Statistical methods for parameter fine-tuning of metaheuristics
Document typeMaster thesis
Rights accessOpen Access
Metaheuristics are an approximate method widely used to solve many hard optimization problems in a multitude of fields. They depend on a variable number of parameters. Despite the fact that they are usually capable of finding good solutions within a reasonable time, the difficulty in selecting appropriate values for their parameters causes a loss of efficiency, as it normally requires much time, skills and experience. This master degree s thesis provides a survey of the main approaches developed in the last decade to tackle the problem of choosing a good set of parameter values, called the Parameter Setting Problem, and compares them from a methodological point of view focusing on the statistical procedures used so far by the scientific community. This analysis is accompanied by a proposal of a general methodology. The results of applying it to fine-tuning the parameters of a hybrid algorithm, which combines Biased Randomization with the Iterated Local Search metaheuristic, for solving the Multi-depot Vehicle Routing Problem are also reported. The computational experiment shows promising results and the need / suitability of further investigations based on a wider range of statistical learning techniques. Along these same lines, different suggestions for future work are described. In addition, this work highlights the importance of statistics in operations research giving a real-world example.