Chapter 5 Robust Performance Tailoring with Tuning - SSL - MIT

a statistical robustness measure such as standard deviation into the objective. The
RPT optimizations are applied to the SCI development model, and it is found that
anti-optimization provides the most robust design when the uncertainty is modeled
as a bounded uniform distribution.
The resulting RPT design is compared to the PT design and it is found that the
worst-case performance of the RPT design is significantly lower than that of the PT
design. However, the increase in robustness comes at the cost of nominal performance,
so that the RPT has a higher RMS performance at the nominal uncertainty values
than the PT design. A comparison of the worst-case performance values of the two
designs over a range of uncertainty bounds shows that RPT designs can tolerate
greater variation in the uncertainty parameters at a specific performance requirement
than the PT design. However, since nominal performance is sacrificed, RPT designs
cannot meet aggressive performance requirements given high levels of uncertainty.
The RPT designs fail when uncertainty is high because it is not possible to find val-
ues of the tailoring parameters that meet requirements under all possible uncertainty
realizations. Therefore, in Chapter 4, the concept of dynamic tuning is introduced.
It is defined as the adjustment of physical tuning parameters on hardware that brings
its performance to within required values. A tuning optimization is formulated and
applied to the PT and RPT designs in the worst-case uncertainty realizations. It
is shown that in both cases, dynamic tuning succeeds in reducing the performance
variance of the worst-case design, however, the effect is more dramatic on the PT
design. The worst-case uncertainty realizations of both designs, over a range of un-
certainty bounds, are tuned, and it is shown that dynamic tuning extends the range
of performance requirements that each design can meet at a given uncertainty level.
The practical issues associated with dynamic tuning are also considered. In reality
only limited performance data is available from the hardware and the actual values
of the uncertainty parameters are not known. Therefore some type of model up-
dating technique or real-time hardware tuning is necessary to improve the hardware
performance. A set of possible hardware tuning methods, ranging from model-only
methods that rely on the uncertain model to find successful tuning configurations to
213