Re: High-dimensional Minimization without analytical derivatives

[Linas Vepstas <linas at austin dot ibm dot com>]

On Sat, Sep 04, 2004 at 09:48:01AM -0500, Anatoliy Belaygorod was heard to remark:
> JDL is right, the main characteristic that I cared about was whether or
> not Simulated Annealing is robust enough to jump from one 'valley' to
> the next in the search of the GLOBAL min, unlike gradient based methods


Simulated annealing does this by going 'uphill' every now and then, 
with a probability of (1-exp(-kT)).  By going uphill, it can hop out of
a local minima.  Its called "annealing" because you slowly lower the
temperature T during the run of the algo.  The idea is at the end of the
algo, you are less likely to hop out of the good minimum you've found,
and back into a bad one.  In real life, "annealing" means "cooling slowly"; 
it allows the atomic dislocations in metal/glass to work themsleves out, 
slowly, making the metal/glass less brittle.

--linas