Publications

@phdthesis{soton509865,

title = {Models and algorithms to optimise team performance in football},

author = {Gregory Alan Everett},

url = {https://eprints.soton.ac.uk/509865/},

year  = {2026},

date = {2026-03-01},

publisher = {University of Southampton},

school = {University of Southampton},

abstract = {With the growing use of Artificial Intelligence (AI) and machine learning, there is increasing potential to model and optimise team behaviour across important domains such as disaster response, security, and team sports. These domains are inherently spatiotemporal, requiring models that capture both spatial and temporal dynamics. This thesis focuses on football, a complex, dynamic sport with rich spatiotemporal data and clear objectives, making it an ideal testbed for developing and validating team-based AI models. Moreover, football analytics is a rapidly expanding industry, with European clubs generating ?38 billion in revenue during the 2023/24 season, driving the demand for models that provide competitive advantages.ensuremath<br/ensuremath>ensuremath<br/ensuremath>This thesis proposes a number of novel methods that utilise spatiotemporal data to advance team prediction, analysis, and decision-making in football and other team-based domains. In particular, we introduce a spatiotemporal agent behaviour imputation model that reduces predictive error by 62% compared to baselines in limited observability settings, significantly improving the accessibility of off-ball football analytics through imputed tracking data. We also present a novel spatial teamwork model, combining Monte Carlo tree search (MCTS) and linear programming, that optimises agent decision-making in real-world football defence, reducing opponent threat by 24%. In addition, we develop new metrics, derived from a graph attention network (GAT), to assign credit to indirect agent contributions in team-based defence. The GAT model predicts football passes with a 6% reduction in loss compared to baselines, and we show how these new metrics can greatly improve off-ball football player evaluation. Finally, we propose a dynamic team formation and agent replacement model that accounts for agent fatigue and unavailability and optimises decision-making using a multi-step MCTS algorithm. Applied to football team selection and substitutions, this model improves long-term team performance by 1% and reduces first-team injuries by 15%. This thesis also highlights key remaining challenges for AI in football, including the use of richer player data (e.g., body pose) and greater use of explainable models to build trust in clubs.},

keywords = {applied machine learning, machine learning, multi-agent systems, sports analytics, team optimisation},

pubstate = {published},

tppubtype = {phdthesis}

}

@phdthesis{soton509865,

title = {Models and algorithms to optimise team performance in football},

author = {Gregory Alan Everett},

url = {https://eprints.soton.ac.uk/509865/},

year  = {2026},

date = {2026-03-01},

publisher = {University of Southampton},

school = {University of Southampton},

abstract = {With the growing use of Artificial Intelligence (AI) and machine learning, there is increasing potential to model and optimise team behaviour across important domains such as disaster response, security, and team sports. These domains are inherently spatiotemporal, requiring models that capture both spatial and temporal dynamics. This thesis focuses on football, a complex, dynamic sport with rich spatiotemporal data and clear objectives, making it an ideal testbed for developing and validating team-based AI models. Moreover, football analytics is a rapidly expanding industry, with European clubs generating ?38 billion in revenue during the 2023/24 season, driving the demand for models that provide competitive advantages.ensuremath<br/ensuremath>ensuremath<br/ensuremath>This thesis proposes a number of novel methods that utilise spatiotemporal data to advance team prediction, analysis, and decision-making in football and other team-based domains. In particular, we introduce a spatiotemporal agent behaviour imputation model that reduces predictive error by 62% compared to baselines in limited observability settings, significantly improving the accessibility of off-ball football analytics through imputed tracking data. We also present a novel spatial teamwork model, combining Monte Carlo tree search (MCTS) and linear programming, that optimises agent decision-making in real-world football defence, reducing opponent threat by 24%. In addition, we develop new metrics, derived from a graph attention network (GAT), to assign credit to indirect agent contributions in team-based defence. The GAT model predicts football passes with a 6% reduction in loss compared to baselines, and we show how these new metrics can greatly improve off-ball football player evaluation. Finally, we propose a dynamic team formation and agent replacement model that accounts for agent fatigue and unavailability and optimises decision-making using a multi-step MCTS algorithm. Applied to football team selection and substitutions, this model improves long-term team performance by 1% and reduces first-team injuries by 15%. This thesis also highlights key remaining challenges for AI in football, including the use of richer player data (e.g., body pose) and greater use of explainable models to build trust in clubs.},

keywords = {},

pubstate = {published},

tppubtype = {phdthesis}

}