Cyrus2D Base: Source Code Base for RoboCup 2D Soccer Simulation League

Nader Zare, Omid Amini, Aref Sayareh, Mahtab Sarvmaili, Arad Firouzkouhi, Saba Ramezani Rad, Stan Matwin, Amílcar Soares

TL;DR

This paper introduces Cyrus2D, an enhanced open-source base code for the RoboCup 2D Soccer Simulation League. By incrementally building upon existing strong baselines and incorporating novel AI techniques like improved blocking, offensive risk evaluation, and a deep learning-based pass prediction system for unmarking, Cyrus2D demonstrates significant improvements in team performance and strategic play.

What Problem Does This Paper Address?

The RoboCup 2D Soccer Simulation League (SS2D) presents a complex multi-agent AI challenge, requiring sophisticated strategies for autonomous decision-making, coordination, and tactical planning. Developing effective base code for agents in this league is crucial for accelerating research and enabling teams to compete at a high level. While several base codes have been released, there's a continuous need for more robust and intelligent frameworks that can adapt to the dynamic and partially observable environment of SS2D. This paper addresses the challenge of creating a more capable and adaptable base code that can improve both defensive and offensive strategies, ultimately leading to better team performance.

Methods

The authors introduce Cyrus2D through an incremental development process, building upon existing successful base codes:

  • Cyrus2D Base Version 0.0: This initial version is a merger of the latest releases of the Helios and Gliders2D bases, with some parameter tuning. It serves as a foundation for subsequent improvements.
  • Cyrus2D Base Version 1.0: This version incorporates three key AI-driven functionalities:
  • Blocking Strategy: An improved multi-agent blocking decision-making system that simulates opponent dribbling behavior to predict optimal defensive positions.
  • Offensive Risk Evaluation (ORE): An enhancement to the existing "Field Evaluator" module that quantifies the risk associated with offensive actions by considering how quickly opponents can intercept the ball.
  • Simple Unmarking Strategy: A basic strategy to help players reposition themselves to receive passes.
  • Cyrus2D Base Version 1.1: This version further refines the unmarking strategy by integrating a Pass Prediction Deep Neural Network (DNN). This DNN, trained on game states, predicts the most likely pass receiver, enabling more intelligent unmarking decisions and the construction of a probable pass tree.

Datasets

The paper describes the creation of a dataset for training the Pass Prediction DNN. This dataset was generated by running 500 games against various established teams (Helios Base v3.11, Gliders2D v1.6 and v2.6, Cyrus 2021, Helios 2021, and YuShan 2021). The authors implemented a "Data Extractor" module within their agents to record game states and corresponding labels (the uniform number of the best pass receiver). This resulted in a dataset of 1,429,032 data instances.

Key Insights and Ideas

  • Incremental Development is Effective: The paper demonstrates that building upon existing strong baselines and incrementally adding new features leads to significant improvements, rather than starting from scratch.
  • Deep Learning for Pass Prediction: The integration of a DNN for pass prediction is a key innovation, allowing for more sophisticated and proactive unmarking strategies by anticipating future game states and optimal pass targets.
  • Offensive Risk as a Crucial Metric: Quantifying offensive risk by considering opponent interception capabilities is vital for making better strategic decisions in dynamic game environments.
  • Coordinated Defense Matters: The improved blocking strategy highlights the importance of multi-agent coordination in defense, moving beyond individual actions to a more strategic team-level approach.

Key Results

The authors evaluated Cyrus2D base v1.1 against several other leading SS2D bases and teams. The results show significant improvements:

  • Win Rate: Cyrus2D v1.1 consistently outperforms Helios and Gliders2D bases, winning against them in over 99% and 84% of games, respectively. Against the top three RoboCup teams (Helios2021, YuShan2021, Cyrus2021), Cyrus2D v1.1 achieved an average win rate 2.86% to 3.76% higher than Gliders2D and Helios bases.
  • Goals Scored/Conceded: Table 2 indicates that Cyrus2D v1.1 generally scores more goals and concedes fewer goals compared to other versions and bases, demonstrating improved offensive and defensive capabilities. For instance, Cyrus2D v1.1 scored an average of 4.4 goals against Helios Base v3.11 while conceding only 0.2.

Limitations and Future Work

  • Limitations: The paper mentions that the "Pass Prediction DNN" achieved a validation accuracy of 68.1% on the test data, suggesting room for further improvement in prediction accuracy. The unmarking strategy, while improved, still relies on simulating a limited number of future states.
  • Future Work: The authors plan to enhance Cyrus2D by improving chain action movement prediction and incorporating multi-agent decision-making into the marking strategy. They also aim to further optimize the pass prediction module and explore more advanced AI techniques.

Summary updated

Abstract

Soccer Simulation 2D League is one of the major leagues of RoboCup competitions. In a Soccer Simulation 2D (SS2D) game, two teams of 11 players and one coach compete against each other. Several base codes have been released for the RoboCup soccer simulation 2D (RCSS2D) community that have promoted the application of multi-agent and AI algorithms in this field. In this paper, we introduce "Cyrus2D Base", which is derived from the base code of the RCSS2D 2021 champion. We merged Gliders2D base V2.6 with the newest version of the Helios base. We applied several features of Cyrus2021 to improve the performance and capabilities of this base alongside a Data Extractor to facilitate the implementation of machine learning in the field. We have tested this base code in different teams and scenarios, and the obtained results demonstrate significant improvements in the defensive and offensive strategy of the team.
BibTeX
@inproceedings{zare2022cyrus,
  author = {Nader Zare and Omid Amini and Aref Sayareh and Mahtab Sarvmaili and Arad Firouzkouhi and Saba Ramezani Rad and Stan Matwin and Amílcar Soares},
  title = {{Cyrus2D Base: Source Code Base for RoboCup 2D Soccer Simulation League}},
  booktitle = {Robocup},
  year = {2022},
  doi = {10.1007/978-3-031-28469-4\_12},
}
Figure