Mid Sweden University

Parameter tuning is an optimization problem with the objective of finding good static pa-rameter settings before the execution of a metaheuristic on a problem at hand. The requirementof tuning multiple control parameters, combined with the stochastic nature of the algorithms,make parameter tuning a non-trivial problem. To make things worse, one parameter vector allowing the algorithm to solve all optimization problems to the best of its potential is verifiable non-existent, as can be inferred from the no free lunch theorem of optimization. Manual tuning can be conducted, with the drawback of being very time consuming and failure prone. Hence, means for automated parameter tuning are required. This paper serves as an overview about recent work within the field of automated parameter tuning.

In this paper, a framework for the simplification andstandardization of metaheuristic related parameter tuning by applyinga four phase methodology, utilizing Design of Experiments andArtificial Neural Networks, is presented. Metaheuristics are multipurposeproblem solvers that are utilized on computational optimizationproblems for which no efficient problem-specific algorithmexists. Their successful application to concrete problems requires thefinding of a good initial parameter setting, which is a tedious andtime-consuming task. Recent research reveals the lack of approachwhen it comes to this so called parameter tuning process. In themajority of publications, researchers do have a weak motivation fortheir respective choices, if any. Because initial parameter settingshave a significant impact on the solutions quality, this course ofaction could lead to suboptimal experimental results, and therebya fraudulent basis for the drawing of conclusions.

That there is no best initial parameter setting for a metaheuristicon all optimization problems is a proven fact (nofree lunch theorem). This paper studies the applicability ofso called robust parameter settings for combinatorial optimizationproblems. Design of Experiments supported parameterscreening had been carried out, analyzing a discreteParticle Swarm Optimization algorithm on three demographicallyvery dissimilar instances of the Traveling SalesmenProblem. First experimental results indicate that parametersettings produce varying performance quality forthe three instances. The robust parameter setting is outperformedin two out of three cases. The results are evensignicantly worse when considering quality/time trade-o.A methodology for problem generalization is referred to asa possible solution.

Adjusting the control parameters of population-based algorithms is a means for improving the quality of these algorithms' result when solving optimization problems. The difficulty lies in determining when to assign individual values to specific parameters during the run. This paper investigates the possible implications of a generic and computationally cheap approach towards parameter analysis for population-based algorithms. The effect of parameter settings was analyzed in the application of a genetic algorithm to a set of traveling salesman problem instances. The findings suggest that statistics about local changes of a search from iteration i to iteration i + 1 can provide valuable insight into the sensitivity of the algorithm to parameter values. A simple method for choosing static parameter settings has been shown to recommend settings competitive to those extracted from a state-of-the-art parameter tuner, paramlLS, with major time and setup advantages.