A Meta-Learning Approach to Select Meta-Heuristics for the Traveling Salesman Problem Using MLP-Based Label Ranking

Different meta-heuristics (MHs) may find the best solutions for different traveling salesman problem (TSP) instances. The a priori selection of the best MH for a given instance is a difficult task. We address this task by using a meta-learning based approach, which ranks different MHs according to their expected performance. Our approach uses Multilayer Perceptrons (MLPs) for label ranking. It is tested on two different TSP scenarios, namely: re-visiting customers and visiting prospects. The experimental results show that: 1) MLPs can accurately predict MH rankings for TSP, 2) better TSP solutions can be obtained from a label ranking compared to multilabel classification approach, and 3) it is important to consider different TSP application scenarios when using meta-learning for MH selection.