This presentation will share progress from ongoing research for FHWA’s Travel Model Improvement Program (TMIP) on improving travel forecasting using artificial intelligence (AI).  The project will ultimately result in the development of a “Playbook” for how to incorporate AI in travel models and promotion of the Playbook through TMIP webinars, etc.  While acknowledging some research aimed towards entirely replacing existing travel model frameworks, this effort is focused on improving existing models with AI-enhanced components.  In particular, the initial focus of the project is on improving destination choice.  Destination choice is the worst performing component of existing models, both trip-based and activity-based, and therefore offers the greatest opportunity for improvement from AI. 

Given the project’s focus on improving/replacing existing model components, together with the understanding that models used for public policy analysis need to be behaviorally explainable and defensible (not total black boxes), the study is particularly focused on Artificial Intelligence – Discrete Choice Models (AI-DCMs).  AI-DCMs combine the attractive features of AI and traditional discrete choice (logit) models.  On the one hand, they can capitalize on the predictive power of AI, while at the same time, preserving the interpretability and behavioral realism of traditional discrete choice models (DCMs) such as logit models.  Roughly half a dozen different AI-DCM frameworks for combining logit models and deep neural networks have been proposed.  The different proposed architectures for AI-DCM models present different advantages and disadvantages for various choice modeling applications.  For the application of destination choice, however, the study is recommending testing TB-ResNet and L-MNL architectures which present clear interpretability in terms of bounded rationality.  

The project has begun with a review and meta-analysis of the published literature.  There has been some, but relatively limited exploration of AI for destination choice in the transportation literature, but there has been considerable work on the topic in the data science literature where the application is location-based marketing rather than transportation planning.  Over 325 research papers have been identified from 1993 to the present as well as nearly 20 surveys/reviews.  As of now over 100 papers have received at least a cursory review, 40 have received an in-depth review and been summarized.  Over 100 methods have been included in a preliminary meta-analysis based on 472 comparisons in 51 published papers using 112 datasets.  Preliminary results suggest that generative AI methods including generative adversarial networks and large language models perform best, but also indicate that semi-supervised learning methods, and convolution contribute to higher performance. 

In the next phase of the project, which should be well underway by MoMo, selected methods will be tested and compared using real MPO/DOT datasets.  Required data and runtime and computing requirements will be considered together with accuracy in evaluating the methods and recommending which should be carried forward for implementation.  Following preliminary testing, pilot projects will be undertaken to implement the most promising methods and case studies developed for inclusion in the Playbook.  The presentation will share the project’s findings to date at MoMo. 

Abstract background:

In traditional agent-based modeling (ABM) frameworks and simulations for transportation systems, agents follow a set of ``a-priori" models of travel behaviors and use these rules to learn, adapt, and improve their travel plans and actions. The emergence of large language model (LLM) agents presents a novel opportunity to supplement and enhance these frameworks by enabling LLM-based autonomous agents to generate travel decisions dynamically based on contextual inputs. Trained on vast amounts of human-generated data, LLMs can capture intricate patterns in decision-making, language use, and reasoning, allowing them to mimic human-like behavior with remarkable contextual awareness. The integration of LLM agents into ABMs has the potential to improve behavioral realism and modeling flexibility in transportation planning and forecasting applications. This presentation examines the methodological integration of LLM agents into the ABM framework, with a focus on travel behavior representation and hybrid modeling integration.

Description of the abstract:

The presentation centers on two key questions:

1. How can LLM agents be designed to more accurately replicate human travel behavior?
2. How can we integrate LLM agents effectively in an ABM modeling framework?

We first assess the extent to which current LLMs can approximate human travel behavior. By treating LLMs as synthetic survey respondents, we prompt them with socio-demographic attributes and travel scenarios to elicit travel decisions. We then compare their responses against real-world travel behavior datasets to evaluate consistency with observed human decision-making patterns in real-world mode choice and route choice contexts. Our findings indicate that while LLMs demonstrate a strong capability to interpret travel contexts and socio-demographic factors, their choice preferences and valuation of attributes deviate from human travelers. 

To enhance the behavioral realism of LLM agents, we integrate structures grounded in behavioral theories in the agent designs. Specifically, we equip the agents with a memory and perception mechanism that emulates human cognitive processes, allowing for more context-aware decision-making. Additionally, we implement a learning and matching procedure to align the agents' attribute valuations with those observed in human travelers, ensuring more representative and behaviorally consistent travel choices. Further evaluation suggests that these enhancements improve the ability of LLMs to replicate realistic travel behavior patterns.

