Abstract
Optimizing modern wireless networks requires high-fidelity models of human mobility. Traditional mathematical models often fail to capture realistic spatial constraints, while modern generative models typically require vast amounts of site-specific data. This paper introduces Map2Traj, a diffusion-based framework that generates realistic human trajectories for unobserved urban areas using only street map data. By learning the intrinsic relationship between city layouts and mobility patterns, Map2Traj enables "zero-shot" trajectory generation, providing high-fidelity environments for network optimization without the privacy risks or costs associated with real-world data collection.
Introduction: The Mobility Data Paradox
Wireless network optimization tasks, such as load balancing and resource allocation, are increasingly handled by Deep Reinforcement Learning (DRL) agents. These agents require accurate simulations of user movement to learn effective policies. However, researchers face a "data paradox": traditional models like the Random Waypoint (RWP) are mathematically simple but fail to reflect how users cluster on streets or avoid obstacles. Conversely, while AI-driven models like GANs produce realistic results, they require massive datasets from the specific target area. If such data were already available, the need for simulation would be diminished.
Map2Traj breaks this cycle by treating street maps—globally available via platforms like OpenStreetMap—as a "pilot" to guide movement patterns in areas where no trajectory data has ever been collected.
Methodology
The authors frame trajectory generation as a conditional image-to-image task, drawing inspiration from zero-shot text-to-image diffusion models.
Map Representation and Conditioning
Rather than using a simple image, Map2Traj decomposes street maps into multiple channels based on road types (e.g., Trunk, Primary, Secondary, Residential). This multi-channel approach allows the model to differentiate movement behaviors across various road categories.
Conditional Diffusion Model
The core of the framework is a Conditional Diffusion Model utilizing a modified U-Net architecture. During training, the model learns to iteratively "denoise" a blurry trajectory image into a clean representation, guided by the multi-channel map condition. This process captures complex spatial dependencies between the physical layout of the city and the resulting user activity heatmaps.
Post-Processing
Because the diffusion model outputs a spatial image (heatmap), a search algorithm is employed to reconstruct sequential coordinates. A secondary deep neural network estimates "sojourn time" at specific points to derive realistic speeds and temporal dynamics.
Experimental Results
The model was trained on trajectory data from Xi’an, China, and tested on entirely different sectors of the city, as well as the city of Chengdu, to evaluate its zero-shot capabilities.
- Pattern Fidelity: Map2Traj achieved an Edit Distance on Real Sequences (EDR) of 21.47, significantly outperforming the map-restricted RWP model (192.0).
- Spatial Accuracy: Using the Wasserstein Distance to measure distributional similarity, Map2Traj scored 6.096, coming remarkably close to the real-world baseline of 2.569.
- Network Optimization Impact: In a load-balancing case study, a DRL agent trained in a Map2Traj-generated environment achieved a 7.688 Mbps edge user rate. This performance is nearly identical to the 8.503 Mbps achieved using real trajectories and far superior to the 2.098 Mbps resulting from standard RWP models.
Conclusion
Map2Traj demonstrates that street maps can serve as effective auxiliary data for generating high-fidelity mobility patterns in a zero-shot manner. By bridging the gap between simulation and reality, it provides a robust tool for wireless network optimization. Future work will focus on increasing spatial resolution and incorporating time-of-day conditioning to simulate varying traffic loads throughout the day.
Summary updated