nathanielmeyer
commented
10 years ago Neighborhood size is a good variable to look at with simulated annealing. When the neighborhood size is small, if the high entropy phase has taken the instance far from global optimum, then the search may not make it back before exhausting the neighborhood.
Stopping criteria is another important consideration. If the module repeatedly calling train doesn't know about the temperature, it doesn't know whether the temperature has cooled enough to start looking for improvements. I need to carefully choose temperature and cooling factor then set a delay long enough for the process to cool before I start checking the stopping criteria.
the results you obtained running the algorithms on the networks: why did you get the results you did? what sort of changes might you make to each of those algorithms to improve performance? Feel free to include any supporting graphs or tables. And by "feel free to", of course, I mean "do".