Analysis And Determination of Key Events in Dota 2 Matches

DotA 2 is a multiplayer online session game, which involves two teams of five gamers. Each team has a base on which the main building is placed — the throne; the aim of the game is to break the enemy throne: as soon as one of the teams does this, it wins the match. Also in the game there is a boss named Roshan, whose murder gives a certain advantage to the team that did it. Each player at the beginning of the match makes a choice of one hero, which he will fight during the fight. The paper deals with two tasks: the prediction of the team that will win the match, as well as which of the teams next kill Roshan, or the fact that it will not be killed by anyone — it is said about the next murder, because it is assumed that the prediction algorithm receives input matches in which not all information is known — part of the fight deliberately hidden from the algorithm to learn to predict events at any time. The algorithms presented earlier in the articles are mainly based on information about the choice of characters, while many participants of the competition in machine learning in such competitions spend a lot of time generating features, the interpretability of which is not always easy, and fine-tuning of learning algorithms. The main purpose of this work is to create the architecture of the neural network, which is in the most successful way combines information about the choice of characters, the performance of characters in time during the fight, and the overall performance of the characters for the fight: inspired by the idea of Factorization Machines, the author offers the architecture of the neural network block, effectively able to account gamers’ choice of the characters for the match. The paper demonstrates the successful implementation of this idea, which is the most successful in terms of quality among all the algorithms that were took into consideration.