Monte Carlo Tree Search based Space Transfer for Black Box Optimization

Official implementation of NeurIPS'24 paper "Monte Carlo Tree Search based Space Transfer for Black Box Optimization".

This repository contains the Python code for MCTS-Transfer , an search space transfer algorithm for expensive Black-Box Optimization. The code is implemented based on LA-MCTS. Data generation code is based on RIBBO.

Requirements

Ubuntu == 20.04

Python == 3.8.0

pip install -r requirements.txt

Download data from the link to the directory /data.

Pretrained MCTS models can be downloaded from link or generated automatically.

Download HPO-B surrogates from the link to directory /functions/hpob/saved-surrogates.

Usage

# test on Sphere2D
bash experiments/run_sphere.sh

# test on BBOB
bash experiments/run_bbob.sh

# test on real-world problem
bash experiments/run_real.sh

# test on design-bench
bash experiments/run_design_bench.sh

# test on hpob
bash experiments/run_hpob.sh

Citation

@inproceedings{mcts-transfer,
    author = {Shu-kuan Wang , Ke Xue, Song Lei, Xiao-bin Huang, Chao Qian},
    title = {Monte Carlo Tree Search based Space Transfer for Black Box Optimization},
    booktitle = {Advances in Neural Information Processing Systems 38 (NeurIPS’24)},
    year = {2024},
    address={Vancouver, Canada}
}