Taguchi Design with ANOVA - 3 factors each with 3 levels

S

ShervinA

Hi,
I have a GA code and I want to use the Taguchi Technique as a parameter tuning tool. I have 3 factors each with 3 levels. I used L9 array and I ran 9 experiments and entered the results.
The problem is that I don't get any ANOVA result and no interaction graphs. Would you please help me with this. :)

Thank you so much in advance.
Shervin

P.S. I attached the Minitab file.
View attachment Taguchi.MPJ
 

Miner

Forum Moderator
Leader
Admin
Re: Taguchi Design with ANOVA

Minitab's Taguchi analysis does not provide much flexibility, so I recommend that you analyze this using the GLM analysis under the ANOVA menu. Another option is to define it as a custom General Full Factorial DOE.

Unfortunately, you will find that you cannot analyze any interactions because there are insufficient degrees of freedom to create the ANOVA table.
 
S

ShervinA

Re: Taguchi Design with ANOVA

Thank you so much for your help. It means a lot.
I have another question I have projects from a couple of years ago that I used Taguchi in Minitab and the ANOVA is reported during the Taguchi method process, but in this case, nothing more than signal to noise ratios are reported. I thought maybe something is wrong with my data.
I would be really grateful if you help me with this too.
 

Miner

Forum Moderator
Leader
Admin
Re: Taguchi Design with ANOVA

I have another question I have projects from a couple of years ago that I used Taguchi in Minitab and the ANOVA is reported during the Taguchi method process, but in this case, nothing more than signal to noise ratios are reported. I thought maybe something is wrong with my data.

If you attach the Minitab project file, I'll take a look. While I am quite familiar with Taguchi methods, I rarely use them anymore since I have found classical approaches to be be much better. I will have to review Minitab's capabilities in this area.
 
S

ShervinA

Re: Taguchi Design with ANOVA

If you attach the Minitab project file, I'll take a look. While I am quite familiar with Taguchi methods, I rarely use them anymore since I have found classical approaches to be be much better. I will have to review Minitab's capabilities in this area.
I guess I found the option under Taguchi design analysis as 'Fit linear model'. It didn't work so I used GLM as you mentioned and it worked. Thank you so much. However the r- sq value is equal to 63% and I guess I can't rely on Taguchi result. Do you think considering additional factors will help or Taguchi is not suitable at all for parameter tuning?!... (the Minitab file is attached in the first post)
I appreciate any ideas and thank you in advance for your brilliant ideas.
Thanks a ton!

Sent from my C5303 using Tapatalk
 

Miner

Forum Moderator
Leader
Admin
I don't like the reliance on S/N ratio. I understand why Taguchi went that route. His methodology was easier in the pre-computer days, and the engineers he was working with were in the telecommunications industry where S/N ratio was prevalent.

I now use classical methods and analyze the mean and standard deviations separately. Minitab adheres strictly to Taguchi's original approach and does not provide the options available under their classical designs.

I recommend using either classical designs, possibly including response surface methods, or an EVOP approach to tuning.
 
S

ShervinA

I don't like the reliance on S/N ratio. I understand why Taguchi went that route. His methodology was easier in the pre-computer days, and the engineers he was working with were in the telecommunications industry where S/N ratio was prevalent.

I now use classical methods and analyze the mean and standard deviations separately. Minitab adheres strictly to Taguchi's original approach and does not provide the options available under their classical designs.

I recommend using either classical designs, possibly including response surface methods, or an EVOP approach to tuning.

Thank you so much for your ideas[emoji4]
To be honest the first time you mentioned classical designs I thought you are talking about DOE full factorial[emoji16] but now that you named response surface and EVOP approach you got me so interested, and I want to learn them. So I'll do some research.
Thanks a lot [emoji120]

Sent from my C5303 using Tapatalk
 
S

ShervinA

Don't forget that classical designs include fractional factorials and center points.
Sounds like I have a long way [emoji16]
Thank you for reminding [emoji120]

Sent from my C5303 using Tapatalk
 
S

ShervinA

Don't forget that classical designs include fractional factorials and center points.
Dear Mr.Miner

Sorry to ask too many questions. I did some research and as you said I found EVOP much more reasonable and efficient. Thank you so much.
However I have another question that I can't find any answers to. My GA code takes a really long time to converge (2000 generations in 3-4 days) and I don't have enough time and resources to manage too many runs. Can I use EVOP or RSM in limited number of generations? I mean instead of running the algorithm till convergence every time I change a parameter value I'd like to compare the best answer found by the algorithm with different levels of crossover and mutation rate in just 100-200 generations? Will the best found parameter levels be reliable?
 
Last edited by a moderator:
Top Bottom