Dominion AI
—Training an agent to play Dominion through reinforcement learning
——— Work in Progress ———
Dominion is a deck-building card game with a complex state space. This project applies reinforcement learning to train bots to play the game efficiently. Key design decisions include whether to use a fixed or dynamic action space and how to structure the Q-learning architecture for the distinct action and buy phases of the game. The work is built on top of the open—source Pyminion engine.
Project Overview
Goal- Develop a reinforcement learning agent capable of playing the card game Dominion effectively.
- Explore strategies for handling the game's large, partially observable state space.
- Investigate how evolving strategies emerge in a non-deterministic, multi-agent environment.
- Built an environment using an open-source digital implementation of Dominion for training and evaluation.
- Experimented with different reward-shaping methods to guide agent learning.
- Refine the agent with advanced methods such as curriculum learning and multi-head Q-networks.