CHAPTER 8

Conclusions

This NCHRP report provides an evaluation of the current state of the art and state of the practice of Machine Learning (ML) at state DOTs and presents a roadmap for how state DOTs and other transportation agencies could develop and leverage ML solutions for their needs. As state DOTs are exploring to build or expand their ML capabilities, the content of this report provides a practical starting point for developing a general understanding of implementing ML solutions for transportation. The potential applications of ML in transportation are broad ranging from performing multiple critical functions in autonomous driving to asset management and system planning. The level of effort can vary significantly depending on the type of application – see the callout box below. As indicated in the accompanying ML guide that is developed as part of this NCHRP project (Samach et al. 2024), while ML offers a promising future, it is not suitable for all types of problems and sometimes brings new challenges and risks. This report and the referenced ML guide highlight some of the key considerations for state DOTs when embarking on creating and deploying ML solutions. This section of the report provides some of the conclusions derived from the literature review, state DOT surveys, case studies, and the development of the ML guide.

Level of Effort in Implementing ML Applications

Literature Review

The review conducted based on the academic literature and a scan of relevant websites revealed that a diverse range of applications is being explored and implemented across various facets of the transportation sector. The integration of ML and deep learning (DL) technologies has emerged as an important factor in enhancing transportation system efficiency, safety, and operations, driven by the vast data generated by connected and automated vehicles (CAVs) and multimodal transportation systems. This surge in data has outpaced traditional statistical methods, necessitating the adoption of ML techniques to process and extract actionable insights efficiently. Several of the key conclusions from the literature review are summarized below.

Widespread Applications Across Transportation Domains: ML technologies are being applied to a broad spectrum of transportation problems, including traffic prediction, incident detection, vehicle identification, traffic signal timing, public transportation enhancements, and visual recognition tasks. The advancements in DL have particularly shown superior results in traffic state prediction (e.g., travel times, traffic volumes), vehicle identification, and visual recognition tasks.

Rapid Increase in ML-Related Publications: The number of ML-related publications in the transportation field has seen a remarkable increase, particularly after 2017, indicating a growing interest and investment in ML technologies. This trend aligns with the global recognition of DL’s potential following its success in international data competitions and the availability of DL tools for model training.

State of the Practice at Transportation Agencies: Transportation agencies are increasingly considering, leveraging, and implementing ML, with a focus on supervised ML and computer vision for object, incident, and pedestrian detection. Transportation systems management and operations (TSMO) emerges as a popular domain for ML applications, with intelligent traffic signal systems and work zone management being notable areas of focus. There also appears to be a lot of research emphasis on safety-related applications, such as hazard detection and driver behavior monitoring. Asset management is an area that is seeing increased interest and utilization as well.

Collaboration and Procurement: Most agencies procure ML capabilities from vendors or collaborate with university partners, often starting with pre-trained baseline models and open-source tools. Data sources for ML applications are diverse, with agencies using existing data as well as acquiring additional data to support their ML initiatives.

Infrastructure and Funding: The foundational elements for supporting ML applications, including data storage, computational resources, and software, are critical. Subscription-based software procurement is becoming a common practice, and some agencies are using federal grants to kickstart their ML programs.

Challenges: Despite the successes demonstrated in various academic publications, there are areas requiring further improvement, such as enhancing DL models for incident detection to reduce false alarms and creating an established ML method for signal timing that is validated with field data. The need for benchmark datasets to enable fair comparison of models is also highlighted as a critical requirement for progressing ML applications in transportation.

Economic, Social, and Business Impacts of ML: The potential economic, social, and business impacts of widespread AI and ML adoption in transportation are significant. However, challenges such as forecasting under unexpected events, computational complexity, model transparency, and potential biases in manual labeling processes need to be addressed to fully realize these benefits.

Surveys with State DOTs

A web-based survey was conducted to gather data on the adoption and familiarity of ML methods within state DOTs across the United States. The key findings based on the 43 survey responses received from 29 different states are summarized below.

Familiarity and Adoption of ML: The survey results indicate a varied level of familiarity with ML among state DOTs, with more than half of the respondents reporting not being very familiar with ML methods and tools. Despite this, there is a notable inclination toward the adoption of ML, with 13 respondents reporting ML deployments or plans for developing ML applications. The presence of data science professionals in over half of the agencies underscores an increasing recognition of the importance of data science skills in effectively leveraging ML technologies.

Application Areas and ML Methods: The primary application areas for ML within transportation agencies include asset management and maintenance, and TSMO. The prevalent use of advanced ML techniques, such as artificial neural networks and deep learning, reflects a trend toward employing sophisticated methods to tackle complex transportation issues. However, the survey also indicates that many ML applications are still in the nascent stages of research and development or prototyping, with a majority of respondents expressing only moderate satisfaction with the current ML applications, which could perhaps be attributed to the fact that some of these technologies are new, and the benefits have not been fully realized yet. This suggests there is considerable scope for improvement in aligning ML applications with the agencies’ objectives and performance expectations.

Prospects and Challenges: Looking forward, the survey highlights both motivations and challenges influencing the adoption of ML in transportation. Key motivators and opportunities include the effectiveness of ML methods over traditional approaches, the ability to process large data volumes, and potential reductions in labor costs. Conversely, significant barriers to further ML adoption include the shortage of AI/ML skilled workforce and difficulties in integrating ML technologies with existing systems and processes. This highlights a pressing need for enhancing skill sets and adapting organizational structures to fully capitalize on ML technologies. Despite these challenges, the interest in exploring new ML applications, such as automated incident detection and predictive models for various transportation applications, indicates a proactive stance among DOTs toward incorporating advanced technologies, albeit with a realistic acknowledgment of existing hurdles.

