Content uploaded by Younghun Bahk
Author content
All content in this area was uploaded by Younghun Bahk on Aug 23, 2022
Content may be subject to copyright.
Research Article
Transportation Research Record
2022, Vol. 2676(7) 276–295
ÓNational Academy of Sciences:
Transportation Research Board 2022
Article reuse guidelines:
sagepub.com/journals-permissions
DOI: 10.1177/03611981221077982
journals.sagepub.com/home/trr
Private Autonomous Vehicles and Their
Impacts on Near-Activity Location Travel
Patterns:IntegratedModeChoiceand
Parking Assignment Model
Younghun Bahk
1
, Michael F. Hyland
1
, and Sunghi An
1
Abstract
The goal of this study was to analyze the impact of private autonomous vehicles (PAVs), specifically their near-activity location
travel patterns, on vehicle miles traveled (VMT). The study proposes an integrated mode choice and simulation-based parking
assignment model, along with an iterative solution approach, to analyze the impacts of PAVs on VMT, mode choice, parking
lot usage, and other system performance measures. The dynamic simulation-based parking assignment model determines the
parking location choice of each traveler as a function of the spatial–temporal demand for parking from the mode choice
model, whereas the multinomial logit mode choice model determines mode splits based on the costs and service quality of
each travel mode coming, in part, from the parking assignment model. The paper presents a case study to illustrate the
power of the modeling framework. The case study varies the percentage of persons with a private vehicle (PV) who own a
PAV versus a private conventional vehicle (PCV). The results indicated that PAV owners traveled an extra 0.11 to 1.51 mi
compared with PCV owners on average, and the PV mode share was significantly higher for PAVowners. Therefore, as PCVs
are converted into PAVs in the future, the results indicate substantial increases in VMT near activity destinations. However,
the results also indicated that adjusting parking fees and redistributing parking lot capacities could reduce VMT. The significant
increase in VMT from PAVs implies that planners should develop policies to reduce PAV deadheading miles near activity loca-
tions, as the automated era comes closer.
Keywords
planning and analysis, mathematical modeling, simulation modeling, systems modeling, transportation supply, impact analysis,
mode choices
Over the last 10 years, a large volume of research has
focused on modeling and predicting the impacts of
autonomous vehicles (AVs) on travel behavior, travel
demand, and transportation systems broadly. Although
AVs are expected to result in more efficient vehicle oper-
ations that improve traffic flow, most studies suggest
that AVs will also increase overall vehicle miles traveled
(VMT) (1). Given that AVs are not yet widely available,
their overall impact on travel demand and traffic conges-
tion is still uncertain (2). However, to plan for AVs,
including allocating resources for infrastructure invest-
ments and setting policies and regulations, it is important
to model, understand, and forecast the potential impacts
of AVs on transportation systems under a variety of dif-
ferent conditions.
One particular concern about AVs is that they are
expected to drastically increase overall VMT and thereby
increase congestion, energy consumption, and vehicle
emissions. The existing literature identifies a variety of
behavioral changes stemming from the introduction of
AVs that may increase private vehicle (PV) usage and
overall VMT. For example, AVs are expected to decrease
the burden or disutility of PV travel as AVs do not
require a traveler to drive the vehicle, an onerous and
1
Department of Civil and Environmental Engineering, University of
California, Irvine, Irvine, CA
Corresponding Author:
Michael F. Hyland, hylandm@uci.edu
unproductive task, thereby making PV travel less costly
and increasing overall travel and travel distances for a
variety of trip purposes (3–5). As another example, peo-
ple without a driver’s license, seniors, and people with
medical conditions preventing them from driving are
expected to make more trips and increase their vehicle-
based travel when AVs enter the market (6). From a
long-term land use perspective, some people may change
their home locations and work locations as a result of
AVs reducing travel costs to/from major activity loca-
tions (7–9). Also, the improved convenience of PVs will
attract current transit users to switch trips to PAVs,
thereby increasing VMT (10,11).
Additionally, as drivers become riders in PAVs, travel
patterns of PAVs are likely to diverge from travel pat-
terns in private conventional (i.e., non-autonomous)
vehicles (PCVs). PAV travel patterns are likely to involve
dropping off travelers at their exact activity locations
and traveling empty (i.e., deadheading) to another loca-
tion to park during the activity. Although deadheading
in PAVs is similar to current taxi and ride-hailing ser-
vices, in the case of taxis and ride-hailing, the next loca-
tion is likely to be a traveler pickup spot, whereas in the
case of PAVs, the next location is likely a parking spot.
Both PAVs and conventional ride-hailing will inevitably
generate deadheading miles, that is, vehicles driving with-
out passengers. However, the degree of deadheading in
both cases depends on a variety of factors. Recent studies
show that deadheading miles from ride-hailing services,
unsurprisingly, increase road network congestion (12,
13).
This study focused on the near-activity location travel
associated with PAVs and their impact on VMT relative
to the current world with only PCVs. Figure 1 displays
potential travel patterns for PCVs and PAVs for the same
person trip from a home location to an activity location.
Figure 1 shows that PCV travel typically involves a trave-
ler driving to a parking lot and then walking to the activ-
ity location from the parking lot. However, in the case of
PAVs, the AV drives the traveler directly to the activity
location, negating walking, and then deadheads to a
parking location. Notably, because the traveler does not
need to walk from the parking location to the activity
location, the traveler is more willing to choose parking
locations farther away from the activity location (or
allow the AV itself to choose more distant parking loca-
tions) if they are cheaper.
The goal of this study was to develop a modeling
framework to analyze the potential impacts of PAVs on
near-activity travel patterns and overall VMT. Near-
activity travel patterns for PVs denote the travel between
activity locations and parking lots by vehicles and peo-
ple, in cases in which the parking lot is not at the same
location as the activity. To model this problem, this
paper presents an integrated parking location choice and
mode choice model. The parking location choice model
considers factors such as parking fee, parking lot capac-
ity and congestion, driving cost per mile, walking dis-
tance for PCV travelers, and waiting time for PAVs to
pick up travelers for their return home trip. The mode
choice model captures the potential shifts between tran-
sit, shared vehicles like ridesourcing and taxi, and PVs as
a function of the cost and service quality provided by
each of these modes. Moreover, by integrating the mode
choice and parking choice model, the framework cap-
tures the balancing effects of mode shifts toward PAVs
(and to a lesser extent PCVs) and parking lot capacity
and congestion impacts on the attractiveness of PAVs
and PCVs.
The study also presents an iterative solution approach
to solve the integrated mode choice and parking location
choice problem. The output of the model and solution
algorithm includes mode shares, VMT, parking lot occu-
pancy, traveler wait times, traveler walk distances, and
traveler in-vehicle travel time (IVTT). By varying the
percentage of PAVs and PCVs in various scenarios, the
study aimed to analyze the impact of PAVs on overall
VMT. The authors believe that integrating mode choice
Figure 1. Travel pattern of PCVs and PAVs.
Note: PCV = private conventional vehicle; PAV = private autonomous vehicle.
Bahk et al 277
with parking location choice is critical for assessing the
impacts of PAVs on near-activity VMT, since PV travel
is likely to increase in a future with AVs compared with
the current transportation system without AVs.
This paper makes several contributions to the existing
literature. First, it introduces an integrated mode choice
and parking assignment problem with PAVs, and formu-
lates it as a fixed-point problem, to analyze the impacts
of PAVs on near-activity travel patterns and VMT in
particular. Previous research has aimed to analyze the
impacts of PAVs on travel patterns and VMT, but those
studies did not explicitly integrate mode choice and park-
ing assignment. Second, this paper proposes a novel
simulation-based parking assignment model to evaluate
near-activity travel patterns, VMT, parking lot conges-
tion, traveler walking distance, and other important
travel attributes. Third, the paper presents an efficient,
iterative solution approach to solve the integrated mode
choice and parking assignment problem. Fourth, the
paper presents valuable insights into the tradeoffs
between VMT, travel time, and travel costs when com-
paring a system with PCVs versus a system with PAVs.
Fifth, the paper provides insights into the role of parking
lot prices and the spatial distribution of parking lot
capacity can have on VMT.
The remainder of this paper is structured as follows.
The next section provides a brief review of the existing
literature. The Problem Formulation section presents the
mathematical formulation of the integrated mode choice
and parking location choice problem. The section that
follows this presents an iterative solution approach to
solve the integrated model. A case study based on an
artificial central business district is outlined in the subse-
quent section. The computational results from the case
study and associated scenario analyses are then pre-
sented. The penultimate section discusses the implica-
tions of the model results, and the final section presents
the conclusions.
Literature Review
Although many studies have analyzed the factors related
to AVs that affect travel behavior, relatively few have
analyzed the impact of AVs on near-activity location
travel and parking. Moreover, most parking studies
related to AVs focus on microscopic topics such as opti-
mizing parking lot configurations and how to find a
parking location more efficiently (14–17). Conversely,
the current study focused on parking and AVs across a
transportation network to understand and forecast the
potential impacts of AVs on VMT, parking lot usage,
and other relevant metrics for transportation planning
purposes. This section provides a brief review of studies
that analyze the relationship between parking, travel
behavior, and transportation system performance for
PCVs before reviewing the small set of recent studies that
incorporate PAVs alongside the other factors.
The parking location choice problem for PCVs is well
established in the literature. Feeney provides a review of
studies in the 1970s and early 1980s covering the impact
of parking policy measures on travel demand (17). The
behavioral models (mostly logit models) show that fac-
tors such as parking fees and time costs (e.g., walking
time) affect mode choice and travel behavior (18).
