Google DeepMind, Alphabet's artificial intelligence subsidiary, has been working with the UK’s Liverpool FC football club for several years on the development of an artificial intelligence system capable of analysing and advising coaches on game tactics. It has now developed a specialist AI system for optimising corner kicks.

DeepMind recently presented TacticAI, an artificial intelligence system capable of providing practical information when taking a corner kick in a soccer match. It combines predictive and generative models to analyse what has happened in previous matches and make adjustments for even more effective shots. It allows coaches to create alternative player configurations for each routine of interest, then directly evaluate the possible outcomes of these choices.

In this way, the coach can find out who is most likely to receive the ball and whether they will have a clear scoring opportunity, as well as whether similar tactics have worked in the past. It's worth noting that data on corners is relatively limited, since an average of just 10 corner kicks are played per match in the English Premier League.

In concrete terms, DeepMind's tool offers a graphical representation of the action, in which each player is meticulously placed on the pitch. A graph neural network then calculates and adjusts their position to optimize goal chances, taking into account the placement of opposing defenders. The aim of these predictive models is to help the coach develop more effective play patterns for future corner kicks.

The idea now is to extend this technology to other sequences of play. For penalty kicks, for example, the algorithms will take into account the individual statistics of the players, i.e. the way they take kicks, their power, their success rate, etc. This will then be compared to data for the goalkeeper facing them.

DeepMind believes that coaches will be using virtual assistants of this kind within the next few years, whether to prepare for an upcoming match or to make half-time adjustments.

What's more, DeepMind recently succeeded in getting small humanoid robots to play soccer, one-on-one. This achievement uses deep reinforcement learning to teach bipedal robots a simplified version of the sport.

DeepMind is already using this learning technique in other fields, such as for board games like chess and Go. Currently, these robots can handle the ball, shoot at the goal and even get up after coming into contact with the other player, although they might not yet be ready for the RoboCup league.

The firm's work training these relatively agile robots is outlined in a paper published in Science Robotics.

 - AFP Relaxnews