is a curvature effect, add axial points and analyze as a response surface design

Another approach to consider is EVOP (EVolutionary OPtimization).

 

[DavidSt]

Hi Miner,

Thanks for the response and sorry for the delay.

How much better would you suggest the sequential experimentation approach is compared to the Taguchi method for finding the optimal (or near optimal) factor values as described in my original post?

Using the Taguchi method, I could use a 32 trial design for the six factors of varying levels (8 to 4). When you say there are more efficient approaches, do you mean that the 32 trial Taguchi design would not likely yield parameter values that are in fact optimal due to the non-linear, interacting nature of the factors? (since if you are referring to the number of trials, 32 trials is quite efficient for my needs).

Assuming the former, could you recommend any good resources (webpages, books, ...) for a beginner to response surface design or 2^k screening designs? A resource that explains the concepts well, with straightforward examples, that would allow me to use the methods for my own design?

Thanks,
David