A Machine Learning Approach to Detect Strategic Behavior from Large-Population Observational Data Applied to Game Mode Prediction on a Team-Based Video Game
Boshen Wang, Luis E. Ortiz
TL;DR
This study tackles the problem of determining whether players act strategically in a real-world multi-agent setting when only observed actions are available. It introduces a general ML framework that compares baseline neural predictors using per-player state features against augmented predictors that incorporate historical co-play signals, aiming to detect strategic behavior without access to payoffs. On a large Heroes of the Storm dataset, the authors find statistically significant predictive gains when including historical co-play features, indicating strategic decision making by a subset of players and justifying more advanced game-theoretic modeling in future work. The work demonstrates a practical approach to signal detection in noisy observational data and lays groundwork for scalable integration of strategic reasoning into predictive models for complex multi-agent systems.
Abstract
Modeling the strategic behavior of agents in a real-world multi-agent system using existing state-of-the-art computational game-theoretic tools can be a daunting task, especially when only the actions taken by the agents can be observed. Before attempting such a task, it would be useful to gain insight into whether or not agents are in fact acting strategically at all, from a game-theoretic perspective. In this paper, we present an initial step toward addressing this problem by proposing a general approach based on machine learning fundamentals for detecting potentially strategic behavior. We instantiate the approach by applying state-of-the-art machine learning tools for model selection and performance evaluation of prediction models in the context of detecting the strategic behavior of players for game mode selection in the multiplayer online video game Heroes of the Storm. Specifically, as a baseline, we first train neural networks to predict players' game mode selections using only information about the state of the player themselves. Then, we train a new set of neural networks using the same architectures, this time incorporating "historical co-play" features that encode players' past interactions with other players. We find that including these new features led to statistically significant improvements in game mode prediction accuracy, providing a sufficiently strong signal that players indeed make decisions strategically, which justifies the development of more complex computational game-theoretic tools in the hope of improving modeling and predictive power. We discuss remaining research work about potential approaches to validate the effectiveness of this initial step to detect strategic behavior.