AI-Driven Football Strategy: Introducing TacticAI, an AI Assistant for Football Coaches

TacticAI, an innovative AI football tactics assistant developed in collaboration with experts from Liverpool FC, employing predictive and generative components to analyze corner kicks.

Apr 10, 2024

Introduction

The paper starts by highlighting the significance of data-driven insights in modern professional football leagues, where teams strive to capitalize on every opportunity to gain an advantage on the pitch. While player execution on the field is dynamic and influenced by various factors, set pieces, such as corner kicks, offer a strategic opportunity to exert control and influence the game’s outcome. Given the importance of corner kicks and their potential to modify game outcomes, the paper proposes TacticAI, an AI football assistant designed to support experts in set piece analysis, specifically focusing on corner kicks.

TacticAI leverages raw spatiotemporal player tracking data to efficiently represent corner kick tactics. It adopts a graph-based representation to model player relationships, which are deemed crucial in corner kick situations. By employing graph machine learning models, TacticAI generates high-dimensional latent player representations, enabling it to predict outcomes such as the player most likely to make first contact with the ball or the likelihood of a shot occurring.

“**Figure 1 | A “bird’s eye” overview of TacticAI.** **(A)**, how corner kick situations are converted to a graph representation. Each player is treated as a node in a graph, with node, edge and graph features extracted as detailed in the main text. Then, a graph neural network operates over this graph by performing message passing; each node’s representation is updated using the messages sent to it from its neighbouring nodes. **(B)**, how TacticAI processes a given corner kick. To ensure that TacticAI’s answers are robust in the face of horizontal or vertical reflections, all possible combinations of reflections are applied to the input corner, and these four views are then fed to the core TacticAI model, where they are able to interact with each other to compute the final player representations — each “internal blue arrow” corresponds to a single message passing layer from (A). Once player representations are computed, they can be used to predict the corner’s receiver, whether a shot has been taken, as well as assistive adjustments to player positions and velocities, which increase or decrease the probability of a shot being taken.

A unique aspect of TacticAI is its use of geometric deep learning to produce player representations that respect symmetries of the football pitch. This ensures that representations remain consistent under transformations like horizontal or vertical reflections, mitigating the need to learn these symmetries directly from data. This is particularly valuable given the limited availability of high-quality tracking data.

TacticAI offers functionalities beyond prediction, serving as a retrieval system to mine similar corner kick situations and as a generative recommendation system, suggesting adjustments to player positions and velocities to optimize shot probability. Through experiments conducted with domain expert coaches and analysts from Liverpool FC, TacticAI demonstrates its utility by providing realistic and accurate tactical suggestions, thereby validating its effectiveness in supporting tactical decision-making in football.

Results and Analysis

The following section outlines the design of TacticAI with three distinct components: receiver prediction, shot prediction, and tactic recommendation through guided generation. These components are benchmarked tasks for quantitatively evaluating TacticAI’s performance in corner kick analysis. By integrating predictive and generative components, TacticAI enables coaches to explore alternative player setups for corner kick routines and assess potential outcomes.

The section proceeds with a description of the quantitative analysis conducted to evaluate TacticAI’s predictive accuracy in predicting corner kick receivers and shot situations on held-out test corners. The results indicate that TacticAI’s predictive components perform well and that the proposed player adjustments align closely with ground-truth situations.

However, while quantitative analysis provides valuable insights, it may not directly reflect TacticAI’s utility in real-world scenarios. To address this, a comprehensive case study in collaboration with Liverpool FC is conducted. Human expert raters from the club evaluate the utility of TacticAI’s predictions and player adjustments. The section promises to provide detailed results and analysis from this case study to offer a more nuanced understanding of TacticAI’s effectiveness in supporting tactical decision-making in football.

Furthermore, the section briefly outlines the minimal level of description necessary to comprehend TacticAI’s components for the evaluation. Detailed explanations of TacticAI’s components are deferred to the Methods section, ensuring clarity and transparency in the evaluation process.

