Unsupervised In-Context Operator Learning for Mean Field Games

Learn about a novel framework for solving high-dimensional mean-field games using unsupervised in-context operator learning in this 39-minute conference talk. Discover how recent deep learning advances have created innovative approaches to mean-field games (MFGs) but remain limited to single-instance solutions with extensive computational requirements. Explore the presenter's breakthrough method that uses in-context learning to create a model that takes MFG instances as input and directly outputs solutions in a single forward pass, dramatically improving computational efficiency. Understand the two key advantages of this approach: its discretization-free nature that makes it particularly effective for high-dimensional MFGs, and its ability to train without supervised labels, reducing the computational burden of preparing training datasets common in existing operator learning methods. Examine the generalization-error analysis on this transformer-based model that bridges the framework to emerging theory on in-context learning, highlighting broader implications and future research directions for scientific machine learning applications.