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.
and added to the objective function:<br />
yk =argmin<br />
y∈S {f (y)+µkB (y)} k =0, 1,... (4.8)<br />
The subscript k indicates the iteration number, and the weight of the barrier function<br />
relative to the objective decreases at each iteration allowing the search to approach the<br />
constraint boundary. The search direction can be determined through any standard<br />
gradient search procedure such as steepest descent, Newton’s method or conjugate<br />
gradient (See Appendix A). However, the step-size must be properly selected to<br />
ensure that all iterates lie <strong>with</strong>in the feasible region, S. An interior point method is<br />
attractive for the problem of hardware tuning since all iterates are feasible.<br />
The barrier method is implemented in MATLAB as outlined in Figure 4-8. The<br />
initial iterate, x0, termination condition tolerances and barrier parameters, ɛ and µ0,<br />
are inputs to the algorithm. The iteration loop begins by initializing the algorithm<br />
as shown. Then the constraint equations, their derivatives and the barrier function<br />
(Equation 4.6) are calculated analytically at the current iterate. The tuning param-<br />
eters are set to the current iterate on the hardware and a test is run. This data is<br />
saved as f (yk), and the objective cost at this iterate is evaluated. Next, the gradients<br />
of the objective are needed to generate a new search direction:<br />
∂J (yk)<br />
∂y<br />
= ∂f (yk) ∂B (yk)<br />
+ µk<br />
∂y ∂y<br />
∇J (yk) = ∇f (yk)+µk∇B (yk) (4.9)<br />
The performance gradients, ∇f (yk), are determined <strong>with</strong> Equation 4.3, requiring<br />
an additional hardware setup and test, while the barrier gradients are calculated<br />
analytically:<br />
∇B (yk) =−<br />
r�<br />
j=1<br />
�<br />
− 1<br />
g (yk)<br />
�<br />
∂g (yk)<br />
∂y<br />
(4.10)<br />
The objective gradients are then used to find the new search direction. The MATLAB<br />
implementation developed for this thesis allows either a steepest descent or conjugate<br />
gradient directions. Once the search direction is obtained the step-size is calculated<br />
125