AI-Based 2D Soccer Simulation

Generating and Learning from Players' and Ball Trajectory Data

Abstract

In Mizuno, Fujimoto, and Ishikawa (Front. Phys., 2022), it was demonstrated that an AI capable of generating individual movement trajectories could be constructed by training the GPT architecture from scratch on historical location data. However, the challenge of generating movement trajectories with interactions among multiple individuals remained. In this study, we construct an AI that generates trajectories of moving players and the ball while maintaining and sometimes disrupting formations, by simultaneously learning from scratch two-dimensional tracking data of players and the ball in a soccer match using the GPT architecture. By sequentially generating the positions of all players and the ball on the field, we simulate parts of a soccer game.