Start with one offensive scheme. Watch how the defense is predicted to react, then compare that response with the recorded defense and a team of your choice.
Each category uses one recorded offensive sequence. The offense stays identical across all three courts, so you can focus on the defensive response.
All five defenders move together. A shared AI model predicts each defender’s movement separately from the same court snapshot, using their movement history, the offense, the ball, the basket, and the positions of the other defenders. It does not predict all ten defensive coordinates in a single output.
After the initial recorded defensive history, the model advances the defense in 0.4-second steps. Its predicted positions become the starting point for the next prediction, while the offense follows the recorded play.
Pooled defense: the model’s learned response across teams.
Recorded defense: how the original opponent actually played this possession.
Your chosen team: the model’s predicted response using that team’s input.
The ghosts show predicted reactions—not recorded possessions for the selected team.
With limited data per team, team predictions tend to stay close to the pooled defense. The model blends each team’s response with the pooled prediction, giving teams with fewer examples less influence.
A compact multilayer perceptron (MLP) learns defender movement from tracking data. Its architecture is 55 → 64 → 32 → 2, with 5,730 parameters.