TeamSkill: Modeling Team Chemistry in Online Multi-player Games

摘要

In this paper, we introduce a framework for modeling elements of "team chemistry" in the skill assessment process using the performances of subsets of teams and four approaches which make use of this framework to estimate the collective skill of a team. A new dataset based on the Xbox 360 video game, Halo 3, is used for evaluation. The dataset is comprised of online scrimmage and tournament games played between professional Halo 3 teams competing in the Major League Gaming (MLG) Pro Circuit during the 2008 and 2009 seasons. Using the Elo, Glicko, and TrueSkill rating systems as "base learners" for our approaches, we predict the outcomes of games based on subsets of the overall dataset in order to investigate their performance given differing game histories and playing environments. We find that Glicko and TrueSkill benefit greatly from our approaches (TeamSkill-A11K-EV in particular), significantly boosting prediction accuracy in close games and improving performance overall, while Elo performs better without them. We also find that the ways in which each rating system handles skill variance largely determines whether or not it will benefit from our techniques.