Offline multi-agent reinforcement learning for evaluating and optimizing football attacking strategies against low-block defences
Rattachement africain : cn. Niveau de preuve : code pays fourni par la source.
Le résumé fourni par la source
Applying reinforcement learning to real-world multi-agent systems remains challenging due to coordination complexity, partial observability, and the constraints of learning from static datasets. Football serves as a premier and fascinating testbed for these issues, as exemplified by the ‘low block’, a representative and research-worthy scenario that demands sophisticated multi-agent coordination. In this paper, we formulate the low-block attacking play in football as a multi-agent offline reinforcement learning problem. By using large-scale event and tracking data, we model each attacking outfield player as a heterogeneous agent and construct synchronized spatiotemporal state representations that preserve the cooperation between players. To bridge data-driven learning and football domain knowledge, we introduce an interpretable abstraction of off-ball movement strategies, validated with professional coaches and analysts, ensuring alignment with human tactical reasoning. Building on these representations, we propose a transformer-based multi-agent Q-network with temporal graph embeddings to jointly evaluate both on-ball and off-ball actions. Transcending the inherent conservatism of human play, which often favours low-risk actions, our model recommends proactive, space-creating, and coordinated behaviours. These strategies remain tactically sound while introducing creative attacking solutions that are typically absent in real-world matches. Empirical case studies demonstrate that the proposed framework generates tactically coherent yet more creative attacking strategies, offering actionable decision support for coaches and advancing offline multi-agent reinforcement learning in complex real-world environments.
Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.
Le contrôle bibliographique ouvert
DOI retrouvé dans Crossref DOI retrouvé ; titre concordant.
- Titre Crossref
- Offline multi-agent reinforcement learning for evaluating and optimizing football attacking strategies against low-block defences
- Date Crossref
- 01/04/2026
- Éditeur
- Elsevier BV
- Type
- journal-article
Ce recoupement confirme des métadonnées liées au DOI. Il ne confirme ni la méthode ni les conclusions de l’étude, et il ne compte pas comme une seconde source scientifique indépendante.
Où se fait cette recherche
-
Shandong Institute of Automation pays non établi dans la noticeStructure de recherche
-
Institute of Automation pays non établi dans la noticeStructure de recherche
-
University of Chinese Academy of Sciences pays non établi dans la noticeUniversité ou école supérieure
-
Tianjin University pays non établi dans la noticeUniversité ou école supérieure
-
Beijing Academy of Artificial Intelligence pays non établi dans la noticeInstitution
-
The Key Laboratory of Cognition and Decision Intelligence for Complex Systems pays non établi dans la noticeStructure de recherche
-
School of Artificial Intelligence pays non établi dans la noticeUniversité ou école supérieure
-
The Nanjing Artificial Intelligence Research of IA pays non établi dans la noticeInstitution
-
Association de la Jeunesse Auxerroise pays non établi dans la noticeOrganisation à but non lucratif
Shandong Institute of Automation, Institute of Automation et University of Chinese Academy of Sciences, avec 6 autres affiliations.
Une affiliation ne permet pas de déduire la nationalité d’un auteur.