Computational Approaches to Strategy and Tactics in Sports

Competition between athletes is central to sport. Athletes need to determine how and when to invest their limited energy resources to optimise and self-regulate their pace depending on their physiological and biomechanical capacity as well as environmental factors, such as the presence and behaviour of another athlete. Remote technology can be used to explore mechanisms involved in exercise regulation and competition, for example via the use of avatars, which can be a graphical representation of an athlete’s own performance or that of another athlete. I will overview a series of studies using avatar scenario’s to better understand decision-making and pacing in sport and competition.

In this paper, we develop a Markov chain representation of football. States consist of different locations on the pitch and set-pieces, and different actions, such as passes, shots, and tackles, move the chain between these states. Novel variables which describe the abilities of the opposing teams in different aspects of football over different zones of the pitch drive the transition probabilities. By simulating a match numerous times using different scenarios, we can identify the optimal choice of tactics: who should play and in what formation, or what style of football to implement. A case-study using a recent match between Everton and Chelsea demonstrates the usefulness of the model, as well as the detailed predictions one can obtain.

A remarkable development of data measurement and collection technology in baseball has provided us with opportunities to gather comprehensive performance data. Our research team conducted detailed analyses of batter and fielder performances in the official professional baseball games. Specifically, the kinematic characteristics of batted balls were analysed collectively to evaluate how fast, how high, and in which direction the batters should aim to hit the ball to increase the chance of making base hits and home runs. Additionally, we introduced a novel method for fielder performance evaluation, addressing the challenge of comparing abilities fairly due to asymmetrical distributions of batted ball characteristics across the baseball field. A machine learning algorithm was used to predict the probability of flyout from the kinematic characteristics of fly balls and compared the probability score for a systematically constructed set of fly balls, uniformly distributed across the field. We discovered that hitting towards the same side of the field increased the chance of a base hit and argued that the advantage was attributable to the large variance in both the direction and magnitude of deflection seen in the trajectories of the pulled fly balls. Based on these studies, using classical statistical methods as well as simulation with predictive modelling with vast amounts of batted ball data will help us conduct fair and rigorous comparisons of players’ performance and provide target values with a given aim in baseball. The future development of this methodology in baseball and its potential applications in other sports will be discussed to provide findings and implications in practical situations effectively.

In this talk, I will give a short intro one a new line of research in 1 vs 1 actions. To study them and to quantify action in football in general, one needs a valid success measure. Expected Thread is one of the suggested measures. However, models that circulate right now might have different short comings and are not validated properly. I like to give an overview on the different models in the literature and propose some standards and guidelines to validate them.

In fluid team sports such as soccer and basketball, analyzing team formation is one of the most intuitive ways to understand tactics from domain participants’ point of view. However, existing approaches either assume that the team formation is consistent throughout a match or assign formations frame-by-frame, which disagree with real situations. To tackle this issue, we propose a change-point detection framework named SoccerCPD that distinguishes tactically intended formation and role changes from temporary changes in soccer matches. We first assign roles to players frame-by-frame and perform two-step change-point detections: (1) formation change-point detection based on the sequence of role-adjacency matrices and (2) role change-point detection based on the sequence of role permutations. The evaluation of SoccerCPD using the ground truth annotated by domain experts shows that our method accurately detects the points of tactical changes and estimates the formation and role assignment per segment. Lastly, we introduce practical use-cases that domain participants can easily interpret and utilize.

I try to classify computational approaches to strategy and tactics according to their usefulness for sports practice. Basic research in informatics based on sports data, Messi-and-Ronaldo-look-good-in-my-data or I-found-a-new-performance-indicator typed studies may be criticized in this respect. In the narrow sense of support for coaches and teams (not in the wider sense of sports analytics), practical hints for training should be the ultimate aim. The analysis process in sports practice is presented and more concrete requirements for support are derived.

This talk introduces the game observation concept used by the German national teams in beach volleyball. First, the talk discusses the type of statements in match analysis from a conceptual perspective and explores questions that relevant for beach volleyball coaches. The methods section outlines the logical work steps of match analysis, including: i) querying game scenes according to a classifier; ii) the quantitative preliminary analysis supported by descriptive statistics and graphical reports; and iii) the main qualitative analysis based on video-recordings. It also shows how results are communicated to the German athletes by using a custom made presentation tool and how they use the data for anticipation training.

