Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!sdd.hp.com!zaphod.mps.ohio-state.edu!sol.ctr.columbia.edu!ira.uka.de!ifistg!ifi!reczko
From: reczko@is6.ifistg.uucp (Martin Reczko)
Newsgroups: comp.ai.neural-nets
Subject: Re: simulated annealing vs genetic algs.
Message-ID: <RECZKO.90Nov7132348@is6.ifistg.uucp>
Date: 7 Nov 90 12:23:48 GMT
Sender: news@ifistg.uucp
Distribution: comp.ai.neural-nets
Organization: IfI, Univ. Stuttgart, W Germany
Lines: 29

In article <1990Oct31.172833.18197@ecn.purdue.edu> androula@cb.ecn.purdue.edu (Ioannis Androulakis) writes:
>   2) A lot of people are using simulated annealing techniques
>   to search continuous domains. From what I know, the
>   proof of asymptotic convergence of SA, under suitable
>   annealing schedules etc., was established under the assumption
>   that the state transitions where performed in a discrete
>   domain. If this is so, how legitimate is it to use a convergence
>   result, derived for discrete state transitions, to problems
>   where the state transitions are performed in a continous
>   space ? Or, in other words, how can we prove convergence of
>   continuous SA ?

I compared SA vs GA for small feedforward nets. The following reference
describes a modified SA-algorithm for continuous optimization (using
an interesting, but for NN-applications too space consuming (O(n^2),
n=problem-dimension=#weights) continuous state generation method).
I think the `continuous convergence proof' was contained.

@article(Vanderbilt:84,
author={Vanderbilt, D. and S.C. Louie},
title={A Monte Carlo Simulated Annealing Approach
to Optimization over Continuous Variables},
journal={Journal of Computational Physics},
volume={36},
year={1984},
pages={259-271})

-- Martin
-- reczko@is.informatik.uni-stuttgart.de