GLNS (C++)

GLNS is a solver for the Generalized Traveling Salesman Problem (GTSP). This repository contains a C++ implementation of GLNS, with a command-line solver, C and C++ library interfaces, and Python bindings. Both symmetric and asymmetric instances are supported.

The original Julia implementation is available at stephenlsmith/GLNS.jl. This port follows the same large neighborhood search algorithm and mode defaults. It can be used independently of Julia.

More information on the solver is given at https://ece.uwaterloo.ca/~sl2smith/GLNS/.

The current release is version 0.1.0. The installation instructions below build from source; see the changelog for release notes. The Julia implementation remains the reference implementation; this C++ port is maintained on a best-effort basis.

Citing this work

The GLNS solver and its settings are described in the following paper [DOI] [PDF]:

@Article{Smith2017GLNS,
    author =    {S. L. Smith and F. Imeson},
    title =     {{GLNS}: An Effective Large Neighborhood Search Heuristic
                 for the Generalized Traveling Salesman Problem},
    journal =   {Computers \& Operations Research},
    volume =    87,
    pages =     {1-19},
    year =      2017,
    doi =       {10.1016/j.cor.2017.05.010},
}

Please cite this paper when using GLNS.

Using the solver

The solver can be run from the command line, called from Python, or linked into a C or C++ application. GLNS is a heuristic: it searches for good feasible tours without generally certifying optimality.

Installation

To build the command-line solver and C/C++ library, install CMake 3.18 or newer and a C++17 compiler. Run the following from the repository root:

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release --parallel

The command-line executable is build/glns. With Visual Studio, it is usually build/Release/glns.exe; use that path in the examples below.

Run the C/C++ tests with:

cd build
ctest -C Release --output-on-failure
cd ..

For Python, use Python 3.9 or newer. Install directly from GitHub:

python -m pip install "git+https://github.com/stephenlsmith/glns-cpp.git@v0.1.0"

Or, from a clone of the repository, run from its root:

python -m pip install .

This builds the Python extension and installs its NumPy dependency. A C++17 compiler is required; the Python build backend can obtain CMake and Ninja as needed. Installing the Python package builds the library bindings; use the CMake commands above to build the command-line executable.

Input instances

The file input is a text file in GTSPLIB format, an extension of the TSPLIB format. Example inputs are provided in examples/.

Supported distance types are EUC_2D, CEIL_2D, MAN_2D, GEO, ATT, and EXPLICIT. Supported explicit formats are FULL_MATRIX, LOWER_ROW, LOWER_DIAG_ROW, UPPER_ROW, and UPPER_DIAG_ROW. The parser also accepts the Julia solver's simple matrix format.

Instances must contain at least two nonempty sets that partition the vertices: each vertex belongs to exactly one set. A tour selects one vertex from each set and returns to its starting vertex. Individual edge costs must be integers in the range 0 to 2,147,483,647. Tour costs use 64-bit integers and may exceed that range. Scale and round floating-point costs before passing a matrix to the library.

Indexing: vertices in the C, C++, and Python APIs are 0-indexed. GTSPLIB files and the CLI's printed tours and tour files use 1-indexed vertex IDs.

Solver usage and examples

GLNS has three modes: default, fast, and slow. These provide different tradeoffs between search effort and solution quality. Options can also set a time limit or stop the search when a cost threshold is met.

Example 1: Solve 39rat195.gtsp with the default settings and write the tour and its cost to tour.txt:

./build/glns examples/39rat195.gtsp -output=tour.txt

Example 2: Run with the slow setting:

./build/glns examples/39rat195.gtsp -mode=slow

Example 3: Allow repeated trials with a 60-second time limit, stopping earlier if a tour of cost 13,505 or less is found:

./build/glns examples/107si535.gtsp -max_time=60 -budget=13505 -trials=100000

Example 4: Supply a seed for reproducible runs:

./build/glns examples/39rat195.gtsp -seed=42

The command-line syntax follows the Julia GLNScmd.jl interface:

./build/glns filename.gtsp -option=value
./build/glns --help

Common options are:

CLI option	Meaning
-mode=default	Choose default, fast, or slow.
-max_time=60	Set the solver time limit in seconds.
-budget=13505	Stop when a tour with cost at most this value is found.
-trials=5	Number of cold trials, each starting from a new tour.
-restarts=3	Warm restart attempts within each cold trial.
-num_iterations=60	Iteration scale per set; controls cooling and stagnation thresholds, rather than imposing a hard iteration cap.
-seed=42	Use an unsigned 64-bit RNG seed.
-verbose=0	Set verbosity from 0 to 3; the CLI defaults to 3.
-output=tour.txt	Write the tour summary to a file.

