Mathematical Approaches for Learning From Gaming Data

Project began May 2020.

Games, as open ended constructs, tend to create massive action spaces. This makes theory-driven analysis of player behavior extremely challenging for non-domain experts. Data-driven machine learning modeling, while a relatively new tool in the field of game analytics, shows tremendous promise in this respect. [1] provides an excellent survey of machine learning techniques used for modeling player behavior. The authors argue that, despite the challenges stemming from quality and quantity of available gaming data and performance of existing algorithms, data-driven modeling will play a large role in game development.

With our increasing computational capabilities and development of novel virtual reality (VR) frameworks, the games are becoming more realistic. This has allowed games to evolve beyond simple entertainment tools - they are now also being used for educational purposes, such as historical teaching and enhancing museum visits [2]. The US Army routinely employs realistic games for training soldiers in a low-risk environment. With such a wide array of practical applications, learning from game data is becoming a key challenge.

In this project, we aim to build a general framework in which we can pose and answer the question of learning optimal strategies from available gameplay data. We also seek to quantify the uncertainties of the success of these policies to make this framework more robust.

We seek to answer the following fundamental question: ​Can we mathematically characterize the features of a game (e.g. an optimal strategy) from detailed records of different players, each playing the same game several times?

This project leverages the research being performed in another ARC project.

Publications:

• Chen B, Tandon S, Gorsich D, Gorodetsky A & Veerapaneni S. Behavioral Cloning in Atari Games Using a Combined Variational Autoencoder and Predictor Model, IEEE Congress on Evolutionary Computation, 2021
• Aksoy D, Gorsich D, Veerapaneni S, Gorodetsky A. An Incremental Tensor Train Decomposition Algorithm, SIAM Journal on Scientific Computing, 2022 (In review)
• Chen B, Aksoy D, Gorsich D, Veerapaneni S, Gorodetsky A. Low-Rank Tensor Network Encodings for Video-to-Action Behavioral Cloning, IEEE Conference on Games, 2023 (Submitted)

References:

[1] D. Hooshyar, M. Yousefi, and H. Lim, “Data-driven approaches to game player modeling: A systematic literature review,” ​ACM Computing Surveys (CSUR)​, vol. 50, no. 6, pp. 1–19, 2018.

[2] E. F. Anderson, L. McLoughlin, F. Liarokapis, C. Peters, P. Petridis, and S. De Freitas, “Developing serious games for cultural heritage: A state-of-the-art review,” ​Virtual reality​, vol. 14, no. 4, pp. 255–275, 2010.

#1.A81