Finally, we demonstrate the application of our hybrid modeling framework by integrating traditional behavioral models with LLM agents within a simulation environment to assess transportation policies. Using a mini-city case study, we analyze simulation outcomes in terms of agent travel behavior and network-wide traffic flow dynamics. Our results illustrate the proposed framework's potential for enhancing transportation planning and policy evaluation.

Statement on why abstract is worthy:

This work is notable as it is the first work to integrate LLMs as traveler agents into the ABM framework in transportation planning applications. Our results demonstrate the potential of using LLM agents to enhance the ABM paradigm.

Custom GPT-based tools, using LLMs as their foundation, are emerging as innovative instruments to bridge the gap between complex technical analyses and the broader needs of non-technical stakeholders—including executive management, decision makers, and the public. Insight Transportation Consulting has integrated GPT tools with the Central Florida Regional Planning Model (CFRPM 7). The CFRPM is one of the larger models in Florida geographically: it includes 11 counties, covers 20,000 square miles, and evaluates trips of 5 million residents and 75 million visitors annually.

Our presentation will provide both a technical overview and a practical demonstration of three GPTs developed by Insight for the CFRPM7: Ask Me About CFRPM 7 (https://chatgpt.com/g/g-q7pvbsw0q-ask-me-about-cfrpm-7-gpt), CFRPM 7 Traffic Data GPT (https://chatgpt.com/g/g-PIAh6nTAV-ask-me-about-cfrpm-7-data-gpt), and GTFS GPT (https://chatgpt.com/g/g-67a1694f44648191a0777c9c728008c8-gtfs-gpt).

Ask Me About CFRPM7 provides natural language access to the detailed CFRPM 7 documentation, streamlining processes such as model installation, setup, calibration, and scenario analysis (800 pages in total). Users can, for example, inquire, “How was the employment data verified?”, “Please describe the data fields in the roadway network”, “How were speed and capacity values assigned and validated in the CFRPM 7 network?”, and “Please summarize the traffic assignment validation results”. Users receive concise, contextually relevant answers that obviate the need to consult extensive technical manuals. This GPT has been released publicly.

The CFRPM 7 Traffic Data GPT provides natural language access to CFRPM 7’s traffic modeling results. This GPT offers the intersection and roadway segment volumes in 5-year increments between 2015 and 2045, inclusive, for roadways in the 11-county region. Users can, for example, inquire, “What roadway segments have the top 5 largest volumes in Flagler County in 2025?”, “How do the traffic volumes in these top 5 segments compare to the volumes in 2015 and 2045?”,  and “Please provide a bar chart of the intersection volume at the intersection of SR 434 and Hammock Ln in the AM and PM peak periods between 2015 and 2045.” Users can request downloadable bar charts and data files for presentations or other analysis. This GPT is currently undergoing public beta testing and is expected to be released publicly in summer 2025.

The GTFS GPT (https://chatgpt.com/g/g-67a1694f44648191a0777c9c728008c8-gtfs-gpt) provides basic editing of GTFS files with natural language query. CFRPM 7 uses STOPS as its regional transit model, so this GPT will help users modify its GTFS transit network. Users can upload a compressed (zipped) file of the GTFS files to request information or modify the GTFS files. “Please modify the headway of Route #2 to 15 minutes all day”, “List all stops used by Route #101”, and “Add a new Route #101” are some of the commands that the GPT can answer accurately. This GPT is currently undergoing public beta testing and is expected to be released publicly in spring 2025.

Our presentation will include Interface Screenshots, illustrating the natural language query fields and sample responses, and Live Use-Case Visualizations showcasing real-world applications.

Abstract Background

Efficient traffic monitoring is crucial for congestion management, policy decisions, and urban planning. Existing surveillance systems often suffer from poor camera placement or resolution and environmental factors, limiting vehicle detection and trajectory analysis. Traditional solutions to these constraints rely on specialized hardware or extensive recalibration, making large-scale implementation costly and impractical. Notably, acquiring high quality data requires infrastructure upgrades, such as optimal camera positioning, enhanced lens quality, and environmental control. We evaluated the application of a machine-learning algorithm, You Only Look Once (YOLO), with image processing techniques to address these problems. This study proposes a generalizable approach that adapts to diverse camera conditions for accurate vehicle detection and counting. By integrating advanced preprocessing with object detection and tracking models, our method transforms standard traffic cameras into intelligent, multi-purpose traffic sensors without the need for hardware modifications or additional infrastructure modifications.

