Advanced Matchmaking for Online First Person Shooter Games using Machine Learning

Abstract: Matchmaking is an essential part of many modern online multiplayer games and is used by game developers to give the players the best possible online gaming experience. However, since video games have become more complex, traditional matchmaking systems like TrueSkill have reached their limits in predicting match outcomes. An extensive literature survey on engagement and balance in video games as well as an analysis of Battlefield 4 player data showed that balance can have a significant impact on player engagement. This lays the basis for the new matchmaking approach that is presented in this thesis. It is developed for the online First Person Shooter game Battlefield 4, with the goal of increasing player engagement by balancing online multiplayer matches. The developed matchmaking system is based on regression models, which use player performance metrics to predict the balance of online multiplayer matches. The experimental evaluations of the developed models show that the quality of the prediction results are influenced by the complexity of the different game modes available in Battlefield 4. Furthermore the historical Battlefield 4 game report data, which is used for building the predictive models, shows that this complexity as well as imbalances in the game design add significant noise to balance predictions. Both evaluated regression models – Linear Regression and Multivariate Adaptive Regression Splines – showed similar prediction errors within statistically expected deviation. Additionally it is shown that both methods have significantly smaller errors than the TrueSkill system, when predicting the outcome of games in Team Death Match or Conquest mode. The features that resulted in the lowest errors are commonly used in online First Person Shooter games. Hence the findings of this thesis can not only improve the matchmaking of Battlefield 4, but also benefit other video games of the same genre.