Basketball coaches often sketch plays on a whiteboard to help players get the ball through the net. A new AI model predicts how opponents would respond to these tactics.
What’s new: A team of researchers in Taiwan trained a conditional generative adversarial network on data from National Basketball Association games. They trained their network to show how players on the opposing team likely would move in response to human-drafted plays.
How it works: The researchers built a two-dimensional simulation of a half court complete with a three-point line and a net. A coach can draw motion paths for five players represented by dots, as well as ball movements including passes and shots. No dunking, however.
- Once a coach has drawn a play, a generator determines how the five defensive players would react.
- A discriminator evaluates these movements to make sure they match realistic gameplay.
- The model then displays the coach’s play and the defensive maneuvers.
Results: A cohort of NBA pros, basketball fans, and basketball non-fans evaluated the generated defenses for realism. While the non-pro fans and non-fans had a hard time spotting the computer’s defensive plays, the NBA pros could tell they were not designed by a human coach.
Behind the news: Stat SportVU has collected real time player motion data for the NBA since 2011. The system uses six cameras to collect each player’s position and track who has possession of the ball, 25 times per second. It uses machine learning to identify events like dribbles and passes, and play types like drives, isolations, and screens.
Why it matters: Pro sports is a high-stakes industry that has embraced technology to optimize performance. It’s conceivable that a neural network someday might generate AlphaGo-like winning tactics that no human had envisioned.
We’re thinking: This model isn’t a slam-dunk, given that the pros weren’t fooled. However, it appears to be sophisticated enough to help teach beginners how to think strategically off the court.