Quantitative Action Evaluation Metrics in Football Based on Action Selection and State Transition Probability Estimation through Language Modeling

Abstract

In applied sports analytics, especially in goal-oriented sports such as soccer, the necessity to process continuous data involving multiple agents poses a significant challenge for comprehensive analysis. Existing metrics, which often rely on static and simple inputs such as the ball's coordinates and nearby statistical data, fail to account for the broader match context. Recently, there has been growing interest in Large Language Models (LLMs), recognized for their potential not only in natural language processing but also in reinforcement learning and multi-agent trajectory forecasting. However, their application in soccer analytics is still in its infancy. This study introduces an innovative approach that employs language models to predict sequences tokenizing both the observable states and selected actions of players on the field. This method aims to model the probabilities associated with players' decision-making and state transitions. Furthermore, this research pioneers a framework to calculate and evaluate the state-action value (Q-value) of each player and team, leveraging the model's outputs within a reinforcement learning context. This marks the first attempt to establish such a quantitative evaluation framework in this domain.