The greatest achievements of Taguchi were:
- developing a methodology for experimental
design that could be used by people who
were not statisticians
- raising awareness of the importance of
minimizing process and product variation
- identifying the importance of robust
design; designing products and processes
that are as tolerant as possible to 'noise'
factors
The methodology of Taguchi is often criticized
as being statistically unsound, or at least
open to question. For example, it is not
hard to demonstrate that the Signal to Noise
approach does not necessarily give the best
solution.
The dangers of ignoring interactions, or
using arrays with confounded columns are
troubling. Taguchi does stress the importance
of carrying out conformation experiments
at the selected settings, and so this can
be regarded as a calculated risk.
A summary of some of the differences between
the classical and Taguchi approaches:
Classical
|
Taguchi
|
| Mainly Resolution
IV or V and all potential interactions
considered |
Mainly Resolution III (screening designs),
and only consider selected 2 way interactions |
| Use ANOVA analysis
and hypothesis testing to create a regression
equation |
Use mainly graphical methods, and base
the analysis on the largest S/N value |
| Randomization considered
important |
Randomization not important, replaced
by the 'outer array' |
| Confirmation runs
not considered essential |
Confirmation runs stressed |
| Little attention
paid to dispersion |
Dispersion is the key consideration |
| Emphasis on complying
with assumptions, particularly the normality
assumption |
Less emphasis on the finer points,
the S/N ratio is less sensitive to the
normality assumption |
The logic behind using the
Signal to Noise ratio to minimize or maximize
the response make sense, although probably
don't affect the outcome in most cases. The
idea of identifying 'control factors' when
trying to achieve a specific average response
is important.
The outer array concept is
effective if a small number of noise factors
that have a significant effect on the response
can be identified, and if the 'common cause
variation' is relatively small. This is often
the case when a product is being designed,
the conditions under which the product will
be used can usually be identified and simulated.
|