The values in the table are examples. Unset search options use the chosen mode's defaults. --help also lists the reoptimization, noise, insertion, removal, and adaptation options. With -verbose=0, the CLI prints just the cost unless an output file is requested.

Time and budget limits are checked after building the initial feasible tour, at restart boundaries, and after each search iteration. An operation already in progress completes before the time limit is checked, so the limit is not a strict wall-clock deadline. max_time=0 returns an initial tour with zero search iterations. A zero-cost tour stops the search immediately, since nonnegative edge costs prove it is optimal.

A fixed seed reproduces a run with the same instance, options, and build when time does not determine termination. Runs cut short by a time limit can differ with machine speed and load; the seed does not synchronize the C++ and Julia random-number streams.

Calling from Python

Solve an instance from a file. The example instances are in the repository's examples/ directory, not in the installed package, so run this from a clone or adjust the path:

import glns

sol = glns.solve_file("examples/39rat195.gtsp", mode="default", seed=42)
print(sol.cost)
print(sol.tour)  # 0-indexed vertex IDs

Or supply a distance matrix and the sets directly:

import numpy as np
import glns

dist = np.array([
    [0, 3, 4, 7],
    [3, 0, 6, 5],
    [4, 6, 0, 4],
    [7, 5, 4, 0],
], dtype=np.int64)
sets = [[0, 1], [2], [3]]

sol = glns.solve(dist, sets, mode="fast", seed=42)
print(sol.cost, sol.tour)

dist[i, j] is the cost of traveling from vertex i to vertex j. Nested Python lists are also accepted. Costs and vertex IDs must be integers; floating-point arrays and fractional vertex IDs are rejected.

Solver options are keyword arguments with the same names as the CLI, except that -output becomes output_file. For example:

sol = glns.solve_file(
    "examples/107si535.gtsp",
    max_time=60, budget=13505, trials=100000, output_file="tour.txt",
)

The returned Solution has tour, cost, solve_time (seconds), timeout, budget_met, and total_iterations (completed search iterations). The Python API is silent by default (verbose=0) and releases the GIL while the C++ solver runs.

Calling from C++

The C++ API is declared in include/glns/glns.hpp:

#include <iostream>
#include <glns/glns.hpp>

int main() {
    glns::Instance instance = glns::read_instance("examples/39rat195.gtsp");
    glns::Params params;
    params.mode = "default";
    params.seed = 42;

    const glns::Solution sol = glns::solve(instance, params);
    std::cout << "Cost: " << sol.cost << '\n';
    for (int v : sol.tour) std::cout << v << ' ';
    std::cout << '\n';
}

To construct an instance in memory, replace the read_instance call with:

glns::Instance instance;
instance.num_vertices = 4;
instance.num_sets = 3;
instance.sets = {{0, 1}, {2}, {3}};
instance.dist = glns::Matrix(4, 0);
const glns::Cost costs[4][4] = {
    {0, 3, 4, 7}, {3, 0, 6, 5}, {4, 6, 0, 4}, {7, 5, 4, 0}
};
for (int i = 0; i < 4; ++i)
    for (int j = 0; j < 4; ++j)
        instance.dist.set(i, j, costs[i][j]);
instance.finalize();

Call finalize() after constructing or modifying an instance, before calling solve(). read_instance() does this automatically. Parsing and validation errors are reported as C++ exceptions. As with Python, the library is silent by default and returns the solution in memory.

