Closed timholy closed 9 years ago
Wow, I think this could be a big breakthrough!!
It also could be unusably slow, if we want to register each stack. The mismatch computation is fast, it's the optimization I'm worried about---optimizing 10^7 parameters simultaneously is not child's play. We'll see how it works out.
This prepares the way for testing an older strategy---explicitly finding the minimum over an entire experiment, not just on a stack-by-stack basis---with the new registration code. The key commit is 2698f2b. It's mildly outrageous how easy this was to add, a possible sign that the Driver/Worker architecture is well-designed :smile:.
Still working on actually using this information for optimization (that is much harder).
While this PR will fail due to the julia bug in the module paths, locally (on cannon) it should now be possible to run
runtests.jl
.