Modeling Private Autonomous Vehicles

We're exploring how to model demand for autonomous vehicles (AV's) with a passive data model like CityCast. You can read about the background and motivation for our approach in the first post on the subject.

We've structured our investigations following the thoughts of Lauren Isaac (2016), who proposes that the positive or negative effects of AV's largely rest on whether AV's end up being privately owned or shared among several users. We covered the shared scenarios in our last post, and we're now back as promised with private scenarios.

In these scenarios, individuals own vehicles must as they always have, with the exception that the vehicles can drive themselves, allowing more than one person in a household to use them. A vehicle can carry someone to work, and then return to take another household individual on their daily activity.


  • No AV's People drive their own (conventional) vehicles or take transit.
  • 10% Random Fleet Replacement 1:1 ("Random") Ten percent of private conventional vehicles are randomly replaced with private AV's in a 1:1 ratio.
  • 10% Higher-Income Fleet Replacement 1:1 ("High-Income") Ten percent of conventional vehicles are replaced with higher AV's in a 1:1 ratio. Rather than assuming that the replacement is random, we assume that higher- and upper middle-income households are more likely to acquire private vehicles, particularly in the early stages of the technology when the vehicles are more expensive.
  • 10% Higher-Income Replacement 1:1 with Better Utility ("Better Utility") Ten percent of conventional vehicles owned by higher-income households are replaced with private AV's in a 1:1 ratio. Rather than assuming that the disutility of travel is about equal between conventional vehicles and AV's, we assume that the private AV trips have a better utility. <! data-preserve-html-node="true" data-preserve-html-node="true"-- By how much? -->
  • 10% Lower-Income Replacement 1:1 with Better Utility ("Shared") Ten percent of private conventional vehicles owned by lower- and middle-income households are randomly replaced with shared AV's in a 1:1 ratio. The same utility is used in this scenario as in the previous one. This scenario provides comparative context for the previous shared AV scenarios we ran.

Technical details

In our passive data model, open source MATSim optimizes each person's daily activity pattern, mode choices, and route choices. In MATSim, a person "requests" an AV from a dispatcher. In the shared scenarios, all of the AV's in the system see all of the requests, and the request is handled by the closest available vehicle. In the private scenarios, the AV's only see the requests from members of their own household. This can lead to some situations where an individual becomes "stuck," where the only AV they can access is currently transporting another member of their household. In this case, the stuck individual will walk, take transit, or re-plan their day so that their AV is available when they need it.

In these scenarios, we used the dvrp and drt extensions which are developed to handle demand-responsive services. We wrote our own vehicle search algorithm to filter the list of available vehicles to those whose ID matched the household ID of the person making the trip. Our code is available as the privateAT branch of the MATSim project.


The table below shows the mode share in each of the scenarios.

Scenario Auto AV Transit Walk
No AV’s 87.11 0.00 12.89 0.00
Random 76.77 4.24 15.76 3.23
High-Income 76.55 4.28 15.93 3.24
Better Utility 75.85 10.93 13.22 0.00
Shared 79.73 9.12 11.15 0.00

In the Random and High-Income scenarios, the number of people who walk and who take transit is substantially higher; this is because these individuals get "stuck" and decide that their best available option is to take another mode instead of replanning their schedules. In the Better Utility option, the utility of taking a private AV is so preferred, on the other hand, that people end up replanning their days in order to use their AV when it's available.

One thing that's perhaps a little surprising is that under the Better Utility scenario, the number of trips made on public transit increases relative to the No AV's scenario. This is perhaps a consequence of individuals who take transit to work in the morning and then have their AV collect them in the evening; though this daily pattern is impossible or difficult now, it is something that AV researchers have considered as a possible outcome.

The table below shows the Vehicle Miles Traveled (VMT) by all agents in the simulation (in million VMT). Unsurprisingly, the highest AV VMT is seen in the scenarios where AVs have the lowest cost of use. We're going to be exploring the effects of utility and value of time in a future scenario.

Scenario Auto AV Total
No AV’s 10.027 0.000 10.027
Random 9.233 0.174 9.406
High-Income 9.221 0.176 9.397
Better Utility 8.748 0.374 9.121
Shared 9.590 0.412 10.002

An unexpected outcome of this analysis is that the Better Utility scenario (with private vehicles) sees a lower total VMT than the Shared scenario. The two scenarios use the same utility coefficients, so the discrepancy appears to be largely due to the larger transit mode share in the private scenario.

In a way, this makes some sense. If an AV is available for all trips at all times of the day, there is less reason for individuals to choose something else. If, however, there is some scarcity in the AV's that are available to you then other options might be better.


This does not mean that a private scenario is more "utopian" under Isaac's definition. Remember that in a shared future, the total number of vehicles required to move the residents of Asheville might be substantially lower, reducing infrastructure needs and expenses past the tradeoff point with slightly higher VMT. This is worth consideration.

Above all else, an increase in utility seems to have the largest affect on the mode splits, and so we will explore different utility assumptions in the next blog post in more detail.

In these analyses we have assumed that the future of AV's would be entire private or entirely shared. There is reason to believe that a likely outcome would be a mix of the two. Some who own private AV's may choose to lease their cars to fare-paying passengers when they would be otherwise parked; others might fear to leave their tennis racquets in a vehicle that will be used by strangers. The future is likely to be mixed, determined by both public policy and the success of private endeavors.


In configuring MATSim to run these scenarios, we had a great deal of help from two members of the MATSim team at Technische Universität Berlin, Michal Maciejewski and Joschka Bischoff.