Offline Model Based Optimization

Offline model-based optimization (MBO) tackles the challenge of optimizing complex functions using only pre-existing datasets, avoiding costly or risky online evaluations. Current research heavily focuses on leveraging generative models, particularly diffusion models, and energy-based models, often within a framework of inverse or forward mappings, to generate improved designs or actions. These methods address limitations of traditional surrogate models by incorporating uncertainty quantification and techniques to avoid out-of-distribution predictions, improving robustness and efficiency in diverse applications like materials science, robotics, and industrial control. The resulting advancements promise more efficient and reliable optimization in scenarios where online experimentation is impractical or impossible.