ZEROSHOTOPT
收藏资源简介:
ZEROSHOTOPT数据集是一个为连续黑盒优化任务而设计的预训练模型,通过在12种贝叶斯优化变体的大规模优化轨迹上使用离线强化学习进行训练。该数据集包含了约2000万个合成的高斯过程函数,旨在帮助模型学习可迁移的优化策略。数据集的创建过程涉及使用高斯过程生成具有多样景观的合成函数,并从12种贝叶斯优化变体中收集大规模的优化轨迹。这些轨迹用于训练一个可扩展的transformer模型,该模型能够在2D到20D的维度范围内进行高效的黑盒优化。ZEROSHOTOPT数据集旨在解决在严格评估预算下进行黑盒优化的问题,为未来的扩展和改进提供了一个可重用的基础。
The ZEROSHOTOPT dataset is a pretrained model designed for continuous black-box optimization tasks, trained via offline reinforcement learning on large-scale optimization trajectories of 12 Bayesian optimization variants. This dataset contains approximately 20 million synthetic Gaussian process functions, aiming to help models learn transferable optimization strategies. The creation of the dataset involves generating synthetic functions with diverse landscapes using Gaussian processes, and collecting large-scale optimization trajectories from 12 Bayesian optimization variants. These trajectories are used to train a scalable Transformer model that can perform efficient black-box optimization within the dimensionality range of 2D to 20D. The ZEROSHOTOPT dataset is intended to address the problem of black-box optimization under strict evaluation budgets, providing a reusable foundation for future expansions and improvements.

- 1ZeroShotOpt: Towards Zero-Shot Pretrained Models for Efficient Black-Box Optimization麻省理工学院 · 2025年



