Stirring the Pot - Teaching Reinforcement Learning Agents a ”Push-Your-Luck” board game

Recent successes in AI research concerning traditional games like GO, have led to increased interest in the field of reinforcement learning. Modern board game design, however, has risen in complexity. This paper introduces a novel task for reinforcement learning: “Quacks of Quedlinburg”. A modern board game with risk management, deck building, and the option to choose a specific rule set out of thousands of possible combinations for every game. We provide an environment based on the game and perform initial experiments. In these, we found that Deep Q-Learning agents can significantly outperform simple heuristics.