Learning to Play

Published by admin on

Where was this when we were grad students absconding from labs to watch March Madness ?

Sicilia et al. have a charming paper ( https://arxiv.org/abs/1902.08081 ) on machine learning for basketball. They begin with an annotated database of player and ball positions at 1/25 second resolution for 650 NBA games in 2016-2017. They segment the action into time windows ending in a set of labelled outcomes. They use an LSTM net for learning. There are many twists, please read the paper for detail.

I attach Fig 6, to show what is possible. The system can tell expected probabilities of shots and expected points as plays develop. I fully expect the NBA to license this to broadcasters for live TV.

Here is an article about it:


Categories: DataScience