Unlike most of the literature that relies on revealed pre-
ference data, Axhausen and Polak employ stated prefer-
ence data to estimate a parking choice model (19).
Specifically, they create a parking type choice set that
includes off-street, surface lot, and multistory parking.
Two other studies develop and use agent-based parking
choice models within MATSim (20,21). Bischoff and
Nagel found that incorporating parking choice in
MATSim for Klausenerplatz in Berlin increased total
VMT estimates by almost 20% (21). Habib et al. incor-
porate parking type choice alongside activity scheduling
decisions within an activity-based travel demand model
(22).
More recently, several studies have analyzed changes
in parking behavior related to PAVs. Table 1 provides a
summary of these studies alongside a summary of the
current study. Levin and Boyles adopt the conventional
multiclass, four-step, trip-based model to predict PAV
travel patterns, assuming some PAVs will drive a traveler
to their activity location before deadheading to the same
traveler’s origin (home) to avoid parking fees near the
high-demand activity center (23). In a PAV-only sce-
nario, Childress et al. found a 50% discount in parking
fees results in a significant increase in VMT (24). Zhang
et al. suggest that PAVs will generate unoccupied VMT
because of the reduction of household vehicle ownership
and deadheading (25). Zhang et al. develop an integrated
parking choice and route choice model (26). Harper
et al. predict that some PAVs will greedily search for
more distant and economical parking spots including
unrestricted parking areas rather than downtown park-
ing lots, thereby increasing VMT (27). On the other
hand, Zhao et al. proposed a centrally controlled parking
system that collects travelers’ destination information
and dispatches the vehicles to the parking lots and found
that this can reduce VMT (28).
It is not possible to compare the results of those stud-
ies directly since they each make different assumptions and
employ different modeling approaches. However, there are
several emerging key factors that illustrate the relationship
between AVs, travel behavior, and VMT. For example,
parking fees and walking time are the most important fac-
tors in parking location choice (17–19,24,27). A vehicle’s
cost per mile is a factor as well. For PAVs, waiting time
278 Transportation Research Record 2676(7)
should be included in behavioral models since travelers
need to wait for pickup after calling the AV, unless the tra-
veler summons the PAV to arrive at the pickup point first,
in which case the PAV may have to wait for the traveler.
The model in this study incorporates all of these factors
into a utility maximization framework for mode choice
and parking location choice.
Problem Formulation
This study presents the integrated mode choice and park-
ing assignment problem, wherein the parking assignment
model captures congestion and capacity constraints in
parking lots throughout the analysis region. Since the
demand for parking is a function of mode choice (i.e.,
higher PV demand increases parking lot congestion), and
mode choice is a function of parking congestion (i.e.,
congestion in parking lots reduces demand for PVs), this
study models the integrated mode choice and parking
assignment problem using a fixed-point problem formu-
lation. In general, a fixed point of a function, fðÞ,isa
value, p, such that fpðÞ¼p, or put another way, the
value pis unchanged by function fðÞ(29). The variable
pcan be a scalar or a vector.
Equation 1 displays the general form of the integrated
mode choice and parking assignment model in the form
of a fixed-point problem. A solution to Equation 1 is a
multidimensional array of probabilities, p, that when
input into fmfpðÞ
remain unchanged. The parking func-
tion fpðÞ in this study does not have a straightforward
functional form, rather, this study employs a dynamic
simulation-based parking assignment model that is
detailed in the next section. Equation 2 shows that the
function fpðÞ is nonseparable because the mode splits
(podt )between each origin o2Oand destination d2D
at time interval t2Timpact the service quality, price,
and therefore parking location choice for travelers using
the parking system between all origins, all destinations,
and all future time periods. Conversely, the mode choice
function, fmðÞ, displayed in Equation 3, which returns
mode splits for travelers going from origin o2Oto des-
tination d2Dat time interval t2T, is separable by ori-
gin, destination, and time interval.
p¼fmfppðÞ
ð1Þ
q¼fppðÞ ð2Þ
podt ¼fmqodt
ðÞ ð3Þ
where
M= set of modes in the transportation system,
indexed by m2M;
O= set of origin zones in the transportation network,
indexed by o2O;
Table 1. Comparison of PAV Studies on Parking Behavior
Study Purpose Approach Parking-related findings
Levin and Boyles (23) Analyze impact of AVs on travel
behavior and network
performance
Four-step, trip-based travel
forecasting model
PAVs increase VMT owing to
deadheading to cheap parking
Childress et al. (24) Quantify impacts of AVs on
travel behavior and network
performance
Activity-based travel forecasting
model
Improved road capacity, reduced
VOT, and discounted parking
fees increase PAV demand and
VMT
Zhang et al. (25) Quantify excess VMT stemming
from vehicle deadheading
Household travel model.
Greedy scheduling algorithm
for required household
vehicles. Mixed-integer
program for unoccupied VMT
Reduction of household vehicles
increases VMT because of
unoccupied PAV travel
Zhang et al. (26) Quantify network equilibrium
patterns under AV parking
behavior
Integrated route choice and
parking assignment choice
model and solution approach
PAVs increase traffic congestion
resulting from parking search.
Some PAVs will park at home
Harper et al. (27) Evaluate impact of AVs on VMT,
emissions, and parking
revenues
Agent-based parking simulation
model on grid network with
greedy parking lot selection
PAVs park at distant and
economical parking locations
and increase VMT
Zhao et al. (28) Analyze improvements in
congestion under centralized
parking dispatch
Optimization of parking control
with macroscopic fundamental
diagram
Optimized parking assignment
reduces cruising VMT for
parking
This study Estimate impacts of PAV parking
travel patterns on VMT and PV
demand
Integrated mode and parking
location choice model.
Iterative solution approach
PAVs increase demand for PV
travel and as a result, VMT
increases
Note: Note: PV = private vehicle; AV = autonomous vehicle; PAV = private autonomous vehicle; VMT = vehicle miles traveled; VOT = value of time.
Bahk et al 279
D= set of destination zones in the transportation net-
work, indexed by d2D;
T= set of time intervals in the analysis period,
indexed by t2T;
Km= set of service quality and cost/price attributes
associated with mode m, indexed by k2Km;
podtm = choice probability for mode m, for a traveler
going from origin o, to destination d, and departing at
time interval t;
podt = vector of mode choice probabilities for origin
o, destination d, and departing time interval t, with
dimension M
jj
;
p= multidimensional array of mode choice probabil-
ities for all modes, origins, destinations, and time inter-
vals, with dimension M
jj
3O
jj
3D
jj
3T
jj
; and
q= multidimensional array of service quality and
price attributes for all modes, origins, destinations, and
time intervals, with dimension M
jj
3O
jj
3D
jj
3
Tjj3Kjj.
The next section describes the detailed agent-based
parking simulation model, fpðÞ; it also provides the func-
tional form and the parameters for the mode choice func-
tion, fmðÞ, which is a standard multinomial logit model.
Solution Approach
Figure 2 displays the proposed iterative solution
approach to solve the integrated mode choice and park-
ing assignment problem. The remainder of the section
describes the iterative solution approach along with the
model input and output.
Model Inputs
The left-most box labeled ‘‘Input’’ in Figure 2 includes a
scenario setting box that leads into an input parameters
box. This study performed sensitivity and scenario
analyses based on changes in a variety of model para-
meters. These parameters and the changes to them are
detailed in later sections. The input data and parameters
in this study include the available travel modes, mode
choice model parameters, fixed modal attributes for
nonPV modes, the location, capacity and price of park-
ing lots, parameters for the parking congestion model,
the transportation network, and demand data including
trip origins and destinations. The following subsections
provide details about the available travel models and the
mode choice model parameters.
Travel Modes. This study incorporates three types of high-
level travel modes, PVs, shared vehicles (SVs), and public
transit. PVs include PCVs and PAVs. SVs include
shared-use automated vehicles (SAVs,) ride-hailing and
ride-sharing services, and taxis. SV travelers wait for a
vehicle, travel inside an SV, pay a fare, and receive door-
to-door service. Public transit effectively refers to high-
capacity buses. Transit riders walk to a bus stop, wait
for a bus, pay a fare, travel inside the bus as a rider, and
walk to their destination—they may also need to transfer
between routes, but this study assumes transfers are not
necessary.
Specific scenario details are given in the Case Study
section; however, it is important to note that each trave-
ler has access to a single PV—either a PCV or PAV but
not both—in this study. Additionally of note, in the sce-
narios with all PCVs, SVs are conventional vehicles
(SCVs); conversely in the scenarios with all PAVs, the
SVs are all SAVs.
Mode Choice Model Parameters. Important mode choice
model parameters include the disutility of travel time for
in-vehicle travel and out-of-vehicle travel (walking and
waiting) and the disutility of travel costs. Combining
the disutility of travel time and travel costs produces
Figure 2. Solution approach.
Note: PV = private vehicle; SV = shared vehicle.
280 Transportation Research Record 2676(7)
estimates of a user’s value of time (VOT). According to
previous studies and reports, VOT varies widely depend-
ing on a variety of factors (19,30). Axhausen and Polak
found a wide range of walking VOT estimates ranging
from $1.35/h to $47.43/h in the mode choice context and
$7.67/h to $58.21/h in the parking choice context (19).
