Efficient and Robust Parameter Tuning for Heuristic Algorithms

The main advantage of heuristic or metaheuristic algorithms compared to exact optimization methods is their ability in handling large-scale instances within a reasonable time, albeit at the expense of losing a guarantee for achieving the optimal solution. Therefore, metaheuristic techniques are appropriate choices for solving NP-hard problems to near optimality. Since the parameters of heuristic and metaheuristic algorithms have a great influence on their effectiveness and efficiency, parameter tuning and calibration has gained importance. In this paper a new approach for robust parameter tuning of heuristics and metaheuristics is proposed, which is based on a combination of Design of Experiments (DOE), Signal to Noise (S/N) ratio, Shannon entropy, and VIKOR methods, which not only considers the solution quality or the number of fitness function evaluations, but also aims to minimize the running time. In order to evaluate the performance of the suggested approach, a computational analysis has been performed on the Simulated Annealing (SA) and Genetic Algorithms (GA) methods, which have been successfully applied in solving respectively the n-queens and the Uncapacitated Single Allocation Hub Location combinatorial problems. Extensive experimental results showed that by using the presented approach the average number of iterations and the average running time of the SA were respectively improved 12 and 10.2 times compared to the un-tuned SA. Also, the quality of certain solutions was improved in the tuned GA, while the average running time was 2.5 times faster compared to the un-tuned GA.