TSPLIB

TSPLIB
Name	TSPLIB
Type	Data library
Subject	Combinatorial optimization
Creator	Gerhard Reinelt
Country	Germany
Established	1991
Discipline	Operations research
Format	Text files

TSPLIB TSPLIB is a curated library of sample instances for the traveling salesman problem and related combinatorial optimization problems, widely used by researchers in operations research, computer science, and applied mathematics. It supports benchmarking of exact algorithms and heuristics by providing standardized instances that are referenced in publications by authors affiliated with institutions such as MIT, Stanford University, ETH Zurich, University of Waterloo, and University of California, Berkeley. The library has influenced experimental studies in venues including SIAM Journal on Computing, Operations Research, Mathematical Programming, INFORMS Journal on Computing, and Journal of the ACM.

Overview

TSPLIB provides labeled instances for problem types including traveling salesman, vehicle routing, and related network optimization problems, enabling comparison across methodologies like branch-and-bound, dynamic programming, and polyhedral methods. The repository's instances are routinely employed in computational experiments by researchers at IBM Research, Microsoft Research, Google Research, Bell Labs, and Bellcore, and cited in textbooks from authors at Princeton University, Harvard University, Columbia University, Cornell University, and Caltech. Benchmark studies using TSPLIB often appear alongside algorithmic frameworks such as Christofides' algorithm, Lin–Kernighan heuristics, and Held–Karp bounds developed and discussed in works from RAND Corporation, Carnegie Mellon University, University of Pennsylvania, University of Oxford, and Cambridge University.

History and Development

The collection was initiated in the early 1990s by Gerhard Reinelt while affiliated with institutions in Heidelberg and in collaboration with researchers from Karlsruhe Institute of Technology, University of Bonn, University of Erlangen–Nuremberg, and groups connected to EURO, the Association of European Operational Research Societies. Early distributions were exchanged among conference attendees at meetings of IFORS, INFORMS, ESA and demonstrated at workshops in Munich, Zurich, Paris, and Boston. Over time the dataset expanded through contributions inspired by problem instances from VLSI design groups at Bell Labs, logistics problems studied by firms like DHL and FedEx, and geographic instances derived from maps produced by United States Geological Survey, Ordnance Survey, and National Geographic Society.

File Formats and Problem Types

TSPLIB uses plain text formats with headers specifying NAME, TYPE, COMMENT, and DIMENSION fields, similar to instance conventions used in repositories from UCI Machine Learning Repository and Netlib. Problem TYPES in the library include TSP, ATSP, SOP, HCP, CVRP, and VRP variants referenced in algorithmic literature from IBM T.J. Watson Research Center, Los Alamos National Laboratory, and Sandia National Laboratories. Coordinate systems in instances draw on geographic conventions from EPSG, with edge weight computations employing Euclidean, Geographical, and P-metric models discussed in papers from SIAM, IEEE, and ACM SIGGRAPH authors. Commonly used sections like NODE_COORD_SECTION and EDGE_WEIGHT_SECTION echo formats in datasets curated by Stanford Large Network Dataset Collection and Kaggle competitions.

Representative Instances and Benchmarks

Canonical instances in the collection include city-based datasets derived from coordinates of locations in Berlin, Prague, London, New York City, and San Francisco, and large-scale instances inspired by routing problems in Germany, Switzerland, France, United Kingdom, and United States. Famous benchmark instances such as those modeled after Att48, pcb442, and lin318 have been used in comparative studies by research groups at ETH Zurich, University of Leuven, University of Melbourne, and National University of Singapore. Results on these instances are commonly reported alongside algorithmic developments like Concorde from researchers at Rice University, University of Texas at Austin, and collaborators from DIMACS challenges and competitions hosted by SIAM and INFORMS.

Usage and Applications

Researchers employ TSPLIB for performance evaluation of heuristics (e.g., Lin–Kernighan) and exact solvers (e.g., branch-and-cut) in studies originating at Columbia University, Brown University, Duke University, University of Michigan, and Johns Hopkins University. Practitioners in logistics and transportation planning at companies such as UPS, Amazon, Maersk, Siemens, and Volkswagen use similar instance sets for testing routing and scheduling systems, often integrating results with solvers from Gurobi, CPLEX, and open-source projects like COIN-OR. Educational use appears in coursework at Massachusetts Institute of Technology, Imperial College London, University College London, Nanyang Technological University, and Tsinghua University.

Limitations and Criticisms

Critics note that TSPLIB's static, relatively small set of instances can bias algorithmic evaluation compared with modern distributions arising in contexts like last-mile delivery studied at Amazon Logistics or large-scale urban networks analyzed by researchers at MIT Senseable City Lab and Singapore-MIT Alliance. Concerns have been raised by authors publishing in Journal of Heuristics, Computers & Operations Research, and Transportation Science about representativeness, reproducibility, and the need for larger, more diverse benchmarks inspired by datasets from OpenStreetMap, Google Maps, and national statistical agencies such as U.S. Census Bureau and Eurostat.

Category:Combinatorial optimization