Caltrans uses the following VOTs: $13.65/h for automo-
bile and transit in-vehicle VOT and $27.30/h for transit
out-of-vehicle VOT in 2016 dollars (31). Kolarova et al.
estimated VOT from the German national household
travel survey data, segmented by mode and income class
(3). Based on the middle-income class’s PCV commuting
trips ($8.18/h), the other values of in-vehicle VOT were
$5.26/h, $8.72/h, and $4.89/h for PAV, SAV, and public
transit, respectively. The walking VOT was $12.03/h,
whereas the AV and public transit waiting VOT were
$9.49/h and $7.45/h, respectively. Zhong et al. provide
VOT ranges for PCV, PAV, and SAV in the United
States by place of residence: $9.36/h (rural) to $53.71/h
(urban) for PCV; $7.71/h to $40.89/h for PAV; and
$8.64/h to $46.53/h for SAV (5). This study mostly used
the values in Kolarova et al. (3).
Moreover, this study used $ 0.50/mi as the cost per
vehicle mile of travel, based on the 2020 electric vehicle
cost provide by the American Automobile Association
(32).
According to several studies, about 40% of a ride-
hailing service travel is deadheading miles (13,33). In
other words, when 1 mi of PCV travel from an origin to
a destination (except the parking travel distance) is chan-
ged to ride-hailing vehicle travel, the travel distance
becomes 1.67 mi (67% extra travel). Considering that
PAVs do not cruise to find and then pick up another
passenger, the PAV’s VMT increase only depends on the
empty miles driven from activity location to parking lot.
Iterative Solution Approach
The middle portion of Figure 2 displays an overview of
the proposed solution approach that involves iterating
between the mode choice model and the dynamic
simulation-based parking assignment model. In the itera-
tive process, the output of the parking assignment model
is the performance of the transportation system, specifi-
cally the costs and service quality attributes associated
with PAV and/or PCV travel. Given that the parking
model is agent-based, these cost and service quality attri-
butes are available at the agent level and can easily be
aggregated over time and space (e.g., travel analysis
zones). The costs and service quality attributes for the
other modes—transit and SV—were fixed in this study.
The costs and service quality modal attributes for PVs
from the parking assignment model are the inputs to the
mode choice model, alongside the fixed modal attributes
for SVs and transit. The outputs of the mode choice
model are the modal splits, which are the inputs for the
next iteration of the parking assignment model. This
iterative process repeats until there is consistency
between the mode choice model and the parking assign-
ment model in relation to modal service quality/costs
and modal splits.
The following two subsections describe the dynamic
simulation-based parking assignment model and the mul-
tinomial logit mode choice model, respectively.
Dynamic Simulation-Based Parking Assignment Model. The
mode choice model returns modal splits, sn, where the n
superscript denotes the current iteration number. Given
that the modal attributes for SV and transit are fixed,
and these modes do not use parking lots, only the modal
splits for PV are needed as input for the parking assignment
model, sn
m¼PV . The formula for the spatial (origin zone to
destination zone) and temporal demand for PVs in the cur-
rent iteration, sn
o;d;t;PV , is displayed in Equation 4,
sn
o;d;t;PV ¼pn
o;d;t;PV 3Dodt 8o2O;8d2D;8t2Tð4Þ
where Dodt denotes the total trip demand from origin
zone, o, to destination zone, d, departing at time t, which
is exogenous to the integrated model system, meaning it
is independent of the iteration. Notably, the demand for
origin zone, o, to destination zone, d, is based on aggre-
gating traveler agents with origin nodes that are inside
origin zone, o, and destination nodes that are inside des-
tination zone, d.
Each traveler agent in the dynamic simulation-based
parking assignment model must choose a parking lot,
where Ais the set of parking lots, indexed by a2A.In
this study, each traveler creates an ordered list of parking
lot preferences, based on their own expected generalized
cost of travel. Each traveler with a PCV drives from their
origin to a parking lot before walking from the parking
lot to their activity location. Each traveler with a PAV
rides from their origin to their activity location (i.e., des-
tination node), after which the vehicle deadheads to a
parking lot. Equations 5 and 6 display the expected gen-
eralized cost functions for PCV travelers and PAV trave-
lers respectively.
ECPCV;a¼VOTtrv 3ttrv;a+VOTwlk 3twlk;a+CPM 3da8a
ð5Þ
ECPAV ¼VOTwt 3twt;a+CPM 3d8að6Þ
where
VOTtrv = in-vehicle VOT;
ttrv;a= travel time between the origin and parking lot, a;
VOTwlk = walking VOT;
Bahk et al 281
twlk;a= walking time between parking lot, a, and the
traveler’s destination;
VOTwt = waiting VOT;
twt;a= length of time the traveler must wait at the
activity location to be picked for their return home trip,
when their PAV is in parking lot, a;
CPM = vehicle’s cost per mile; and
da= travel distance between the traveler’s origin and
parking lot, a(via the destination in the case of PAVs).
The parking assignment model simulates the move-
ments of PAV and PCV travelers and the vehicles
themselves as well as the occupancy of parking lots, in a
time-driven simulation. Therefore, the simulation cap-
tures the current location of travelers, PAVs, and PCVs
as well as the current occupancy of all parking lots in the
transportation network, every time step, which is denoted
Dt (and equal to 6 s in this study).
Each traveler has an ordered list of parking lots
because it is possible that a parking lot is full when the
PV arrives at the parking lot entrance in the simulation,
in which case the traveler or the traveler’s PAV needs to
travel to the next parking lot on their ordered list. Of
note, this study assumes a traveler only becomes aware
of a parking lot’s occupancy when they arrive at the
parking lot—future studies may assume travelers always
have full knowledge of parking lot occupancies.
Additionally, since travelers can go from parking lot to
parking lot in the simulation, the expected costs for a
parking lot a,ECa, in a traveler’s ordered list does not
reflect the detour travel time and distance that occurs in
the simulation. Therefore, the ordered list of parking lots
for an agent is fixed within the current iteration of the
model, that is, an agent does not update their ordered
parking list during a simulation.
In addition to capturing hard capacity constraints at
each parking lot in the transportation network, the park-
ing assignment model also captures in-lot parking search
time. This is an important model feature for dense urban
areas with limited parking supply, as drivers can spend
considerable time inside parking lots finding an open
parking spot. In this study, the parking time after enter-
ing the parking lot (in-lot parking time) depends on the
volume to capacity ratio of the parking lot. For example,
this study used a Bureau of Public Roads function to
reflect the in-lot parking time, expressed as Equation 7,
tprk va
ðÞ¼t031+ava
Ca
b
()
ð7Þ
where
tprk = in-lot parking time;
va= number of vehicles currently in parking lot a
(parking and searching for parking),
Ca= capacity of parking lot a, and
a,b, and t0are model parameters to be calibrated based
on data.
The parking assignment model also captures network
IVTT and network walking time. The simulation model
assumes both vehicles and pedestrians travel along the
shortest network path. The model does not currently cap-
ture congestion in the road network, as the assumption is
that parking lot capacity is the limiting constraint on PV
mode demand; nor does it capture congestion or capacity
at drop-off points (i.e., activity locations).
As noted in Figure 2, the simulation-based parking
assignment model returns the service quality and costs
for PV modes. It does so by taking the average values for
service quality and cost from all traveler agents with ori-
gin o, destination d, departure time t, and PV mode m,as
denoted in Equation 8,
qn
o;d;t;m;k¼Pr2Rdr;n
o;d;t;mqr;n
k
Pr2Rdr;n
o;d;t;m
8o2O;8d2D;8t2T
8m2M;8k2Km
ð8Þ
where
dr
o;d;t;mis an indicator variable equal to 1 if agent rhas
origin o, destination d, departure time t, and was assigned
to mode min iteration n; and
qr;n
kis agent r’s experienced service quality or cost
metric k’s value in iteration n.
The set of experienced service quality or cost metrics, Km,
vary by PV mode, m. For PCV, KPCV includes IVTT from
origin to parking lot, in-lot parking time, parking fee,
walking time from/to the parking lot, the opposite direc-
tion IVTT, and vehicle travel distance to calculate vehicle
parking cost. Conversely, for PAV, KPAV includes include
IVTT from origin to destination, total vehicle travel dis-
tance, parking fee, and the waiting time for the PAV to
pick up the traveler for the return home trip. The values
in Equation 8 are then fed into the mode choice model.
Multinomial Logit Mode Choice Model. This section describes
the mode choice model. The study employed the random
utility maximization framework to model mode choice.
The utility function for each mode can be written as
Equations 9–13,
UPCV ¼bIVTT;PCV ttrv +tprk
+bwlk twlk +bcost coprd+cprk tdur
+2ð9Þ
UPAV ¼bIVTT;PAV ttrv +bwt twt +bcost coprd+cprktdur
+2
ð10Þ
USCV ¼bSV +bIVTT;SCV ttrv +bwt twt +bcost cfr +2ð11Þ
USAV ¼bSV +bIVTT;SAV ttrv +bwt twt +bcostcfr +2ð12Þ
UTransit ¼bTransit +bIVTT;Transitttrv +bwlk twlk +bwt twt
+bcostcfr +2ð13Þ
282 Transportation Research Record 2676(7)
where
ttrv = path travel time (origin to the final parking lot
entrance);
tprk = in-lot parking time from Equation 7,
twlk = walking time,
twt = waiting time,
copr = vehicle operating cost per mile,
d= vehicle driving distance (including parking lot
searching travel),
cprk = parking fee per hour,
tdur = parking duration time,
cfr = fare of shared vehicle or transit,
bSAV and bTransit = mode-specific coefficients,
bIVTT = IVTT coefficient,
bcost = cost coefficient, and
bwlk and bwt = coefficients for each variable.
