Online Per-Instance Algorithm Selection for Constrained Multi-Objective Optimization Problems
Abstract: Landscape-aware algorithm selection relies on statistical landscape features that are usually extracted offline, independently of the optimization process. However, the computational costs when using this offline method can be excessive and typically the approach ignores information accumulated by the optimization algorithm. In this paper, we investigate online per-instance algorithm selection (PIAS) in black-box constrained multi-objective optimization problems (CMOPs). Two complementary aims guide our work: (1) to reduce the cost of the sampling strategy by using solutions that have already been evaluated by the optimization algorithm as sample points, and (2) investigate whether there are actual performance gains when switching to the selected algorithm. The experimental results show that the feature extraction step can be effectively incorporated into the optimization process, reducing the computational costs of the PIAS model. Moreover, the performance of the Online-PIAS model outperforms the best single algorithm on average and is comparable to the Offline-PIAS model. Additionally, the choice of the starting algorithm significantly impacts the overall performance.
Loading