This simulation demonstrates that our model with macro-goals has learned long-term team formations for basketball offense and stochastic policies that imitate expert behavior.
We provide 10 preset initializations from the test set because our model performs better with a burn-in period. The 5 visible players are the only inputs to the model (ball and defense excluded).
See below for a free-form simulation.
If the simulation isn't loading, try refreshing the page.
In this free-form simulation, you can set the starting location for each player and their respective macro-goals.
Note that the model might make errors because a combination of starting locations and macro-goals can be vastly different from what was encountered during training. We also do not burn-in with ground-truth states.