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.
where the notation yij indicates the j th tuning parameter at the i th uncertainty vertex.<br />
Keep in mind, that the �yi vectors are only the tuning configuration at the vertices of<br />
the uncertainty space and are not necessarily the final values of the tuning paramters.<br />
The hardware tuning process discussed in <strong>Chapter</strong> 4 must still be employed if the<br />
hardware realization does not meet requirements. However, the assumption is that if<br />
�x is chosen during tailoring such that the system is tunable at the uncertainty vertices<br />
then the actual hardware will be tunable across the entire uncertainty space. In this<br />
formulation, JTT and JRP T T are almost identical except that the tuning parameters<br />
are allowed to change <strong>with</strong> the uncertainty vertices in JTT.<br />
The formulation in Equation 5.3 improves the computational efficiency of the<br />
tailoring optimization since tuning optimizations are no longer required at each eval-<br />
uation of the objective function. However the optimization is still a combination of<br />
min-max formulations, and it is unclear how to obtain the analytical gradients. It is<br />
possible to further simplify the RPTT formulation by posing it as a simple minimiza-<br />
tion problem similar to the minimization form of anti-optimization (Equation 3.5):<br />
min<br />
�x,�yi,�z<br />
�<br />
JTT<br />
����<br />
(1 − α) z1 +α<br />
s.t �g (�x, �yi) ≤ �0<br />
JRP T<br />
����<br />
z2<br />
h1i (z1,�x, �yi,�pi) ≤ 0<br />
h2i (z2,�x, �pi) ≤ 0<br />
∀i =1...npv<br />
�<br />
(5.5)<br />
In this formulation the cost function consists only of the weighted sum of two dummy<br />
variables, z1 and z2. The tailoring for tuning and robustness metrics are included<br />
through the augmented constraints:<br />
h1i (z1,�x, �yi,�pi) = −z1 + f (�x, �yi,�pi) (5.6)<br />
h2i (z2,�x, �pi) = −z2 + f (�x, �y0,�pi) (5.7)<br />
155