Game-theoretic Modeling of Traffic in Unsignalized Intersection Network for Autonomous Vehicle Control Verification and Validation

Ran Tian (田然), Ran Tian, Nan Li, Ilya V. Kolmanovsky, Yildiray Yildiz, Anouck GirardarXiv 2019 · arXiv

Abstract

Testing Autonomous Vehicles (AVs) in real-world environments is a slow and costly process. This paper presents a framework for simulating realistic, heterogeneous traffic at unsignalized intersections to enhance AV control system verification and validation. By integrating level-$k$ game theory with Dataset Aggregation (DAgger) imitation learning, the researchers developed a computationally efficient model that captures complex human-like interactions. The results demonstrate high fidelity to real-world datasets and scalability to large-scale urban networks, providing a robust environment for testing AV software against diverse driving behaviors.

The Challenge of Unsignalized Intersections

The navigation of unsignalized intersections represents a significant challenge for autonomous systems. Unlike signalized junctions, these scenarios require drivers to negotiate right-of-way through subtle cues and strategic predictions. Existing simulators often rely on macroscopic models that treat traffic like a fluid or microscopic models that lack realistic social interaction. This research addresses the gap by modeling the strategic decision-making processes inherent in human driving, allowing for more realistic "wild west" urban simulation.

Methodology: Modeling Strategic Behavior

The proposed model utilizes a multi-layered approach to simulate driver behavior:

  1. Kinematic Foundation: Vehicles are modeled using a discrete-time "unicycle model," balancing computational speed with the accuracy needed for low-to-medium speed urban maneuvers.
  2. Level-$k$ Game Theory: This framework defines driver "personalities." Level-0 drivers are aggressive and non-strategic; Level-1 drivers are cautious, assuming others are Level-0; and Level-2 drivers are strategic, assuming others are Level-1.
  3. Receding-Horizon Optimization: Drivers solve optimization problems at each time step to maximize a reward function balancing safety, speed, and lane discipline.
  4. Imitation Learning (DAgger): To overcome the computational burden of real-time game-theoretic optimization, the authors trained neural networks using the DAgger algorithm. Unlike standard behavioral cloning, DAgger iteratively incorporates "failure states" into the training set, preventing the model from drifting when small errors occur.
Figure 5: Time-series snapshots of multi-vehicle interactions at a four-way intersection, illustrating how different strategic levels influence velocity and positioning.

Key Insights: The "Nice Guy" Deadlock

A critical finding of the simulation is the "Nice Guy Deadlock." If all vehicles at an intersection are cautious (Level-1), the system can reach a standstill where everyone yields and nobody moves. Conversely, a purely aggressive (Level-2) population leads to frequent collisions. Realistic traffic flow requires a heterogeneous mix of these strategic levels. Furthermore, the study highlights that an adaptive AV controller, which estimates the strategic level of surrounding drivers on the fly, performs significantly better than rule-based baselines.

Results and Scalability

The model was validated against the INTERACTION dataset, showing that simulated trajectories and average speeds statistically align with real-world human patterns. The neural network approximations achieved a 97.8% accuracy in replicating the complex game-theoretic expert logic.

Figure 4: Large-scale simulation of a complex urban network featuring a heterogeneous mix of Level-1 and Level-2 cars across various intersection types.

As shown in the large-scale simulation (Figure 4), the framework successfully manages a complex network with 30 interacting vehicles across four-way intersections, roundabouts, and T-junctions. In 2,000 simulation episodes, the adaptive controller demonstrated lower collision and deadlock rates compared to standard baselines, particularly as traffic density increased.

Future Directions

While the current model provides a robust foundation, several areas for improvement remain. Future work aims to integrate higher-fidelity physics (e.g., into the CARLA simulator), utilize Inverse Reinforcement Learning (IRL) to learn driver reward weights directly from human data, and account for real-world sensor limitations such as occlusions.

Summary updated

BibTeX
@misc{2019game,
  author = {Ran Tian (田然) and Ran Tian and Nan Li and Ilya V. Kolmanovsky and Yildiray Yildiz and Anouck Girard},
  title = {{Game-theoretic Modeling of Traffic in Unsignalized Intersection Network for Autonomous Vehicle Control Verification and Validation}},
  year = {2019},
  eprint = {1910.07141},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1910.07141},
}
Figure