Date of Award
5-1-2026
Document Type
Thesis
Degree Name
M.S. in Engineering Science
First Advisor
David Harrison
Second Advisor
Hong Xiao
Third Advisor
Bo Wang
School
University of Mississippi
Relational Format
dissertation/thesis
Abstract
This thesis examines how different multi-agent prediction architectures perform when forecasting future player positions in soccer using only ball and player coordinate data, with-out relying on pretrained knowledge of soccer rules or tactical concepts. Spatial positioning is a fundamental part of soccer and is often a defining characteristic of high-level players. However, there is still no fully deterministic method for evaluating how effectively a player uses space. As multi-agent trajectory prediction methods have improved over the last decade, it has become increasingly feasible to develop systems that can analyze and eventually grade player movement. A necessary first step toward that goal is the ability to predict where players are likely to move next.
This work compares the performance of Gated Recurrent Unit (GRU) and Transformer-based architectures for future player-location prediction. The models are evaluated using multiple metrics, including Root Mean Square Error (RMSE), the percentage of successful predictions within a defined distance threshold, and the percentage of catastrophic failures where predicted positions deviate substantially from the ground truth. These metrics pro-vide a broader understanding of model behavior by measuring not only average error, but also practical prediction usefulness and severe failure cases.
Recommended Citation
Heskett, Andrew John, "Predicting Future Positions Based on Spatiotemporal Data" (2026). Electronic Theses and Dissertations. 8830.
https://egrove.olemiss.edu/etd/8830