The purpose of this presentation is three-fold: i) measuring team tactical performance, ii) using small-sided games, and iii) highlighting applications to the football practice. Team tactical performance in football (soccer) refers to the cooperation of players within a team and the competition between teams. Positional data from tracking technology is used for team tactical performance metrics representing the positioning and dispersion of players on the pitch. Knowledge of match performance allows for developing and improving performance through training. Small-sided games (SSGs) are training formats representing the match (phases) and therefore a popular training drill. Usually, changing the pitch size and number of players in SSGs affects the individual performance (physical and technical) and team tactical performance compared to match performance, compromising the usefulness and representativeness of SSGs. In this presentation, I shared insights of optimising SSGs by using a similar relative pitch area (total pitch surface / number of players) as the match. This leads to similar team tactical performance in SSGs and the match. These insights can be applied to the football practice with examples of new designs of grassroots competitions, coach dashboards, and a range of games for training programs. Involving coaches in this process is crucial for successful applications.

This talk will present several challenges and opportunities within the area of team sports analytics and key research areas within multiagent systems (MAS). We specifically consider invasion games, where players invade the opposing team’s territory and can interact anywhere on a playing surface (ice hockey or soccer). We discuss how MAS is well-equipped to study invasion games and will benefit both MAS and sports analytics fields. We highlight topics along two axes: short-term strategy (coaching) and long-term team planning (management).

In the fast-evolving landscape of soccer analytics, leveraging the power of deep learning has become imperative for gaining a competitive edge. However, the opacity of these models often leaves coaches and players in the dark, hindering the practical application of insights. We need to delve into the significance of using interpretable models in soccer analytics, shedding light on the “why” behind every “what.”

Automatically labeling trajectories of multiple agents is key to behavioral analyses but usually requires a large amount of manual annotations. This also applies to the domain of team sport analyses. In this paper, we specifically show how pretraining transformer models improve the classification performance on tracking data from professional soccer. For this purpose, we propose a novel self-supervised masked autoencoder for multiagent trajectories to effectively learn from only a few labeled sequences. Our approach employs a masking scheme on the level of individual agent trajectories and makes novel use of a factorized transformer architecture for multiagent trajectory data. As a result, our model allows for a reconstruction of masked trajectory segments while being permutation equivariant with respect to the agent trajectories. In contrast to related work, our approach is conceptually much simpler, does not require handcrafted features and naturally allows for permutation invariance in downstream tasks.

The transformation into a data-driven culture presents significant challenges for sports organizations, notably the issue of data silos that impede the free flow of information necessary for innovation in sports data analytics. This talk aims to tackle these barriers by offering practical solutions and discussing their implications. Firstly, we will introduce a data-centric approach to break down internal data silos within sports clubs, ensuring optimized access to data for analytics. Secondly, for industry-wide data silos, we will present the blockchain-based protocol OCEAN, illustrated through a case study from Swiss ice hockey, to privacy conserving data sharing across different clubs. Furthermore, the talk will address the necessity of seamless integration of academic research and solutions into the sports industry’s existing technological frameworks. Strategies to facilitate this integration, enhancing the flow of innovative ideas from universities to the field, will be explored. Additionally, the growing complexity of sports information systems is identified as a challenge to innovation; we propose the adoption of microservices architecture as a scalable and flexible solution to this problem. Conclusively, the presentation will outline actionable steps that organizations can take to overcome these innovation barriers, paving the way for a more integrated, efficient, and innovative future in sports data analytics.

We conducted an experiment with the Brigham Young University Men’s Tennis Team to investigate the spatial element of serving strategies in tennis. Serve data, including precise spatial coordinates, were collected for 12 players, with known targets for each serve. Leveraging this data, we estimate player-specific optimal aim locations, accounting for factors such as first vs second serve, speed, and the distribution of their serves around the intended targets, termed “execution error”. Our experiment also provides insights on the interplay between conscious beliefs and on-court performance. Our preliminary results reveal apparent differences between players’ subconscious behavior and their explicit articulation of optimal aiming locations.

Data analytics in tennis is a fast growing topic of interest, often related to fan engagement at tournaments. Higher ranked players are also able to benefit from the statistics and data collected by these tournaments. However, the collection of spatiotemporal (and even event) data has historically been complicated and expensive, and so developing players at all levels still do not have access to these types of resources. Recent advances in computing power and AI algorithms have allowed for cheaper and more efficient data collection for individual tennis matches. This promises to expose more players to the idea of analytics, with the hope that it can level the playing field. But there are still challenges to consider when building for even more scalability and versatility. This talk will explore some of the issues that arise when trying to fully automate the data collection process of a tennis match. This includes many facets of the approach from court detection, player identification, bounce detection etc. and includes the edge cases that one needs to consider when aiming for fully automatic data collection.

