Real Time In-Game Playstyle Classification Using A Hybrid Probabilistic Supervised Learning Approach
Abstract
In interactive digital media, such as video games, bringing about an adaptive or personalised experience requires a mechanism for correctly classifying or identifying the player style, before attempting to modify the experience in some way that improves player interest and immersion. This work presents a framework for solving this problem of in-game real time playstyle classification. We propose a hybrid probabilistic supervised learning approach, using Bayesian Inference informed by a K-Nearest Neighbors based likelihood, that is able to classify players in real time at every step within a given game level using only the latest player action or state observation. This improves on current approaches dependent on previous episodic player action trajectories in order to classify the player. Furthermore, we highlight the effect that this representation of the player state-action observation has on the in-game playstyle classification’s accuracy, prediction stability, and generalisability. We apply and test our framework using MiniDungeons, a rogue-like dungeon exploration game, and further evaluate our framework using a natural dataset containing human player action data from the platforming game Super Mario Bros. The experimental results obtained from our approach outperforms existing work in both domains. Furthermore, the evaluation results highlights the ability of our framework to generalise to unseen levels, without the need for additional retraining. Additionally, the Super Mario evaluation results illustrates the scalability of our framework to a more complex game environment with human player data.
Benjamin Rosman
Lab Director
I am a Professor in the School of Computer Science and Applied Mathematics at the University of the Witwatersrand in Johannesburg. I work in robotics, artificial intelligence, decision theory and machine learning.