GENETIC ALGORITHM PARAMETER TUNING USING EXPLORATORY LANDSCAPE ANALYSIS AND MACHINE LEARNING
Maxim Pikalov, Aleksei Pismerov · Известия Южного федерального университета. Технические науки · 2024
The choice of parameter values in evolutionary algorithms greatly affects their performance. Manypopular parameter tuning methods are constrained by the maximum number of fitness function evaluationsto find a good set of parameter values. Recently, an approach to algorithm selection for optimizationproblems has been proposed, which uses the analysis of the fitness function landscape and machine learningto select the optimal algorithm based on the characteristics of its landscape. Such application of fitnesslandscape analysis motivates further research, particularly in the context of parameter tuning in evolutionaryalgorithms. The use of landscape features allows for the identification of similar problems andthe use of parameter tuning data obtained from testing on benchmark problems, significantly reducing thenumber of required fitness function evaluations during tuning. This work considers an approach to automaticparameter selection using landscape analysis of the objective function and machine learning, using a genetic algorithm as an example. The proposed solution evaluates the characteristics of thelandscape of the optimization problem's objective function and suggests optimal parameter values for thealgorithm using a neural network. This network was trained on a dataset of landscape features expressedas numerical features and their corresponding optimal algorithm parameter sets. In contrast to approachesfor automatic algorithm selection for a specific problem, this work addresses the problem of regressingalgorithm parameters instead of classifying the most suitable algorithm from a given set. The results ofexperiments on different configurations of the W-model problem, as well as on the MAX-3SAT problem,show that the proposed approach to automatic parameter selection considering the landscape of the objectivefunction can help determine appropriate values for the static parameters of the geneticalgorithm. The algorithm with the proposed parameter values outperforms other consideredoptions on average, requiring fewer evaluations of the objective function to find the optimumcompared to the other algorithms considered.