• Citer

Automated Algorithm Configuration for Hard Optimization Problems

par Yuan, Zhi
Président du jury Bersini, Hugues
Promoteur Stützle, Thomas
Co-Promoteur Birattari, Mauro
Publication Non publié, 2019-10-18

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

Algorithms for solving hard optimization problems usually have a number of parameters that greatly influence the algorithm performance. Instead of manual or ad hoc methods to adjust the parameters, using automatic tools for configuring such parameters have been proved to be crucial for deriving high-performing algorithms. Automatic algorithm configuration can be regarded as a black-box mixed-discrete-continuous variable optimization problem. Solving such problem usually requires two components for the two subtasks: a black-box search method that generates candidate configurations, and an evaluation method that evaluate the quality of the generated configuration under stochasticity.In this work, we define two frameworks, namely, iterated selection and post-selection, for combining search method and evaluation method for algorithm configuration. Through extensive literature review, all the established algorithm configurators can be identified to be under the iterated selection framework. We have extended an established evaluation method, the statistical racing method, with an ad-hoc iterated search method for configuring categorical and conditional parameters, as well as combined the racing method with established black-box search methods, such as MADS and CMA-ES, following the iterated selection framework. We have introduced state-of-the-art black-box optimizers such as CMA-ES, BOBYQA, and Nelder-Mead simplex to solving the problem of automatic algorithm configuration, and demonstrated how state-of-the-art configurator can be obtained by hybridizing these black-box search method with statistical racing method following the post-selection framework. The best settings for devising post-selection configurators are empirically analyzed in depth. We have identified and proposed a decreasing population CMA-ES with post-selection as the state-of-the-art algorithm configurator across extensive benchmark configuration problems with various dimensions and configuration budget.