Engineering Features to Improve Pass Prediction in Soccer Simulation 2D Games

Nader Zare, Mahtab Sarvmaili, Aref Sayareh, Omid Amini, Stan Matwin, Amílcar SoaresRobocup 2021 · DOI

TL;DR

This paper introduces a novel data extraction module for Soccer Simulation 2D (SS2D) that captures rich contextual information to improve pass prediction. By engineering new features and exploring various data sorting techniques, the authors achieve significant accuracy gains in predicting teammate passes, demonstrating the importance of detailed feature engineering beyond basic positional data.

What Problem Does This Paper Address?

In the dynamic environment of robotic soccer, specifically in the Soccer Simulation 2D (SS2D) league, predicting passing behavior is crucial for strategic gameplay. Effective pass prediction allows teams to optimize possession, create scoring opportunities, manage player stamina, and enhance overall decision-making. While previous research has focused on positional data, this paper addresses the need for more comprehensive feature engineering to capture the nuances of passing decisions, especially in complex scenarios involving multiple agents and dynamic game states. The challenge lies in extracting and utilizing the right information to build accurate predictive models that can adapt to different playing styles and opponent strategies.

Methods

The core contribution of this paper is the development of an embedded Data Extractor module within the SS2D agent's decision-making process. This module operates online, capturing game events and transforming them into training data. It consists of two sub-modules:

  • Feature Extractor: This module generates a comprehensive set of features, going beyond simple positions. These features are categorized into nine subgroups: Position, Kicker (relative to the ball holder), Velocity, Body angle, Team information, Player Type, Top k-th riskiest opponents, Top k-th nearest opponents, and Goal proximity.
  • Label Generator: This module assigns labels to the extracted features, primarily the uniform number (UNUM) of the intended receiver or an index number based on sorted features.

The paper also proposes four distinct sorting methods for preparing the training data: Uniform Number Sorting, X Sorting, Field Evaluator Sorting, and a "Kicker be First" binary attribute. Combining these sorting methods with the "Kicker be First" attribute results in six different data ordering strategies.

The trained models used are Deep Neural Networks (DNNs) and Random Forests (RFs). The DNN architecture includes multiple layers with ReLU activation and a softmax output, while dropout is employed for regularization.

Datasets

The primary dataset was generated by playing 3,000 games within the SS2D environment using the Helios base sample team equipped with the Data Extractor module. This resulted in 1,361,126 data instances. The dataset was split into training (1,088,237 instances) and testing (272,889 instances) subsets. The authors also conducted experiments by playing against 6 top teams from RoboCup 2019 to evaluate the models' performance against more sophisticated opponents, generating an additional 16,672 data samples.

Key Insights and Ideas

  • Beyond Position: The most significant insight is that comprehensive feature engineering, particularly incorporating relative positions (Kicker features) and opponent risk assessment (Top k-th riskiest opponents), dramatically improves pass prediction accuracy compared to relying solely on positional data.
  • Relative Features Matter: Features that describe the spatial relationships between the ball holder, potential receivers, and opponents (e.g., Kicker features) are more impactful than absolute positions. This suggests that understanding the immediate tactical context around the ball is paramount.
  • Opponent Awareness is Key: Identifying and quantifying the "risk" posed by opponents is a highly influential feature for predicting passes. This highlights the importance of an agent's ability to assess threats when deciding where to pass.
  • Sorting Strategies Impact Performance: Different data sorting methods can significantly influence model performance, with "X Sorting" and "Field Evaluator Sorting" showing robustness to changes in opponent uniform numbers.
  • Index Prediction is More Robust: Predicting the index of the pass receiver (based on sorted features) is more accurate than predicting their uniform number, especially when dealing with potential uniform number changes or variations in team compositions.

Key Results

  • The proposed methodology achieved an accuracy of 84% on the test dataset for predicting teammate passes.
  • Using all engineered features improved prediction accuracy by 2.03% to 4.65% compared to using only position features.
  • Predicting the Index of the pass receiver was 3.41% more accurate than predicting its UNUM.
  • Placing the ball holder's position as the first feature increased prediction accuracy by 0.83% on average, and up to 1.94% with specific sorting methods.
  • When playing against real opponents (top teams from RoboCup 2019), the presence of all features improved prediction probability by almost 10% (65.22% vs. 55.91% baseline).
  • The "Top Riskiest Opponents" feature group was found to have the greatest influence on prediction accuracy for both DNN and RF models.
  • The "Kicker" feature group was also found to be more important than the "Position" group, especially for DNN models.

Limitations and Future Work

  • Full State Assumption: The current work relies on the "full state option" in the server, which provides exact observations. The authors plan to explore scenarios with noisy information by disabling this option to better simulate real-world conditions.
  • Model Exploration: While DNN and RF were used, the authors suggest exploring other machine learning models, such as recurrent neural networks, with the newly engineered features.
  • Game Outcome Impact: Future work will investigate how the newly engineered features impact the actual game outcomes (win, tie, loss), not just prediction accuracy.
  • Robustness to Noise: The paper tested robustness to uniform number changes, but further exploration of robustness to other forms of noise or environmental variations could be beneficial.

Summary updated

BibTeX
@inproceedings{zare2021engineering,
  author = {Nader Zare and Mahtab Sarvmaili and Aref Sayareh and Omid Amini and Stan Matwin and Amílcar Soares},
  title = {{Engineering Features to Improve Pass Prediction in Soccer Simulation 2D Games}},
  booktitle = {Robocup},
  year = {2021},
  doi = {10.1007/978-3-030-98682-7\_12},
}

Loading related content...

Figure