Chapter 5 Robust Performance Tailoring with Tuning - SSL - MIT
Chapter 5 Robust Performance Tailoring with Tuning - SSL - MIT
Chapter 5 Robust Performance Tailoring with Tuning - SSL - MIT
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
Table 4.6: <strong>Tuning</strong> results on fifty hardware simulations: performance and number of<br />
required tests.<br />
# of Sims # Hardware tests<br />
Method Success Failure Maximum Average<br />
Iso <strong>Tuning</strong> 50 0 4 2.5<br />
Hardware: BSD 50 0 231 46.2<br />
Hardware: SA 50 0 2618 790<br />
Worst-Case Model 29 21 2 2<br />
Nominal Model 29 21 2 2<br />
Optimized Model 24 26 2 2<br />
AO <strong>Tuning</strong> 0 50 2 2<br />
successful on only 50% of the trials. The only exception is AO tuning, which does not<br />
successfully tune any of the simulations. It is worthwhile to note that, although the<br />
model-only methods failed in the example discussed previously, there are some hard-<br />
ware realizations for which these methods are adequate. Since each of these methods<br />
require only two hardware tests, the initial test to assess nominal hardware perfor-<br />
mance and a second one to try the tuning configuration, there is little cost in starting<br />
the tuning process <strong>with</strong> these methods as much time and effort can be avoided if they<br />
succeed. If the model-only method fails, the tuning configuration is a good starting<br />
point for a hardware tuning scheme. Optimized model tuning and isoperformance<br />
tuning are particularly compatible in this way since the uncertainty parameters used<br />
for the model tuning also serve as an initial point on the isoperformance contour. It<br />
is also interesting to note that the nominal model and worst-case model tuning meth-<br />
ods result in the exact same numbers of successes. It turns out that these methods<br />
succeed and fail on the same hardware realizations indicating that these methods are<br />
interchangeable for this particular problem.<br />
The hardware tuning methods and isoperformance tuning, on the other hand, are<br />
consistently successful across the entire sample space. There is quite a difference in the<br />
number of hardware tests that are required by the three methods, however. Real-time<br />
simulated annealing is by far the least efficient requiring an average of 790 hardware<br />
tests across the sample space. This result is not surprising since the method is a<br />
random search of the tuning space. It is equivalent to simply trying random tuning<br />
147