AI
Using LLMs and representative agents to scale realistic traffic models
Researchers combine large language models with reinforcement learning to simulate traveler behavior at scale, addressing cost and stability issues in agent-based traffic prediction.
Agent-based traffic models that use large language models to mimic traveler decision-making offer greater realism than traditional algorithms, but face practical barriers when deployed at realistic scales. A new approach addresses these constraints by deploying a single representative LLM agent per traveler demographic, dramatically cutting computational expense while maintaining behavioral fidelity.
The scalability challenge#
Simulating urban traffic with LLMs has traditionally required querying a language model for every individual traveler—an approach that becomes prohibitively expensive at city scale. Beyond cost, prior LLM-based traffic agents have demonstrated two troubling properties: their routing choices often remain unexplained to human observers, and their day-to-day learning patterns tend to oscillate unpredictably rather than converge toward stable equilibrium.
How representative agents work#
The proposed method assigns a single LLM agent to represent each homogeneous group of travelers facing similar choice contexts. This agent maintains and updates a mixed strategy—a probabilistic distribution over available routes—that mirrors the group's aggregate travel patterns. Each day, the LLM reviews its travel experience and identifies routes it would prefer to use more frequently based on observed performance. An interpretable, tunable update rule then translates these preferences into concrete adjustments to the strategy distribution, with a progressively decaying step size controlling learning speed.
This separation of reasoning from strategy updating clarifies how decisions arise while simultaneously stabilizing convergence behavior.
Testing and real-world patterns#
In standard traffic assignment problems, the approach converges rapidly to user equilibrium—the stable state where no traveler can improve by unilaterally changing routes. Testing with richer scenarios incorporating income variation, multi-criteria costs, and multiple transport modes yielded stable and interpretable dynamics. The model successfully reproduced well-documented behavioral phenomena, including the decoy effect (how a third option shifts preferences between toll and non-toll roads) and income-dependent willingness to pay for convenience when choosing among driving, transit, and park-and-ride options.
