Real-Time Predictor in Two-Players Fighting Game via Vision Transformer

Kittimate Chulajata, Sean Wu, Eric Laukien, Fabien Scalzo, Eun Sang Cha

Abstract: Forecasting the outcome in competitive gaming tournaments with real-time analytics has the potential to further engage audiences. However, there has yet to be a robust, real-time forecasting method specifically tailored for two-player fighting games. Live predictions are difficult due to the fast pace and unpredictable nature of these games, involving character positioning, movesets, decision-making, and containing few numerical indicators. In an attempt to address these issues, we present ArcadeViT, a novel real-time prediction algorithm utilizing Vision Transformer. ArcadeViT enables real-time and robust predictions of win-lose outcomes using live gameplay footage. As a proof of concept, we evaluate our model’s performance within a classic, two-player arcade game, Street Fighter II. We also benchmark our method against baseline models, we open-source our dataset and code in hopes of furthering work in predictive analysis for two-player fighting games.

External IDs:dblp:conf/isvc/ChulajataWLSC24