A General Framework Based on Machine Learning for Algorithm Selection in Constraint Satisfaction Problems

• Main Authors: José C. Ortiz-Bayliss, Ivan Amaya, Jorge M. Cruz-Duarte, Andres E. Gutierrez-Rodriguez, Santiago E. Conant-Pablos, Hugo Terashima-Marín
• Format: Article
• Language: English
• Published: MDPI AG 2021-03-01
• Series: Applied Sciences
• Subjects: algorithm selection; machine learning; constraint satisfaction; heuristic
• Online Access: https://www.mdpi.com/2076-3417/11/6/2749

Many of the works conducted on algorithm selection strategies—methods that choose a suitable solving method for a particular problem—start from scratch since only a few investigations on reusable components of such methods are found in the literature. Additionally, researchers might unintentionally omit some implementation details when documenting the algorithm selection strategy. This makes it difficult for others to reproduce the behavior obtained by such an approach. To address these problems, we propose to rely on existing techniques from the Machine Learning realm to speed-up the generation of algorithm selection strategies while improving the modularity and reproducibility of the research. The proposed solution model is implemented on a domain-independent Machine Learning module that executes the core mechanism of the algorithm selection task. The algorithm selection strategies produced in this work are implemented and tested rapidly compared against the time it would take to build a similar approach from scratch. We produce four novel algorithm selectors based on Machine Learning for constraint satisfaction problems to verify our approach. Our data suggest that these algorithms outperform the best performing algorithm on a set of test instances. For example, the algorithm selectors Multiclass Neural Network (MNN) and Multiclass Logistic Regression (MLR), powered by a neural network and linear regression, respectively, reduced the search cost (in terms of consistency checks) of the best performing heuristic (KAPPA), on average, by 49% for the instances considered for this work.