Evaluating Reinforcement Learning as a Technique for Optimizing and Automating NFL Play Calling

Abstract

Offensive coaches in the NFL are tasked with one of the sport’s most difficult jobs: making successful play calling decisions in real time. While play calling requires acute tactical nous and a thorough understanding of the game, some of the league’s brightest minds have turned to analytics in order to improve. We propose the use of reinforcement learning methods, namely off-policy approaches such as Q-learning, to augment and ultimately automate certain aspects of play calling. We design a Markov Decision Process to simulate offensive drives by sampling existing plays from the 2017 season and train optimal policies for deciding between running and passing. Among our findings is that, across the board, teams should be more aggressive than they currently are by adopting pass-heavy strategies. In addition, we find that considering the time remaining and score is important in order to achieve better offensive outcomes. Ultimately, we are confident that improved reinforcement learning methods can serve as a platform for generating better play calls on average and the approach’s flexibility can be tailored to fit the needs of specific personnel and coaching philosophies.