3/11/2022: An overview of per-instance algorithm selection, Roberto Asín, University of Concepcion

Presentation description:

In this talk, we reviewed important concepts and problems related to Automatic Algorithm Selection. For instance, a conceptual analysis of: Per-set algorithm selection, algorithm configuration and parallel portfolio, per-instance algorithm selection, algorithm configuration and parallel portfolio, algorithm schedules and their relation with algorithm selection systems, was performed. Also, we reviewed the characteristics and evaluation metrics proposed for the two International Competitions on Algorithm Selection that were held so-far, as well as the main ideas behind the structure of the meta-solvers that participate on such tasks. We contrasted this fine-tuned systems with the more general deep-learning-based proposal of [3] which used Convolutional Neural Networks for performing generic algorithm selection from images created from any input file for a given problem, without the need of proposing handcrafted features.

References: 

[1] Kerschke, P., Hoos, H. H., Neumann, F., & Trautmann, H. (2019). Automated algorithm selection: Survey and perspectives. Evolutionary computation, 27(1), 3-45.

[2] Lindauer, M., van Rijn, J. N., & Kotthoff, L. (2019). The algorithm selection competitions 2015 and 2017. Artificial Intelligence, 272, 86-100.

[3] Loreggia, A., Malitsky, Y., Samulowitz, H., & Saraswat, V. (2016, February). Deep learning for algorithm portfolios. In Thirtieth AAAI Conference on Artificial Intelligence.