Title: Markov decision processes in sports analytics
Authors: Oliver Schulte - Simon Fraser University (Canada) [presenting]
Abstract: Markov decision processes are a fundamental framework for optimizing sequential decisions. We describe how they can be applied in sports analytics, to provide a powerful statistical analysis of dynamic sports data. The main idea is to build a Markov decision process model of a sport and estimate a value function for it using reinforcement learning. Large-scale models have been built for several sports, including hockey, soccer, golf, basketball, and American football. The focus is on player evaluation, a fundamental problem of sports analytics. We will describe a neural net model of a value function trained on over 3M play-by-play events in the National Hockey League, the leading ice hockey league. We give several natural definitions of player performance derived from the value function, and prove their equivalence. Empirical evaluation shows that the resulting player ranking is consistent throughout a play season, and correlates highly with standard success measures and future salary.