Data-driven analysis of decision-making in football

Football, one of the most popular sports in the world, attracts millions of people’s attention with its blend of strategic depth, physical prowess, and moments of unpredictable brilliance. The unpredictability of football stems from the myriad contextual factors that determine the outcome of a game. These factors can be the intricate interactions of cooperation among teammates and the competitive dynamics between opposing sides. This makes the game perpetually thrilling to watch yet difficult to analyse simultaneously. This complexity and unpredictability pose significant challenges for teams aiming to optimise performance and strategy. In this context, football analytics emerges as an essential tool for processing complex information in football games and generating valuable insights that coaches and analysts can use to analyse the game more efficiently. Football analytics’ essential aspect is evaluating players’ performance and decision-making, which is often the key to winning the game. While existing methodologies have provided valuable insights into player performance and team strategies, there are several challenges in fully capturing the intricate dynamics of the game to solve problems in football analytics.

The most common way to make the evaluation is to use a metric that has been proven efficient for that context or to use subjective measurement techniques. Nonetheless, this can be a subjective and inconsistent process since these measures can hardly capture all the intricate dynamics in complex high-dimensional data, leading to incomprehension results. To solve this challenge, we build a novel deep generative model contextualised with on-pitch information like the movement of players to predict the teams’ interactions in space-control and propose a quantification method to translate the prediction to a computational benchmark, which is applied in the assessment of team’s and individual’s performance. Compared to benchmark models, the experimental results on the Premier League football dataset show that the model predicts the Pitch Control Map sequences more accurately with a higher Structural Similarity Index Measure (SSIM) score. Besides, it is shown that our model effectively captures the difference between adjacent frames from the ground truth in the prediction regarding the three patterns: pushing, backing, and staying, outperforming the benchmark models. Additionally, as shown in the qualitative results from example game scenarios, our model is objective and more precise than the existing metric in evaluating team and individual performance and decision-making. Our models will enable the coaches to analyse the performance of the whole team and each player more efficiently by providing them with the quantification of contribution to the attacking or defending regarding each decision and action taken in the game.

In football game analysis, pressure, one of the key factors that should be considered when assessing how the team or the player performs, is often manually annotated in video analysis, which is costly, time-consuming, and error-prone. To address this challenge, we propose two novel methods to quantify individual pressure and team pressure, respectively. i For individual pressure quantification, we introduce a 2D measurement of the pressure on off-ball players with contextual information like the positions and speed of players and create an incisive 3D amplifier to fine-tune the 2D pressure into a more precise metric to quantify the pressure on the ball carrier. For team pressure, we build a contextualised Graph Neural Network with the Player Pressure Map (PPM) as input, a novel representation of the game scene with the embedding of individual pressure vectors, to predict the likelihood of the team losing control of the ball in the near future. The prediction is used to measure the pressure that the attacking team receives in a given game context. In the experiment, our Team Pressure Model (TPM) is proven to be not only more accurate in predicting the likelihood of possession being lost than the benchmark models, but also effective in quantifying team pressure according to the qualitative results from an example game scenario. Overall, the experiment results have shown our quantification methods are good at detecting the pressure on the pitch and accurate in computing the pressure at both individual and team levels. The proposed pressure quantification models can be used to enhance the performance evaluation with extra contextualisation, achieving more reliable results that the coaches can use to improve their team performance.

In the aforementioned approaches, the predictions or benchmarks cover just one or a few aspects of the game. For a specific game scenario, if there are 22 AI ghosting players behaving like players in real life, the 22 players’ actions can be used as a much more detailed benchmark to evaluate every facet of the game and answer most of the challenging questions in football analytics. To this end, we build a novel Multi-agent Reinforcement Learning (MARL) model to train the ghosting agents to play football. The experiment results show that our novel reward-shaping technique efficiently resolves the credit assignment issue of MARL in the football environment by outperforming the benchmark Value Decomposition Network (VDN) in the number of goals conceded per episode in all three game difficulties. The qualitative results from the testing game episodes show that our trained agents’ decision-making capabilities in the game are on par with those of real football players. The effectiveness of the proposed method in the 5-vs-5 game setting highlights its potential applicability in the more complex 11-vs-11 game scenario. The ghosting players will help answer the what-if questions in football game analysis by simulating the possible variations of the game-play and also be useful benchmarks to evaluate players’ performance in any aspect of the game.

Overall, in this thesis, we address the three challenges in the current football analytics technologies: lacking benchmarks in performance evaluation, lacking pressure quantification that is essential to precise performance assessment, lacking ghosting players who can be used to answer what-if questions in the game analysis and to be benchmarks in performance evaluation. We contribute to the field by solving these challenges and introducing novel Machine Learning techniques, which can be used to evaluate on-pitch decisions more accurately with contextualised modelling.