Evaluating Training Data for Applying Learning to Rank to Soccer Agents

Abstract

In competitive team sports, individual player decisions significantly impact overall team performance. Designing an appropriate evaluation function for scoring player behavior in complex games like soccer is challenging. Incorporating the team supervisor’s guidance into decision-making processes is essential, but accurately scoring numerous trials poses difficulties. This paper employs a learning to rank method to obtain an evaluation function for action selection, focusing on ball-chasing behavior in soccer. The RoboCup soccer simulator is used for experiments, deriving a ranking model from players’ action logs. Gradient boosting trees are applied, and CatBoost is employed as the implementation of the learning-to-rank model. The study demonstrates that a model with satisfactory performance can be learned when the number of situations exceeds approximately 1,000, even with training data generated by humans.