This is a review of the deep imitation learning research conducted by Thomas Seidl, Aditya Cherukumudi, Andrew Hartnett, Peter Carr and Patrick Lucey.
Tracking data in sports has become a staple for teams and coaches as they can now study their upcoming opponents in order to determine the best strategy possible. However, coaches do not have access to this data during the game itself. Instead, coaches typically rely on their experience and instincts.
Play sketching might be a solution. In play sketching, coaches can sketch plays and instantly see how their opponent is likely to respond.
Player tracking data from the 2016-2017 NBA season was used. Games were divided into possessions. Each possession had ten two-dimensional trajectories (one for each player, referred to as ‘ghosts’) and one three-dimensional trajectory for the ball. Information such as game clock, shot clock, player fouls, and the number of seconds each player has played in the game to this point was included. The data consisted of 30,764 possessions.
When tested against an actual game, the positions of the players on the court and the ghosts on the screen were not identical but the expected outcomes were very similar indicating that the ghost behavior was similar to the behavior of the players on the court.
Inferred tracking data makes it possible to visualize a sketched play as an animation. In this format, the user gets immediate feedback about the design and timing of a play. Additionally, the animation makes it easy to communicate the intention of the set play to those without the required experience to interpret a sketch directly.
Because sketches can be generated directly from tracking data, a user can edit any offensive play that was run in any game. For example, players can erase the actual pass that took place, and visualize how the defense would have reacted if the ball had been passed to a different teammate. Similarly, the routes of the teammates can be modified to better spread the defense.
Up to this point insights could only be gained from player tracking data after the game was finished. This framework allows coaches access to the data for use in in-game decisions by combining ghosting with a digital sketching interface. This framework is highly intuitive, allowing anyone to draw a play and easily understand how a team is likely to defend against it. It is also very quick in responding to questions.
Analysts, coaches and fans can all use this tool to explore an endless variety of scenarios, looking for weaknesses and strengths, and determining if they can find a better play for that scenario.
How mature is your team’s analytics program? Take the Sports Analytics Maturity Assessment.
Learn about the Groundbreaking Sports Analytics Model coaches and sports analysts are talking about!
Learn all about Sports Analytics here.