Association Rule Mining for Unknown Video Games

Computational intelligence agents can reach expert levels in many known games, such as Chess, Go, and Morris. Those systems incorporate powerful machine learning algorithms, which are able to learn from, e.g., observations, play-traces, or by reinforcement learning. While many black box systems, such as deep neural networks, are able to achieve high performance in a wide range of applications, they generally lack interpretability. Additionally, previous systems often focused on a single or a small set of games, which makes it a cumbersome task to rebuild and retrain the agent for each possible application. This paper proposes a method, which extracts an interpretable set of game rules for previously unknown games. Frequent pattern mining is used to find common observation patterns in the game environment. Finally, game rules as well as winning-/losing-conditions are extracted via association rule analysis. Our evaluation shows that a wide range of game rules can be successfully extracted from previously unknown games. We further highlight how the application of fuzzy methods can advance our efforts in generating explainable artifical intelligence (AI) agents.