Amazon Web Services has unveiled a machine learning model that predicts in real time how likely a National Hockey League (NHL) player is to win a faceoff. It takes into account the context of the game, the characteristics of the players and data from hundreds of thousands of face-offs over more than a decade.
Predictive analytics in professional sports, including hockey, not only evaluates the actions of a particular player, but also predicts their future behavior. Machine learning of this level requires a lot of data. Statistics of goals, shots, deletions, and so on is added to the standard metrics, for example, tracking movement on the site through the analysis of the coordinates of the puck and hockey players.
This technology allows you to calculate the optimal combination of players on the court and the best substitution strategy. In 2018, the Russian company ICEBERG used it to select the optimal line-up for the Olympic team for performances in the Korean Pyeongchang, and the NHL has been using it for several years during its matches. Sensors in the puck and on the players transmit a signal 200 times per second, tracking the puck with centimeter accuracy, as well as the direction of serves, shots and other parameters, says N+1.
Amazon Web Services on Tuesday introduced the Face-off Probability machine learning model. It will allow real-time prediction of how likely an NHL player will be to win a face-off – putting the puck into play at the start of each period or after it has stopped.
Even before the puck is dropped, Face-off Probability determines where it will occur on the field, who will fight in it, and what is the probability of victory for each of the players. The model is based on data from sensors located in the puck and on the players, as well as NHL statistics on puck face-offs for more than ten years. It includes puck put-in statistics by player during home and away games, player characteristics such as height, weight and dominant hand, and game context. The latter includes, for example, in which part of the field the throw-in took place, how long it lasted, what was the score.
Model predictions change depending on the situation in the game, notes NIX Solutions. If a player is sent off for a foul during a faceoff, it updates its prediction based on puck tracking data and players with a delay of less than a second. Face-off Probability can also determine whether a team’s chances of winning increase or decrease when a center forward is replaced on a face-off.
Statistics will be displayed simultaneously with the broadcast of the game. It will appear on Canada’s Sportsnet this week, and later in March on US ESPN and Turner.