Predicting Opponent’s Movement in Dota 2

Please enable javascript in your browser.

Bashiri, Vahid | 2021

269 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 54073 (19)
University: Sharif University of Technology
Department: Computer Engineering
Advisor(s): Fazli, Mohammad Amin
Abstract:
Video games with respect to an ever Increasing player pool and Industry growth, have attracted a lot of attention in recent years. Dota 2, as one of the most successful games both in casual gamers’ community and E-sport community, is considered as a proper case study, however, most of the research done was limited to predicting games’ outcome. Despite The popularity, the rather unintelligent AI of the game has made quite a frustrating experience for new players. In this research, with a novel approach, hero features are used to predict their future positions. For this purpose, 35 professional games are collected and analyzed and 601 features are extracted. Then, suitable features are selected and a deep neural network is trained. The proposed method not only outperforms base-line methods in professional games by 14 percent, but also have better performance in public games by 20 percent. Finally, a comparison between proposed method and expert players has been made. The result shows, the proposed method has 14 percent more accuracy than participant average players’ predictions. It could be used to improve AI decision-making and performance in Dota 2
Keywords:
Opponent Modeling ; Machine Learning ; Strategy Prediction ; Trajectory Prediction ; Dota 2 Game ; Video Games

No TOC