The article mentions the problem of "local minimums". The author suggests occasionally accepting worse solutions, in order to escape from such. "The hope here is that by following this worse solution, we can eventually get to the global minimum." This is good but one may wish to take it further.
I once messed with something called "metropolis coupling", where you run multiple threads in parallel, with different thresholds for how often they accept worse solutions, i.e. one thread is "cold", one is slightly "hotter", and so on. If a hot thread's current solution is better than the solution of the neighbouring colder thread, the solutions are swapped. In this way the coldest thread (which is the one we're paying attention to) gets pulled out of the local minimum.
As I say, I messed with this once and it did seem to help. It seems to have been developed for inference in evolutionary phylogenetics, and is perhaps a bit obscure?
I once messed with something called "metropolis coupling", where you run multiple threads in parallel, with different thresholds for how often they accept worse solutions, i.e. one thread is "cold", one is slightly "hotter", and so on. If a hot thread's current solution is better than the solution of the neighbouring colder thread, the solutions are swapped. In this way the coldest thread (which is the one we're paying attention to) gets pulled out of the local minimum.
As I say, I messed with this once and it did seem to help. It seems to have been developed for inference in evolutionary phylogenetics, and is perhaps a bit obscure?