Benchmarking TacticAI

In the process of benchmarking TacticAI, the evaluation focuses on its three components using a dataset of corner kicks from the 2020–2021 Premier League seasons. The dataset comprises 7,176 corner kicks, randomly split into training (80%) and test (20%) sets. TacticAI operates on graphs, with each corner kick situation represented as a graph where each node represents a player. The features associated with each node encode player movements (velocities and positions) and simple profiles (heights and weights) at the time of the corner kick. No information regarding ball movement is included in the graph representation.

**“Table 1** | Summary of the features used in the corresponding tasks.”

The graphs are fully connected, meaning that there is an edge connecting every pair of players, with each edge encoding a binary feature indicating whether the two players are on opposing teams. For each task (receiver prediction, shot prediction, tactic recommendation), the relevant dataset of node, edge, and graph features along with corresponding labels are generated.

**“Table 2** | Summary of the details of the features used to construct graphs.”

During training, the components are trained separately with their respective corner kick graphs. Notably, only a minimal set of features is used to construct the corner kick graphs, without explicitly encoding ball movements or distances between players. A consistent training-test split is maintained across all benchmark tasks to facilitate benchmarking of both individual components and their interactions, providing a comprehensive evaluation of TacticAI’s performance.

Accurate receiver and shot prediction through geometric deep learning

TacticAI’s predictive models aim to accurately forecast the receiver among the 22 on-pitch players and predict shot events following corner kicks. For receiver prediction, TacticAI employs a deep graph attention network leveraging geometric deep learning techniques, particularly 𝐷2 group convolutions. This model achieved a top-3 test accuracy of 0.782 ± 0.039, demonstrating significant improvement over other architectures. An ablation study confirmed the effectiveness of the chosen base architecture and group convolutions for receiver prediction, despite the inherent challenges of the task due to factors like player fatigue and ball trajectory. The high predictive power of TacticAI’s receiver prediction model was further validated through a case study involving human raters.

In contrast, directly predicting shot events proved challenging, yielding a test F1 score of 0.52±0.03. However, leveraging the receiver prediction model, TacticAI decomposed the shot probability into two components: the probability of a receiver and the conditional shot probability given the receiver. This two-phased approach significantly improved shot prediction, achieving a test F1 score of 0.64 ± 0.02. By incorporating receiver predictions, TacticAI captured more signal compared to unconditional shot predictors, despite the inherent uncertainties associated with predicting shot events.

**“Figure 2 | Corner kicks represented in the latent space shaped by TacticAI.** We visualise the latent representations of attacking and defending teams in 1024 corner kicks using 𝑡-SNE. A latent team embedding in one corner kick sample is the mean of the latent player representations on the same attacking ((A), (B) and (C) or defending (D) team. Given the reference corner kick sample (A), we retrieve another corner kick sample (B) with respect to the closest distance of their representations in the latent space. We observe that (A) and (B) are both out-swing corner kicks and share similar patterns of their attacking tactics, which are highlighted with rectangles having the same colours, although they bear differences with respect to the absolute positions and velocities of the players. All the while, the latent representation of an in-swing attack (C)is distant from both (A) and (B) in the latent space. The red arrows are only used to demonstrate the difference between in- and out-swing corner kicks, not the actual ball trajectories.”

Furthermore, TacticAI learned generalizable representations of corner kick data solely through predicting receivers, allowing it to cluster similar team setups in its latent space. This suggests that TacticAI can serve as a valuable corner kick retrieval system, which will be further evaluated in the case study section. Overall, TacticAI’s predictive models demonstrate strong performance in accurately forecasting receivers and predicting shot events following corner kicks, providing valuable insights for tactical analysis and decision-making in football.

Controlled tactic refinement using class-conditional generative models

The integration of class-conditional generative models into TacticAI enables the refinement of tactics to enhance or diminish the likelihood of specific outcomes following corner kicks. The goal is to adjust player movements, including positions and velocities, to optimize the probability of a shot event based on the initial corner setup. The refinement process involves generating adjustments for one team while keeping the other team fixed, simplifying the tactic refinement task. The generative model is trained using an auto-encoding objective, where the ground-truth shot outcome is fed as an additional feature, allowing the model to learn to reconstruct a probability distribution of player coordinates.

