New AI Paper Learns Policy Embeddings for Imperfect-Information Games

A new paper by Kevin Wang, Kevin Yang, Arjun Prakash, and Amy Greenwald tackles the problem of learning useful policy representations in two-player zero-sum imperfect-information games. The authors propose three contributions: first, they introduce methods for creating datasets of policies tailored to specific games; second, they develop self-supervised learning techniques to embed those policies into low-dimensional representations; and third, they define downstream tasks—such as policy clustering or similarity search—to evaluate the quality of the embeddings. The experiments are conducted on two classic poker benchmarks: Kuhn Poker and Leduc Poker, which are standard testbeds for imperfect-information games because of their hidden cards and sequential betting.

The results show that even with very basic self-supervised methods, the learned embeddings capture meaningful behavioral patterns. The authors claim this is among the first systematic comparisons of different self-supervised learning approaches for policy representations in games. While the techniques are still primitive, the work opens the door to transferring insights across policies, detecting opponent styles, and building more robust AI agents. The full code is available on GitHub, allowing other researchers to extend the methods to more complex games like no-limit poker or strategic board games. This research bridges machine learning and game theory, promising smarter AI for negotiations, auctions, and security domains.