ml4co-kit 0.3.3

ml4co-kit provides convenient dataset generators for the combinatorial optimization problem

Combinatorial Optimization (CO) is a mathematical optimization area that involves finding the best solution from a large set of discrete possibilities, often under constraints. Widely applied in routing, logistics, hardware design, and biology, CO addresses NP-hard problems critical to computer science and industrial engineering.

ML4CO-Kit aims to provide foundational support for machine learning practices on CO problems, including the follow aspects.

• algorithm: common post-processing algorithms.
• data: common test datasets and our generated traning dataset.
• draw: visualization of problems and solutions.
• evaluate: evaluator for problems and solvers.
• generator: data generation of various distributions.
• learning: implemented base classes that facilitate method development for ML4CO.
• solver: solvers' base classes and mainstream traditional solvers.
• utils: general or commonly used functions and classes.

⭐ Official Documentation: https://ml4co-kit.readthedocs.io/en/latest/

⭐ Source Code: https://github.com/Thinklab-SJTU/ML4CO-Kit

Development status

Basic

Problem	Generator	Basic Solver (IO)	Traditional Solver
ATSP	sat, hcp, uniform	tsplib, txt	LKH
CVRP	uniform, gaussian	vrplib, txt	LKH, HGS, PyVRP
LP	uniform	txt	Gurobi
MCl	er, ba, hk, ws, rb	gpickle, txt, networkx	Gurobi
MCut	er, ba, hk, ws, rb	gpickle, txt, networkx	Gurobi
MIS	er, ba, hk, ws, rb	gpickle, txt, networkx	Gurobi, KaMIS
MVC	er, ba, hk, ws, rb	gpickle, txt, networkx	Gurobi
TSP	uniform, gaussian, cluster	tsplib, txt	LKH, Concorde, GA-EAX

Extension

Problem	Visualization	Algorithm	Test Dataset	Train Dataset
ATSP	📆	2	📆	📆
CVRP	✔	1	vrplib, uniform	📆
LP	📆	📆	📆	📆
MCl	✔	6	📆	📆
MCut	✔	3	📆	📆
MIS	✔	6	📆	📆
MVC	✔	5	📆	📆
TSP	✔	5	satlib, uniform	uniform

1~9: Number of supports; ✔: Supported; 📆: Planned for future versions (contributions welcomed!).

ML4CO Organization:

We are still enriching the library and we welcome any contributions/ideas/suggestions from the community. A comprehensive modular framework built upon this library that integrates core ML4CO technologies is coming soon.

Installation

You can install the stable release on PyPI:

$ pip install ml4co-kit

or get the latest version by running:

$ pip install -U https://github.com/Thinklab-SJTU/ML4CO-Kit/archive/master.zip # with --user for user install (no root)

The following packages are required and shall be automatically installed by pip:

Python>=3.8
numpy>=1.24.4
networkx>=2.8.8
tqdm>=4.66.3
pulp>=2.8.0, 
pandas>=2.0.0,
scipy>=1.10.1
aiohttp>=3.10.11
requests>=2.32.0
async_timeout>=4.0.3
pyvrp>=0.6.3
cython>=3.0.8
gurobipy>=11.0.3
scikit-learn>=1.3.0

To ensure you have access to all functions, such as visualization, you'll need to install the following packages using pip:

matplotlib
pytorch_lightning

Usage Examples

Solve with Traditional Solver Baselines

We provide base classes with a user-friendly approach for implementing traditional and learning-based solvers. Taking TSPSolver as an example, it includes functionalities for data input and output, as well as an evaluation function. The solver supports different data inputs, such as Numpy arrays and .txt and .tsp files. The outputs can be saved to corresponding types of files as needed. Additionally, the solver offers an evaluation function, by which users can quickly obtain the average tour length, average gap, and standard deviation of the test dataset. Traditional solvers are directly incorporated in our library inheriting TSPSolver.

>>> from ml4co_kit.solver import TSPLKHSolver

# initialization
>>> tsp_lkh_solver = TSPLKHSolver(lkh_max_trials=500)

# input instances and reference solutions by a .txt file
>>> tsp_lkh_solver.from_txt("path/to/read/tsp500_concorde.txt", ref=True)

# lkh solving
>>> tsp_lkh_solver.solve(num_threads=10, show_time=True)

# evaluate
>>> tsp_lkh_solver.evaluate(calculate_gap=True)
(16.546304871926175, 16.545805334644392, 0.0030213676759083515, 0.009747905769875538)

# save solving results
>>> tsp_lkh_solver.to_txt("path/to/write/tsp500_lkh.txt")

Data Generation

from ml4co_kit import TSPDataGenerator

# initialization
tsp_data_lkh = TSPDataGenerator(
    num_threads=8,
    nodes_num=50,
    data_type="uniform",
    solver="LKH",
    train_samples_num=16,
    val_samples_num=16,
    test_samples_num=16,
    save_path="path/to/save/"
)

# generate
tsp_data_lkh.generate()

Evaluate

>>> from ml4co_kit.evaluate import TSPLIBOriEvaluator
>>> from ml4co_kit.solver import TSPLKHSolver

