Decision Making in American Football under State Uncertainty by Stochastic Inverse Reinforcement Leaning

Abstract

Reinforcement Learning (RL) techniques are often used to analyze and evaluate the strategies of virtual game to maximize a well-defined preset reward. However, in realistic sports game such as American football the reward functions are hardly definable. Meanwhile, how many yards are gainable on the next offence in real American football is also usually uncertain during strategy planning. In order to tackle these issues, we propose a stochastic inverse reinforcement leaning (IRL) algorithm. The expert data for IRL are built by using the American football 2017 season event data in National Football League (NFL). The stochastic state transition distribution is extracted from the same dataset. A mixture density network is used to learn the probabilistic distribution. At the last, simulation results from the maximum entropy IRL are compared with the ones from mathematical two-stage stochastic optimization.