Reinforced In-Context Black-Box Optimization

This repository contains the Python code for Reinforced In-Context Black-Box Optimization (RIBBO), a method to reinforce-learn a BBO algorithm from offline data in an end-to-end fashion.

Requirements

• Python == 3.10
• PyTorch == 2.0.1
• offlinerllib==0.1.1
• utilsrl==0.6.3
• google-vizier==0.1.9
• gpytorch==1.11
• botorch=0.9.4

File Structure

• algorithms directory is the main implement of RIBBO, BC, BC Filter, and OptFormer
• data_gen directory is the implement of behavior algorithms and data collection
• datasets directory provides the interface of the offline datasets
• problems directory is the implement of the benchmark problems
• scripts directory provides some scripts for reproduction

Usage

Run bash scripts/run_main.sh to evaluate RIBBO and other baselines

Datasets

The datasets are generated by using codes in data_gen directory. Run bash scripts/run_data_gen.sh to generate the datasets parallely.

Citation

@inproceedings{RIBBO,
    author = {Lei Song, Chenxiao Gao, Ke Xue, Chenyang Wu, Dong Li, Jianye Hao, Zongzhang Zhang, Chao Qian},
    title = {Reinforced In-Context Black-Box Optimization},
    booktitle = {Proceedings of the 34th International Joint Conference on Artificial Intelligence (IJCAI'25)},
    year = {2025},
    address={Montreal, Canada}
}