United States Department of Transportation
i
An Exploration of the Use of Artificial Intelligence for Enhanced Traffic Management, Operations and Safety
-
2024-01-24
-
[PDF-27.19 MB]
Details:
-
Creators:
-
Corporate Creators:
-
Corporate Contributors:
-
Subject/TRT Terms:
-
Publication/ Report Number:
-
Resource Type:
-
Geographical Coverage:
-
Edition:September 2019 – October 2023
-
Corporate Publisher:
-
Abstract:This project explored the use and value of artificial intelligence and machine learning (ML) in transportation taking a multi-pronged approach that includes a literature review, a workshop, a survey, the development of3 prototype ML models for four high-priority use cases, and the field testing of one of the prototyped models. Initial tasks were exploratory in nature and led to a better understanding of the current and prospective uses of ML in transportation, corresponding data needs, and specific use cases of interest to TxDOT. Prototype model development was used to assess the value, challenges, and limitations of implementing several types of machine learning models to support the use cases prioritized by TxDOT. The prototypes leveraged emerging and traditional data sources: Wejo event data was used along CRIS data to build supervised and unsupervised learning models for understanding safety hot-spots and evaluating the effects of the pandemic on safety and traffic patterns; a microsimulation environment was used to explore the feasibility of adjusting traffic signal timing plans in real-time in a frontage road setting using reinforcement learning models; probe-based speeds from INRIX were combined with traffic volume data from TxDOT’s ITS to generate short term travel time predictions on I-35, which could lead to more accurate driver information. Short -term travel time prediction models were selected for field testing given the promising results found during the prototyping and the maturity and widespread availability of the involved data sources. In our preliminary models ML experienced travel time predictions were 40% more accurate during peak periods that those from traditional approaches. Field testing included the training of additional models in Austin and El Paso, and the development of a framework to expedite model training, testing, and evaluation, and to support real-time deployment. Real-time predictions were shared with TxDOT through a web-based application that also facilitated model evaluation. The performance evaluation of the newly trained models considered the ability to correctly identify the fastest route among competing alternatives, and predictive ML models were found to be correct more often than traditional approaches.
-
Format:
-
Funding:
-
Collection(s):
-
Main Document Checksum:
-
Download URL:
-
File Type:
