Identifying Factors Influencing Electronic Gaming Machine Player Behavior Using Interpretable AI and Mimicking Player Behavior Using Reinforcement Learning

Abstract

People in venues like casinos play games on Electronic Gaming Machines (EGMs). These machines do not record information about players, such as their playing experience. The gaming industry is keen on learning about the different types of player behaviors and what leads to these behaviors. In this thesis, we justify an assumption and define whether or not a player had a positive experience during the session based on play session attributes. Furthermore, we identify the factors and their importance that contribute to a positive experience using Interpretable AI. We classify player sessions using a Decision Tree, Logistic Regression, and Explainable Boosting Machine (EBM) to gain insights into the factors used for prediction. EBM gave a comparable performance to that of a state-of-art model with high interpretability and accuracy of 95%. This understanding will provide insights into game performance as well as responsible gaming behaviors. Moreover, to acquire a good evaluation of a machine learning model as a viable alternative for real-world players, we train models that mimic player behavior. We use K-means to cluster different playing behaviors and determine termination states for one of the playing behaviors for Reinforcement Learning. We implemented PPO and ACKTR models to generate the playing behavior, with the agents being rewarded based on their proximity to the termination states. ACKTR performed well as the playing behavior generated by this model were statistically matching the real-world players behavior within the selected cluster.