Abstract

In this paper, we present a new approach for Automatic Algorithm Selection. In this new procedure, we feed the predictor of the best algorithm choice with a runtime limit for the solvers. Hence, the machine learning model should consider and learn from the Anytime Behavior of the solvers, together with features characterizing each instance. For this purpose, we propose a general Framework and apply it to the Knapsack problem. Thus, we created a large and diverse dataset of 15, 000 instances, recorded the anytime behavior of 8 solvers on them and trained and tested three machine learning strategies, collecting the results for different machine learning algorithms. Our results show that, for the majority of the tuples , time>, the solver that computes the best objective value can be predicted. We also make this data publicly available, as a challenge for the community to work in this problem and propose new and better machine learning models and solvers. (C) 2020 Elsevier Ltd. All rights reserved.