Matthew Caron and Prof. Oliver Müller presented their research project TacticalGPT at the StatsBomb Conference 2023 at Wembley Stadium in London. The conference was attended by researchers as well as representatives from FIFA and leading clubs from various leagues, including the Bundesliga, Premier League and Serie A.
TacticalGPT is a Large Language Model (LLM) that has been trained on event data from Premier League matches and is based on a GPT-3-like model. By fine-tuning on the event data, it can answer questions such as: Who will take the free kick? Where will he shoot it? What will the player do next? Where will a player position himself?
To enable the training of such a large model with 20 billion parameters, techniques such as 4-bit quantization and low-rank adapters (QLoRA) were used. The model could be further improved by Reinforcement Learning from Human Feedback (RLHF) and become an indispensable tool for professional soccer clubs.
The research paper can be found at the following link: https://statsbomb.com/wp-content/uploads/2023/10/TacticalGPT-Uncovering-the-Potential-of-LLMs-for-Predicting-Tactical-Decisions-in-Professional-Football.pdf.