Projects | Xiaohui's Academic Website

Traffic Flow Theory and Simulation

Thu, 01 Sep 2022 00:00:00 +0000

My research in traffic flow theory and simulation focuses on bridging the fundamental gap between microscopic driving behaviors and macroscopic traffic flow properties through analytical modeling approaches and advance data-driven techniques to accurately reproduce human driving behaviors. The central challenge I address is understanding and quantifying the inherent stochasticity in traffic systems, where individual driver behaviors create uncertainty that propagates through the entire traffic network. Rather than treating traffic flow as deterministic, my work develops mathematical frameworks that explicitly account for the probabilistic nature of driving decisions and their collective impact on system-level performance.

Integrated Behavior Modeling Through Deep Learning

I have developed Long Short-Term Memory neural networks that simultaneously model car-following and lane-changing behaviors by observing only the positions of surrounding vehicles. This represents a fundamental shift from conventional approaches that model these behaviors separately. The LSTM architecture automatically extracts relevant features and captures the temporal dynamics of traffic interactions, while my Hybrid Retraining Constrained training method further optimizes performance.
Validation with real-world trajectory data demonstrates superior accuracy compared to classical models and excellent transferability across different traffic environments.

Analytical Micro-Macroscopic Modeling

Traditional traffic flow models have struggled to connect individual driving behaviors with aggregate traffic patterns in a mathematically rigorous way. I have developed the Leader-Follower Conditional Distribution-based Stochastic Traffic Modeling framework that provides this crucial analytical link. Through probabilistic modeling of leader-follower interactions using Brownian dynamics, I establish a general mathematical representation that encompasses the longitudinal interactions between vehicles while remaining compatible with classical car-following theories.
This approach describes the joint distribution of vehicle speeds within traffic platoons through Markov chain modeling, creating a mathematical pathway from individual behavioral uncertainty to system-wide flow characteristics. The framework provides analytical derivations of macroscopic traffic relations, including both mean flow-density relationships and their associated variances under equilibrium conditions. This represents a fundamental advancement in traffic flow theory by providing closed-form mathematical expressions that connect microscopic stochasticity to macroscopic uncertainty.

Stochastic Fundamental Diagram

My primary contribution lies in developing analytical models for the Stochastic Fundamental Diagram, which captures the inherent variability in traffic flow relationships that traditional deterministic models cannot explain. While previous research has attempted to reproduce stochastic traffic patterns through various empirical methods, my work provides the first comprehensive analytical framework for understanding why and how this stochasticity emerges from underlying driving behaviors.
Using maximum entropy principles within my micro-macroscopic framework, I derive theoretical expressions for both the mean and variance of fundamental traffic flow relationships. This maximum entropy approach provides specific conditional distributions for longitudinal vehicle interactions, enabling the solution of analytical functions that describe traffic flow uncertainty. The resulting model quantifies how individual driver variability compounds into system-level unpredictability, providing insights into the fundamental nature of traffic flow stochasticity.

Theoretical Validation and Practical Implications

Through comprehensive validation using multiple real-world datasets including NGSIM I-80, US-101, and HighD trajectories, I demonstrate remarkable consistency between theoretical predictions and empirical observations. This validation confirms both the mathematical soundness of the Leader-Follower Conditional Distribution framework and the practical accuracy of the maximum entropy-based Stochastic Fundamental Diagram model.
The theoretical insights from this work have direct implications for traffic management and control strategies. By understanding how microscopic behavioral variations propagate to macroscopic flow uncertainty, the model provides guidance for developing interventions that promote smoother driving behaviors to suppress stochasticity and improve overall traffic flow efficiency. This represents a shift from reactive traffic management toward predictive strategies based on fundamental understanding of traffic flow uncertainty.

Research Impact and Contribution

My work establishes a new paradigm in traffic flow theory by providing the mathematical tools to analytically connect individual driving behaviors with system-level traffic performance under uncertainty. This bridges a critical gap that has limited the development of robust traffic control strategies and provides a theoretical foundation for understanding complex traffic phenomena. The analytical nature of the framework enables deeper insights into traffic dynamics than purely empirical approaches, while the validation with real-world data ensures practical relevance.
The Leader-Follower Conditional Distribution framework represents a general analytical tool that can be applied to various traffic scenarios and extended to incorporate different behavioral models. This flexibility, combined with the rigorous mathematical foundation, positions the work as a significant contribution to both theoretical traffic flow research and practical transportation engineering applications.

References

Zhang, X., Sun, Jie, Sun, Jian, 2025. On the stochastic fundamental diagram: A general micro-macroscopic traffic flow modeling framework. Communications in Transportation Research 5, 100163.
Zhang, X., Sun, Jie, Qi, X., Sun, Jian, 2019. Simultaneous modeling of car-following and lane-changing behaviors using deep learning. Transportation Research Part C: Emerging Technologies 104, 287–304.

Autonomous Driving's Impacts on Traffic

Wed, 01 Sep 2021 00:00:00 +0000

Understanding the impacts of autonomous vehicles (AVs) on traffic requires bridging the gap between vehicle-level behaviors and system-level outcomes. My research approaches this challenge through two complementary studies: the stability of AV car-following dynamics and the fundamental diagram of mixed traffic flow.

