AI learned to generate tennis matches

Scientists from the United States have presented an AI that can generate tennis matches. Despite the weak graphics, the model can already be used to analyze sports matches or training.

Due to the coronavirus pandemic, Wimbledon, one of the most famous international tennis tournaments on the planet, was not held in 2020. Therefore, researchers at Stanford University decided to simulate this annual competition using artificial intelligence (AI).

The team trained the AI ​​using a database with annotated records. The cyclical nature of the matches helped them create a statistical model that predicts how stars such as Novak Djokovic and Roger Federer would play in certain situations.

The model takes into account the strategies and playstyle of each player. The AI ​​has been trained in thousands of tennis matches, so it knows which player likes to direct the ball towards the opponent’s weak side and takes into account the peculiarities of receiving the ball. According to the team that created the AI, it is this behavioral aspect that separates them from their counterparts, which can also simulate the game of tennis.

The system can create endless “what if” scenarios – it predicts what action will happen next. For example, it can simulate a match that cannot actually take place – when a tennis player is playing against an opponent of the opposite sex or against himself. Researchers can change where the ball lands during a particular rally – then the course of the entire match can change. In addition, the system allows you to control the position of the player during the serves, so the model can be used for training athletes.

Scientists note that the lack of their model in weak visual effects – while they look like clips from video games of the 90s. In the future, the researchers will improve the graphics, as well as add shadows to the players and the ball. However, they are confident that other aspects of the model are correct and can already be used to analyze sports matches or training.