Among those variables, ttrv,tprk ,cprk, and dchange in
the parking assignment model, and the other non-beta
parameters and variables remain unchanged based on
scenario settings. bIVTT;PCV ,bIVTT;PAV ,bwlk ,bwt , and bcost
are frequently used variables in the mode choice model
and can be found in many studies in the literature.
Wardman (34) and a Transit Cooperative Research
Program Report (35) collect and list relative time valua-
tions for transit travel from decades of studies in the UK
and the United States, respectively.
The study also assumes that the error terms, E, are
independent (across modes and agents) and identically
distributed. Therefore, the functional form for mode
choice is the multinomial logit model. Given the modal
attributes from the previous iteration of the parking
assignment model, the exogenous modal attributes and
other parameter values, as well as the beta coefficients,
determining the mode choice probabilities, p, from the
multinomial logit model is straightforward and computa-
tionally inexpensive.
Model Output
After the iterative solution approach converges to a solu-
tion, there are a variety of system-level and agent-level
performance metrics that can be output for analysis pur-
poses. The system-level metrics include VMT, empty
VMT, final mode splits, parking lot occupancy, and
parking lot revenue. The agent-level metrics include
travel time, walk time, travel cost, generalized cost, and
systematic utility.
Case Study
Network Configuration
This study used a grid network describing an imaginary
central business district (CBD). The network, displayed
in Figure 3, has 8 external origin nodes (Nodes 1 to 8),
22 activity locations (Nodes 9 to 30), and 10 in-network
parking lots (Nodes 31 to 40) with 1 out-of-network
parking lot (Node 41) that accommodates unassigned
vehicles. The size of a block is 600 3500 ft and the
width of the road is 60ft. The main road links are unidir-
ectional with a uniform vehicle speed (25 ft/s) and a uni-
form walking speed (4 ft/s).
Parking assignment requires a fine spatial resolution,
particularly in the CBD. Each intersection is divided into
four nodes to reflect intersection delays. Each internal
short link in each intersection has additional travel times:
12 s for the through direction and 24 s for a left turn and
a U-turn. Each activity location and parking lot has two
bidirectional links connected with the main road that
take 18 s to traverse and are accessible only from the
adjacent direction links (i.e., only right turn is available),
and a short detour, such as a U-turn, at the downstream
intersection is required for the opposite direction travel.
For example, assume a PAV with External Origin 3 and
Activity Location 21 parks in Lot 36 in Figure 3, the
node sequence of the path would be [3, 539, 537, 122,
535, 533, 531, 529, 530, 549, 550, 212, 21, 211, 550, 552,
362, 36].
Trip Generation and Distribution
Vehicle trips were generated every 6 s (Dt ¼6s) and the
simulation ran for 4 h (i.e., there were 2,400 time steps
during the process). To collect enough samples, each
simulation ran three times (i.e., 3 days). There were
12,000 entering trips including PV and nonPV (SV or
public transit) users per scenario. To balance the parking
location availability throughout the day, 3,000 PVs ran-
domly exited the parking lot during the analysis period.
The 12,000 entering trips had uniformly distributed ori-
gin and destination nodes (and zones) and departure
times. Additionally, the model did not explicitly model
travel from activity location or parking lot to external
origin. Rather, the study used fixed values for PAV user
pickup wait time and IVTT to external origin.
Parking Lots
Each parking lot had a fixed parking capacity and a
fixed parking fee. The total parking capacity across the
10 parking lots was 4,000 and 15% of parking spots
(600) were vacant at the beginning in the base scenario.
The Results section includes scenario analyses with
respect to changes in parking fees and parking lot capaci-
ties. Parking fees ranged from $1.5 to $5.5/h with mean
(median) values of $3.65/h ($3.75/h). Parking lot fees
were based on lots in major cities in Germany and the
United States (36). When all parking lots were full, vehi-
cles had to go to the out-of-network parking lot (Lot
Bahk et al 283
41), which is 0.5 mi away, costs $5.5/h, and has a capac-
ity of 10,000.
For the in-lot parking space search time function
(Equation 7), the study used the following parameter val-
ues for all parking lots: t0¼1min and a¼b¼2.
According to the function, parking time is 60 s when the
parking lot is empty, 90 s at 50% vacancy, 120 s at 30%
vacancy, and 180 s at 1% vacancy.
Model Parameters and Values of Time
The model parameters and VOTs used in this study were
based on those in Kolarova et al. (3). Since there is no
experience of AV travel yet, the value of AV travel time
in most studies relies on SP survey or assumptions. The
San Diego Association of Governments multiplies 0.75
from the PCV in-vehicle VOT as a modifier considering
the improved convenience (37), which is the same as
Figure 3. Grid network for parking assignment simulation.
284 Transportation Research Record 2676(7)
Correia et al. (38). Conversely, Kolarova et al. estimate
that in-vehicle VOT in PAVs is 0.64 of in-vehicle VOT in
PCVs (3). According to a review paper by Singleton, sev-
eral simulation studies assume various VOTs of AVs,
and the value ranges from 0% to 100% of PCV VOT
(39). This study applied Kolarova et al.’s survey-based
number in the base scenario and adjusted the number in
alternative scenarios with different modifiers (3). Note
that Kolarova et al.’s SAV refers to a ‘‘driverless taxi’’ in
their survey (3). The coefficient values used in this study
are shown in Table 2.
Travel Costs for Mode Choice
The mode choice model includes out-of-network IVTT
since the mode choice is not only based on the travel in
the simulated network, but is also affected by the whole
travel path. For each traveler, the out-of-network IVTT
time for two directions were added to the in-network
IVTTs (including the parking lot searching time) deter-
mined by the parking assignment model. PV and SAV
users’ out-of-network IVTT were set to 20 min per one
way, and transit users’ out-of-network IVTT was set to
30 min per one way. Including the in-network IVTT, the
total IVTT becomes around the U.S. average (27.6min
for one-way commute) according to recent data (40).
Assuming the average speed is 24 mph, the out-of-
network one-way travel distance is 8 mi.
This study assumed 10 min (5 min in each direction)
of waiting time for SAV and 20 min (10 + 10) of waiting
time and 10 min (5 + 5) of walking time for transit.
Transit fare was $5 (thus, $10 for two-way trips). Uber
fares consist of a base fare ($2), cost per minute ($ 0.4/
min), and cost per mile ($1/mi), which can be changed
when the company starts to run AVs (41). For SAVs,
Chen and Kockelman use $ 0.75 to 1.00/mi (42),
Kaddoura et al. assume $ 0.64 to 0.84/mi (e0.35 to 0.46
per kilometer) (43), and An et al. estimate $ 0.66/min
(44), which is a simplified cost estimation of the current
Uber service. Considering those studies, this study used
$1.2/mi for SCV fares and $ 0.8/mi for SAV fares.
Scenarios
This study analyzed several scenarios that reflect various
possible future conditions at different points in time.
Table 2. Model Parameters
Variable PCV PAV SCV SAV Transit
Mode-specific constant 0 0 20.927 20.927 23.23
Time-related In-vehicle time (minutes) 20.0966 20.0621 20.11 20.103 20.0577
Walking time (minutes) 20.142 na na na 20.031
Waiting time (minutes) na 20.112 20.112 20.112 20.088
Cost-related Operating cost and parking fee (USD) 20.709 20.709 na na na
Fare (USD) na na 20.709 20.709 20.709
Note: PCV = private conventional vehicle; PAV = private autonomous vehicle; SCV = shared conventional vehicle; SAV = shared autonomous vehicle; na =
not applicable.
Table 3. Scenarios
Scenario Modes PAV VOT
Parking
fee
Parking
lot capacity
PAV
percentage Description
One origin and one destination spatial aggregation in mode choice
Base PCV, SCV, Transit na Varied Uneven 0 CV default
A PAV, SAV, Transit 64.3% of PCV Varied Uneven 100 AV default
B1 PAV, SAV, Transit 90.0% of PCV Varied Uneven 100 Variations in PAV
VOT parameterB2 PAV, SAV, Transit 50.0% of PCV Varied Uneven 100
C1 PAV, SAV, Transit 64.3% of PCV Uniform Uneven 100 Variations in cost variable
C2 PAV, SAV, Transit 64.3% of PCV Uniform Even 100
One origin and four destinations spatial aggregation in mode choice
D1 PCV, SCV, Transit 64.3% of PCV Varied Uneven 0 Variations in PAV
ownership percentageD2 PCV, PAV, SCV, SAV, Transit 64.3% of PCV Varied Uneven 25
D3 PCV, PAV, SCV, SAV, Transit 64.3% of PCV Varied Uneven 50
D4 PCV, PAV, SCV, SAV, Transit 64.3% of PCV Varied Uneven 75
D5 PAV, SAV, Transit 64.3% of PCV Varied Uneven 100
Note: CV = conventional vehicle; AV = autonomous vehicle; PCV = private conventional vehicle; PAV = private autonomous vehicle; SCV = shared
conventional vehicle; SAV = shared autonomous vehicle; VOT = value of time; na = not applicable.
Bahk et al 285
Table 3 displays the full set of scenarios. In the base sce-
nario, all PVs were PCVs. Those PCVs were all converted
into PAVs in Scenario A. Scenarios B1 and B2 were all
PAV scenarios, but they applied different in-vehicle
VOTs for PAV, 50% and 90% of PCV in-vehicle VOT,
respectively. Scenarios C1 and C2 were also all PAVs but
uniformly applied a $3.5/h fee to all parking lots, and
Scenario C2 additionally attempted to evenly distribute
parking lot capacity across the network. Table 4 displays
the parking lot fees, capacity, and initial vacancy across a
variety of scenarios.
