The evolution of sports analytics facilitated by AI research
Developing test environments to take AI research from the lab and into the real world presents a major challenge. AI has long been connected with sports, and games, and as a result – it’s not shocking that sports is indicative of an immense opportunity, providing analysts a testbed in which AI-based systems can help human beings in complicated, real-time decision making in a multiagent environment with various dynamic and interacting persons.
The rapid increase of sports data mining means we are in the middle of a considerably critical age for sports analytics. Sports data is available at increasing quantities and granularity, evolving from the days of aggregate high-level stats and sabermetrics to more sophisticated data like event stream data, for instance, annotated passes or shots, high-fidelity player positional data, and physical sensors.
But the domain of sports analytics has only of late begun to leverage artificial intelligence and machine learning for both comprehending and advising human decision-makers in sporting. In a latest research paper produced in collaboration with Liverpool FC, the famous Premier League club in JAIR, the future landscape of sports analytics has been envisioned leveraging a combo of statistical learning, video understanding, and game theory. Specifically, soccer is a good microcosm for the study of AI research, providing advantages over the long-term to decision-makers in sports through automated video assistant coaches (AVAC) systems.
Soccer – an immense avenue for AI
Contrasted with various other sports, soccer has been somewhat late to the party with beginning to systematically gather large sets of data for scientific analytics reasons intending to evolve the game quality of teams. This is due to various reasons, with the most noteworthy being that there are a lot less manageable settings of the beautiful game in contrast with other sports (big outdoor pitch, dynamic gameplay, etc.) and also the dominating credo to depend primarily on human experts with track records and experience in professional soccer.
Arrigo Sacchi, a famous Italian soccer coach and manager who wasn’t a professional player during his career, reacted to critique over his lacking in-game experience with his famous quotation when appointed the coach of AC Milan 34 years ago: “I never realized that to be a jockey, you had to become a horse first.”
Soccer analytics puts forth challenges that are apt for a broad array of AI strategies, stemming from the confluence of three domains: statistical learning, computer vision, and game theory. While these domains are on their own, useful for soccer analytics, their advantages become particularly tangible when they are taken together: players are required to make decisions in sequence, in the presence of other players – both in their own team and the opponent’s – and game theory, a theory concerning decision making in such scenarios involving interaction, increases in relevance. Additionally, tactical solutions to specific in-game scenarios can be learned on the basis of in-game and particular players representations, which makes statistical learning of increasing relevance. Lastly, players can be tracked and game situations can be identified automatically from broadly-available audio/video inputs.
The AVAC framework that was visualized is located within the microcosm that is at the confluence of these three research domains. The research into this promising field, will not just provide a roadmap with regards to scientific and engineering challenges that can be managed in the time to come, but also provides new and original outcomes at the crossroads of game theoretic analyses, statistical learning, and computer vision, to demonstrate what this promising sphere has to provide to soccer.
How can AI assist soccer?
Game theory exerts a critical influence in the research of sports, facilitating theoretical grounding of agent’s behavioral techniques. With regards to soccer, many of its situations can consistently be modelled as zero-sum games, which have been researched comprehensively ever since the formulation of game theory. For instance, the penalty kick in soccer can be modelled as a scenario of a two-player asymmetric game, where the penalty taker’s techniques may be conveniently classified as right, center, or left kicks. To analyze this issue, game-theoretic analyses is augmented in the penalty kick situation with player vectors, which summarizes the play strategies of individual soccer players. With these portrayals of individual players, we have the capacity to classify kickers with play styles that resemble each other, and then execute game theoretic analyses at the group level. The findings demonstrate that the detected kicking strategies of various groups are statistically unique.
For instance, we identify that a group has a preference to take shots to the left corner of the goal, while another group has a preference to target both corners more or less the same. These insights can assist goalkeepers vary their defending techniques during play against varying types of players. Adding on to this game-theoretic perspective, one can take up the durative quality of soccer through analysis of it in the form of temporally-extended games, leverage this to inform strategies to individual players, or go one step further in optimizing the cumulative team strategies.
With regards to statistical learning, representation learning is yet to be completely leveraged within sports analytics, which would facilitate informative summaries of the disposition of individual players and soccer teams. Furthermore, the interaction between statistical learning and game theory would spur progression in sports analytics even more. In the aforementioned penalty kick situation, for example, augmentation of the analyses with statistics particular to players (Player Vectors) imparted in-depth insights into how several variants of players act or go about decision-making concerning their actions in the penalty kick situation.
Another instance of this, one can analyze ‘ghosting’ which is in reference to a specific data driven analysis of how players should have behaved in retrospect in sports analytics (which is connected to the concept of regret in game theory and online learning.)
The ghosting framework indicates alternate player trajectories for a particular play, for example, on the basis of the league aggregates or a particular team. Forecasted trajectories are typically conceptualized as a translucent layer over the original play, therefore the term ghosting. Generative trajectory forecasting frameworks enable us to obtain insights through analysis of critical scenarios of a game and how they might have occurred differently.
These frameworks also hold potential in forecasting the implications of tactical modifications, a critical player’s injuries, or of substitutes on the own team’s performance combined with the opponent team’s reaction to these modifications.
Lastly, computer vision was considered to be one of the most powerful avenues for progressing the limits of state-of-the-art sports analytics research. By identifying events directly from video, a subject that has been well-researched in the computer vision society, the possible scope of application is massive. By connecting events with specific frames, videos are searchable and more relevant. Soccer video, provides a fascinating field for computer vision. The huge amount of soccer videos fulfills a prerequisite for sophisticated AI strategies.
While every soccer video is unique, the settings do not demonstrate great variance, which renders the activity ideal for honing AI algorithms. 3rd party providers also provide manually-labelled event information that can be leveraged to train video models which are time intensive to produce, so both supervised and unsupervised algorithms can be leveraged for soccer event detection.
The leveraging of sophisticated AI strategies in soccer has the possibility to incite a revolution in the game across various axes, for players, decision-makers, broadcasters, and fans alike. These advancements will also be critical as they also hold the possibility to further democratize the sport itself, (for instance, over being dependent on gut instinct from in-person experts and scouts, one may leverage strategies like computer vision to quantify skills of players from under-represented areas, like those from lower-tier leagues.)
The generation of increasingly sophisticated AI strategies provided by the soccer microcosm might be relevant to wider domains.