Share Your Spoils: Reinforcement Learning For Card Games
Stage 12 - Confex Level 2
~All~Game DesignTech & Coding
Information
Card games such as Poker or Bridge often present a significant challenge for AI opponents in games given their imperfect information, since one has to guess the cards in the hands of other players. In this presentation, we demonstrate how deep reinforcement learning was employed in the novel trick-taking card game, "Spoils," set within “Saltsea Chronicles” by Die Gute Fabrik, in collaboration with modl.ai. The discussion unveils an innovative approach that harnesses self-play and adaptability to novel game rules through machine learning.
The presentation will detail how to approach games with imperfect information and why reinforcement learning can be an effective solution. It will also showcase methods to address challenges such as the absence of training data, rule variants, and the indirect communication and strategy between the AI-controlled player and the human player in a team setting. This will be accomplished by outlining the tailored AI models used in Spoils and how the same architecture can be applied to various other card games or games with imperfect information.
Target Audiences
This talk is aimed at an audience with a more technical background and interest in using machine learning for game development. Basic knowledge of deep learning and reinforcement learning is encouraged, but not required, as a short introduction to th
Experience Level
Intermediate
Key Take Aways
Attendees will discover advanced insights into integrating deep reinforcement learning in imperfect information games using Spoils from Saltsea Chronicles, a trick-taking card game, as a prime example. The detailed exploration covers adapting AI to novel rules, indirect communication, self-play, and teamplay.
Session Type
Talk
