🚀 Scalable Neural Network-based Blackbox Optimization

This repository provides implementation for SNBO (Scalable Neural Network-based Blackbox Optimization) — a novel method for efficient blackbox optimization using neural networks. It also includes code for benchmark algorithms and a suite of test problems used in the paper.

📝 Note: This work is currently under review but a preprint version is available at: https://arxiv.org/abs/2508.03827

📌 Features

This repository includes implementation for the following optimization algorithms:

• SNBO: Scalable neural network-based blackbox optimization (proposed method)
• BO+LogEI: Bayesian optimization with LogEI acquisition
• IBNN: Bayesian optimization with IBNN kernel and LogEI acquisition
• TuRBO: Trust region Bayesian optimization
• DYCORS: Dynamic coordinate search with repsonse surface model

This repository also contains following test problems:

• Ackley function
• Rastrigin function
• Levy function
• Rover trajectory optimization
• Half-Cheetah problem

🛠 Installation

Before running any example, you need to set up the environment and install the dependencies.

🐍 1. Create a conda environment

conda create -n "snbo" python=3.12
conda activate snbo

📦 2. Install dependecies

Install the required Python packages using pip:

pip install numpy==2.3.1 scipy==1.16.0 matplotlib==3.10.3 torch==2.7.1
gpytorch==1.14 botorch==0.14.0 gymnasium["mujoco"]==1.2.0

If you want to use DYCORS, then please install it using following commands:

git clone https://github.com/aquirosr/DyCors
cd DyCors
pip install .

▶️ Running examples

📥 Important: Before running any example, download the latest released or tagged version of this repository from the releases page.

The main file for running experiments is optimize.py, found in algorithms folder. This file provides following arguments:

Argument	Type	Description
method	str	Optimization method: "bo", "ibnn", "turbo", "dycors", "snbo"
problem	str	Test problem: "ackley", "rastrigin", "levy", "rover", "halfcheetah"
dim	int	Dimensionality of the problem
n_init	int	Number of initial samples
max_evals	int	Maximum number of function evaluations
neurons	list	Neurons per layer (only for snbo)
act_funcs	list	Activation functions per layer (only for snbo)
seed	int	Random seed (optional)

There are two ways to run this file:

✅ Option 1: You can directly run the file from the terminal. For example, to solve the rover problem using SNBO method, you can use following command:

python algorithms/optimize.py --method "snbo" --problem "rover" --dim 100 --n_init 200 --max_evals 2000 --neurons 256 256 --act_funcs "GELU" "GELU"

⚠️ NOTE: Before running the file, you need to append the path of the root folder to PYTHONPATH variable. This can be done by running following command in the terminal before running the python file:

export PYTHONPATH=PATH_TO_SNBO_FOLDER:$PYTHONPATH

Or, you can add this line to your shell configuration file (.bashrc or.zshrc) and reload the terminal.

✅ Option 2 (Recommended): Instead of directly running the optimize.py file, you can use one of the ready-to-use scripts available in runscript folder. These scripts already include the export statement requried for appending the PYTHONPATH variable at the start of the script, you just need to ensure that correct path is defined.

To solve a test problem using SNBO or any of the benchmark methods, you can use single_run.sh file under runscripts folder. To execute the file, run:

bash runscripts/single_run.sh

If you want to run a batch of optimization, you can use batch_run.sh file under runscripts folder. To execute the file, run:

bash runscripts/batch_run.sh

⚠️ NOTE: It is recommended to run the python or the bash file from the root folder and NOT from within the subfolder.

When you run the optimize.py file or any of the bash script, a folder named results will be created that consists of different subfolders, depending on the problem you are solving and the method you selected. A mat file will be saved within appropriate subfoler that contains entire optimization history.

📊 Results from paper

To reproduce the data reported in the paper, you can use paper_results.sh script. Use the following command to run this script:

bash runscripts/paper_results.sh

For each seed value, this script loops through each problem and solves it using all the methods.

⏳ Warning: This script will take a long time to run, depending on the resources used

🧾 Citation

If you use SNBO method in your research, please cite the original work (citation coming soon, paper under review).