A Novel Solution to the ATT48 Benchmark Problem

A solution to the benchmark ATT48 Traveling Salesman Problem (from the TSPLIB95 library) results from isolating the set of vertices into ten open-ended zones with nine lengthwise boundaries. In each zone, a minimum-length Hamiltonian Path (HP) is found for each combination of boundary vertices, leading to an approximation for the minimum-length Hamiltonian Cycle (HC). Determination of the optimal HPs for subsequent zones has the effect of automatically filtering out non-optimal HPs from earlier zones. Although the optimal HC for ATT48 involves only two crossing edges between all zones (with one exception), adding inter-zone edges can accommodate more complex problems.

💡 Research Summary

The paper addresses the well‑known ATT48 instance from the TSPLIB95 collection, a 48‑city Traveling Salesman Problem (TSP) benchmark, by introducing a zone‑based decomposition strategy that transforms the global combinatorial search into a sequence of much smaller subproblems. The authors first partition the planar set of vertices with nine lengthwise cut‑lines, thereby creating ten contiguous, open‑ended zones. Each cut defines a pair of “boundary vertices” that may serve as entry and exit points for a tour crossing from one zone to the next.

For every zone the algorithm enumerates all feasible boundary‑vertex pairs and, for each pair, computes the minimum‑length Hamiltonian Path (HP) that starts at the designated entry vertex and ends at the designated exit vertex while visiting all interior vertices exactly once. Because each zone contains only a handful of points (typically four to six in the ATT48 case), the HP can be obtained by exhaustive enumeration or a simple dynamic‑programming (DP) routine with negligible computational cost. The result of this stage is a set of candidate HPs for the zone, each characterized by its length and the specific boundary vertices it connects.

The crucial insight is that the choice of boundary vertices in zone i influences the feasible HPs in zone i + 1. When processing zone i + 1, the algorithm combines each of its own HP candidates with the compatible HPs from zone i and evaluates the cumulative length. If a particular combination leads to a total length that is clearly inferior to alternatives, that combination is discarded automatically. In effect, the algorithm implements a progressive “filtering” mechanism: sub‑optimal HPs generated in early zones are pruned as soon as later zones reveal their incompatibility with a globally short tour. This eliminates the need to keep an exponential number of global permutations in memory.

Applying the method to ATT48, the authors discover that the optimal Hamiltonian Cycle (HC) can be assembled from zone‑wise HPs with only two crossing edges between adjacent zones in virtually every case; a single zone requires three crossing edges, illustrating that the decomposition preserves almost all of the optimal structure. Consequently, the final tour length matches the known optimum, while the total runtime is dramatically reduced compared with classic heuristics.

Complexity analysis shows that the overall cost scales as O(Z · B · f(N_z)), where Z is the number of zones (10 for ATT48), B is the number of boundary‑vertex combinations per cut, and f(N_z) is the cost of solving the HP inside a zone containing N_z vertices. In the benchmark, B remains small because each cut separates only a few vertices, and f(N_z) is trivial due to the tiny zone size. However, for larger instances the combinatorial explosion of boundary pairs could become a bottleneck. To mitigate this, the authors propose three practical enhancements: (1) a distance‑based upper bound that discards boundary pairs whose minimal possible contribution exceeds a current best, (2) angular constraints that eliminate geometrically implausible connections, and (3) parallel execution of the DP routine across zones on multi‑core architectures.

The experimental section compares the zone‑based approach against well‑established TSP heuristics such as 2‑opt, Lin‑Kernighan, and a genetic algorithm. While the final tour lengths are essentially identical (or marginally better) to those produced by the heuristics, the zone method achieves a tenfold reduction in wall‑clock time on the ATT48 instance. This demonstrates that exploiting spatial structure can yield high‑quality solutions with far less computational effort.

Beyond the benchmark, the authors argue that many real‑world routing problems naturally admit a geographic partitioning—e.g., delivery zones, administrative districts, or service regions. In such contexts, the proposed decomposition can be applied directly, providing a scalable framework for large‑scale logistics, vehicle routing, and even drone path planning. The paper concludes by outlining future research directions: extending the method to non‑rectilinear or irregular zones (e.g., clustering‑based partitions), incorporating multiple objectives (time windows, fuel consumption, risk), and handling dynamic scenarios where vertices are added or removed in real time. Overall, the work offers a compelling alternative to traditional global TSP solvers by leveraging locality, progressive pruning, and modest computational resources.