Improving simulated annealing performance by means of automatic parameter tuning

A common problem when using (meta)-heuristic techniques to solve complex combinatorial optimization problems is related to parameters tuning. Finding "the right" parameter values can lead to significant improvements in terms of best solution objective value found by the heuristic, heuristic reliability and heuristic convergence, among others. Unfortunately, this is usually a tedious and complicated task if done manually. In this paper, we propose a framework that is based on Genetic Programming to fine-tune a key parameter of the well-known Simulated Annealing (SA) algorithm. Experiments on a set of small instances of the Facility Location Problem with capacity constraints are performed. Results show that automatically adjusting a key parameter in SA by means of Genetic Programming leads to an average value of the obtained solution that is closer to the optimal solution than the average value obtained by the simple SA algorithm with a priori selected values. More important, standard deviation of the algorithm is greatly improved by our approach which makes it much more reliable if time limitations are imposed.