Case Studies

Five case studies were conducted with state DOTs - Nebraska, California, Iowa, Missouri, and Delaware - focusing on their experiences with developing and implementing ML solutions for various transportation challenges. These projects aimed at enhancing asset management, safety, and TSMO through advanced ML technologies. The case studies explore different aspects of ML deployment, including application areas, data management, costs, evaluation metrics, workforce capacity, collaboration with stakeholders, and lessons learned. Each DOT approached ML deployment with unique objectives and encountered diverse challenges, yet shared common goals of improving safety, efficiency, and decision-making capabilities. From the Nebraska DOT’s focus on asset management using image recognition to Delaware’s comprehensive AI-integrated transportation management system (ITMS) program for real-time traffic management, these case studies highlight the critical role of data quality, stakeholder collaboration, and continuous learning in successful ML implementations.

A recurring theme across the case studies is the need for robust data infrastructure and the careful evaluation of data sources to ensure the accuracy and effectiveness of ML models. The challenges of integrating ML solutions with existing systems and the importance of fostering internal expertise and stakeholder buy-in were also emphasized. The experiences of these DOTs demonstrate that while the path to leveraging ML in transportation is complex and multifaceted, the potential benefits in terms of enhanced safety, operational efficiency, and proactive management are significant.

As DOTs are enhancing their ML capabilities, it is important for them to continue sharing insights and lessons learned to collectively advance the adoption of ML technologies. Emphasizing the development of in-house technical capabilities, fostering partnerships with academic and industry experts, and prioritizing scalability and adaptability in ML solutions will be key to addressing the evolving challenges of transportation systems management. As these case studies show, the journey toward fully integrating ML into transportation is ongoing, but the progress made by these DOTs and others offers valuable guidance for future projects in this highly dynamic field.

ML Tools and Methods

There are numerous tools and frameworks that have been developed to aid in the design, implementation, and deployment of ML models. The ML tool landscape is vast and continuously evolving, with tools for data preprocessing, model development, deployment, monitoring, collaboration, and specialized applications. The emergence of tools requiring minimal or no coding skills, facilitated by advancements in automated machine learning (AutoML) and intuitive user interfaces, is making ML more accessible to a broader audience. Also, many tools and various DL codes are made available to the public on open-source platforms. This democratization of ML is expected to accelerate adoption across various sectors. Furthermore, powerful ML capabilities (e.g., those offered by large language models) are being integrated into easy-to-use interfaces (e.g., chatbots) and common office software (e.g., Microsoft Copilot) to facilitate a seamless transition into practice and to leverage the power of ML for diverse applications. State DOTs could leverage these capabilities not only for specific transportation applications but also for agency operations, workforce development, and office productivity.

The ML field’s rapid expansion necessitates ongoing learning and adaptation by DOTs. New tools and improved versions of existing tools are continuously developed, offering enhanced capabilities and potentially better alignment with project goals. Generative AI models, like ChatGPT and DALL-E, are becoming more versatile and powerful. Although their applications in DOT operations are currently limited, they hold potential for future use in areas such as autonomous driving research and creating synthetic data for model testing and validation. This constant evolution calls for DOTs and other practitioners to remain agile and open to exploring new technologies. As these innovations are mostly being developed by technology companies, state DOTs need to have a trained workforce and a strategy to leverage these innovations for DOT applications.

ML Guide

As part of this NCHRP project, a guide (Samach et al. 2024) comprised of ten main steps was prepared to help state DOTs and other transportation agencies in identifying promising ML applications as well as assessing costs, benefits, risks and limitations of different deployment approaches. The guide is designed to help agencies capitalize on and expand ML capabilities in a broad spectrum of transportation applications. Some of the key insights gleaned while preparing the ML guide are summarized below.

Since ML solutions are data-driven and data-dependent, their effectiveness relies heavily on the availability and quality of data. Unlike traditional predictive methods, ML algorithms learn from data, making it essential for agencies to have a robust digital infrastructure capable of handling big data, storage, and computing needs. Agencies must be mindful of the costs associated with data processing and transmission, as they can significantly impact the overall expenses of ML projects. DOTs have multiple pathways to incorporating ML, including developing custom models in-house or purchasing ML as a service. Each approach comes with its own set of benefits and risks. A best practice identified for ML deployment is to start with small-scale projects to demonstrate value, secure stakeholder buy-in, and work out potential issues before scaling up. This approach helps mitigate risks and manage resources more

effectively. Some ML projects may have an iterative nature as enhancements can be made to the ML models when new data and resources become available. A trained workforce and a basic understanding of ML principles are important factors for the successful deployment of ML projects. This knowledge is crucial for being aware of potential deployment pitfalls, such as model drift and bias, overseeing vendors or consultants, and ensuring the successful application of ML technologies.

State DOTs have an opportunity to build communities of practice within their DOT as well as with other state and local agencies, industry, and academic partners to advance their ML capabilities, share data and trained models, and exchange insights on risks, impacts, and promising ML applications. As numerous problems and needs faced by state DOTs are common (e.g., detecting pavement cracks, detecting and geolocating standard traffic signs), sharing data, models, and expertise can significantly accelerate the development and refinement of ML solutions across the nation. By establishing communities of practice, DOTs can create a dynamic ecosystem that fosters collaborative innovation, maximizes resource utilization, and minimizes duplication of efforts.

Moreover, the partnership with industry and academic entities can introduce state DOTs to the latest ML technologies and methodologies, enhancing their capabilities beyond what would be possible through internal development alone. Industry partners can offer insights into practical applications and scalability of ML solutions, while academic institutions can contribute cutting-edge research findings and theoretical advancements. Together, these collaborations can propel the practical application of ML in transportation, leading to more efficient operations, enhanced safety measures, and improved public services.