Abstract:
|
With the rapid progress in machine learning (ML) and reinforcement learning (RL), there is a growing interest in applying these methods to sports. There are several reasons for optimism on this front: ML methods can deal well with the complicated feature spaces found in sports data, while RL methods are well-suited to characterizing and optimizing sequential decisions of the type that players make during games. However, ML and RL methods do not necessarily construct causal representations of the systems they model, which can make them ill-suited for problems such as team construction and strategy optimization that concern interventions. In this talk, I will review some ML and RL approaches that have been taken in sports, and suggest some strategies for aligning them with hypotheses about causal structure.
|