SPO: Single Pass Optimization for Soccer Simulation 2D

• Walber M. Rodrigues UFPE
• Pedro V. Cunha UFPE
• Rafael R. De Labio UFPE
• Mateus G. Machado UFPE
• Paulo S. G. De Mattos Neto UFPE
• Edna N. S. Barros UFPE

Abstract

The RoboCup Soccer Simulation 2D league uses autonomous agents to compete in a simulated soccer environment. The agent uses action generators and evaluator functions to create a pool of actions and select the best action. However, a complex evaluation function increases computational demands when evaluating multiple passes, impacting real-time decision-making during the game. We propose a machine learning architecture for single-pass generation for the Soccer Simulation 2D environment, independent of position bias and player role. Our model performs a single-step target point generation for passing, avoiding multiple inferences of a costly evaluation function, making it up to 3.07 times faster than the current approach. The effectiveness of the Learnable Field Evaluator model is measured by a Root Mean Square Error (RMSE) of 309.2768, Kendall- τ of 0.9902, and Spearman- ρ of 0.9998. Also, the pass generated by the Target Point Generator is selected in 99.98% of the cases when compared to the passes generated by the heuristic pass generator. In conclusion, SPO learns the heuristic field evaluation function and how to generate optimized passes.

Keywords: Training, Neural networks, Measurement uncertainty, Machine learning, Generators, Real-time systems, Root mean square, Optimization, Robots, Sports, Ball Passing, Machine Learning, Neural Networks, Soccer Simulation 2D