Algorithm Configuration via Continuously Racing: Preliminary Results
Yunshuang Xiao, Leslie Pérez Cáceres, Manuel López‐Ibáñez, Thomas Stützle · 2023
Automatic algorithm configuration procedures aim at supporting the design and application of optimization algorithms by providing specialized tools to automatically adjust their parameters to use the available computational resources effectively. The irace configurator is a state-of-the-art method implementing an iterated racing procedure that, in its current implementation, allows parallel executions of the target algorithm. Parallel evaluation is crucial for the efficient use of computational resources, reducing the wall-clock time that the configuration process needs to find a good configuration. In this work, we propose an alternative to irace that performs a single race instead of iterative races. The continuously racing configurator (crace) evaluates, removes and generates new configurations asynchronously, granting a high level of flexibility regarding the configuration process when compared to the previous iterative scheme. In this paper, we provide a general description of the crace configuration procedure and perform initial exploratory experiments on five configuration scenarios. These experiments focus on evaluating the effect of the minimum number of configurations required to be involved in the race and the number of parallel evaluations. The first results are encouraging, showing that crace can be a competitive procedure on the scenarios evaluated.