A frequent impediment to applied causal analysis is the identification and quantification of confounding variables. With the advent of tracking data in sport, there is often a realistic chance of dealing with the confounding variable problem. In this presentation, we consider three questions involving soccer tactics that are each approached using causal methods (a) what is the benefit of crossing the ball? (b) what is the benefit of playing with pace? and (c) what is the benefit associated with throw-in decisions? The problems each have a common structure, and we provide a template for approaching such problems. The differences between the problems lie in the nature of the independent variables, the dependent variables and the confounding variables.

There has been an explosion of data collected about sports. Because such data is extremely rich and complicated, machine learning is increasingly being used to extract actionable insights from it. Typically, machine learning is used to build models and indicators that capture the skills, capabilities, and tendencies of athletes and teams. Such indicators and models are in turn used to inform decision-making at professional clubs. Unfortunately, how to evaluate the use of machine learning in the context of sports remains extremely challenging. On the one hand, it is necessary to evaluate the developed indicators themselves, where one is confronted by a lack of labels and small sample sizes. On the other hand, it is necessary to evaluate the models themselves, which is complicated by the noisy and non-stationary nature of sports data. The goal of this presentation is three-fold. First, we detail some aspects about how analytics are used within a club environment. Second, we discuss pitfalls and best practices for evaluating models learned from sports data. Third, we overview various ways to validate developed indicators, which requires assessing if they can provide value to the workflow of practitioners.

Strategy-optimization is a fundamental element of dynamic and complex team sports such as soccer, American football, and basketball. As the amount of data that is collected from matches in these sports has increased, so has the demand for data-driven decision-making support. If alternative strategies need to be balanced, a data-driven approach can uncover insights that are not available from qualitative analysis. This could tremendously aid teams in their match preparations. In this talk, I present a novel Markov model-based framework for soccer that allows reasoning about the specific strategies teams use in order to gain insights into the efficiency of each strategy. The framework consists of two components: (1) a learning component, which entails modeling a team’s offensive behavior by learning a Markov decision process (MDP) from event data that is collected from the team’s matches, and (2) a reasoning component, which involves a novel application of probabilistic model checking to reason about the efficacy of the learned strategies of each team. I will illustrate the framework on one use case, namely that it can be used to optimise a team’s defensive strategies when playing against a particular team.

Team sports tactics are about who (which player and/or player role) does what (in terms of legal actions, which includes moving around) where on the pitch, court or whatever the playing field is called, potentially modified by when (in the shot clock, game clock etc).I used my presentation to quickly touch on those different aspects (excluding time) by summarizing a number of works from the literature. The lines are not always clearly drawn: some papers mix what athletes do with where it happens or with who does it. The “what” part is arguably the most complex one and the papers I touched on all use what I’ve called the “vocabulary” of possible actions/movements to describe what teams and individual athletes do. Some do so rather explicitly, using topic models originally developed for text document analysis, others learn or predefine patterns that occur in certain situations or for certain teams. Some learn the vocabulary from the available data, others predefine it and “only” learn the “phrases” that are being formed. When it comes to the “who”, finally, network-based modeling proves to be very powerful and allows to explore “what if” scenarios that would occur if one put different line-ups in the field.

There was no final agreement on the definition of tactics and strategy after this break-out session and we also realised that the vocabulary is used differently across the world. For example, in North America the word tactics is barely used while in Europe the word tactics is more common. Most participants agreed that tactics have a shorter time-line than strategies. The outcome of the groups is as follows.

Group 1 defined tactics as a set of unconstrained actions and reactions that are anchored in a respective strategy and take the environment into account. In contrast, strategy is planned out and ranges on a longer term. This group also questioned whether it is possible to quantify the quality of a strategy.

Group 2 viewed strategy also as a long term, pre-planned idea of how a team aims to play. Tactics are decisions made on the field, where it makes a difference who takes the decision. While the strategy is decided by coaches, the tactics is rather a joint effort by players and coaches. The difference is also dependent on the actual sport and there may be fluent transitions but all strategies are pre-planned (like load management and player rotation) while this does not hold for all tactics.

Group 3 focused on athletics and skill action plans as tactical knowledge. The group also agreed that tactical decisions are very short term and considered a continuum that ranges from one player, via groups of players, and the team to a game and finally to the entire season and the philosophy of the coach. There are certainly action plans with various time scales, however, strategy can be viewed as long term tactics.

Group 4 also aimed to draw a line between tactics and strategy. The long term strategy was placed at a business level, however discussions went about where to draw the line for in-game decisions? The group agreed that strategy is decided on before the game while tactics are adjustments of the strategy during the game.

