marsopt

Mixed Adaptive Random Search for Optimization

marsopt is a Python library designed to simplify and accelerate hyperparameter and black-box optimization. It supports continuous, integer, and categorical variables. Based on Mixed Adaptive Random Search (MARS) algorithm, marsopt dynamically balances exploration and exploitation through:

• adaptive noise for sampling,
• elite selection to guide the search toward promising regions,
• flexible handling of log-scale and categorical parameters,
• minimization or maximization of any user-defined objective.

MARS iteratively refines a population of “elite” solutions and generates new candidates by perturbing those elites with gradually decreasing noise. For a more detailed explanation of MARS, see our algorithm overview in the documentation.

Features

• Mixed Variable Support: Optimize integer, float (with optional log-scale), and categorical variables in the same study.
• Adaptive Sampling: Early iterations explore widely, while later iterations exploit the best regions found, thanks to a built-in cosine annealing scheme for noise.
• Easy Setup: Simply define an objective function, specify a variable search space, and run study.optimize().
• Resume & Extend: Continue a Study with more trials at any time without losing past information.
• Rich Tracking: Inspect all trial details (objective values, parameters, times, and so on) for deeper analysis.

Installation

Install marsopt from PyPI:

pip install marsopt

Getting Started

Below is a simplified example of how to use marsopt to tune a few common hyperparameters:

from marsopt import Study, Trial
import numpy as np

def objective(trial: Trial) -> float:
    lr = trial.suggest_float("learning_rate", 1e-4, 1e-1, log=True)
    layers = trial.suggest_int("num_layers", 1, 5)
    optimizer = trial.suggest_categorical("optimizer", ["adam", "sgd", "rmsprop"])

    score = -5 * (np.log10(lr) + 3) ** 2  
    score += np.log1p(layers) * 10  
    score += {"adam": 15, "sgd": 5, "rmsprop": 20}[optimizer]

    return -score

# Run optimization
study = Study(direction="minimize", random_state=42)
study.optimize(objective, n_trials=50)

Documentation

For more detailed information about the API and advanced usage, please refer to the full documentation.

Contributing

Contributions are welcome! If you'd like to improve marsopt or suggest new features, feel free to fork the repository and submit a pull request.

License

This project is licensed under the MIT License.