Machine learning methods for data-driven microscopic traffic simulation modelling and calibration

Abstract

Microscopic traffic modelling and simulation plays a crucial role in understanding and managing the complexities of modern road networks. However, existing approaches often struggle to capture the intricacies of real-world traffic behavior. Recent advancements in computing power and the abundance of data provide an opportunity to refine existing modelling and calibration approaches. Therefore, this thesis embarks on a systematic exploration of the synergy between modern machine learning (ML) techniques and classic microscopic traffic simulation (MTS) models, aiming to enhance the accuracy and realism of simulations. It introduces novel approaches for the integration of data-driven insights into MTS, particularly via car-following modelling and calibration as well as vehicle trajectory reconstruction that tackles challenging real-world missing patterns.

Research in MTS modeling has traditionally relied on physics-based models, which offer great analytical insights but often fall short in simulation accuracy. Conversely, learning-based (data-driven) models, while accurate, function as black boxes with limited explainability. Despite significant progress in data-driven car-following modeling and calibration, there has been limited emphasis on the integration of physics and learning, despite its potential advantages. First, there is a crucial research gap in the development of a new methodology that combines the analytical strengths of physics-based models with the accuracy of learning-based models. Second, it is desirable to combine them in a manner that can effectively balance simulation accuracy and physical consistency. Third, there is also the need for integrating data-driven car-following models (CFM) into real-time simulations while being capable of learning from online data. Last, the calibration of these models, particularly in dynamic environments and with real-world incomplete and irregular trajectory data also presents significant challenges.

Addressing these gaps, this thesis makes new major contributions to the field in several ways as follows.

Firstly, a deep reinforcement learning approach for CFM calibration is proposed. This model maintains analytical explainability while leveraging learning-based methods to enhance accuracy through dynamic calibration.

Secondly, a physics-guided machine learning approach is proposed to develop a graph learning-based CFM. This model improves learning capabilities and addresses the vulnerability of behavior cloning-based CFMs when they are subject to distributional shift.

Thirdly, a jointly trained physics-guided CFM is proposed. This model is designed for efficiency and physical consistency and forms part of a new dynamic data-driven microscopic traffic simulation system while overcoming the lack of online learning capability in existing data-driven CFMs.

Finally, the thesis underscores the necessity of reconstructing noisy, incomplete real-world trajectory data. It proposes a fine-grained trajectory reconstruction framework using MTS calibrated with a surrogate-assisted evolutionary optimization, achieving accurate and consistent reconstruction results under various missing patterns.

In summary, this thesis bridges the gap between modern ML methods and classic MTS models by integrating physics and learning, developing efficient real-time simulation models, and enabling fine-grained trajectory reconstruction. These contributions enhance the accuracy and realism of traffic simulations, reflecting the complexities of real-world road traffic in the era of big data analytics. They mark a significant step forward in the field of traffic simulation, offering a systematic framework for future research and practical applications.