Overview
Research addresses the computational expense associated with exactly solving large-scale instances of the Traveling Salesman Problem (TSP). Traditional graph sparsification methods, often employed to enhance computational efficiency in such problems, typically rely on fixed heuristics. These methods do not fully leverage instance-specific structural information. A novel approach, termed Graph Edge Sparsification (GES), was proposed as a learning-based sparsification technique specifically for Euclidean TSP.
GES integrates both geometric structural information and combinatorial optimization technology. This integration allows the method to adaptively generate a sparsification graph tailored to different problem instances. The primary objective of GES is to significantly reduce the size of the graph, thereby accelerating the solving process for TSP instances.
Research Context
The Traveling Salesman Problem (TSP) is recognized for the computational challenges involved in finding exact solutions for large-scale instances. This computational burden frequently necessitates the use of graph sparsification methods. Existing sparsification techniques primarily operate on fixed heuristics. A limitation identified with these traditional methods is their inability to fully exploit the unique structural characteristics inherent to individual problem instances.
Approach
The proposed method, Graph Edge Sparsification (GES), is a learning-based approach. It is designed for application to the Euclidean Traveling Salesman Problem (TSP). The methodology explicitly incorporates geometric structural information. Additionally, it integrates combinatorial optimization technology. This combined approach enables GES to adaptively generate a sparsification graph. The adaptive nature of the method means it can produce a distinct sparsification graph for varying problem instances. The explicit aim of this process is to reduce the overall graph size.
Findings
- Experimental results demonstrated that the Graph Edge Sparsification (GES) method achieved a pruning rate of up to 95% of edges on the MATILDA dataset.
- For experiments on the MATILDA dataset, the solution gap remained within 1% of the optimal value.
- The approach exhibited strong generalization capability when tested on the TSPLIB benchmark.
- On some large-scale instances within the TSPLIB benchmark, the pruning rate exceeded 99%.
- For these large-scale TSPLIB instances, the optimality gap remained below 1%.
- The reduction in graph size achieved through sparsification contributes to accelerating the solving process for TSP instances.
Why This Matters
The development of Graph Edge Sparsification (GES) offers a method to address the computational expense associated with solving large-scale Traveling Salesman Problem (TSP) instances. By adaptively reducing graph size through significant edge pruning while maintaining solution quality within a 1% gap from optimal, the approach provides a mechanism for accelerating problem-solving processes. This adaptability and efficiency contribute to managing the inherent computational challenges of large-scale combinatorial optimization problems.