In the base scenario and Scenarios A to C2, the trave-
ler agents were aggregated into a single origin zone and
single destination zone for the mode choice model.
However, in Scenarios D1 to D5, the traveler agents
were aggregated into four destination zones in the mode
choice model. Scenarios D1 through D5 varied the pro-
portion of PVs that were PAVs, as opposed to PCVs,
between 0 and 1 in increments of 0.25.
Results
No Spatial Disaggregation Scenarios
The Solution Approach section and Figure 2 describe an
iterative solution approach to solve the fixed-point inte-
grated mode choice and parking location choice problem,
p¼fmfppðÞ
:However, when the variable pis a scalar or
low-dimensional vector, and the function fmfppðÞ
is rel-
atively easy to evaluate, it is possible to use enumeration
to solve the fixed-point problem. The base scenario and
Scenarios A through C2 met these criteria because the
mode choice model did not include any spatial or tem-
poral disaggregation; therefore, the dimension of pis
131313M
jj
. Using an enumeration method also has
the added benefit of ensuring that all fixed points are
identified, whereas the iterative solution approach may
not identify all possible fixed point solutions.
Figure 4 shows the results of the enumeration
approach for the base scenario and Scenarios A through
C2. The x-axis displays the input values for the PV mode
share, pm¼PV :The y-axis displays the evaluation of the
integrated parking assignment and mode choice func-
tion, fmfppm¼PV
ðÞ
. Values along the diagonal represent
solutions to the fixed-point problem.
Using an increment of 1 %, Figure 4 shows that there
is a unique solution for the base scenario and Scenarios
A through C2. Unsurprisingly, the lines are all downward
sloping. Moreover, the relative flatness of Scenarios C1
and C2 likely stems from the parking fees across the net-
work being uniform. The existence and uniqueness of a
solution for all scenarios engenders a straightforward
analysis of the fixed-point solutions across scenarios.
Table 4. Parking Lot Information Across Scenarios
Parking lot
Default Scenarios C1 and C2 Scenario C2
Parking fee (USD/h) Capacity (veh) Initial vacancy (veh) Parking fee (USD/h) Capacity (veh) Initial vacancy (veh)
31 3.5 450 60 3.5 250 38
32 2 1,100 200 3.5 250 37
33 3 250 75 3.5 750 112
34 4 190 45 3.5 400 60
35 5 140 20 3.5 500 75
36 5.5 420 25 3.5 550 83
37 5 320 40 3.5 350 52
38 1.5 550 30 3.5 300 45
39 3 380 60 3.5 300 45
40 4 200 45 3.5 350 53
41 5.5 10,000 10,000 5.5 10,000 10,000
Figure 4. Fixed-point solutions for private vehicle mode choice
probability, Pr(PV).
Note: Pr(PV) = choice probability of private vehicle.
286 Transportation Research Record 2676(7)
Mode Share Metrics. Figure 5 shows the mode shares for
all modes in each scenario. The mode share for PV was
lowest in the base scenario in which the PVs were PCVs,
and the mode share was 50%. In Scenario A, in which all
PVs were PAVs, PV mode share significantly increased
to 87% owing to eliminating walking time, potentially
reducing parking fees, and the reduction in IVTT
disutility.
In Scenario B2, the assumption was that PAV in-
vehicle VOT was 50% of PCV in-vehicle VOT, and the
PAV mode share increased all the way to 92%. In
Scenario B1, when PAV in-vehicle VOT was 90% of
PCV in-vehicle VOT, the PAV mode share was 74%.
Taken together, Scenarios A, B1, and B2 unsurprisingly
indicated that PAV IVTT disutility had a significant
impact on mode share.
The properties of parking lots also affected the choice
probability. Instead of the varied parking fees that ran-
ged from $1.5 to $5.5/h in the base scenario, all parking
fees were set to $3.5/h in Scenarios C1 and C2. In addi-
tion, Scenario C2 redistributed the parking lot capacities
to be more even in the network. In Scenarios C1 and C2,
the PAV mode shares were 67% and 69%, respectively.
This represents a notable reduction in mode share com-
pared with Scenario A, in which the larger parking lots
had lower fees.
VMT Metrics. In addition to the increase in travel demand
(Figure 6a), total PV VMT substantially increased in the
PAV scenarios (Figure 6b). Note that we only considered
in-network VMT (starting from external origin nodes),
and did not include the VMT from the actual origin to
external nodes. In-network VMT increased by 15,000 to
22,000 mi in Scenarios A, B1, and B2 compared with the
base scenario. Conversely, the increases were reduced
when there was no difference in parking fees in Scenarios
C1 and C2.
As shown in Table 5 and Figure 6c, the average VMT
for a PCV was 1.07 mi in the base scenario, and the aver-
age VMT for a PAV stretched from 1.18 to 2.57mi in the
other scenarios. The VMT per vehicle increased by 1.38
to 1.51 mi/veh in default parking lot settings compared
with the base scenario. In Scenarios C1 and C2, VMT
per vehicle increased by only 0.11 to 0.35mi/veh. This
clearly indicated that the spatial distribution of parking
prices and parking supply had a significant impact on
average VMT per vehicle. Therefore, if policy makers
and planners are interested in reducing VMT in a future
era with PAVs, parking supply and pricing must be con-
sidered alongside other policy measures.
The increase in in-network VMT from PAVs shown in
Figure 6bstemmed from both an increase in PV trips
(shown in Figure 6a) and an increase in network VMT
per vehicle (shown in Figure 6c). Therefore, VMT in a
future with AVs is likely to increase because of travelers
switching to PVs and also driving more miles in PAVs
than they did or would have done in PCVs. Policy mak-
ers interested in decreasing VMT are likely to need a mul-
tipronged approach to address these two factors that are
expected to increase VMT.
Travel Time and Travel Cost Metrics. Table 5 shows the aver-
age travel time components for travelers along several
dimensions, along with average total travel time, and
Figure 5. Mode share by scenarios.
Note: PCV = private conventional vehicle; PAV = private
autonomous vehicle; SCV = shared conventional vehicle; SAV =
shared autonomous vehicle.
Table 5. Average PV Traveler Distances and Times.
Scenario
Number
of PVs (veh)
Travel distance
(mi/veh)
In-lot parking
time (min/veh)
IVTT
(min/prs)
Walking time
(min/prs)
Waiting time
(min/prs)
Total time
(min/prs)
Base 6,000 1.07 2.45 8.19 12.73 – 20.9
A 10,440 2.53 2.71 4.83 – 2.37 7.2
B1 8,880 2.45 2.70 4.83 – 2.40 7.2
B2 11,040 2.57 2.72 4.84 – 2.35 7.2
C1 8,040 1.42 2.61 4.84 – 1.39 6.2
C2 8,280 1.18 2.45 4.85 – 1.22 6.1
Note: Note: PV = private vehicle; IVTT = in-vehicle travel time; prs = person; na = not applicable.
Bahk et al 287
average travel distance. Travel distance is the distance in
the grid network (counted from the external origin node)
and includes the deadheading travel distance. The travel
distance in every PAV scenario is longer than the dis-
tance in the PCV base scenario.
Since PCV travelers need to travel to parking lots and
search for parking, whereas PAV travelers do not, the
average IVTT of PCV travelers was 3.3 min longer than
that of PAV travelers on average. On the other hand,
PAVs spent more time searching for parking than PCVs.
The reasons for this were twofold: first, there were more
vehicles in the PAV scenarios and, second, PAVs had
more homogeneous parking lot preferences—they want
cheap parking and are less sensitive to distance from
activity location and parking spot search time—making
cheaper parking lots more crowded.
The PCV users’ average (one-way) walking time from
parking lot to activity location was about 6 min in one
direction, and nearly 13min in total including the activ-
ity location to parking lot return walk. Of course, walk-
ing time was zero minutes for the PAV scenarios. The
average waiting time for PAVs to pick up PAV users was
1.2 to 2.4 min (Table 5). The variation across scenarios
comes from the distance between parking lots and activ-
ity locations.
The final column sums average traveler IVTT, walking
time, and waiting time to determine total travel time. The
results showed that the total roundtrip in-network travel
time for PCV was significantly higher than total round-
trip in-network travel time for PAV users. Therefore,
there were significant time benefits associated with PAVs
compared with PCVs.
Table 6 presents an even more holistic comparison of
the travel experiences of PV users across scenarios; it
includes average monetary costs and monetized travel
time components based on the values of IVTT, walking
time, and waiting time in the mode choice model. The
final column of Table 6 displays the total generalized
cost per traveler.
Table 6 shows that the in-network vehicle operating
cost was $ 0.06 to 0.75 higher for PAVs than PCVs,
depending on the scenario. This result stemmed from the
deadheading distance that the PAVs traveled after drop-
ping off travelers at their activity locations.
Scenarios A, B1, and B2 had higher average parking
fees for travelers compared with the base scenario.
Therefore, despite PAVs being able to travel further to
cheap parking lots, the increase in total PV demand in
the PAV scenarios forced some travelers to pay for park-
ing at the high-cost parking lots, which more than offset
their ability to access cheap parking lots. Since the park-
ing fees were unified in Scenarios C1 and C2, the popular
cheaper-than-average parking lots were not cheap any-
more. Thus, the average parking fee increased in those
scenarios.
The monetized IVTT, monetized walking time, and
monetized waiting time columns of Table 6 parallel the
IVTT, walking, and waiting time columns in Table 5.
IVTT was higher and walking time was significantly
higher for PCVs than PAVs, whereas waiting time was
higher for PAVs.
