1 présentation

• 
  14h00 - 15h00

  Automated Design of Metaheuristic Algorithms: Methods, Applications and Perspectives

  • Manuel López-Ibáñez, prés., University of Manchester
  • Thomas Stützle, Université Libre de Bruxelles

  Metaheuristic algorithms are often defined as high-level recipes that must be adapted and configured for particular problems. Unfortunately, designing a metaheuristic and configuring its parameter settings can be a time-consuming and difficult task due to the large number of possible algorithmic components and parameter values, the diversity of problem instances and computational environment, and the stochasticity of the algorithms. Traditionally, design and configuration are carried out by means of a manual, human-intensive process guided mainly by the expertise and intuition of a human expert. In recent years, several methods for automatic algorithm configuration have shown that it is possible to identify high-performing designs and parameter settings given large and complex parameter spaces. The role of automatic configuration methods goes beyond the classical concept of parameter tuning. When combined with carefully-designed algorithmic frameworks, these methods allow a more effective exploitation of the space of algorithm designs. This talk will review the most prominent automatic configuration methods and discuss several applications, ranging from minimizing the time required by exact solvers to find the optimum, to automatically designing multi-objective evolutionary algorithms according to multiple quality metrics. Recent promising results of automatic designed metaheuristics will be presented, giving a window into the potential future of metaheuristics research and development.