GTA: Generative Traffic Agents for Simulating Realistic Mobility Behavior

Paper Title

GTA: Generative Traffic Agents for Simulating Realistic Mobility Behavior

Publication Info

• Topic area: Simulation of urban mobility using generative agents grounded in population data.
• Keywords: Generative agents, urban mobility, traffic simulation, LLM, SUMO, census data, modal split, transportation modeling, HCI, empirical validation.

Background and Problem

• Problem / challenge: Traditional traffic simulations rely on fixed rules and costly data collection, making them unsuitable for early-stage evaluation of mobility innovations. Existing LLM-based generative agents lack empirical validation and fail to embed agents in realistic urban environments.
• Significance: Predicting transportation choices at scale is critical for urban planning and sustainable transport. Scalable, realistic simulations can help policymakers and designers evaluate the impact of new technologies and policies.
• Motivation and related work: Prior work on agent-based models and LLM-driven simulations has shown promise but remains limited by handcrafted assumptions, lack of sociodemographic grounding, and insufficient validation against real-world data. This paper addresses these gaps by introducing GTA.

Solution

• Proposed approach: Generative Traffic Agents (GTA), a framework that uses LLM-powered, persona-based agents grounded in census data to simulate realistic urban mobility behavior.
• Novelty:
  1. Integration of census-aligned microdata with LLM-driven personas to generate heterogeneous, realistic mobility behavior.
  2. Embedding agents in a city-scale traffic simulation (SUMO) to model mobility under real-world constraints.
  3. Empirical validation of simulated behaviors against survey data and traffic counts.
  4. Modular architecture enabling interoperability with standard datasets and tools.
• Procedure and key techniques:
  • Profile Module: Generates agent personas using census-calibrated socio-demographic data and LLM-based descriptions.
  • Planning Module: Creates daily activity schedules and assigns transport modes based on persona attributes and network-informed options.
  • Action Module: Executes traffic simulations using SUMO, incorporating dynamic user equilibrium for realistic routing.

Results

• Concrete findings:
  • GTA closely replicates modal split distributions from the "Mobility in Germany 2017" survey, with an RMSE of 4.07.
  • Systematic deviations include underrepresentation of very short trips (<1 km) and overestimation of medium-range trips (2–20 km).
  • Traffic flow simulations capture key temporal patterns but underestimate nighttime activity and show sharper morning peaks.
• Advantage over baselines: GTA achieves behavioral realism at both individual and aggregate levels, outperforming prior LLM-based systems in empirical validation and city-scale applicability.
• Experiments / evaluation:
  • Simulated 1% of Berlin's population (35,769 agents) and a 10% sample for the Wedding district (8,680 agents).
  • Compared simulation outputs against survey data (modal split, trip lengths, durations) and traffic counts.
  • Conducted ablation studies to test sensitivity to LLM models, prompting strategies, and system configurations.
• Limitations and future work:
  • Behavioral biases toward socially desirable actions (e.g., frequent biking, low nighttime activity).
  • High computational cost for large-scale simulations (e.g., 43 hours for 10% of Berlin's population).
  • Future plans include integrating with alternative simulators (e.g., MATSim), modeling social networks, and adding multi-day memory for agents.

Summary

GTA introduces a novel framework for simulating urban mobility using LLM-driven, persona-based agents grounded in census data. By integrating with SUMO and validating against empirical data, GTA achieves realistic, scalable simulations of transportation behavior. The system captures key trends in modal split and traffic flow but exhibits biases and computational challenges. GTA offers a practical tool for early-stage evaluation of mobility innovations, with potential extensions to improve diversity, efficiency, and long-term behavioral modeling.