This generative model operates independently of TacticAI’s predictive systems, sharing the encoder architecture but utilizing different decoders. During inference, desired shot outcomes can be fed into the model to sample new player positions and velocities, enabling flexible downstream use for coaches to optimize corner kick setups according to specific outcomes of interest.

“**Figure 3 | Example of refining a corner kick tactic with TacticAI.** TacticAI makes it possible for human coaches to redesign corner kick tactics in ways that help maximise the probability of a positive outcome for either the attacking or the defending team by identifying key players, as well as by providing temporally coordinated tactic recommendations which take all players into consideration. As demonstrated in this example, for a corner kick in which there was a shot attempt in reality, TacticAI can generate a tactically-adjusted setting in which the shot probability has been reduced, by adjusting the positioning of the defenders. The suggested defender positions result in reduced receiver probability for attacking players 2–5 (see bottom row), while the receiver probability of Attacker 1, who is distant from the goalpost, has been increased. The model is capable of generating multiple such scenarios. Coaches can inspect the different options visually and additionally consult TacticAI’s quantitative analysis of the presented tactics.”

Quantitative evaluation of the generated adjustments demonstrates their realism, as they are indistinguishable from real corner kicks based on classifier analysis. Additionally, leveraging TacticAI’s shot predictor indicates the saliency of proposed adjustments. Analysis of corner kick samples with threatening shots shows that defensive refinements significantly decrease shot probability, while attacking team refinements increase it. Human raters also assess the utility of TacticAI’s proposed adjustments in a case study, providing further insight into their effectiveness and practical applicability.

Case study with expert raters

The case study conducted with expert raters from Liverpool FC provides a comprehensive evaluation of TacticAI’s practical utility in football tactic assistance. The study encompasses four tasks aimed at assessing various aspects of TacticAI’s components, including the realism of generated adjustments, the plausibility of receiver predictions, the effectiveness of corner kick retrieval, and the usefulness of recommended adjustments.

In evaluating the realism of adjusted corner kicks and the plausibility of receiver predictions, raters were tasked with distinguishing between real and TacticAI-generated samples and identifying likely receivers in corner kick scenarios. Results indicate that raters had difficulty distinguishing between real and generated samples, suggesting that TacticAI’s adjustments closely resemble real corners. Furthermore, TacticAI’s receiver predictions aligned closely with human raters’ assessments, demonstrating the model’s accuracy in predicting receivers.

Analysis of the variability in receiver prediction ratings among raters suggests that different raters may focus on different features when evaluating likely receivers. However, TacticAI’s high performance in predicting receivers across different raters indicates that it effectively captures salient patterns of corner kick strategies.

“**Figure 4 | Statistical analysis for the case study tasks.** In task 1, we tested the statistical difference between the real corner kick samples and the synthetic ones generated by TacticAI from two aspects. (A) the distributions of the ratings of the two types of samples, and (B) the distributions of the top-3 accuracy of receiver prediction using those samples. No statistical difference was observed in either cases ((A)(𝑧 = 0.34, 𝑝 > 0.05), and (B)(𝑧 = 0.968, 𝑝 > 0.05)). On the other hand, for task 2 (receiver prediction), we observed a statistically significant difference between the ratings of different raters, with three clear clusters emerging (C). Specifically, Raters A and E had similar ratings (𝑧 = 1.41, 𝑝 > 0.05), and Raters B and D also rated in similar ways (𝑧 = 2.53, 𝑝 > 0.05), while Rater C responded differently from all other raters. This suggests a good level of variety of the human raters with respect to their perceptions of corner kicks. In task 3 of identifying similar corners retrieved in terms of salient strategic setups, there was no significant difference among the distributions of the ratings by different raters (D), suggesting a high level of agreement on the usefulness of TacticAI’s capability of retrieving similar corners (𝐹1,4 = 1.01, 𝑝 > 0.1). Finally, in task 4, we compared the ratings of the strategic refinements by the human raters (E), and the raters also agreed on the general effectiveness of the refinements recommended by TacticAI (𝐹1,4 = 0.45, 𝑝 > 0.05).”