>>> lkh_solver = TSPLKHSolver(scale=1)
>>> evaluator = TSPLIBOriEvaluator()
>>> evaluator.evaluate(lkh_solver, norm="EUC_2D")
           solved_costs      ref_costs          gaps
eil51        429.983312     429.983312  0.000000e+00
berlin52    7544.365902    7544.365902  3.616585e-14
st70         677.881928     678.597452 -1.054416e-01
eil76        544.837795     545.387552 -1.008012e-01
pr76      108159.438274  108159.438274 -1.345413e-14
kroA100    21285.443182   21285.443182  0.000000e+00
kroC100    20750.762504   20750.762504  0.000000e+00
kroD100    21294.290821   21294.290821  3.416858e-14
rd100       7910.396210    7910.396210  0.000000e+00
eil101       642.244814     642.309536 -1.007642e-02
lin105     14382.995933   14382.995933  0.000000e+00
ch130       6110.739012    6110.860950 -1.995428e-03
ch150       6532.280933    6532.280933 -2.784616e-14
tsp225      3859.000000    3859.000000  0.000000e+00
a280        2587.930486    2586.769648  4.487600e-02
pr1002    259066.663053  259066.663053  0.000000e+00
pr2392    378062.826191  378062.826191  0.000000e+00
AVG        50578.945903   50578.963027 -1.020227e-02

>>> evaluator.evaluate(lkh_solver, norm="GEO")
           solved_costs    ref_costs          gaps
ulysses16     74.108736    74.108736  1.917568e-14
ulysses22     75.665149    75.665149  3.756248e-14
gr96         512.309380   512.309380  0.000000e+00
gr202        549.998070   549.998070 -8.268163e-14
gr666       3843.137961  3952.535702 -2.767786e+00
AVG         1011.043859  1032.923407 -5.535573e-01

Algorithm

>>> from ml4co-kit import TSPSolver, tsp_insertion_decoder

# create solver and load data
>>> solver = TSPSolver()
>>> solver.from_txt("your/txt/data/path", ref=True)

# use insertion algorithm to solve the problems
>>> tours = tsp_insertion_decoder(points=solver.points)
>>> solver.from_data(tours=tours, ref=False)

# evaluate (average length, ground truth, gap, std)
>>> solver.evaluate(calculate_gap=True)
(6.299320133465173, 5.790133693543183, 8.816004478345556, 3.605743337834312)

Visualization

Below, we use TSP, MIS, and CVRP as representative examples for illustration.

TSP

>>> from ml4co_kit.solver import TSPConcordeSolver
>>> from ml4co_kit.draw.tsp import draw_tsp_solution, draw_tsp_problem

# use TSPConcordeSolver to solve the problem
>>> solver = TSPConcordeSolver(scale=1)
>>> solver.from_tsplib(tsp_file_path="examples/tsp/tsplib/kroA150.tsp")
>>> solver.solve(norm="EUC_2D")

# draw images of problem and solution
>>> draw_tsp_problem(
        save_path="docs/assets/tsp_problem.png",
        points=solver.ori_points,
    )
>>> draw_tsp_solution(
        save_path="docs/assets/tsp_solution.png",
        points=solver.ori_points,
        tours=solver.tours
    )

Visualization Results:

MIS

>>> from ml4co_kit.solver import KaMISSolver
>>> from ml4co_kit import draw_mis_problem, draw_mis_solution

# use KaMISSolver to solve the problem
>>> mis_solver = KaMISSolver(time_limit=10)
>>> mis_solver.solve(
        src="examples/mis/instance",
        out="examples/mis/solution"
    )

# draw images of problem and solution
>>> draw_mis_problem(
        save_path="docs/assets/mis_problem.png", 
        graph_data=mis_solver.graph_data[0],
        self_loop=False
    )
>>> draw_mis_solution(
        save_path="docs/mis_solution.png",
        graph_data=mis_solver.graph_data[0],
        self_loop=False
    )

Visualization Results:

CVRP

>>> from ml4co_kit import CVRPHGSSolver
>>> from ml4co_kit import draw_cvrp_problem, draw_cvrp_solution

# use CVRPHGSSolver to solve the problem
>>> solver = CVRPHGSSolver(
    depots_scale=1,
    points_scale=1,
    time_limit=1.0
)
>>> solver.from_vrplib("examples/cvrp/vrplib_1/problem/A-n32-k5.vrp")
>>> solver.solve()

# draw images of problem and solution
>>> draw_cvrp_problem(
        save_path="docs/assets/cvrp_problem.png",
        depots=solver.depots[0],
        points=solver.points[0]
    )
>>> draw_cvrp_solution(
        save_path="docs/assets/cvrp_solution.png",
        depots=solver.depots[0],
        points=solver.points[0],
        tour=solver.tours
    )

Visualization Results:

Develop ML4CO Methods

Please refer to ml4co_kit/learning for the base classes that facilitate a quick establishment of a ML4CO project. You can easily build a project by inheriting the base classes and additionally implement task-specific and methodology-specific functions according to [ML4CO Organization](#ML4CO Organization:).

Citation

If you find our code helpful in your research, please cite

@inproceedings{li2025streamlining,
    title={Streamlining the Design Space of ML4TSP Suggests Principles for Learning and Search}, 
    author={Yang Li and Jiale Ma and Wenzheng Pan and Runzhong Wang and Haoyu Geng and Nianzu Yang and Junchi Yan},
    year={2025},
    booktitle={International Conference on Learning Representations}
}