Des approches d'apprentissage automatique pour la modélisation et la prédiction de la performance dans les sports collectifs / Machine Learning approaches to model and predict performance in Team Sports
par
Amphithéâtre Laurent Schwartz, bâtiment 1R3
Institut de Mathématiques de Toulouse
Jury composé de :
Léo GERVILLE-REACHE, Rapporteur, Université de Bordeaux
Christophe LEY, Rapporteur, Université du Luxembourg
Brigitte GELEIN, Examinatrice, ENSAI Rennes
Béatrice LAURENT-BONNEAU, Examinatrice, INSA Toulouse
Sébastien DEJEAN, Directeur de thèse, Université de Toulouse
Cristian PASQUARETTA, Co-directeur de thèse, Université de Toulouse
Il sera également possible d'assister à la soutenance en visio :
https://rendez-vous.renater.fr/Soutenance_Arnaud_Odet_623fa6-73d9ac-b3c9e0
Résumé :
As sporting events and competitions occupy an ever-increasing place in contemporary society, with professional athletes’ salaries, media deals, and related revenues reaching unprecedented heights, sport analytics research has blossomed accordingly in recent years. As the stakes grow higher, the demand for performance optimization in sports has never been more pressing. In parallel, the latest advances in computer science, combined with the ever-increasing capacity of computing hardware, have propelled Machine Learning into the spotlight as a potential new industrial and societal revolution. Deep generative models are interfering more and more with our daily lives, and Artificial General Intelligence can hardly be dismissed as a mere fantasy anymore.
In this context, this thesis aims to leverage recent advances in Machine Learning to provide practical and actionable tools to ultimately improve collective performance in team sports.
At first, we advocate that combining Machine Learning predictive power to the task of forecasting games outcomes and algorithms explainability techniques yields promising results. Specifically, we present a framework providing multiscale diagnostic analyses that offer valuable insights regarding : (i) the drivers of a game outcome, (ii) the identification of strengths and weaknesses of a given team, and (iii) a general strategic understanding of the game. We illustrate this framework on men’s rugby union and women’s basketball. We then demonstrate that incorporating players’ individual tracking data into deep
learning time series models increases our ability to forecast the short-term performance of their team. Applying feature permutation methods, we highlight which players and features bear the greatest importance to this forecasting task, thereby characterizing their respective roles in play development.
Finally, we adopt unsupervised machine learning to abstract player archetypes from individual player statistics, and leverage these archetypes to identify the best-performing lineups (i.e. combinations of players). Our results suggest that players’ complementarity prevails over juxtaposition of individual talents. Adopting a broader angle and focusing on rosters (i.e. registered players for a team) instead of lineups, we further emphasize that
stability is positively correlated with long-term performance.
Our work opens up several promising research directions, including the use of reinforcement learning to construct optimal rosters, the identification, and potential design, of the most effective set plays and tactical systems, the modelling of tracking and event data through graph networks, and the integration of emerging technologies. These perspectives are extensively discussed in the dedicated section. We conclude that while Machine Learning brings tremendous opportunities to improve performance in team sports, interdisciplinary collaboration is required to fully unveil and achieve its potential.