Knowledge Based... For now
As a physicist, I used to resent this approach. It's not physical! The idea is to make use of known information, and convert it to some form of potential, or scoring function, or set of building blocks to build the RNA from scratch. It works, sometimes, particularly for smaller molecules (for larger ones, part of the big problem is sampling, or being able to find a good solution since there are so many possible ones...). In fact, it appears to work really well also for larger RNAs if you input enough constraints (determined experimentally).
Using a physics based approach is slow, and takes a long time for the solution to converge. And "physics based" right now refers to using parameterised force-fields. In reality, we don't quite know how well they work for RNAs. In fact, I was just reading a paper in a journal that these force-fields don't quite represent reality just yet (see http://www.citeulike.org/user/qshao/article/5491395 for instance). So the physics based approach just isn't feasible right now, and neither can we guarantee that it is good.
So for now, I'll just stick to the knowledge based business... And learn not to resent it (as much). There's so much information we can gather from high resolution crystal structures, and we're constantly moulding our understanding from these information. Looks like we really are in the era of information overload... I guess the key is to make sure we don't read in to much to it... (see http://personas.media.mit.edu/)