| 成龙 |
2009-09-15 14:42 |
How Many Rays Do I Need for Monte Carlo Optimization?
4 z^7T
While it is important to ensure that a sufficient number of rays are traced to I(uM` g
distinguish the merit function value from the noise floor, it is often not necessary to nSp O TQ
trace as many rays during optimization as you might to obtain a given level of yA~1$sA1
accuracy for analysis purposes. What matters during optimization is that the p]rV\,Yss
changes the optimizer makes to the model affect the merit function in the same way ]jSRO30H3<
that the overall performance is affected. It is possible to define the merit function so :"'*1S*
that it has less accuracy and/or coarser mesh resolution than meshes used for cJ#n<Rsz
analysis and yet produce improvements during optimization, especially in the early :u`gjj$:s
stages of a design. 2(k m ]H^
A rule of thumb for the first Monte Carlo run on a system is to have an average of at z:oi @q
least 40 rays per receiver data mesh bin. Thus, for 20 bins, you would need 800 rays 1h3`y
on the receiver to achieve uniform distribution. It is likely that you will need to s\ e b
define more rays than 800 in a simulation in order to get 800 rays on the receiver. L| ]fc9W:
When using simplified meshes as merit functions, you should check the before and d)kO W!5\
after performance of a design to verify that the changes correlate to the changes of -_BX\iP{
the merit function during optimization. As a design reaches its final performance PQ 2rNY6
level, you will have to add rays to the simulation to reduce the noise floor so that ; sAe#b
sufficient accuracy and mesh resolution are available for the optimizer to find the )Lg~2]'?j
best solution.
|
|