peaRL: Spatio-Temporal Graph Learning for Predictive and Exploratory Analytics in Esports using Rocket League

Under the supervision of Prof. Suzan Üsküdarlı.

Abstract

Improving strategy and player development in esports is a critical goal for teams. Given the abundance of replay data, analytics tools are increasingly used to find insights. However, for Rocket League these tools focus on post-hoc descriptive analysis, failing to capture the complex and dynamic nature of the game. To address this, we introduce peaRL, a system that combines a high-frequency dataset of over 10 million game states with a predictive machine learning pipeline. We first establish a statistical ledger based on spatial discretization. We then implement and compare a hierarchy of predictive architectures, ranging from primitive baselines to temporal graph models. We also validate this pipeline on lower-ranked gameplay. Finally, we operationalize these models within an immersive 3D interactive portal, demonstrating how this system functions as a counterfactual exploration tool that allows users to manipulate game states and receive immediate, quantitative feedback on their strategic decisions.