Rating Systems for Winner Prediction in eSports

Dota 2 is a popular game where two teams of 5 players compete each other in a real time strategic action by controlling a single hero chosen out of 110+. Predicting a winner in games like Dota with the help of machine learning methods and mathematical models is currently a hot topic of data analysts’ interest and both: amateur and academic research. This study aimed to come up with a modified version of TrueSkill rank- ing system in order to check whether an alternative approach to predicting posterior distributions of players’ ratings would yield a more precise prediction. It turned out that the direct minimization of Kullback-Leibler divergence between the true posterior distribution of ratings and its approximation yields roughly the same albeit slightly more accurate prediction than the original model. At the same time, this approach conceptually simplifies the way the algorithm works.