A Model for Collective Strategy Diffusion in Agent Social Law Evolution

Yichuan Jiang, Toru IshidaIJCAI 2007

Abstract

This paper introduces a dynamic model for the emergence of "social laws"—the rules governing coordination in multi-agent systems (MAS)—from the bottom up. By balancing the influence of social rank against the collective power of numbers, the authors demonstrate how social norms evolve through competition and diffusion. The model highlights the phenomenon of "collective insurgent diffusion," where a majority of junior agents can influence superiors, and explains why certain outliers persist even after a global law is established.

Introduction

In multi-agent systems, coordination is essential for system stability and efficiency. Traditionally, researchers utilized "social laws"—pre-defined sets of behavioral restrictions—to ensure agents interact predictably. However, these laws were typically designed "off-line" and remained static throughout a system's lifecycle.

The authors argue that static laws are unrealistic for complex, evolving environments. In the real world, social norms are dynamic; they emerge and change through social competition. While previous research focused on "hierarchical diffusion" (top-down influence), it often ignored the "power of the people." This paper seeks to bridge that gap by modeling how collective group pressure can drive strategy adoption across different social ranks.

The Collective Strategy Diffusion Model

The proposed model is built upon the interplay of two opposing forces that dictate whether an agent will change its behavior:

  1. Diffusion Impact Force ($IF$): This represents the pressure a group (an overlay group sharing the same strategy) exerts on an individual. It is calculated based on the group's collective social rank and the "distance" or similarity between the group's strategy and the agent's current state.
  2. Counteracting Force ($CF$): This represents an agent’s internal resistance to change. It is determined by the agent’s own social rank, the average rank of the influencing group, and the radicalness of the proposed change.

Decision Logic

The model evaluates the ratio of $IF/CF$ against two thresholds, $\xi$ (high) and $\eta$ (low):

  • Total Adoption: If $IF/CF \ge \xi$, the agent adopts the new strategy entirely.
  • Partial Shift: If $\eta \le IF/CF < \xi$, the agent moves partially toward the new strategy.
  • Resistance: If $IF/CF < \eta$, the agent maintains its current position.

Experimental Setup: The Queue Orientation Scenario

To test the model, the authors utilize a synthetic simulation environment called the "Queue Orientation" scenario. In this setup, agents are placed on a grid and assigned randomized initial orientations (one of eight directions) and social ranks.

The Queue Orientation scenario setup, showing possible directions (a), agent positions on a grid (b), and their initial social strategies (c).

As shown in Figure 2, agents $a$ through $e$ start with different orientations and ranks. The simulation tracks how these individual strategies diffuse across the grid as agents interact and form collective groups.

Results and Analysis

The authors conducted case studies on a 12x12 grid to observe how strategies converge. A key finding is the emergence of a dominant strategy that eventually functions as a social law.

Simulation of strategy diffusion on a 12x12 grid, illustrating the emergence of Social Strategy 3 (red arrows) as a dominant law.

Key Insights

  • Collective Insurgent Diffusion: The simulation proves that if enough low-ranking agents align on a strategy, their collective Impact Force can overwhelm the Counteracting Force of a high-ranking individual, forcing the leader to conform to the crowd.
  • Persistence of Outliers: The model explains why some individuals never conform. If an agent has a sufficiently high rank, their internal resistance ($CF$) can permanently outweigh the collective pressure ($IF$) of the majority.
  • Convergence to Uniformity: In the primary case study (Figure 6), Strategy 3 (indicated by red arrows) eventually took over almost the entire grid. The speed of this convergence is highly sensitive to the $\xi$ and $\eta$ parameters; higher thresholds lead to slower diffusion and a higher likelihood of persistent outliers.

Conclusion

This research frames a social law not as a permanent rule, but as a dominant strategy that has reached a state of equilibrium. The model provides a robust framework for understanding how bottom-up collective action can reshape social structures in multi-agent systems. Future work will aim to address the computational complexity of many-to-many diffusion and explore multi-dimensional strategy spaces beyond simple orientations.

Summary updated

BibTeX
@inproceedings{jiang2007model,
  author = {Yichuan Jiang and Toru Ishida},
  title = {{A Model for Collective Strategy Diffusion in Agent Social Law Evolution}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2007},
}
Figure