Towards maximizing expected possession outcome in soccer

Abstract

Soccer players need to make many decisions throughout a match in order to maximize their team’s chances of winning. Unfortunately, these decisions are challenging to measure and evaluate due to the low-scoring, complex, and highly dynamic nature of soccer. This article proposes an end-to-end deep reinforcement learning framework that receives raw tracking data for each situation in a game, and yields optimal ball destination location on the full surface of the pitch. Using the proposed approach, soccer players and coaches are able to analyze the actual behavior in their historical games, obtain the optimal behavior and plan for future games, and evaluate the outcome of the optimal decisions prior to deployment in a match. Concisely, the results of our optimization model propose more short passes (Tiki-Taka playing style) in all phases of a ball possession, and higher propensity of low distance shots (i.e. shots in attack phase). Such a modification will let the typical teams to increase their likelihood of possession ending in a goal by 0.025.