[2C6-OS-7c] OS-7 (3)

[2C6-OS-7c-04] Trajectory generation with imitation learning reflecting evaluation of defensive teams in soccer

〇Masakiyo Teranishi¹, Keisuke Fujii¹, Kazuya Takeda¹ (1. Nagoya University)

Keywords:Machine Learning, Imitation Learning, Sports

Soccer is a complex sport with the interaction among 22 players and the ball. Player position data has been recently measured. However, the usage is limited to individual movements, and videos are mainly used for decision making of team tactics. Most of previous studies in team trajectory generation have been evaluated only on prediction errors and did not take tactical evaluations (e.g., good defense) into consideration. In this study, we focus on team defense, which has relatively few individual differences, and explicitly add the index whether the defensive team protects the goal to the feature vector, and perform multi-agent imitation learning reflecting the evaluation. The results showed that the proposed method had the similar prediction performance to the existing method (Le et al. 2017), but generated an improved trajectory in terms of defensive evaluation. This indicates that it is effective for the learning to reflect the tactical evaluation to generate a tactically meaningful trajectory in team sports.