Modeling Strategic Drafting in Esports: A Generative AI Approach Using BERT for Ban/Pick Prediction in DotA 2

Abstract

Despite the rapid growth of the video game industry, particularly in the realm of esports, the comprehension of the increasingly intricate strategic behaviors, such as drafting, ban/pick, item builds, and others, within these competitive environments continues to be largely neglected. This study proposes the utilization of generative AI (GenAI) for the purposes of draft prediction and analysis in esports. We introduce a BERT-based model to predict ban/pick behavior in professional DotA 2, trained on 2,295 matches. Our proposed method not only serves as an innovative and more accurate framework to predict strategic behaviors in esports but can also be used to analyze ban/pick suggestions and identify key interventions, serving as a tool for practitioners to analyze the metagame. This work contributes both a methodological advance in modeling sequential strategic decisions and a practical framework that can be extended to other complex multi-agent behaviors in games.