Learning Long-Term Dependencies to Predict an Opponent’s Behavior in Robot Soccer^*

Abstract

The Small Size League (SSL) robot soccer category stands out for its dynamic matches, in which multiple agents must collaborate and make real-time decisions to execute offensive and defensive actions. Understanding the opponent’s behavior is a skill that can enhance a team’s performance by anticipating moves and making strategic decisions. Although this is a common topic of study in the league, advances in Machine Learning and the increasing amount of accumulated data enable the adoption of data-driven approaches. In this study, a Deep Learning (DL) Long Short-Term Memory Recurrent Neural Network (LSTM) was implemented to extract spatial and temporal patterns and predict future plays based on features collected from past matches. The impact of key parameters, such as Reading Window (RW) and stride, was analyzed and evaluated to predict multiple actions of interest, including kicks, passes, and dribbles. The model used in this study achieved accuracy and F1-Score greater than 60% in the macro average and 80% in the weighted average, demonstrating the effectiveness of this approach in extracting features and capturing temporal dependencies to analyze and predict opponent behavior in robot soccer.