Newsgroups: comp.ai.neural-nets
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!pipex!uunet!zib-berlin.de!math.fu-berlin.de!dp
From: dp@pip.fpms.ac.be (Davor Pavisic)
Subject: simulated annealing
Message-ID: <JUMUB36O@math.fu-berlin.de>
Sender: news@math.fu-berlin.de (Math Department)
Nntp-Posting-Host: pip.fpms.ac.be
Organization: Free University of Berlin, Germany
X-Newsreader: TIN [version 1.2 PL2]
Date: Mon, 21 Nov 1994 10:30:13 GMT
Lines: 17

When using simulated annealing (sa) algorithms one simply draws at random
a solution x in the search space V(x_(n)).   If F(x) <= F(x_(n)), x becomes
the new current solution. Otherwise, one of the two following alternatives
is selected according to some probabilistic law: x becomes the current
solution with probability p(n) or x_(n) remains the current solution with
the complementary probability 1-p(n).   Typically, p(n) decreases with time
(n) and with the size of the deterioration of F(=F(x)-F(x_(n))).

The idea comes from thermodynamics and metallurgy: when a metal in fusion is
cooled slowly  enough it tends to solidify in a structure of minimal energy.
The same principle is at work in s.a.:   at the beginning, almost all moves
are accepted.   This allows to explore the solution space.   Then gradually,
temperature is decreased which means that one becomes more and more
selective in accepting new solutions.   At the end, only the moves that
improve F are accepted.

-[davor]-
