M-TRAIL|CEE|UMD - AI/ML for Transportation

AI/ML for Transportation

Physics-Informed Machine Learning

Physics-informed machine Learning (PIML) represents a transformative approach that integrates physical laws and principles with ML models to enhance predictive accuracy and robustness against data noise. By incorporating established physics constraints into the learning process, PIML provides a unique advantage in scenarios where traditional data-driven approaches may struggle with noisy data. This integration allows PIML to capture intricate system behaviors, producing models that are more reliable, interpretable, and computationally efficient. In engineering research, PIML shows particular promise as it facilitates the development of models that can simulate real-world phenomena with improved fidelity, thereby accelerating advancements across various domains, including transportation research.

Self-supervised Learning

We pioneer self-supervised learning techniques to address critical challenges in transportation data estimation, particularly when ground-truth data is unavailable or incomplete. Our recent work focuses on developing a novel generative adversarial network (GAN) architecture that integrates long short-term memory (LSTM), attention mechanisms, and convolutional neural networks (CNN) to estimate real-time origin-destination (OD) flows along signalized arterials. By treating CV-derived OD matrices as flawed observations and applying a custom-designed self-supervised loss function, our approach enables accurate OD flow estimation without relying on historical data or costly sensor infrastructure. This framework significantly enhances the scalability and practicality of traffic flow modeling in real-world, data-sparse environments.

Reinforcement Learning

We harness Reinforcement Learning (RL) to tackle complex, real-time decision-making challenges in traffic signal control. Our previous work develops a customized Deep Q-Learning (DQL) framework—an advanced form of value-based RL—that learns optimal traffic signal policies directly from field-collected Automatic Number-Plate Recognition (ANPR) data, even under partial observability and sensor imperfections. To overcome the limitations of simulator-dependent training, we introduce a deep neural network to model traffic state transitions and enable efficient sample generation. The RL agent is initially trained via imitation learning using real signal timing data from commercial systems such as SCOOT, and is further refined through self-adaptive exploration of the state-action space. Our approach demonstrates that RL can achieve superior or comparable performance to industry-standard adaptive controls, offering a scalable and data-efficient alternative for urban traffic management.

Large Language Model

We explore the transformative potential of Large Language Models (LLMs) for intelligent transportation systems, particularly in automating the interpretation of complex regulatory and planning documents. Our recent work develops a domain-adapted LLM pipeline by applying low-rank adaptation (LoRA) techniques to continue pretraining state-of-the-art models such as LLAMA-3.1 and Qwen2.5 on U.S. transportation policy documents. By fine-tuning these transformer-based models on curated datasets from FHWA and other sources, we enable advanced capabilities in policy-aware question answering, document summarization, and technical standard retrieval. This research demonstrates that customized LLMs can significantly reduce manual workloads for urban planners and engineers, enhance regulatory compliance, and improve the efficiency of data-driven decision-making across transportation domains.

Page updated

Report abuse