Search

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. [ABSTRACT FROM AUTHOR]

• Features of initial solutions predict the performance of perturbation algorithms. • Low performance variation within portfolios makes algorithm selection difficult. • Hybrid selection models are useful for portfolios with low performance variation. • Accuracy is not a fair measure to evaluate the performance of algorithm selectors. In this paper, we present the construction process of a per-instance algorithm selection model to improve the initial solutions of Curriculum-Based Course Timetabling (CB-CTT) instances. Following the meta-learning framework, we apply a hybrid approach that integrates the predictions of a classifier and linear regression models to estimate and compare the performance of four meta-heuristics across different problem sub-spaces described by seven types of features. Rather than reporting the average accuracy, we evaluate the model using the closed SBS-VBS gap, a performance measure used at international algorithm selection competitions. The experimental results show that our model obtains a performance of 0.386, within the range obtained by per-instance algorithm selection models in other combinatorial problems. As a result of the process, we conclude that the performance variation between the meta-heuristics has a significant role in the effectiveness of the model. Therefore, we introduce statistical analyses to evaluate this factor within per-instance algorithm portfolios. [ABSTRACT FROM AUTHOR]