Description of Abstract

We developed an automated pipeline for real-time traffic data acquisition, preprocessing, object detection, and multi-vehicle tracking. The system uses GPU-accelerated Fast Forward Moving Picture Experts Group tools and available machine learning libraries to process video feeds. To ensure continuous data collection, real-time recording eliminates the disruptions caused by network instability or server limitations, thus providing an uninterrupted video stream. Captured video frames undergo a preprocessing routine to enhance image clarity under varying ambient conditions, to address issues like mist or reduced visibility during dawn or dusk. The enhanced frames are then processed using YOLOv8, for vehicle detection for various scenarios. Vehicle tracking is performed with multi-object tracking algorithm BoT-SORT. One key contribution of this work is a trajectory analysis framework that monitors the evolution of bounding-box center points across consecutive frames, directly computing vehicle movement without reference lines, eliminating the need for predefined geometrical markers and simplifying deployment across traffic scenarios. Validations through side-by-side comparisons with manually confirmed vehicle counts revealed an overall detection accuracy of up to 92% in optimal daylight conditions, 71% in adverse weather (e.g., rainfall), and 78% in low-contrast dusk settings.

Statement on Why Abstract is Noteworthy

The proposed framework offers a scalable, cost-effective solution without extensive retraining or hyperparameter tuning. This work significantly advances the intelligent transportation system (ITS) field, reducing barriers to adoption for city planners, transportation agencies, and researchers. Its ability to adapt to varying camera orientations, resolutions, and environmental conditions illustrates both economic and operational advantages, positioning it as a foundational technology to enhance urban traffic management and congestion mitigation.

Project Status: Expected completion by Spring 2025. Expected Milestones:

●       Integrate a lane line detection module to separate express lane and general-purpose lane flow by May 2025.

●       Extend the classification scheme to distinguish truck categories for a more granular look at freight traffic behavior by May 2025.

The integration of AI into transportation planning has the transformative potential to enhance efficiency, reduce costs, and enable data-driven decision-making. This advancement benefits 420 MPOs across the U.S. and international agencies. However, many planners face challenges due to limited AI education, computational resources, and data transmission capabilities. This abstract introduces our initiative, developed by a team of Arizona State University students, aimed at equipping transportation planners with AI education and practical tools to bridge these gaps.

Our approach focuses on state-wide, nationwide, and global network planning with a multimodal perspective, equipping planners with scalable modeling tools, open data standards, and training programs to bridge the gap between research and practice, fostering more efficient, equitable, and sustainable transportation networks.

Challenges in Transportation Planning:

1. Data Integration Complexity: Planners struggle to harmonize diverse datasets like OpenStreetMap, traffic zones, and points of interest for comprehensive assessments.

2. Scalability Issues: Traditional models fail to efficiently scale from local to global analysis, limiting their applicability.

3. Limitations of Standards & Software: Open standards don’t always translate into effective learning, while commercial software lacks transparency, hindering model understanding.

4. Bridging AI & Planning: Planners explore ML and RL but face challenges in real-world applications.

5. Multimodal Mobility Gaps: Existing models lack flexibility and interoperability, limiting efficient integration of transit, freight, cycling, walking, and MaaS.

Our Approach: Enabling Planners Through AI & Open-Source Tools

To address these challenges, we propose a statewide and global network planning:

• AI-Focused Educational Programs

• Introduce planners to AI decision-making frameworks, ML, RL, and scenario evaluation.

• Provide hands-on training with real-world datasets from state and global transportation systems.

• Practical AI Learning Pathways

• • Demonstrate AI-driven scenario management for improved multimodal planning.

  • Implement modular AI integration for version control, testing, and policy evaluation.

• Advanced Modeling Tools for Planners

• Develop AI-powered tools that enhance accessibility, usability, and transparency beyond traditional models.

• Enable large-scale multimodal network modeling and efficient scenario evaluation.

• Unified Open Data Standards

• Standardize data structures through GMNS to improve consistency and collaboration.

Implications for Practice:

By equipping planners with AI knowledge and open-source tools, we drive:

• Enhancing Multimodal Planning: Scalable, data-driven analyses for better policy and investment decisions.

• Optimizing Efficiency: AI automation streamlines data integration, scenario evaluation, and model validation.

• Bridging AI & Practice: Hands-on training ensures practical ML and RL applications in transportation planning.

• Driving Innovation: AI adoption fosters scenario-based policy modeling for adaptable, sustainable mobility systems.

AI should not just execute models but actively empower planners. Our initiative ensures AI tools are transparent, practical, and scenario-driven, enabling informed decision-making and sustainable transportation planning.