TrueSkills: rating systems

When evaluating the success of a team is not sophisticated processes, according to an out-of-date Elo rating model is often used, which evaluates only victory or defeat, improperly works with a draw, and ignores individual players in a team with different experience. At the first in 2018 seminar Research Group "Machine Learning and Social Computing" Alexander Sirotkin on the example of the game "What? Where? When?" told about the TrueSkill rating system, and how to improve it, by giving the greater contribution of the leader and estimation of cases where the team played in the incomplete composition.

Alexander Sirotkin spoke about the Bayesian rating as a probabilistic model, which means rating - a linearly ordered result of comparing the probabilities of winning teams. What in human sounds like if X has a higher rating, X will probably win. Then, on an example of a coin toss, answered how to calculate the Beta distribution - a mathematical description of the flip of heads or tails, when there was a forecast. After the probability, the Elo rating and its shortcomings were argued. First, for new players, the rating is manually tweaked to rapid evaluating of their skills. Secondly, the fitness, skills or external factors are deliberately ignored, to simplify the model. TrueSkill does not ignore those circumstances because it was developed to form random, but equal teams in the online game Halo, taking into consideration that no one likes to regularly lose and expects a tight game. So, the player's skills are considered normally distributed with the average ν - the skill itself and σ - random external factors, such as a tasteless breakfast or a spoiled mood. The delta of team rating from the victory or defeat is divided between the players in a proportion of the participation and contribution.

So, TrueSkill calculates the actual rating of each player as part of the team, but also applies to online marketing, to assess the success of contextual advertising. Rating systems are a very important topic for our Research Group. For example, Andrey Shelopugin uses GLICKO model, which similar to TrueSkill, in his work. Also, we hope that members of the Research Group will be able to apply TrueSkill in their research.

 Great seminar! Alexander knows how to talk about non-trivial themes of the Bayesian Ranking in a simple way. This helped not only to get acquainted with the basic models of ratings for the first time but also to understand their structure.

Белялов Ильдар Кафисович
Lecturer