The final column of Table 6 is the sum of all the cost
and monetized cost components in the preceding col-
umns. Interestingly, although Scenarios A and B1 had
the lowest total generalized costs, the base scenario had
a lower generalized cost than Scenarios C1 and C2. This
latter finding stemmed directly from the high parking
cost per person in Scenarios C1 and C2.
Together with the VMT results, Tables 5 and 6 illus-
trate the tradeoffs between PCVs and PAVs in relation
Figure 6. PVs’ VMT in the network: (a) number of PV trips, (b) total PV VMT, and (c) PV VMT per vehicle.
Note: PV = private vehicle; VMT = vehicle miles traveled.
288 Transportation Research Record 2676(7)
to travel time, travel cost, and VMT. Compared with the
base scenario, PAV Scenarios A, B1, and B2 significantly
increased VMT, while reducing average traveler in-
network time considerably and slightly reducing traveler
generalized costs. On the other hand, compared with the
base scenario, PAV Scenarios C1 and C2 only slightly
increased VMT, while significantly reducing average in-
network travel time. However, C1 and C2 had a higher
total generalized cost than the baseline scenario because
of the higher parking costs that were needed to reduce
VMT.
Vehicle Hours Traveled versus Traveler In-Vehicle Travel
Time Results
Figure 7 displays both total vehicle hours traveled (VHT)
and total traveler IVTT under the various scenarios.
Figure 7adisplays the total VHT for PVs and traveler
IVTT. Even though the number of travelers and VHT
increased in the PAV scenarios, there was no significant
increase in total traveler IVTT. Understandably, this was
because the PAVs were empty during the parking search
process. Figure 7bshows that PV VHT per vehicle
increased in Scenarios A, B1, and B2 relative to the base-
line scenario; conversely, PV VHT per vehicle only
increased slightly in Scenario C1, whereas Scenario C2
showed a slight decrease. Figure 7cdisplays the average
IVTT per traveler, with the main result being that IVTT
per traveler was lower in the PAV cases than the baseline
PCV scenario. The results in Figure 7cpartially explained
the increase in PV mode share in the PAV scenarios
despite the increase in VHT with PAVs.
Impact From Shared Autonomous Vehicles. According to
Balding et al. (33) and Conway et al. (46), using data
from the 2017 National Household Travel Survey (45),
the share of for-hire vehicles (taxi and TNC) is around
0.5% across the country and up to 1.7% in San
Francisco and 1.5% in Washington, D.C. Since this
study only considered travelers who have their own vehi-
cles, the base scenario (PCV–SCV) showed an even lower
mode share for SVs, 0.2%. However, this percentage
Figure 7. PV VHT and IVTT in the network: (a) total PV VHT and IVTT, (b) PV VHT per vehicle, and (c) PV IVTT per traveler.
Note: PV = private vehicle; VHT = vehicle hours traveled; IVTT = in-vehicle travel time.
Table 6. Average Monetized Traveler Costs
Scenario
Vehicle operating
cost (USD/prs)
Parking fee
(USD/prs)
Monetized IVTT
(USD/prs)
Monetized walking
time (USD/prs)
Monetized waiting
time (USD/prs)
Total generalized
cost (USD/prs)
Base 0.53 4.72 1.12 2.55 0 8.92
A 1.27 (+0.73) 5.34 (+0.62) 0.42 (–0.69) 0 (–2.55) 0.37 (+0.37) 7.40
B1 1.22 (+0.69) 5.07 (+0.35) 0.42 (–0.69) 0 (–2.55) 0.38 (+0.38) 7.10
B2 1.29 (+0.75) 5.46 (+0.74) 0.42 (–0.69) 0 (–2.55) 0.37 (+0.37) 7.54
C1 0.71 (+0.18) 7.77 (+3.05) 0.42 (–0.69) 0 (–2.55) 0.22 (+0.22) 9.12
C2 0.59 (+0.06) 7.84 (+3.12) 0.42 (–0.69) 0 (–2.55) 0.19 (+0.19) 9.05
Note: IVTT = in-vehicle travel time; prs = person.
Bahk et al 289
increased in the PAV–SAV scenarios as the SV’s travel
cost per mile decreased significantly.
Naturally, SVs, particularly SAVs, affect total net-
work VMT in addition to PAVs. Assuming 40% dead-
heading miles for SVs (13,33), SV travel added 0.67
deadhead miles per in-service mile. Figure 8 illustrates
the impact of SAVs on VMT. Figure 8adisplays the
number of SV trips across the scenarios. Interestingly,
Scenarios C1 and C2 produced the highest number of SV
trips. Figure 8bdisplays the total SV deadheading VMT,
which paralleled the results in Figure 8a. Figure 8cdis-
plays the total PV and SV VMT and found that VMT
increased substantially (5,000 to 22,000 mi, depending on
the scenario) in the AV-based scenarios. However, the
impact of SV VMT (green bars in Figure 8c) was rela-
tively small compared with PV VMT (blue bars in Figure
8c) in nearly all scenarios.
Spatial Disaggregation Scenarios
Although the results in the prior subsection were based on
an enumeration-based solutionapproachtotheintegrated
mode choice and parking assignment problem, this section
presents the results from using the iterative solution
approach proposed in the Solution Approach section.
Notably, the iterative solution approach is necessary in
this section because the mode choice model aggregates the
travelers into four destination zones, rather than just one
destination zone like in the prior subsection. This subsec-
tion illustrates the ability of the iterative solution approach
to identify a solution to the fixed-point problem.
Figure 9 displays the mode choice results under a vari-
ety of different scenarios. The parameter adenotes the
proportion of travelers who own a PAV, as opposed to a
PCV. Each row of graphs in Figure 9 denotes a separate
avalue, whereas adoes not vary across columns. The
figure varies abetween 0 and 1 in increments of 0.25.
Each column in Figure 9 denotes a separate initial start-
ing point for PV mode choice to determine whether the
iterative solution algorithm finds different fixed points as
a function of the initial starting points.
The lines in each of the 15 graphs in Figure 9 indicate
that the iterative solution approach converged to a fixed-
point solution under all cases after less than 20 itera-
tions. Moreover, given that the only thing that changed
between the three graphs in each row was the initial
starting point of PV mode choice, the 15 graphs indicate
that the iterative solution approach found the same fixed
point, independent of the starting point of the mode
choice probabilities. The analysis below assumes a single
fixed-point solution based on the empirical finding in
Figure 9 that the algorithm converges to a single fixed-
point. However, it is important to note that this paper
does not prove that the model system always admits a
unique solution.
The results in Figure 9 indicate that PAV owners were
much more likely to choose PV than PCV owners, in all
scenarios. However, an interesting finding was that as the
proportion of travelers who own a PAV, a, increased,
the mode choice probabilities for PAV owners decreased,
whereas they increased for PCV owners. The reason for
this stemmed from PAV and PCV owners preferring dif-
ferent parking lots. PCV owners were highly sensitive to
the distance between a parking lot and their activity loca-
tion, whereas PAV owners were not. This means that as
the proportion of travelers owning a PAV increases,
PAV owners will have to compete with more travelers
who share their parking lot preferences (i.e., PAV owners
Figure 8. SV VMT in the network: (a) number of SV trips, (b) total SV deadheading VMT, and (c) total PV and SV VMT.
Note: PV = private vehicle; SV = shared vehicle; PCV = private conventional vehicle; PAV = private autonomous vehicle; SCV = shared
conventional vehicle; SAV = shared autonomous vehicle; VMT = vehicle miles traveled.
290 Transportation Research Record 2676(7)
who mainly care about price), whereas PCV owners will
compete with fewer travelers who share their parking lot
preferences (i.e., PCV owners who are sensitive to walk-
ing distance in addition to price).
This logic also explains why the range of modal splits
for PCV owners across zones narrowed as aincreased.
When awas zero, the range of PV mode share across
zones was as broad as 15%, indicating that PCV trave-
lers going to a zone with congested parking lots and high
parking costs were significantly less likely to choose PV
than travelers going to zones with uncongested and lower
cost parking lots. Conversely, the range of PAV mode
shares across zones was quite small under all scenarios,
because a traveler’s destination did not heavily affect
where they preferred to and did park their PAV.
Discussion
Although the case study presented in this paper is based
on a fictional CBD, the Results section hopefully
Figure 9. Mode choice convergence plots varying PV mode share starting points by column from 20% to 80%, and PAVownership
proportion by row from 0.0 to 1.0.
Note: PV = private vehicle; PCV = private conventional vehicle; PAV = private autonomous vehicle.
Bahk et al 291
illustrates the power of the integrated mode choice and
parking location choice model to provide valuable, trans-
ferrable, and generalizable insights into VMT, parking
occupancy, transportation system performance, and user
costs and travel times in a future with PAVs and PCVs.
Moreover, the model can be applied to any region if
detailed data about the road network, parking lots, and
travel demand (or trips) are available. The proposed
solution approach, incorporating the simulation-based
parking assignment model and the multinomial logit
mode choice model, are computationally efficient and
would easily scale to large metropolitan areas given data
availability.
The proposed model should also be quite useful for
policy and planning analysis and decision support. For
example, compared with the current PCV-only case,
redistributing parking spaces appears able to prevent
dramatic increases in VMT while not reducing PV mode
share in a future with PAVs. This suggests the spatial
distribution of parking supply and parking pricing could
significantly affect VMT in a future with PAVs.
Moreover, although not shown explicitly in the
Results section, the model can demonstrate, under cer-
tain scenarios, that parking pricing alone may struggle to
reduce VMT and PV demand. Rather, joint parking
pricing and roadway pricing is likely to be necessary in
an AV future to reduce VMT and PV demand.
