Cooperative Behavior in Coupled Simulated Annealing Processes with Variance Control

In this paper we describe the use of coupling to interconnect different Simulated Annealing (SA) processes. The objective is to allow cooperative behavior among the processes in order to improve performance. Coupled Simulated Annealing (CSA) permits a high degree of parallelization while delivering much better results than a typical Parallel SA (PSA) algorithm. This is possible due to the introduction of coupling in the acceptance probability functions. Moreover, the coupling also allows controling the variance of the acceptance probabilities. This is especially important because it reduces the sensitivity to initial parameters, while guiding optimization to quasioptimal runs. It was observed that the solutions generated by CSA are more concentrated around the global optimum, while PSA has concentrations of solutions often in unfavorable regions of the cost function. Also, the number of iterations per process necessary to reach a given minimum energy tolerance decreases exponentially when the number of optimizers is increased.