String stability analysis of NN-based car-following model

The first line of work addresses string stability. Stability is critical because even minor fluctuations in car-following behavior can propagate and amplify, creating stop-and-go waves that degrade traffic efficiency. While traditional analytical models have been widely studied, no theoretical framework existed for analyzing the string stability of neural network–based car-following models, which are increasingly used to capture realistic AV behavior.
To fill this gap, I developed the EADC framework (Estimation–Approximation–Derivation–Calculation), which derives stability criteria for both feedforward and recurrent neural network models.
Validated on representative NN-based models, the framework demonstrates consistency between theoretical and simulation results, establishing the first rigorous method for assessing the stability of AI-driven vehicle behavior.

Stochastic Fundamental Diagram modeling for mixed traffic

The second line of work focuses on the Stochastic Fundamental Diagram (SFD) for mixed traffic. Current AV impact studies often oversimplify traffic dynamics or ignore stochasticity, limiting their realism. To overcome this, I proposed a data-driven probabilistic modeling framework that learns car-following interactions for all leader–follower types (AV–AV, AV–HV, HV–AV, HV–HV) using mixture density networks. By modeling platoons as Markov chains, the framework derives the SFD from trajectory data, validated against NGSIM and Waymo datasets.
Results show that higher AV penetration decreases mean capacity and critical density while reducing capacity uncertainty, highlighting the conservative yet stabilizing effects of current AVs.

Research Impact

Together, these studies reveal how AVs influence traffic from the perspective of stability and capacity: string stability analysis ensures the reliability of AI-based AV behavior, while SFD modeling quantifies the systemic effects of AV penetration. My contributions establish new theoretical and data-driven foundations for assessing AV impacts, with implications for both traffic management strategies and policy-making in the era of mixed traffic.

References

Zhang, X., Sun, Jie, Zheng, Z., Sun, Jian, 2024. On the string stability of neural network-based car-following models: A generic analysis framework. Transportation Research Part C: Emerging Technologies 160, 104525.
Zhang, X., Yang, K., Sun, Jie, Sun, Jian, 2025. Stochastic fundamental diagram modeling of mixed traffic flow: A data-driven approach. Transportation Research Part C: Emerging Technologies 179, 105279.

Artificial Intelligence in Transportation Research

Sat, 01 Jun 2019 00:00:00 +0000

My research applies advanced artificial intelligence to revolutionize autonomous vehicle behavior and decision-making in transportation systems. The central challenge I address is that traditional approaches treat driving behaviors as isolated processes, failing to capture the interconnected nature of real-world traffic dynamics. Through deep learning and reinforcement learning techniques, I develop AI systems that understand and replicate the integrated, adaptive decision-making that characterizes skilled human driving.

Reinforcement Learning for Decision-Making

While autonomous vehicle perception has advanced rapidly, decision-making capabilities remain underdeveloped. I address this through a comprehensive framework combining deep reinforcement learning with high-fidelity simulation environments. This enables autonomous vehicles to learn optimal decision strategies that achieve nearly eight percent efficiency improvements over established adaptive cruise control systems while maintaining safety and comfort.
Extending to complex multi-lane scenarios, my integrated models achieve additional performance gains by coordinating longitudinal and lateral movements as unified decision processes.

Hierarchical AI Integration

Recognizing the need for both learning capability and interpretability in safety-critical applications, I have developed hierarchical frameworks that strategically combine reinforcement learning with rule-based methods. Through the concept of Instantaneous Desired Speed as an intermediate action, high-level AI algorithms focus on strategic decisions while low-level systems handle tactical execution. This approach maintains the adaptability of machine learning while preserving the transparency of traditional control systems.
My Traffic Flow Adaptive strategy extends this hierarchical approach to system-level optimization, enabling individually controlled autonomous vehicles to coordinate for improved overall traffic flow. This shift from isolated decision-making to collaborative traffic participation represents a crucial advancement toward practical autonomous vehicle deployment.

Research Impact

Through rigorous validation using real-world data and comprehensive benchmarking, my research demonstrates that artificial intelligence can create autonomous driving systems that are both more human-like and more efficient than existing approaches. By developing models that integrate multiple driving behaviors, learn from experience, and maintain interpretability, I am establishing the foundation for autonomous vehicles that can operate effectively in mixed traffic environments alongside human drivers. This work contributes directly to the vision of intelligent transportation systems where AI-enabled vehicles enhance both individual mobility and overall traffic system performance.

References

Ye, Y.*, Zhang, X.*, Sun, J., 2019. Automated vehicle’s behavior decision making using deep reinforcement learning and high-fidelity simulation environment. Transportation Research Part C: Emerging Technologies 107, 155–170.
Zhang, X., Sun, Jie, Wang, Y., Sun, Jian, 2023. A Hierarchical Framework for Multi-Lane Autonomous Driving Based on Reinforcement Learning. IEEE Open Journal of Intelligent Transportation Systems 4, 626–638.