---

OpenBox Documentation | OpenBox中文文档 | 中文README

OpenBox: Generalized and Efficient Blackbox Optimization System

OpenBox is an efficient and generalized blackbox optimization (BBO) system, which supports the following characteristics: 1) BBO with multiple objectives and constraints, 2) BBO with transfer learning, 3) BBO with distributed parallelization, 4) BBO with multi-fidelity acceleration and 5) BBO with early stops. OpenBox is designed and developed by the AutoML team from the DAIR Lab at Peking University, and its goal is to make blackbox optimization easier to apply both in industry and academia, and help facilitate data science.

Software Artifacts

Standalone Python package.

Users can install the released package and use it with Python.

Distributed BBO service.

We adopt the "BBO as a service" paradigm and implement OpenBox as a managed general service for black-box optimization. Users can access this service via REST API conveniently, and do not need to worry about other issues such as environment setup, software maintenance, programming, and optimization of the execution. Moreover, we also provide a Web UI, through which users can easily track and manage the tasks.

Design Goal

The design of OpenBox follows the following principles:

• Ease of use: Minimal user effort, and user-friendly visualization for tracking and managing BBO tasks.
• Consistent performance: Host state-of-the-art optimization algorithms; Choose the proper algorithm automatically.
• Resource-aware management: Give cost-model-based advice to users, e.g., minimal workers or time-budget.
• Scalability: Scale to dimensions on the number of input variables, objectives, tasks, trials, and parallel evaluations.
• High efficiency: Effective use of parallel resources, system optimization with transfer-learning and multi-fidelities, etc.
• Fault tolerance, extensibility, and data privacy protection.

Links

• Documentations | 中文文档
• Examples
• Pypi package
• Conda package: to appear soon
• Blog post: to appear soon

News

• OpenBox based solutions achieved the First Place of ACM CIKM 2021 AnalyticCup (Track - Automated Hyperparameter Optimization of Recommendation System).
• OpenBox team won the Top Prize (special prize) in the open-source innovation competition at 2021 CCF ChinaSoft conference.
• Pasca, which adopts Openbox to support neural architecture search functionality, won the Best Student Paper Award at WWW'22.

OpenBox Capabilities in a Glance

Build-in Optimization Components	Optimization Algorithms	Optimization Services
Surrogate Model Gaussian Process TPE Probabilistic Random Forest LightGBM Acquisition Function EI PI UCB MES EHVI TS Acquisition Optimizer Random Search Local Search Interleaved RS and LS Differential Evolution L-BFGS-B	Bayesian Optimization GP-based BO SMAC TPE LineBO SafeOpt Multi-fidelity Optimization Hyperband BOHB MFES-HB Evolutionary Algorithms Surrogate-assisted EA Regularized EA Adaptive EA Differential EA NSGA-II Others Anneal PSO Random Search	Local Machine Cluster Servers Hybrid mode Software as a Service

Installation

System Requirements

Installation Requirements:

• Python >= 3.8 (Python 3.8 is recommended!)

Supported Systems:

• Linux (Ubuntu, ...)
• macOS
• Windows

We strongly suggest you to create a Python environment via Anaconda:

conda create -n openbox python=3.8
conda activate openbox

Then we recommend you to update your pip, setuptools and wheel as follows:

pip install --upgrade pip setuptools wheel

Installation from PyPI

To install OpenBox from PyPI:

pip install openbox

For advanced features, install SWIG first and then run pip install "openbox[extra]".

Manual Installation from Source

To install the newest OpenBox from the source code, please run the following commands:

git clone https://github.com/PKU-DAIR/open-box.git && cd open-box
pip install .

Also, for advanced features, install SWIG first and then run pip install ".[extra]".

For more details about installation instructions, please refer to the Installation Guide.

Quick Start

A quick start example is given by:

import numpy as np
from openbox import Optimizer, space as sp

# Define Search Space
space = sp.Space()
x1 = sp.Real("x1", -5, 10, default_value=0)
x2 = sp.Real("x2", 0, 15, default_value=0)
space.add_variables([x1, x2])

# Define Objective Function
def branin(config):
    x1, x2 = config['x1'], config['x2']
    y = (x2-5.1/(4*np.pi**2)*x1**2+5/np.pi*x1-6)**2+10*(1-1/(8*np.pi))*np.cos(x1)+10
    return {'objectives': [y]}

# Run
if __name__ == '__main__':
    opt = Optimizer(branin, space, max_runs=50, task_id='quick_start')
    history = opt.run()
    print(history)

The example with multi-objectives and constraints is as follows:

import matplotlib.pyplot as plt
from openbox import Optimizer, space as sp

# Define Search Space
space = sp.Space()
x1 = sp.Real("x1", 0.1, 10.0)
x2 = sp.Real("x2", 0.0, 5.0)
space.add_variables([x1, x2])

# Define Objective Function
def CONSTR(config):
    x1, x2 = config['x1'], config['x2']
    y1, y2 = x1, (1.0 + x2) / x1
    c1, c2 = 6.0 - 9.0 * x1 - x2, 1.0 - 9.0 * x1 + x2
    return dict(objectives=[y1, y2], constraints=[c1, c2])

# Run
if __name__ == "__main__":
    opt = Optimizer(CONSTR, space, num_objectives=2, num_constraints=2,
                    max_runs=50, ref_point=[10.0, 10.0], task_id='moc')
    history = opt.run()
    history.plot_pareto_front()  # plot for 2 or 3 objectives
    plt.show()

We also provide HTML Visualization. Enable it by setting additional options visualization=basic/advanced and auto_open_html=True(optional) in Optimizer:

opt = Optimizer(...,
    visualization='advanced',  # or 'basic'. For 'advanced', run 'pip install "openbox[extra]"' first
    auto_open_html=True,       # open the visualization page in your browser automatically
)
history = opt.run()

For more visualization details, please refer to HTML Visualization.

