About Simulated Annealing Solver in galapagos

Hi David and everyone,

I am trying to get a better understanding of the annealing solver machanism in Galapagos.

As my understanding, simulated annealing usually starts from a high temperature, say, 10000 degree, and the probability of worse result acceptance falls along with the temperature drops. While in the solver, it actually start from an already low temperature, which is about 1 degree. I suppose the reason for this is for an neighborhood selection strategy(a truncated normal distribution). But in this case, how did the acceptance working in a same manner if it still follows the origin, which is a exponential of the absolute value of difference divided by temperature(the probability is already really close to 0 from the begining)?

Do you apply a customized annealing mechanism other than the simplest one?

Please correct me if any of above is off the point.

thanks!

Replies to This Discussion

Permalink Reply by Anders Holden Deleuran on April 1, 2016 at 8:21am

David's paper "NAVIGATING MULTI-DIMENSIONAL LANDSCAPES IN FOGGY WEATHER AS AN ANA... provides an excellent overview of the different methods. Probably a good place to start :)

Permalink Reply by David Rutten on April 1, 2016 at 8:46am

The solver always starts at maximum temperature (where maximum is hot enough to jump most of the way across phase space), except the first iteration, which may start at a lower temperature. There is a setting for this, but it is only used during the first iteration.

The temperature drops at (I think) a logarithmic rate and the jump likelihood is calculated according to standard SA algorithms.

Permalink Reply by Christoph1 on June 1, 2016 at 6:44am

Hi David,

can I ask you about some details of your simulated annealing implementation?

- when using delta Energy for calculating the acceptance probability, do you normalize it? If so, how can you have a generalized knowledge of the normalization factor, without knowing the problem (especially the objective function value bounds)?

- also in an earlier post you said, you have a Cauchy distribution for this acceptance probability. Does it mean, you impose a distribution on the probability? I thought, p depends on the current temperature and delta Energy?

Thanks a lot!

Christoph

RSS