Another implicit finding from this study was that
PAVs searching for parking would often look for the
cheapest possible lot in the area, particularly when the
driving cost per mile was low. Therefore, if all PAVs
want to access the same cheap lot(s) in the periphery of
the CBD, this/these lot(s) will become full, and the other
PAVs will need to search for and drive to the next cheap-
est lot. This finding has important technology, policy,
and modeling implications. From a technology stand-
point, providing accurate real-time information to trave-
lers, PAVs, or both about parking lot occupancy could
be quite useful. From a policy standpoint, setting parking
prices based on disaggregate spatial resolutions in CBDs
may not decrease VMT in a world of PAVs. Moreover,
there is clearly a value in promoting a reservation system
of parking lots and even spaces in parking lots to reduce
both parking lot search time and parking space search
time, respectively. Finally, from a modeling standpoint, a
future extension involves incorporating traveler/PAV
knowledge of parking lot occupancy into the modeling
framework to analyze the benefits of this information on
VMT.
Another future modeling extension involves incorpor-
ating roadway congestion into the modeling framework.
The results in this paper clearly indicated a significant
increase in roadway VMT as a result of the attractive
attributes of PAVs as well as the increase in parking
search distance for PAVs. However, at some point, if
enough vehicles are driving around searching for the
cheapest parking lot with available space, the network is
going to experience congestion. This increase in conges-
tion would normally have a leveling effect on parking
search costs, as human drivers would perceive the time
costs of sitting in congestion and probably choose more
expensive parking locations and leave the roadway net-
work. However, if the vehicles searching for a cheap
parking spot are driverless, they will have much lower
costs per minute in congestion and are much less likely
to choose nearby parking lots and exit the roadway net-
work. This is a particularly troubling insight for cities in
the future. It suggests that congestion pricing in cities
may become even more vital to prevent gridlock, and
vehicles may need to be charged not just per mile but per
minute on the road network to avoid regular gridlock in
CBDs.
A related future model extension includes incorporat-
ing congestion and capacity constraints at pickup and
drop-off spots near activity locations in dense urban
areas. With a large percentage of PAVs, SAVs, or both
in a dense urban area, large queues are likely to build at
pickup and drop-off points associated with activity loca-
tions with high demand, such as large office buildings.
These queues may even spill over into the roadway net-
work, thereby requiring a response for traffic managers,
planners, or regulators.
A final research area includes conducting stated pre-
ference surveys to better estimate the model parameters
used in this study. Parameters associated with willing-
ness-to-pay, willingness-to-wait, and willingness-to-walk
are likely to have a significant impact on model results
related to mode share and VMT.
Conclusion
Modeling, understanding, and forecasting the potential
impacts of AVs and PAVs on travel behavior, travel
demand, and transportation systems under a variety of
possible future scenarios is critical for planning for AVs.
This study focused on the potential transportation sys-
tem implications during the transition from PCVs to
PAVs for near-activity travel in urban areas. Specifically,
given the ability of PAVs to drop off travelers at their
activity location and then deadhead to a parking loca-
tion, under certain assumptions it is conceivable that
PAVs will drive great distances to park and/or drive
around looking for an open parking space. This process
would significantly increase VMT compared with PCVs
that drive directly to a parking location close to the tra-
veler’s activity location.
To analyze the impacts of PAVs on near-activity loca-
tion travel, parking lot usage, overall VMT, and traveler
292 Transportation Research Record 2676(7)
cost and travel time this study proposes an integrated
parking assignment and mode choice modeling frame-
work. The proposed mode choice model form is multino-
mial logit, whereas the parking model is a dynamic
simulation-based model of the temporal dynamics of sup-
ply and demand for a system of urban parking locations.
The study also proposes an iterative solution approach to
solve the integrated mode choice and parking assignment
problem. In the iterative solution approach, the parking
simulation model calculates system performance and
costs for travelers based on the demand for each mode—
determined either by the mode choice model or the initial
modal splits—whereas the mode choice model returns
modal splits based on the travel costs from the parking
simulation model.
The study applied the integrated model and iterative
solution approach to an illustrative CBD network. The
model results indicated that PAVs significantly increased
VMT compared with PCVs. The reason for this result
stemmed from the differential between parking prices
and driving fees in the case study. As such, PAVs did not
simply look at the stations nearby their traveler’s activity
location, instead they considered all parking locations
and were highly price sensitive. Consequently, in cases in
which a few parking locations are particularly attractive
to PAVs, these parking locations may reach capacity,
requiring PAVs to detour and search for other parking
locations, thereby further increasing VMT in dense urban
areas. The Results section also illustrated that PAVs sig-
nificantly reduced IVTT, eliminated walking time, but
required travelers to wait a few minutes to be picked up.
The proposed modeling framework could provide
valuable insights to researchers, planners, policy makers,
and other city officials in relation to the potential impli-
cations of AVs on VMT, parking lot usage, mode share,
and other measures of transportation system perfor-
mance and user costs.
Author Contributions
The authors confirm contribution to the paper as follows: study
conception and design: Y. Bahk, M. Hyland; data collection:
Y. Bahk, M. Hyland; analysis and interpretation of results:
Y. Bahk, M. Hyland, S. An; draft manuscript preparation: Y.
Bahk, M. Hyland. All authors reviewed the results and
approved the final version of the manuscript.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with
respect to the research, authorship, and/or publication of this
article.
Funding
The authors disclosed receipt of the following financial support
for the research, authorship, and/or publication of this article:
The first and second authors would like to acknowledge fund-
ing support from NSF #2125560 ‘‘SCC-IRG Track 1:
Revamping Regional Transportation Modeling and Planning
to Address Unprecedented Community Needs during the
Mobility Revolution.’’
ORCID iDs
Younghun Bahk https://orcid.org/0000-0001-5233-1563
Michael F. Hyland https://orcid.org/0000-0001-8394-8064
Sunghi An https://orcid.org/0000-0002-0951-6105
References
1. Harb, M., A. Stathopoulos, Y. Shiftan, and J. L. Walker.
What Do We (Not) Know about Our Future with Auto-
mated Vehicles? Transportation Research Part C: Emerging
Technologies, Vol. 123, 2021, p. 102948. https://doi.org/10.
1016/J.TRC.2020.102948.
2. Anderson, J. M., N. Kalra, K. D. Stanley, P. Sorensen, C.
Samaras, and O. A. Oluwatola. Autonomous Vehicle Tech-
nology: A Guide for Policymakers. Autonomous Vehicle
Technology: A Guide for Policymakers. RAND Corpora-
tion, 2016.
3. Kolarova, V., F. Steck, and F. J. Bahamonde-Birke. Asses-
sing the Effect of Autonomous Driving on Value of Travel
Time Savings: A Comparison between Current and Future
Preferences. Transportation Research Part A: Policy and
Practice, Vol. 129, 2019, pp. 155–169. https://doi.org/10.
1016/J.TRA.2019.08.011.
4. Auld, J., O. Verbas, M. Javanmardi, and A. Rousseau.
Impact of Privately-Owned Level 4 CAV Technologies on
Travel Demand and Energy. Procedia Computer Science,
Vol. 130, 2018, pp. 914–919. https://doi.org/10.1016/J.
PROCS.2018.04.089.
5. Zhong, H., W. Li, M. W. Burris, A. Talebpour, and K. C.
Sinha. Will Autonomous Vehicles Change Auto Commu-
ters’ Value of Travel Time? Transportation Research Part
D: Transport and Environment, Vol. 83, 2020, P. 102303.
https://doi.org/10.1016/J.TRD.2020.102303.
6. Harper, C. D., C. T. Hendrickson, S. Mangones, and C.
Samaras. Estimating Potential Increases in Travel with
Autonomous Vehicles for the Non-Driving, Elderly and
People with Travel-Restrictive Medical Conditions. Trans-
portation Research Part C: Emerging Technologies, Vol. 72,
2016, pp. 1–9. https://doi.org/10.1016/J.TRC.2016.09.003.
7. Kim, S. H., P. L. Mokhtarian, and G. Circella. Will
Autonomous Vehicles Change Residential Location and
Vehicle Ownership? Glimpses from Georgia. Transportation
Research Part D: Transport and Environment, Vol. 82, 2020,
p. 102291. https://doi.org/10.1016/J.TRD.2020.102291.
8. Bansal, P., and K. M. Kockelman. Are We Ready to
Embrace Connected and Self-Driving Vehicles? A Case
Study of Texans. Transportation, Vol. 45, No. 2, 2018,
pp. 641–675. https://doi.org/10.1007/S11116-016-9745-Z/
TABLES/16.
9. Ahmed, T., M. Hyland, N. J. S. Sarma, S. Mitra, and A.
Ghaffar. Quantifying the Employment Accessibility Bene-
fits of Shared Automated Vehicle Mobility Services:
Bahk et al 293
Consumer Welfare Approach Using Logsums. Transporta-
tion Research Part A: Policy and Practice, Vol. 141, 2020,
pp. 221–247. https://doi.org/10.1016/j.tra.2020.09.002.
10. Kro
¨ger, L., T. Kuhnimhof, and S. Trommer. Does Context
Matter? A Comparative Study Modelling Autonomous
Vehicle Impact on Travel Behaviour for Germany and the
USA. Transportation Research Part A: Policy and Practice,
Vol. 122, 2019, pp. 146–161. https://doi.org/10.1016/J.
TRA.2018.03.033.
11. Huang, Y., K. M. Kockelman, and N. Quarles. How Will
Self-Driving Vehicles Affect U.S. Megaregion Traffic? The
Case of the Texas Triangle. Research in Transportation Eco-
nomics, Vol. 84, 2020, p. 101003. https://doi.org/10.1016/J.
