Chapter 5 Robust Performance Tailoring with Tuning - SSL - MIT

from Chapter 2 that SA is only a stochastic search technique, and therefore optimality,
local or global, is not guaranteed or even expected. It is simply hoped that randomly
searching the design space finds a design that is near-optimal or at least very good. In
addition, the performance of the SA algorithm is sensitive to the cooling schedule, or
the number of designs that are generated. The design is more likely to be near optimal
as the number of iterations increases. However, the complexity of the TPF SCI model
leads to large computational expense since a NASTRAN model is generated and
analyzed with each performance evaluation. This analysis becomes very costly in the
RPT and RPTT optimizations since eight models (one at each uncertainty vertex)
must be generated at each design iteration. In fact, evaluating the RPT and RPTT
costs for 1000 designs requires over fifteen hours of computation time on a Pentium
IV machine with a 3.2 GHz processor. Due to the heavy computational burden the
number of SA iterations are limited and the entire design space is not well searched.
This problem is especially limiting in the case of the RPTT optimization because the
number of design variables is large. A unique set of tuning parameters is allowed for
each uncertainty vertex resulting in 30 design variables (4 tailoring parameters plus 2
tuning parameters at each of the eight uncertainty vertices). Therefore, it is possible
that RPTT designs better than the one presented here exist.
6.5 Summary
A high-fidelity integrated model of a SCI architecture for TPF is presented in detail.
The model consists of a structural model built in NASTRAN, an ACS controller and
passive vibration isolation. The model components are integrated in MATLAB with
the DOCS tool set and the RMS OPD due to realistic reaction wheel disturbances
is evaluated. Appropriate tailoring, tuning and uncertainty parameters are identi-
fied, and the model is used to perform PT, RPT and RPTT design optimizations.
The performance of the resulting designs is evaluated in the nominal and worst-case
uncertainty configurations. The worst-case realizations are then tuned through a dy-
namic tuning optimization, and the tuned performance is evaluated. It is shown that
208