To install the C/C++ library, headers, CLI, and CMake package:

cmake --install build --config Release --prefix /absolute/path/to/install

A downstream application's CMakeLists.txt can then contain:

cmake_minimum_required(VERSION 3.18)
project(my_solver LANGUAGES CXX)
find_package(glns 0.1 CONFIG REQUIRED)
add_executable(my_solver main.cpp)
target_link_libraries(my_solver PRIVATE glns::core)

Set -DCMAKE_PREFIX_PATH=/absolute/path/to/install when configuring that application. Projects using add_subdirectory can also link glns::core.

Calling from C

include/glns/glns.h provides a C interface, also suitable for use through a foreign-function interface:

#include <stdio.h>
#include <glns/glns.h>

int main(void) {
    glns_params params;
    glns_params_init(&params);
    params.seed = 42;
    params.has_seed = 1;

    glns_solution sol;
    char error[256];
    int rc = glns_solve_file("examples/39rat195.gtsp", &params,
                             &sol, error, sizeof(error));
    if (rc != 0) {
        fprintf(stderr, "%s\n", error);
        glns_solution_free(&sol);
        return 1;
    }
    printf("Cost: %lld\n", (long long)sol.cost);
    glns_solution_free(&sol);
    return 0;
}

Use glns_solve() for an in-memory row-major int64_t distance matrix, set sizes, and concatenated vertex IDs. See tests/test_c_api.c for an example. Initialize options with glns_params_init(), and free each returned solution with glns_solution_free() before reusing it. Functions return a nonzero status and an error message on failure.

The implementation requires the C++ runtime. A CMake consumer can use project(my_solver LANGUAGES C CXX) and link glns::core; see tests/install/. The header provides GLNS_VERSION_* macros and glns_version() for version checks. The C and C++ APIs and ABIs may change between minor versions before 1.0.

Relation to the Julia implementation

This C++ port was developed with assistance from AI coding tools, including Claude and OpenAI Codex, for implementation, testing, review, and documentation. The original Julia implementation was written by Stephen L. Smith without AI assistance.

The algorithm and mode defaults follow GLNS.jl. Individual stochastic runs differ. The benchmark comparison uses ten runs per implementation on each of 45 GTSPLIB instances. It found comparable solution quality and 21% less total solver time for C++, with lower mean runtimes on all 45 instances on the benchmark machine. The report includes per-instance results, validation, and reproduction tools.

The C++ version adds library interfaces, seed control, stricter input validation, and stopping checks at initialization and restart boundaries. It also handles zero-cost optima and the LOWER_ROW matrix format. Edge costs are stored as signed 32-bit integers, while tour arithmetic uses 64-bit integers. Library calls are silent by default; the CLI retains the Julia solver's default verbosity.

CMake configures floating-point contraction to preserve coordinate-distance rounding against the Julia parser. For a direct Clang/GCC build, use:

c++ -std=c++17 -O3 -ffp-contract=off -Iinclude -Isrc src/*.cpp cli/main.cpp -o glns

Tests

The CTest suite covers parsing, solver feasibility and cost consistency, stopping conditions, reproducibility, validation, the C API, and CLI options. To run the Python tests after installing the package:

python -m pip install pytest
python -m pytest -q python/tests

The CI workflow configures C/C++ builds on Linux, macOS, and Windows, an AddressSanitizer/UndefinedBehaviorSanitizer build, and Python tests on versions 3.9 and 3.14.

Index of files

• cli/main.cpp — command-line solver.
• include/glns/ — public C and C++ headers.
• src/ — solver, parser, tour optimizations, and supporting routines; the organization follows the Julia source files.
• python/glns/ — Python interface.
• bindings/python/ — bindings to the C++ solver.
• examples/ — sample GTSP instances.
• tests/ and python/tests/ — test suites.
• docs/ — comparison with the Julia implementation.
• CMakeLists.txt, cmake/, and pyproject.toml — build and packaging configuration.

License

Copyright 2018 Stephen L. Smith and Frank Imeson

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0.

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See LICENSE for the specific language governing permissions and limitations under the License.

Contact information

Prof. Stephen L. Smith
Department of Electrical and Computer Engineering
University of Waterloo
Waterloo, ON Canada
web: https://ece.uwaterloo.ca/~sl2smith/
email: stephen.smith@uwaterloo.ca