Newsgroups: comp.constraints
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!howland.reston.ans.net!ix.netcom.com!netcom.com!jabowery
From: jabowery@netcom.com (Jim Bowery)
Subject: Re: Feasibility problem with simple non-linear constraints
Message-ID: <jaboweryD4qvHr.GCv@netcom.com>
Organization: NETCOM On-line Communication Services (408 261-4700 guest)
X-Newsreader: TIN [version 1.2 PL1]
References: <3il0ma$s5h@disunms.epfl.ch> <3ipte2$3gs@gap.cco.caltech.edu> <jaboweryD4nF8v.1K8@netcom.com> <jaboweryD4otHM.CAs@netcom.com> <3iv7ss$h62@gap.cco.caltech.edu>
Date: Wed, 1 Mar 1995 04:37:02 GMT
Lines: 30
Sender: jabowery@netcom9.netcom.com

Things appear to be a bit better now, but still not perfect.

I think the problem was my misunderstanding of the way ASA was supposed 
to work.  If so, I had two sources of confusion:

1)  I thought that the multiple times the algorithm was running had 
something to do with reannealing, when in fact, it was just multiple runs 
(100 of them once I found the for loop in user.c) of the ASA algorithm, each 
time from scratch -- no memory between runs except apparently the 
random_seed.

2)  I thought that the default parameters given in asa_opt would 
guarantee convergence on a global optimum, and I when I used your 
example in the NOTES (Shubert Problem), substituting my own cost.c 
algorithm for the Shubert function, with all the asa_opt parameters 
unaltered, it didn't always converge on every one of the 100 runs, and 
the last of the 100 runs just happened to come up with a suboptimum.

I increased the Temperature_Ratio_Scale by a factor of 100 and it hit the 
global optimum much more often, but it still fails about 1 out of 4 times.
(By the way, this is a relatively simple test problem in my domain.)

I'm not that concerned about making it run fast -- I am more concerned 
about certainty of finding the global optimum.  How can one know what 
configuration values to use to avail one's self of ASA's analytic 
certainty of reaching a global optimum?
-- 
The promotion of politics exterminates apolitical genes in the population.
  The promotion of frontiers gives apolitical genes a route to survival.
                 Change the tools and you change the rules.