The study also evaluates TacticAI’s ability to retrieve similar corners, comparing its performance with a feature-based baseline. Results show that TacticAI significantly outperforms the baseline, indicating its effectiveness in extracting salient features from corners and uncovering opposing team tactics.

Finally, human raters assess the practical utility of TacticAI’s recommended adjustments by rating their effectiveness in improving tactics. Statistical analysis confirms that TacticAI’s recommendations are generally constructive and useful, with high inter-rater agreement indicating broad recognition of their practical usefulness.

“**Figure 5 | Examples of the tactic refinements recommended by TacticAI**. These examples are selected from our case study with human experts, to illustrate the breadth of tactical adjustments that TacticAI suggests to teams defending a corner. The density of the yellow circles coincides with the number of times that the corresponding change is recognised as constructive by the human experts. Instead of optimising the movement of one specific player, TacticAI can recommend the improvements for multiple players in one generation step through suggesting better positions to block the opposing players, or better orientations to track them more efficiently. Some specific comments from expert raters follow. In (A), according to raters, TacticAI suggests more favourable positions for several defenders, and improved tracking runs for several others — further, the goalkeeper is positioned more deeply, which is also beneficial. In (B), TacticAI suggests that the defenders furthest away from the corner make improved covering runs, which was unanimously deemed useful, with several other defenders also positioned more favourably. In (C), TacticAI recommends improved covering runs for a central group of defenders in the penalty box, which was unanimously considered salient by our raters. And in (D), TacticAI suggests substantially better tracking runs for two central defenders, along with a better positioning for two other defenders in the goal area.”

Overall, the case study highlights TacticAI’s strong components for prediction, retrieval, and tactical adjustments in corner kick scenarios, demonstrating its potential as a valuable tool for football tactic analysis and optimization. Examples presented in the study illustrate the salient recommendations provided by TacticAI, further emphasizing its practical relevance in football strategy development.

Discussion and Conclusions

The discussion and conclusions of the research paper highlight the efficacy and potential of TacticAI, an AI assistant for football tactics, as demonstrated through a comprehensive case study with expert human raters from Liverpool FC. The study confirms three main findings: TacticAI accurately predicts the first receiver after a corner kick and the probability of a shot resulting from the corner, produces plausible tactical variations that improve outcomes, and provides a powerful means to retrieve similar set piece tactics.

Unlike previous approaches that focus on individual aspects such as pass prediction, shot prediction, or corner kick tactical classification, this work uniquely combines and evaluates predictive and generative modeling of corner kicks for tactic development. The method leverages geometric deep learning to efficiently incorporate various symmetries of the football pitch, minimizing the need for intricate feature engineering. However, the methodology requires high-quality tracking and event data, which are currently limited to top leagues, and does not explicitly model exogenous uncertainty, which could provide valuable context for football analysts.

While the empirical study primarily focuses on corner kicks in association football, the approach can readily generalize to other set pieces and team sports with suspended play situations. Moreover, the learned representations and overall framing of TacticAI pave the way for future research to integrate a natural language interface, enabling domain-grounded conversations with the assistant and facilitating interactive tactical suggestions.

In conclusion, TacticAI represents a significant advancement in AI assistance for football tactics, laying the groundwork for the next-generation AI assistant in the field. Its combination of predictive and generative modeling, along with its potential for natural language interaction, holds promise for enhancing strategic decision-making in football and possibly other team sports.

Be a Team Player — Pass It On!

Wang, Z., Veličković, P., Hennes, D., Tomašev, N., Prince, L., Kaisers, M., … & Tuyls, K. (2024). TacticAI: an AI assistant for football tactics. Nature Communications, 15(1), 1–13. https://doi.org/10.1038/s41467-024-45965-x