09.10.16 - 14.10.16, Seminar 16412

Automated Algorithm Selection and Configuration

Diese Seminarbeschreibung wurde vor dem Seminar auf unseren Webseiten veröffentlicht und bei der Einladung zum Seminar verwendet.

Motivation

The state of the art in solving an increasing number of important problems - in particular: NP-hard optimisation and decision problems - has been substantially advanced by automated algorithm selection and configuration techniques. Based on advanced methods from machine learning and optimisation, these techniques have been able to capitalise on the performance potential inherent in sets of algorithms with complementary strengths and in highly parameterised algorithms on prominent and important problems, including propositional satisfiability (SAT), mixed integer programming (MIP), supervised machine learning, AI planning and continuous black-box optimisation. State-of-the-art algorithm selection and configuration procedures make use of statistical models that enable performance predictions for previously unseen problem instances or parameter settings, and these empirical performance models also have other important uses, e.g., in load scheduling and distribution on large computer clusters.

This 5-day seminar will bring together a carefully selected group of 30 experts from artificial intelligence, machine learning, statistics, empirical algorithmics, theoretical computer science and industrial applications with the goals of

  1. gaining a deeper, also theoretical, understanding of existing algorithm selection, algorithm configuration and performance prediction techniques;
  2. further improving their efficacy, leveraging empirical approaches and theoretical insights, and
  3. focussing explicitly on practical applications and industry-driven challenges.

The seminar will cover a wide range of topics related to algorithm selection, configuration and performance prediction, with a particular focus on multi-objective optimisation and continuous black-box optimisation problems. A further focus will be on real-world applications of these techniques and their impact on the way algorithms for solving challenging computational problems are developed, deployed and used.

To take the best advantage of the unique opportunities offered by a Dagstuhl Seminar, we have decided on a highly interactive format that complements working groups on specific topics with brainstorming sessions and a limited number of invited presentations, providing ample opportunity for lively, informal discussion.