GES-TSP: Learning-based method prunes 95% of edges for faster TSP
New AI approach cuts graph edges by 99% while keeping solutions within 1% of optimal.
Solving the Traveling Salesman Problem (TSP) exactly for large instances is computationally intensive, often requiring exponential time. Traditional graph sparsification methods apply fixed heuristics that ignore instance-specific structure. In a new paper, Tianfeng Chen and Xianyue Li introduce Graph Edge Sparsification (GES), a learning-based approach tailored for Euclidean TSP. GES incorporates geometric structural information and combinatorial optimization techniques to adaptively generate a sparsified graph for each instance. This significantly reduces the graph size, accelerating subsequent exact solvers without sacrificing solution quality.
Experimental results on the MATILDA dataset show GES can prune up to 95% of edges while keeping the solution gap within 1% of the optimal value. On the TSPLIB benchmark, especially for large-scale instances, the pruning rate exceeds 99% while the optimality gap remains below 1%. The method also demonstrates strong generalization across different problem sizes and geometries. This work opens the door to making exact TSP solvers practical for real-world logistics, routing, and manufacturing applications where optimality is critical.
- Prunes up to 95% of edges on MATILDA dataset with <1% optimality gap
- On large TSPLIB instances, pruning exceeds 99% while maintaining <1% gap
- Learning-based adaptation per instance outperforms fixed heuristic sparsification
Why It Matters
Makes exact TSP solvers viable for large-scale logistics, cutting computation time dramatically without sacrificing optimality.