RETREC.2020.101003.
12. Cramer, J., and A. B. Krueger. Disruptive Change in the
Taxi Business: The Case of Uber. American Economic
Review, Vol. 106, No. 5, 2016, pp. 177–182. https://doi.
org/10.1257/AER.P20161002.
13. Henao, A., and W. E. Marshall. The Impact of Ride-Hail-
ing on Vehicle Miles Traveled. Transportation, Vol. 46,
No. 6, 2019, pp. 2173–2194. https://doi.org/10.1007/s
11116-018-9923-2.
14. Han, Y., J. Shan, M. Wang, and G. Yang. Optimization
Design and Evaluation of Parking Route Based on Auto-
matic Assignment Mechanism of Parking Lot. Advances in
Mechanical Engineering, Vol. 9, No. 7, 2017, p. 2017.
https://doi.org/10.1177/1687814017712416.
15. Abidi, S., S. Krichen, E. Alba, and J. M. Molina. A New
Heuristic for Solving the Parking Assignment Problem.
Procedia Computer Science, Vol. 60, No. 1, 2015,
pp. 312–321. https://doi.org/10.1016/J.PROCS.2015.08.132.
16. Young, W., R. G. Thompson, and M. A. P. Taylor. A
Review of Urban Car Parking Models. Transport Reviews,
Vol. 11, No. 1, 2007, pp. 63–84. https://doi.org/10.1080/
01441649108716773.
17. Feeney, B. P. A Review of the Impact of Parking Policy
Measures on Travel Demand. Transportation Planning and
Technology, Vol. 13, No. 4, 2007, pp. 229–244. https://doi.
org/10.1080/03081068908717403.
18. Gillen, D. W. Parking Policy, Parking Location Decisions
and the Distribution of Congestion. Transportation,Vol.7,
No. 1, 1978, pp. 69–85. https://doi.org/10.1007/BF00148372.
19. Axhausen, K. W., and J. W. Polak. Choice of Parking:
Stated Preference Approach. Transportation, Vol. 18, No.
1, 1991, pp. 59–81. https://doi.org/10.1007/BF00150559.
20. Waraich, R., and K. Axhausen. Agent-Based Parking
Choice Model. Transportation Research Record: Journal of
Transportation Research Board, 2012. 2319: 39–46.
21. Bischoff, J., and K. Nagel. Integrating Explicit Parking
Search into a Transport Simulation. Procedia Computer
Science, Vol. 109, 2017, pp. 881–886. https://doi.org/10.
1016/J.PROCS.2017.05.414.
22. Nurul Habib, K. M., C. Morency, and M. Tre
´panier Mar-
tin. Integrating Parking Behaviour in Activity-Based Travel
Demand Modelling: Investigation of the Relationship
between Parking Type Choice and Activity Scheduling Pro-
cess. Transportation Research Part A: Policy and Practice,
Vol. 46, No. 1, 2012, pp. 154–166. https://doi.org/10.
1016/J.TRA.2011.09.014.
23. Levin, M. W., and S. D. Boyles. Effects of Autonomous
Vehicle Ownership on Trip, Mode, and Route Choice.
Transportation Research Record: Journal of the Transporta-
tion Research Board, 2019. 2493: 29–38.
24. Childress, S., B. Nichols, B. Charlton, and S. Coe. Using
an Activity-Based Model to Explore the Potential Impacts
of Automated Vehicles. Transportation Research Record:
Journal of the Transportation Research Board, 2015. 2493:
99–106.
25. Zhang, W., S. Guhathakurta, and E. B. Khalil. The Impact
of Private Autonomous Vehicles on Vehicle Ownership and
Unoccupied VMT Generation. Transportation Research
Part C: Emerging Technologies, Vol. 90, 2018, pp. 156–165.
https://doi.org/10.1016/J.TRC.2018.03.005.
26. Zhang, X., W. Liu, and S. T. Waller. A Network Traffic
Assignment Model for Autonomous Vehicles with Parking
Choices. Computer-Aided Civil and Infrastructure Engineer-
ing, Vol. 34, No. 12, 2019, pp. 1100–1118. https://doi.
org/10.1111/MICE.12486.
27. Harper, C. D., S. M. Asce, C. T. Hendrickson, M. Asce,
C. Samaras, and A. M. Asce. Exploring the Economic,
Environmental, and Travel Implications of Changes in
Parking Choices Due to Driverless Vehicles: An Agent-
Based Simulation Approach. Journal of Urban Planning
and Development, Vol. 144, No. 4, 2018, p. 04018043.
https://doi.org/10.1061/(ASCE)UP.1943-5444.0000488.
28. Zhao, C., F. Liao, X. Li, and Y. Du. Macroscopic Model-
ing and Dynamic Control of On-Street Cruising-for-Park-
ing of Autonomous Vehicles in a Multi-Region Urban
Road Network. Transportation Research Part C: Emerging
Technologies, Vol. 128, 2021, p. 103176. https://doi.org/10.
1016/J.TRC.2021.103176.
29. Boyles, S. D., N. E. Lownes, and A. Unnikrishnan. Trans-
portation Network Analysis. First Public Beta Version,
2020.
30. U.S. Department of Transportation. The Value of Travel
Time Savings: Departmental Guidance for Conducting Eco-
nomic Evaluations Revision 2 (2016 Update). 2016.
31. Caltrans. Transportation Economics. https://dot.ca.gov/
programs/transportation-planning/economics-data-manage
ment/transportation-economics. Accessed December 15, 2021.
32. AAA. Your Driving Costs. AAA Association Communica-
tion, 2020. https://newsroom.aaa.com/wp-content/upload-
s/2020/12/Your-Driving-Costs-2020-Fact-Sheet-F-
INAL-12-9-20-2.pdf. Accessed December 15, 2021.
33. Balding, M., T. Whinery, E. Leshner, and E. Womeldorff.
Estimated TNC Share of VMT in Six US Metropolitan
Regions (Revision 1). Fehr Peers, 2019. https://www.feh
randpeers.com/what-are-tncs-share-of-vmt/.
34. Wardman, M. Public Transport Values of Time. Transport
Policy, Vol. 11, No. 4, 2004, pp. 363–377. https://doi.
org/10.1016/j.tranpol.2004.05.001.
35. National Academies of Sciences, Engineering, and Medi-
cine. Bus Routing and Coverage. In Traveler Response to
Transportation System Changes Handbook (Pratt, R. H.,
and I. E. Evans, eds.), The National Academies Press,
Washington, D.C., 2004.
36. Parkopedia. https://www.parkopedia.com. Accessed Novem-
ber 26, 2021.
294 Transportation Research Record 2676(7)
37. Resource Systems Group. Project Report: SANDAG Travel
Model Enhancements to Support 2021 Long-Range Trans-
portation Plan. San Diego Association of Governments,
2020.
38. Correia, G. H., A. de E. Looff, S. van Cranenburgh, M.
Snelder, and B. van Arem. On the Impact of Vehicle Auto-
mation on the Value of Travel Time While Performing
Work and Leisure Activities in a Car: Theoretical Insights
and Results from a Stated Preference Survey. Transporta-
tion Research Part A: Policy and Practice, Vol. 119, 2019,
pp. 359–382. https://doi.org/10.1016/J.TRA.2018.11.016.
39. Singleton, P. A. Discussing the ‘‘Positive Utilities’’ of
Autonomous Vehicles: Will Travellers Really Use Their
Time Productively? Transport Reviews, Vol. 39, No. 1,
2018, pp. 50–65. https://doi.org/10.1080/01441647.2018.
1470584.
40. United States Census Bureau. Census Bureau Estimates
Show Average One-Way Travel Time to Work Rises to
All-Time High. https://www.census.gov/newsroom/press-
releases/2021/one-way-travel-time-to-work-rises.html.
Accessed Dec. 15, 2021.
41. Helling, B. Uber Cost: Fare Pricing, Rates, and Cost Esti-
mates. Ridester.Com. Ridester. https://www.ridester.
com/uber-rates-cost/. Accessed December 15, 2021.
42. Chen, T. D., and K. M. Kockelman. Management of a
Shared Autonomous Electric Vehicle Fleet: Implications of
Pricing Schemes. Transportation Research Record: Journal
of the Transportation Research Board, 2016. 2572: 37–46.
43. Kaddoura, I., J. Bischoff, and K. Nagel. Towards Welfare
Optimal Operation of Innovative Mobility Concepts:
External Cost Pricing in a World of Shared Autonomous
Vehicles. Transportation Research Part A: Policy and Prac-
tice, Vol. 136, 2020, pp. 48–63. https://doi.org/10.1016/J.
TRA.2020.03.032.
44. An, S., D. Nam, and R. Jayakrishnan. Impacts of Integrat-
ing Shared Autonomous Vehicles into a Peer-to-Peer Ride-
sharing System. Procedia Computer Science, Vol. 151, 2019,
pp. 511–518. https://doi.org/10.1016/J.PROCS.2019.04.
069.
45. NHTS. National Household Travel Survey. U.S. Depart-
ment of Transportation, Federal Highway Administration,
2003. https://nhts.ornl.gov/.
46. Conway, M. W., D. Id, D. A. Salon, and I. D. King.
Trends in Taxi Use and the Advent of Ridehailing, 1995–
2017: Evidence from the US National Household Travel
Survey. Urban Science, Vol. 2, No. 3, 2018, p. 79.
https://doi.org/10.3390/URBANSCI2030079.
Bahk et al 295