home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.ai.genetic
- Path: sparky!uunet!peora!tarpit!cs.ucf.edu!news
- From: long@next1.acme.ucf.edu (Richard Long)
- Subject: Re: GA versus Simulated Annealing
- Message-ID: <1993Jan21.172459.27732@cs.ucf.edu>
- Sender: news@cs.ucf.edu (News system)
- Organization: University of Central Florida
- References: <1993Jan19.155159.16736@thunder.mcrcim.mcgill.edu>
- Date: Thu, 21 Jan 1993 17:24:59 GMT
- Lines: 38
-
- In article <1993Jan19.155159.16736@thunder.mcrcim.mcgill.edu>
- gblais@McRCIM.McGill.EDU (Gerard Blais) writes:
- >
- >
- > I have been using a genetic algorithm to solve a minimization
- > problem. I performed the same minimization using the "Very Fast
- > Simulated Reannealing" algorithm created by Lester Ingber and
- > Bruce Rosen (I got it from a friend, but I believe it's available
- > from the net). In all cases I tried, the Annealing search was
- > orders of magnitude faster for converging to the global minimum
- > than the GA. My friend observed the same result for a different
- > minimization problem.
- >
- > My question is: Are there any cases that you know of where a
- > genetic algorithm would actually perform better than simulated
- > annealing? Are there any cases where simulated annealing would
- > fail but not genetic algorithms?
-
- One can view simulated annealing and reannealing as a GA with a mutation
- rate set to non-monotonic decreasing and the crossover and inversion rates
- set to zero. From the few cases that I've read about the successes of GA,
- crossover is particularly good for a problem that has a natural
- hierarchical structure to the optimal solution, since crossover exploits
- local fitness maxima in the construction of a global maximum. If there is
- no hierarchical structure to the optimal solution, local fitness
- contributes very little to global fitness, and there is little advantage
- to keeping local good solutions as crossover does. Inversion exploits
- symmetries in the problem.
- Some problems that may be naturally hierarchical and symmetric, such as
- image classification, may yield to GA methods more than simulated
- annealing or simulated reannealing.
- --
- Richard Long
- Institute for Simulation and Training
- University of Central Florida
- 12424 Research Parkway, Suite 300, Orlando, FL 32826
- (407)658-5026, FAX: (407)658-5059
- long@acme.ucf.edu
-