Event data or play-by-play data is a common type of data that is collected about many different team sports such as soccer, ice hockey, and rugby among others. This type of data records specific semantically meaningful events that occur during a match. For example, relevant events in soccer include passes, tackles, and shots. Each event is annotated with information such as the players involved and the location where the event took place. Such data forms the backbone of many different analysis tasks.

However, event data can be challenging to work with because it is collected by multiple different providers in each sport and each provider often records the data in a different format. This makes working with similar data from different providers tedious for practitioners because each data source usually has to be converted or mapped into a unified format. An even bigger challenge is that each provider may use different definitions for certain events, particularly since some events are inherently subjective.

Less critical, but also challenging in the domain of soccer is the comparison of different tracking data sources. While the format of the data-set itself is somehow restricted to x/y/z-coordinates of either the center of mass or different body-points, different vendors share the data also in different formats, and use different preprocessing algorithms.

Gabriel Anzer, Pascal Bauer and Joshua Smith are involved in a project called the Common Data Format (initiated by FIFA and the DFB) that will attempt to address this challenge for soccer. They presented the key ideas underlying the format, which strives for a touch-based model that incorporates both event and positional data. Moreover, they are committed to providing clear definitions of events that companies can use to collect data in the proposed format.

Most participants had experienced some of the aforementioned issues. From an academic and club perspective, there is a clear need for such a unification as academia and clubs cannot reproduce or repeat their experiments on other data formats and comparability of approaches and results is an issue. However, from a provider perspective there is an appreciation that changing definitions is difficult from an operational perspective and creates legacy issues for data that was collected differently. Moreover, the data format also has to adhere to the needs of other user, such as the broadcasting media, which makes reaching an agreement on a joint format difficult.

Communication between analysts and coaches can be difficult as there is a gap between the terms used by analysts and coaches, which can make it challenging for analysts to prepare what the coaches want. Martin Rumo calls this the “semantic gap”. To help bridge this gap, he advocates for a co-creation approach between the analyst and user/coach.

More generally, several challenges exist that hinder the adoption of a common terminology. First, Jan Van Haaren noted that sometimes no term exists and people simply operate by describing the concept. In these cases one may need to invent a term. Second, terminology is often specific to a club and may stay consistent even with changes to personnel. Third, cultural factors affect terminology. For example, players come from different places and players from South American may use different terminology than those from Europe. Sigrid Olthof noted that this an issue in academic work too as papers from the computational literature may use, e.g., artificial intelligence specific terms instead of ones used in sports science. Fourth, misalignment of goals can lead to different terminology. For example, people may build their career around specific terminology and hence are not incentivized to change. One initiative in football (soccer) to help mitigate this problem is FIFA’s common football language that will be used in coach analyst courses.⁵⁵5https://www.fifatrainingcentre.com/en/resources-tools/football-language/ As more people follow these training courses, it will help introduce a more standard vocabulary.

Beyond this, several pieces of advice that were offered included:

• $\mathrm{■}$

  It can be useful to provide a clip library to illustrate good / bad behavior.

• $\mathrm{■}$

  When making visualizations, avoid using colors associated with a rival.

• $\mathrm{■}$

  When discussing probabilities, it can be useful to explicitly state the chances of each open. For example, instead of just saying “There is an 80% chance that we reach the playoffs” also add “and there is a 20% chance that we do not reach the playoffs”.

• $\mathrm{■}$

  The Pysport website contains a list of visualization packages: https://opensource.pysport.org/?categories=Visualization

• $\mathrm{■}$

  The following article summarizes some good advice about visualizations: https://knowablemagazine.org/content/article/mind/2019/science-data-visualization

Currently most analysis is done on data from a few matches, or a few seasons. However, very interesting questions can be posed when looking at data that has been collected over many years. For example women’s soccer has gone through a relatively quick development and is considered to have a very different style than the men’s game. With longitudinal analysis it could be investigated if the women’s game is developing to get closer to the men’s game over time and if this development for example is going faster than how the men’s game developed over the last 20 years. Other questions that could be answered with longitudinal analysis are around variability across seasons, changes in coaching styles and finally talent development. Having longitudinal data from developmental players starting when they join the academy at a very young age until they are retired can answers questions about talent development programs but also about how to maximize a players value/time at the club.

It may be important when doing analysis to normalize data or exclude outliers because it does not help finding large trends in the data. However, if data on the context was collected, such as players injured or weather data new insights might be created from these outliers. Context also matters on a match level: For example, using event data alone allows track whereabouts of the ball, but not whether the ball possessing player is facing one, two, or even more opponents. The context of how many defenders the player faces influences where he/she will pass the ball to.