Skip to content

mila-iqia/Awesome-Offline-Model-Based-Optimization

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

33 Commits
bibtex
bibtex
 
 
README.md
README.md
 
 

Repository files navigation

Awesome-Offline-Model-Based-Optimization

📰 Must-Read Papers on Offline Model-Based Optimization 🔥

This repository collects important papers for our latest survey: "Offline Model-Based Optimization: Comprehensive Review", which is authored by Minsu Kim, Jiayao Gu, Ye Yuan, Taeyoung Yun, Zixuan Liu, Yoshua Bengio, and Can Chen.

Awesome Stars Forks

  • 💻: links to the code
  • 📖: links to the bibtex

Latest Updates

  • [2025/03/23] Our Survey is Publicly Accessible Now: See Our ArXiv Preprint here!
  • [2025/03/04] First Release of Awesome-Offline-Model-Based Optimization!

🔍 Table of Contents

🌟 What is Offline Model-Based Optimization?

In offline optimization, the goal is to discover a new design, denoted by $\boldsymbol{x}^*$, that maximizes the objective(s) $\boldsymbol{f}(\boldsymbol{x})$. This is achieved using an offline dataset $\mathcal{D}$, which consists of $N$ designs paired with their property labels. In particular, the dataset is given by Offline Dataset where each design vector $\boldsymbol{x}_i$ belongs to a design space $\mathcal{X} \subseteq \mathbb{R}^d$, and each property label $\boldsymbol{y}_i \in \mathbb{R}^m$ contains the corresponding $m$ objective values for that design. The function $\boldsymbol{f}: \mathcal{X} \rightarrow \mathbb{R}^m$ maps a design to its $m$-dimensional objective value vector.

In offline single-objective optimization, only one objective is considered (i.e., $m=1$). For instance, the design $\boldsymbol{x}$ might represent a neural network architecture, with $f(\boldsymbol{x})$ denoting the network's accuracy on a given dataset. Offline multi-objective optimization extends the framework to simultaneously address multiple objectives. In this setting, the goal is to find solutions that balance competing objectives effectively. For instance, when designing a neural architecture, one might seek to achieve both high accuracy and high efficiency.

We review recent benchmarks, highlighting key tasks and evaluation metrics. We then categorize existing approaches into two key areas: surrogate modeling, which emphasizes accurate function approximation in out-of-distribution regions, and generative modeling, which explores high-dimensional design spaces to identify high-performing designs.

🔗 Benchmark

Task

Synthetic Function

Real-World System

Scientific Design

Machine Learning Model

Evaluation Metric

Usefulness

Diversity

Novelty

Stability

🎯 Surrogate Modeling

Auxiliary Loss

Data-Drive Adaptation

Collaborative Ensembling

Generative Model Integration

🤔 Generative Modeling

Variational Autoencoder (VAE)

Generative Adversarial Network (GAN)

Autoregressive Model

Diffusion Model

Flow Matching

Energy-Based Model

Control by Generative Flow Network (GFlowNet)

Citation

If you found our survey paper is useful for your research, please consider cite our work:

@misc{kim2025offline,
      title={Offline Model-Based Optimization: Comprehensive Review}, 
      author={Minsu Kim and Jiayao Gu and Ye Yuan and Taeyoung Yun and Zixuan Liu and Yoshua Bengio and Can Chen},
      year={2025},
      eprint={2503.17286},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2503.17286}, 
}

About

📰 Must-Read Papers on Offline Model-Based Optimization 🔥

Topics

optimization generative-model model-based-optimization blackbox-optimization biosequences offline-setting

Resources

Readme
Activity
Custom properties

Stars

23 stars

Watchers

1 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •