This page is open for editing because it is part of the Incubator. Have something to add? Please register so you can contribute. Have an option you would like to share? Please click on the 'Talk' button to enter the dialogue. The TF Resource Volunteers appreciate your feedback and interest.

This page is part of the Category autonomous vehicles.

A variety of modeling tools may be used to understand the impacts of autonomous vehicles. This section introduces specific model changes in four-step, activity-based, and strategic models that can help better capture behavioral impacts of CAV5s. An initial table is provided for each of the three main model types, followed by a more detailed discussion of how these changes are important for capturing behaviors.

Using Activity-Based Models for CAV5 Analyses

A state-of-the-practice activity-based model (ABM) should be well-situated for autonomous vehicle policy analysis after some modifications, depending on the specific set of policies to be studied. Representing the impacts of CAV5s requires models sensitive to the various behavioral impacts discussed above. The following table provides a starting point for considering which features of the model should be modified for CAV5 analysis.


Using Trip-Based Models for CAV5 Analyses


Using Strategic Models for CAV5 Analyses


Modeling Component Details

The tables above introduced specific model alterations to consider for modeling the impact of CAV5s within different types of models. The sections below provide a more detailed discussion of these individual model components to provide some context on how these changes might be introduced and other considerations for making model changes.

Vehicle Ownership & Availability

A critical upstream model consideration is vehicle ownership and availability. CAV5 ownership could be modeled in addition to, or in place of a conventional vehicle ownership model. Most conventional ownership models include variables of income, number of household drivers, and other parameters like transit access and parking availability or cost. CAV5 ownership models would likely depend on income, but may no longer include number of drivers, since any household member could use the vehicle. Especially for early adopters, age may be an important variable for forecasting ownership of automated versus conventional vehicles, along with education level, based on analysis of stated preference data by Laveri et al. (2017). Additional AV-specific variables such as total daily time spent commuting, or an indicator for households with fewer cars than workers may be useful as well. Aside from stated preference analysis and assumed costs, there is currently little basis to describe future CAV5 owners. As vehicle are introduced in the coming years, it will be necessary to update and deepen these models with relevant parameters.

As a variety of different mobility services arise and change over time, the concept of vehicle ownership may change substantially enough that a traditional vehicle ownership model is no longer applicable for many households. For instance, access to shared AVs and leased access to vehicles may be more widespread than privately owned individual vehicles. This could require adding ownership models for unique types of AVs or considering an entirely new “access to services” approach in place of ownership models. Additionally, it may become increasingly unlikely that a household owns a mix of CAV5 and conventional vehicles, but it may be necessary to include both vehicle types to model a transition time with low, but increasing CAV5 adoption. For true representation of behavior, the model should also consider vehicle availability, or whether or not the CAV5 is already in use by another household member. These potential changes should be understand as caveats to forecasts and future areas to watch as AVs appear on the market.

Coordinated Activity Patterns

Daily activity patterns drive the demand for travel and are represented in our models in several model components. These activity patterns impact the overall demand, the coordination among household and non-household members for traveling together (including chauffeuring activities), and trip chaining. Activity patterns may change in the following ways:

  • AVs could optimize travel for multiple household members, reducing overall household VMT. Current models were developed to accommodate non-optimized behavior so need to be adjusted to assume optimization.
  • AVs could provide more opportunity for trip-chaining among household members (or friends). Current trip generation models cannot account for trip-chaining, but activity-based models already account for this behavior. Strategic models do not currently account for trip-chaining either, but a VMT adjustment could be included to represent this.
  • AVs will induce travel (based on empirical evidence) so an induced demand factor can be included in trip generation to include these additional trips. AB models already represent induced demand, but this should be recalibrated based on current TNC usage to ensure that the additional demand is within expected ranges.
  • On-demand AVs will generate extra service miles to reposition vehicles for the next customer. This can be estimated as a post-process for 4-step models or as a VMT adjustment for strategic models. These adjustments will need to be based on empirical data of current TNCs. AB models will need to include a deadheading model to represent these 0-person vehicles.
  • Owned AVs will generate extra miles when the owner sends the vehicle to park nearby, return home, or pick up another household member somewhere else. These deadheading trips could be included in a trip generation model with trip rates per vehicle (empty) added to existing trip rates per household based on empirical evidence from early experiments (Joan’s chauffeur survey). These vehicle movements could be included as part of activity pattern models of the vehicle in addition to those of the household members. Strategic models could represent this with a VMT adjustment factor.

Location Choice and Land Use

Many of the impacts of AVs will be reflected in short- and long-term location choice and impacts on the built environment. For instance, decreases (or perceived decreases) in travel times by CAV5s would likely impact where people travel, work, and live. CAV5-specific skims and lifecycle variables can be introduced to destination choice models in ABMs or advanced four-step models to measure short-term impacts based on immediate travel characteristics. Longer-term impacts like residential and firm location could also be captured with a land-use model that depends on travel times or logsums that include CAV5 travel times and costs.

Additionally, parking choice becomes important as both a primary and secondary impact of CAV5s, which may require additional model considerations. Parking availability and cost at a destination will impact whether users leave the vehicle at the location (if parking is cheap and available) or send the vehicle further away to park, generating unoccupied (deadheading) to a further parking location, or even returning to home. The amount of deadheading influences additional roadway demand as a zero-occupancy trip, but could also impact terminal walk times between parking location and final destination. Many models today assume some average terminal times, based on land use variables, which may become more dynamic with AV parking models. Capturing the magnitude of deadhead trips will be important to understand and potentially control with policy levers such as parking pricing, drop-off policies, and land use decisions.

Mode Share

Updates to mode share models are critical for testing impacts of CAV5 scenarios. Existing model structures can be modified to reduce in-vehicle travel time for CAV5s relative to conventional vehicles to represent the less-onerous nature of riding in a vehicle versus driving. Existing model forms may also be altered to remove restrictions of age on certain vehicle travel. For instance, age indicators may be removed new options added to reflect the fact that these users may travel alone in CAV5s.

Along with changes in value of time, and reduced restrictions on user types, mode share models may also be expanded to include emerging mobility options such as taxi/TNC services, either fully autonomous or operated by a human driver. Additionally, some distinctions may be necessary to differentiate a private CAV5 mode from automated for-hire CAV5s (emulating today’s TNC/rideshare services, but functioning autonomously), which may be further segmented for single-household users versus shared for-hire CAV5 usage with other riders (operating like a rideshare carpool today). These separate mode could be nested in a way that makes the most sense to the existing mode choice model structure. If added, these modes should also be added to park-and-ride and kiss-and-ride models, so that the transit choice set is reflecting the full suite of access options now available.

Impact and Implementation

Ideally, model improvements that are high impact and easy to implement should be prioritized. In the chart below, we've attempted to map potential improvements relative to impact and implementation difficulty axes. We encourage agencies seeking to provide better information to the CAV5 policy debate to engage in the same exercise.

Impact and Implementation.png

Network Supply

There is significant research on the operational improvements that are possible with CAV5s, but most travel demand models cannot represent these operational improvements in any detail. Dynamic traffic assignment and traffic microsimulation models have emerged as the preferred tool to address these operational improvements, but these tools require significant time and resources to develop so may not be possible in all cases. Some of this research is being developed to quantify the detailed operational impacts so that these outcomes can be translated into parameter assumptions for aggregate trip assignment models.

The aggregate impact of AVs of road usage efficiency can be modeled by altering network parameters like capacity. Assumptions about average capacity increases could be assumed and tested, or more sophisticated modeling could consider the impacts at a link or facility-level as a response to different policy interventions like targeted signal priority. More general testing could focus on the impacts of assuming that V2V and V2I are successfully and widely implemented, thus providing overall capacity increases of x% on freeways and y% on arterials.


Content Charrette: Autonomous Vehicles