Reinforcement learning for football player decision making analysis

Abstract

Traditionally, the performance of football players has been evaluated using statistics computed from actions such as goals and assists. However, recent advances in football analytics have yielded Possession Value Models (PVMs) that provide a more granular and objective method that can be used to analyse the decision making abilities of a player. Such models include Expected Threat (xT), Valuing Actions by Estimating Probabilities (VAEP) and On-the-Ball-Value (OBV). Nevertheless, these metrics typically only make use of the data directly related to an event, such as the position of the player with the ball, and the player who receives the ball. These PVMs do not account for the position of teammates and opponents when computing the value of the proposed action. We propose a novel metric called Decision Value (DV) which is computed using Deep Reinforcement Learning. The model is trained on both event and tracking data to allow the model to obtain an optimised decision policy that takes both the positions of the teammates and opponents into account when computing the DV of a particular action. This model can then be used to assess players by their decision making abilities within the context of the game. It can also be applied to help scouts to find players that make the best decisions in particular areas of the pitch, or the ones that make decisions that align the most with a particular team style of play.