Team chemistry: the missing link in skill assessment for teams

Abstract

The task of assessing the skill of players and teams in games is an old problem spanning numerous disciplinary fields and can be traced back to foundational work from the early 20th century. However, in the past 15 years, the arrival and immense popularity of online multi-player gaming has kindled new interest in skill assessment due to the importance of ensuring that automatically-generated competitions (in a process called "matchmaking") between perhaps millions of candidate players and teams are fair - that each player or team competing against one another has a roughly equal probability of winning a given game. Poor matchmaking has the effect of discouraging less-skilled players from continuing to play, which, in games that are increasingly reliant on multiplayer competition, is detrimental to a game's longevity and, therefore, its profitability. Beyond this problem, though, there exists a more general need to better account for attributes present in team-based games specifically, including the notion of "team chemistry" - a latent feature corresponding to the level of cohesion among teammates believed to impact the expected performance of teams not accounted for by the comparatively narrow lens of individual player skill alone. In this thesis, we introduce a skill assessment framework which accounts for the effects of "team chemistry" using the performances of subgroups of players in teams. These subgroups therefore form the atomic unit to which skill ratings are assigned and maintained, standing in stark contrast to the existing practice of assigning skill ratings to individual players only. Further, existing skill assessment algorithms, such as Elo, Glicko, or TrueSkill, can be easily modified to be utilized as "base learners" for the maintenance of these subgroup ratings. The final estimated overall skill of a team is then computed as an aggregation of these subgroup skill ratings, and we describe a number of novel approaches for doing so. Through experimentation, it is shown that several of these aggregation approaches greatly improve the likelihood of correctly predicting the outcomes of unseen games, and we draw a number of interesting conclusions based on evaluations conducted on datasets from online multi-player video games and real-world sports.