Invited Talk on AI and Games!

Join us for this invited talk of the Games Engineering group by Timo Bertram:

Z6, Room 0.001 at 16 c.t. this Monday, 23.09.

Title: 

Contrastive Learning for Imperfect Information Games

Abstract:

Games serve as a critical domain for advancing artificial intelligence (AI) research. While
AI agents have outperformed humans in many classical games, contemporary board and
video games continue to present significant challenges. We investigate the application
of contrastive learning—a method rooted in metric learning that leverages comparisons
rather than direct targets—to modern board games. In its classical use, contrastive learning
utilises knowledge about the similarity of items (such as sentences or images) to project
them into a latent embedding space, where proximity reflects their similarity. Despite
its success in image recognition and multimodal data alignment, contrastive learning
remains largely unexplored in GameAI.
As the backbone of our work, we develop the Contrastive Preference Ranking (CPR),
which uses contrastive learning techniques to models complex game dynamics in specific
contexts. CPR compares sets of items based on human preferences, enabling a situation-
specific understanding of decisions in games. Our method is used for two distinct games:
Magic: The Gathering and Reconnaissance Blind Chess.
In Magic: The Gathering, we demonstrate that CPR effectively predicts human decisions
during card drafts. Later, this is extended to the analysis of unseen cards, an area previ-
ously unaddressed in the literature. For the imperfect information game Reconnaissance
Blind Chess, our contrastive approach assesses potential game states, approximating the
probability of each state occurring in a game. This probabilistic modelling enhances agent
performance by reducing the complexity and pruning unlikely states.
Our findings reveal that CPR not only achieves strong results across these diverse games
and tasks but also highlights the general versatility of contrastive learning in game contexts.
We believe that this might lead to a reevaluation of the potential applications of contrastive
learning, and anticipate our research to be a step towards a broader integration into
GameAI.

Find Timo's CV and publications on his github page!