English | 简体中文

ESBox is an efficient tool for black-box optimization with multiple evolutionary strategy algorithms.

ESBox Capabilities in a Glance

Build-in Problems	Algorithms	User APIS
Function problems Ackley Griewank Rastrigin Rosenbrock StyblinskiTang Zakharov RL problems Mujoco HalfCheetah-v5 Humanoid-v5	OpenAI-ES paper Gaussian sampler (Mirror) OpenAIES learner ARS paper Gaussian sampler (Mirror) ARS learner NSRAES paper Gaussian sampler (Mirror) NSRAES learner CMA-ES paper CMA sampler CMAES learner Sep-CMA-ES paper Sep-CMA sampler Sep-CMAES learner	Config Objects Model (torch, paddlepaddle) List (float) Examples Local training Function (List, Model) Distributed training RL problem: HalfCheetah-v5 (Model) QuickStart RL problem: Cartpole-v1 (optimization Model) local training distributed training Function problem: Quadratic function (optimization float List) local training distributed training

How to use

Install

git clone https://github.com/ShuaibinLi/ESBox.git
cd ESBox
pip install . 

Other Dependencies

• ray
• pytorch
• gymnasium Note: To use mujoco/atari env, use pip install "gymnasium[all]" to install gymnasium.

Start training

QuickStart

• Two training methods: local training and distributed training
• Two forms of optimization: Model optimization (CartPole-v1 as an example) and float List optimization (2-dimensional 2nd degree function as an example)

Examples

• Examples of five algorithms that solve two categories of problems
• The results of the benchmark reproduction

Citation

@misc{2025ESBox,
    title={ESBox},
    author={Shuaibin Li},
    howpublished = {\url{https://github.com/ShuaibinLi/ESBox}},
    year={2025}
}