The paper proposes a new method for algorithm-selection using short probing trajectories. This approach is applied by running a solver on an instance for a brief period. The method provides results that are comparable or superior to feature-based methods, which are more computationally demanding. The research suggests that algorithms may perceive instance similarity differently than humans, indicating a need for an algorithm-centric perspective in algorithm selection. This new method aims to contribute to the development of optimization systems that can learn from past experiences.
Publication date: 23 Jan 2024
Project Page: https://arxiv.org/abs/2401.12745v1
Paper: https://arxiv.org/pdf/2401.12745
