Design and Implementation of a General Decision-making Model in RoboCup Simulation

Changda Wang, Xianyi Chen, Xibin Zhao, Shiguang JuarXiv 2004 · arXiv

Abstract

Traditional approaches to RoboCup Simulation often rely on fragmented, tactic-specific algorithms that are difficult to coordinate and tune. This paper presents a "General Decision-making Model" developed by researchers at Jiangsu University for their team, NOVAURO®. By reducing complex soccer tactics to a unified geometric evaluation function based on distance and visual angle, the model achieves high performance with significantly lower development complexity and cost.

Introduction: The Coordination Dilemma

In RoboCup Simulation, 11 autonomous agents must collaborate in a noisy, real-time environment. Historically, top-tier teams have used a "divide and conquer" strategy, building separate black-box algorithms for passing, shooting, and dribbling. However, this approach often leads to a "coordination dilemma," where tuning one behavior unexpectedly degrades another. The NOVAURO® team sought to overcome this by creating a mathematically grounded, universal model that unifies offensive and defensive play.

The General Decision-making Model

The core innovation is the replacement of heuristic-heavy code with formal mathematical functions. The model evaluates the environment using two primary geometric primitives: the vertical distance to the goal ($d$) and the visual angle of the goalposts ($\alpha$).

The core geometric parameters d and alpha used in the General Decision-making Model.

Shooting and Defensive Evaluation

The Shooting Evaluation Function ($Shooting\_Success$) incorporates $d$, $\alpha$, the individual agent's ability ($f$), and the degree of opponent interference ($\xi$). Rather than training a complex neural network, the team uses this formula to calculate the probability of success for any given action.

Uniquely, the defensive model is derived directly from the offense. Following the philosophy that the best offensive position is the most critical defensive one, the Defensive Evaluation Function ($Defensive(d, \alpha)$) directs agents to positions that maximize the opponent's distance to the goal while minimizing their shooting angle.

Coordination and Communication

To manage coordination without excessive communication overhead, NOVAURO® employs "decision coupling." Agents share a common decision-tree structure, allowing them to infer teammate actions from simple "secret signals" rather than full coordinate data. This structured approach to the field allows for more predictable agent behavior.

Field partitioning and agent distribution used to facilitate coordinated decision-making.

Furthermore, the team introduced a "Believe" function ($1/d^2$) to account for communication noise. As the distance between agents increases, the reliability of shared information decreases, and the agents adjust their reliance on external data accordingly.

Results and Performance

The model was tested at the CRC2003 (Chinese Robot Competition). Despite a development budget of only 20,000 RMB—approximately one-fifth that of rival teams—NOVAURO® achieved a record of 9 wins, 2 ties, and 3 losses, reaching the final stages of the tournament.

The success of the model demonstrates that geometric universality can outperform brute-force engineering. While rival teams managed five or more distinct algorithms for different tactics, NOVAURO® used the same core functions for ball handling, passing, and positioning.

Conclusion and Future Work

The General Decision-making Model offers a streamlined, effective alternative to fragmented AI architectures in multi-agent systems. Future work will focus on optimizing stamina distribution, improving specific pass/interception logic, and transitioning from a rigid hierarchy to a dynamic one to better adapt to fluid match conditions.

Summary updated

BibTeX
@misc{wang2004design,
  author = {Changda Wang and Xianyi Chen and Xibin Zhao and Shiguang Ju},
  title = {{Design and Implementation of a General Decision-making Model in RoboCup Simulation}},
  year = {2004},
  eprint = {cs.ro/0411023},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/cs.ro/0411023},
}
Figure