Monday, March 4, 2013

Monday: That should probably be my default image when I don't have any images


Genetic algorithms are passe, right?  I mean, we've all moved on to MCMC because you get a more robust mathematical/statistical understanding of the underlying parameter space, and you can get out of local minima easily.  I guess I never really understood what all the fuss was.  If I know A is a better solution than B and C, then MCMC seems more reasonable: "jump from B to A most of the time, but occasionally jump to C, just to check that A isn't a local peak."  Saying "make a child between A&B and A&C, and see which one is better" seems dumb, especially since you should know something about the gradients in most cases.


No comments:

Post a Comment