More Examples:

• Single-Objective with Constraints
• Multi-Objective
• Multi-Objective with Constraints
• Ask-and-tell Interface
• Parallel Evaluation on Local
• Distributed Evaluation
• Tuning LightGBM
• Tuning XGBoost

Enterprise Users

• Tencent Inc.

• Alibaba Group

• Kuaishou Technology

Contributing

OpenBox has a frequent release cycle. Please let us know if you encounter a bug by filling an issue.

We appreciate all contributions. If you are planning to contribute any bug-fixes, please create a pull request.

If you plan to contribute new features, new modules, etc. please first open an issue or reuse an existing issue, and discuss the feature with us.

To learn more about making a contribution to OpenBox, please refer to our How-to contribution page.

We appreciate all contributions and thank all the contributors!

Feedback

• Fill an issue on GitHub
• Email us via Yang Li, shenyu@pku.edu.cn or jianghuaijun@pku.edu.cn
• [Q&A] Join the QQ group: 227229622

Related Projects

Targeting at openness and advancing AutoML ecosystems, we had also released few other open-source projects.

• MindWare: an open source system that provides end-to-end ML model training and inference capabilities.
• SGL: a scalable graph learning toolkit for extremely large graph datasets.
• HyperTune: a large-scale multi-fidelity hyper-parameter tuning system.

Related Publications

OpenBox: A Python Toolkit for Generalized Black-box Optimization.

Huaijun Jiang, Yu Shen, Yang Li, Beicheng Xu, Sixian Du, Wentao Zhang, Ce Zhang, Bin Cui; JMLR 2024, CCF-A. [paper] [arxiv]

OpenBox: A Generalized Black-box Optimization Service.

Yang Li, Yu Shen, Wentao Zhang, Yuanwei Chen, Huaijun Jiang, Mingchao Liu, Jiawei Jiang, Jinyang Gao, Wentao Wu, Zhi Yang, Ce Zhang, Bin Cui; KDD 2021, CCF-A. [paper] [arxiv]

MFES-HB: Efficient Hyperband with Multi-Fidelity Quality Measurements.

Yang Li, Yu Shen, Jiawei Jiang, Jinyang Gao, Ce Zhang, Bin Cui; AAAI 2021, CCF-A. [paper] [arxiv]

Transfer Learning based Search Space Design for Hyperparameter Tuning.

Yang Li, Yu Shen, Huaijun Jiang, Tianyi Bai, Wentao Zhang, Ce Zhang, Bin Cui; KDD 2022, CCF-A. [paper] [arxiv]

TransBO: Hyperparameter Optimization via Two-Phase Transfer Learning.

Yang Li, Yu Shen, Huaijun Jiang, Wentao Zhang, Zhi Yang, Ce Zhang, Bin Cui; KDD 2022, CCF-A. [paper] [arxiv]

PaSca: a Graph Neural Architecture Search System under the Scalable Paradigm.

Wentao Zhang, Yu Shen, Zheyu Lin, Yang Li, Xiaosen Li, Wen Ouyang, Yangyu Tao, Zhi Yang, and Bin Cui; WWW 2022, CCF-A, 🏆 Best Student Paper Award. [paper] [arxiv]

Hyper-Tune: Towards Efficient Hyper-parameter Tuning at Scale.

Yang Li, Yu Shen, Huaijun Jiang, Wentao Zhang, Jixiang Li, Ji Liu, Ce Zhang, Bin Cui; VLDB 2022, CCF-A. [paper] [arxiv]

Citing

If you use OpenBox, please consider citing the following articles:

@inproceedings{li2021openbox,
  title={Openbox: A generalized black-box optimization service},
  author={Li, Yang and Shen, Yu and Zhang, Wentao and Chen, Yuanwei and Jiang, Huaijun and Liu, Mingchao and Jiang, Jiawei and Gao, Jinyang and Wu, Wentao and Yang, Zhi and others},
  booktitle={Proceedings of the 27th ACM SIGKDD conference on knowledge discovery \& data mining},
  pages={3209--3219},
  year={2021}
}

@article{JMLR:v25:23-0537,
  author  = {Huaijun Jiang and Yu Shen and Yang Li and Beicheng Xu and Sixian Du and Wentao Zhang and Ce Zhang and Bin Cui},
  title   = {OpenBox: A Python Toolkit for Generalized Black-box Optimization},
  journal = {Journal of Machine Learning Research},
  year    = {2024},
  volume  = {25},
  number  = {120},
  pages   = {1--11},
  url     = {http://jmlr.org/papers/v25/23-0537.html}
}

License

The entire codebase is under MIT license.