TLDR: This research introduces a new AI-powered framework for travel demand modeling that synthesizes realistic daily activity patterns for entire households, accounting for intra-household coordination. It integrates population synthesis, activity generation (using DeepCAM), location assignment, and traffic simulation. Demonstrated in Los Angeles with a 10 million population, the model accurately replicates real-world mobility patterns, matches legacy models’ performance with significantly reduced cost and improved scalability, and efficiently generates travel behavior for large populations.
Understanding how people travel is crucial for urban planning, designing mobility systems, and developing effective policies. Traditional models, known as activity-based models (ABMs), have been foundational but often rely on simplified rules and can be expensive and difficult to adapt to different regions. These models typically focus on individual travel decisions, often overlooking how household members coordinate their activities and trips.
Researchers have introduced a groundbreaking approach to address these limitations with a new travel demand modeling framework. This framework is designed to synthesize daily activity patterns for households based on their socio-demographic profiles, offering a more realistic and comprehensive view of travel behavior. The entire system is generative, meaning it can create new data, is driven by real-world data, is highly scalable, and can be easily transferred to different geographical areas.
A core component of this framework is the Deep Coordinated Activity Model (DeepCAM). While previous work developed a deep generative model for individuals, DeepCAM extends this by specifically incorporating intra-household coordination. This means it can capture shared trips and joint schedules among family members, providing a more accurate representation of real-world travel behavior. DeepCAM learns not only how many people coordinate an activity but also the realistic co-participation structures across different household roles, such as spouses running errands together or children going to school.
The framework operates through a unified pipeline that integrates several key stages. First, a population synthesis module creates a synthetic population that mirrors regional census distributions, assigning attributes to both households and individual members. Next, the activity generation phase begins with the Deep Activity Model (DeepAM) creating an initial activity chain for the household head. DeepCAM then takes this seed schedule and generates activity chains for all other household members, ensuring that shared activities and inter-personal dependencies are captured.
Once activity chains are generated, an activity location assignment module maps each activity to a specific geographical zone. Finally, a large-scale microscopic traffic simulation, utilizing platforms like MATSim, takes these location-annotated trajectories and evaluates mode choices (e.g., car, public transit) and route choices, simulating the resulting traffic dynamics. This comprehensive simulation provides fine-grained trajectories, detailing where, when, how, and why people travel, and allows for validation against observed traffic data.
The effectiveness of this next-generation framework was demonstrated through a full-pipeline implementation in Los Angeles, simulating a population of 10 million residents. The model was trained using data from the 2017 National Household Travel Survey (NHTS) and then transferred to the LA region. Validation efforts compared the model’s outputs against real-world traffic data from Caltrans Performance Measurement System (PeMS) and benchmarked them against the legacy SCAG Activity-Based Model (ABM).
The results were highly promising. The framework closely replicated real-world mobility patterns and matched the performance of legacy ABMs while significantly reducing modeling costs and improving scalability. For instance, the origin-destination matrix achieved a cosine similarity of 0.97 when compared to the SCAG ABM benchmark. Daily vehicle miles traveled (VMT) in the network showed a low Jensen-Shannon Divergence (JSD) of 0.006 and a mean absolute percentage error (MAPE) of 9.8% against the SCAG ABM. When compared to real-world observations from Caltrans PeMS, the evaluation on corridor-level traffic speed and volume reached a JSD of 0.001 and a MAPE of 6.11% for speed, indicating strong consistency with observed traffic dynamics.
Also Read:
- Modeling Digital Societies: A New Approach to Understanding Complex Human Systems
- Advanced AI Improves Autonomous Vehicle Eco-Driving and Traffic Flow
Furthermore, the computational efficiency of the framework is remarkable. Training DeepAM took 3 hours, while DeepCAM trained in just 35 minutes on an NVIDIA A6000 GPU. More impressively, DeepCAM completed inference for one million individuals in only 35 seconds, highlighting its efficiency and scalability for large-scale population synthesis. This research represents a significant leap forward in travel demand modeling, offering a more accurate, flexible, and efficient tool for urban planners and policymakers. For more details, you can refer to the full research paper: Next-Generation Travel Demand Modeling with a Generative Framework for Household Activity Coordination.
