Genetic Programming Based Feature Construction for Automated Algorithm Selection

Le résumé fourni par la source

Automated algorithm selection aims to help users to select the best algorithm for new optimization problems without any expertise. As a machine learning task, algorithm selection maps the problem features to the best algorithm. Exploratory landscape analysis is a feature extraction method for algorithm selection based on random samples, which has been widely applied in algorithm selection methods. However, their instability and redundancy greatly affect the accuracy of algorithm selection. To enhance these features, we apply a feature construction method based on multi-tree genetic programming for algorithm selection, which creates more discriminating features by combining or transforming multiple features and further counteracts the negative correlation noise from random samples. The selection process in genetic programming can reduce the redundancy of features by selecting more relevant features. Experimental results show that this method achieves the best performance compared with genetic programming based feature construction methods and algorithm selection methods on the four datasets.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.