Abstract
Simulating complex systems, such as large-scale urban crowds, presents a significant computational challenge: balancing the high fidelity of individual-based (microscopic) models with the efficiency of aggregate (macroscopic) models. This paper introduces a dynamic, multilevel simulation framework based on Holonic Multi-Agent Systems (HMAS). By treating groups of agents as single "super-entities" (holons) and utilizing physics-inspired energy metrics to monitor simulation accuracy, the proposed system automatically scales its computational load. This allows for real-time performance in urban environments without sacrificing the emergent behaviors and visual realism essential for high-quality simulation.
Introduction
Simulating thousands of pedestrians navigating a city is a computational nightmare. High-fidelity microscopic simulations, where every individual is modeled independently, quickly exceed available CPU power as the population grows. Conversely, macroscopic models that treat crowds as simple fluids lose the interesting, emergent behaviors of individuals.
Existing solutions often rely on a fixed "Level of Detail" (LOD), where designers pre-determine which entities are simplified. This paper addresses the challenge of dynamic multilevel simulation: creating a system that can automatically decide, in real-time, which groups of agents can be simplified into a single unit and which must remain individuals, ensuring the simulation remains accurate enough for the task at hand.
Methodology
The framework is built upon two core pillars: the Holonic Organizational Model and physics-inspired similarity indicators.
The Holonic Organizational Model
A "holon" is an entity that is simultaneously a whole (composed of smaller parts) and a part (of a larger whole)—a concept known as the Janus Effect. The authors organize these entities into a hierarchy called a "holarchy." In this structure, an atomic holon represents a single pedestrian at the Micro Level. These can be dynamically grouped into "super-holons" at the Meso Level, representing clusters of people moving with similar intent and characteristics.
Physics-Inspired Indicators
To determine when it is safe to "merge" individuals into a group or "split" a group back into individuals, the authors utilize the "energy" of the system. Three specific metrics are defined:
- Kinetic Energy ($E_c$): Derived from agent velocity.
- Goal Potential Energy ($E_{pg}$): Based on the distance to the agent's destination.
- Constraints Potential Energy ($E_{pc}$): Based on repulsive forces from obstacles and other agents.
By comparing the energy state of individuals to that of a proposed super-holon, the system calculates a Level Similarity ($s$). If the difference is minimal, the aggregated group serves as a valid approximation, allowing the system to save resources by simulating the group as a single entity.
Key Insights and Results
Unlike traditional LOD systems that merely reduce visual complexity (e.g., polygon counts), this system modifies the underlying logic. It switches from individual pathfinding to group-based behaviors, such as "boids" or fluid dynamics, on the fly. Agents are grouped by "affinity"—similar goals and energy states—ensuring that distinct groups (e.g., people walking in opposite directions) are not erroneously merged.
Performance and Scalability
Experimental evaluations in a custom 3D urban simulation environment demonstrated significant efficiency gains. As the number of agents increased, the performance gap between the microscopic and mesoscopic levels widened, proving the scalability of the holonic approach.
The system maintained a steady frame rate of 18–23 FPS while simulating up to 200 agents on legacy hardware. Crucially, the energy-based similarity index ensured that the mesoscopic level maintained a "good approximation," preserving realistic crowd behavior even as computational demand was reduced.
Conclusion and Future Work
While the holonic framework successfully manages computational load, the authors identify several areas for future research. Improving the smoothness of transitions between levels is necessary to prevent visual "pops" or sudden behavioral jumps during merging and splitting. Additionally, the authors suggest implementing predictive switching, where the system anticipates future states to adjust levels before accuracy drops occur. Modern many-core processors and GPGPU techniques also offer promising avenues for further optimizing the implementation of these complex holarchies.
Summary updated