MetaGen: A Framework for Metaheuristic Development and Hyperparameter Optimization

🚀 Why MetaGen?

MetaGen simplifies the development of metaheuristics and the optimization of hyperparameters in machine learning and deep learning. Whether you're a researcher, developer, or practitioner, MetaGen provides a structured, flexible, and scalable framework.

🔹 Key Features

✔ Metaheuristic Development Framework – A base class with the run loop, elitism and callbacks; you write three methods.
✔ Hyperparameter Optimization Tools – Search spaces with integers, reals, categoricals, permutations, groups and variable-length structures, conditional variables that apply only for some values of another, and a range per position of a structure, for layer and architecture-level tuning.
✔ Standardized Interface – Ensures compatibility between metaheuristic developers and end users.
✔ Dynamic Architecture Optimization – Structures whose length the search itself changes.
✔ Seamless Integration – Compatible with scikit-learn, tensorflow, pytorch, and other ML libraries.
✔ Built-in Metaheuristics – Pre-implemented algorithms ready to use.
✔ Reproducible Runs – A seed parameter controls every random draw, also across Ray workers.
✔ Pause and Resume – Give a run a checkpoint file and it continues where it stopped after a power failure or a job scheduler's time limit, with the same result as an uninterrupted run.
✔ Run History and TensorBoard – Every run keeps a record of each iteration, which it can write as a JSON Lines file for pandas; give it a log_dir and follow it graphically in TensorBoard.
✔ Scalable and Distributed Execution – With distributed=True, metaheuristics run across the CPUs of a Ray cluster.

📌 Built-in Metaheuristics

• Random Search
• Hill Climbing
• Tabu Search
• Simulated Annealing
• Genetic Algorithm and Steady-State Genetic Algorithm
• Memetic Algorithm
• Tree-structured Parzen Estimator: TPE, which evaluates a pool of candidates per iteration, and KernelTPE, which evaluates a single candidate per iteration, for expensive fitness functions
• Coronavirus Optimization Algorithm (CVOA), with two strain classes, CVOA and ProbabilisticCVOA, that differ in how deaths, superspreaders and isolation are decided

The documentation has a guide to choosing one, with what each costs in evaluations.

📦 Installation

MetaGen requires Python 3.10+ and can be installed with:

pip install pymetagen-datalabupo

Two features are optional and installed on demand:

pip install pymetagen-datalabupo[distributed]   # Ray, for distributed=True
pip install pymetagen-datalabupo[tensorboard]   # TensorBoard logging
pip install pymetagen-datalabupo[all]           # both

📖 Documentation

The official API reference and usage guides are available at: MetaGen Documentation

🤖 Example: Hyperparameter Optimization

Optimizing hyperparameters for a regression model:

from metagen.framework import Domain, Solution
from metagen.metaheuristics import RandomSearch
from sklearn.datasets import make_regression
from sklearn.linear_model import SGDRegressor
from sklearn.model_selection import cross_val_score

# Generate synthetic dataset
X, y = make_regression(n_samples=1000, n_features=4, random_state=0)

# Define the search space
regression_domain = Domain()
regression_domain.define_real("alpha", 0.0001, 0.001)
regression_domain.define_integer("iterations", 5, 200)
regression_domain.define_categorical("loss", ["squared_error", "huber", "epsilon_insensitive"])

# Fitness function: MetaGen always minimizes
def regression_fitness(solution: Solution) -> float:
    model = SGDRegressor(
        loss=solution["loss"],
        alpha=solution["alpha"],
        max_iter=solution["iterations"]
    )
    mape = cross_val_score(model, X, y, scoring="neg_mean_absolute_percentage_error").mean() * -1
    return mape

# Run optimization; the seed makes the run reproducible
best_solution = RandomSearch(regression_domain, regression_fitness, seed=0).run()
print(best_solution)

Any other metaheuristic takes the same two arguments. To follow the run in TensorBoard, add log_dir="logs/random_search" and launch tensorboard --logdir=logs; to see progress on the console, call set_metagen_logger_level() from metagen.logging.metagen_logger.

🛠 Example: Developing a Metaheuristic

Creating a simple Random Search metaheuristic:

from copy import deepcopy
from typing import Callable, List
from metagen.framework import Domain, Solution

class RandomSearch:

    def __init__(self, domain: Domain, fitness: Callable[[Solution], float], search_space_size: int = 30,
                iterations: int = 20) -> None:

        self.domain = domain
        self.fitness = fitness
        self.search_space_size = search_space_size
        self.iterations = iterations

    def run(self) -> Solution:

        potential_solutions: List[Solution] = list()

        for _ in range(0, self.search_space_size):
            potential_solutions.append(Solution(self.domain, connector=self.domain.get_connector()))

        solution: Solution = deepcopy(min(potential_solutions))

        for _ in range(0, self.iterations):
            for ps in potential_solutions:
                ps.mutate()

                ps.evaluate(self.fitness)
                if ps < solution:
                    solution = deepcopy(ps)

        return solution

A metaheuristic can also inherit from the Metaheuristic base class, which adds the run loop, elitism, seed, log_dir and distributed execution in exchange for three methods; see Extending the Metaheuristic class.

📝 Citing MetaGen

If you use MetaGen in your research, please cite:

D. Gutiérrez-Avilés, M. J. Jiménez-Navarro, J. F. Torres, F. Martínez-Álvarez. MetaGen: A framework for metaheuristic development and hyperparameter optimization in machine and deep learning. Neurocomputing 637 (2025) 130046. https://doi.org/10.1016/j.neucom.2025.130046

@article{metagen2025,
  title   = {MetaGen: A framework for metaheuristic development and hyperparameter optimization in machine and deep learning},
  author  = {Guti{\'e}rrez-Avil{\'e}s, David and Jim{\'e}nez-Navarro, Manuel Jes{\'u}s and Torres, Jos{\'e} Francisco and Mart{\'i}nez-{\'A}lvarez, Francisco},
  journal = {Neurocomputing},
  volume  = {637},
  pages   = {130046},
  year    = {2025},
  doi     = {10.1016/j.neucom.2025.130046}
}

The article describes version 0.2.0, which remains available on PyPI and as a tagged release.

🤝 Contributing

We welcome contributions from developers of all experience levels! To contribute:

• Open an issue or submit a pull request.

• Install the package for development and run the tests:

  pip install -e .[test]
  pytest test
  mypy src
  

  Both must pass; the continuous integration runs them on Python 3.10 to 3.12.

📌 Resources

• MetaGen paper (Neurocomputing, open access)
• CVOA paper
• Google Colab Notebooks

⚖️ License

MetaGen is free software, distributed under the GNU General Public License v3 or later.

MetaGen is an open-source project developed and maintained by:

• David Gutiérrez-Avilés
• Manuel Jesús Jiménez-Navarro
• Francisco José Torres-Maldonado
• Francisco Martínez-Álvarez

The authors are members of the Minerva AI Lab Group at the University of Seville and of DataLabUPO, the Data Science & Big Data Research Lab at Pablo de Olavide University.