The Impact of Private Autonomous Vehicles on Vehicle Ownership and Unoccupied VMT

Abstract

With 36 ventures testing autonomous vehicles (AVs) in the State of California, commercial deployment of this disruptive technology is almost around the corner (California, 2017). Different business models of AVs, including Shared AVs (SAVs) and Private AVs (PAVs), will lead to significantly different changes in regional vehicle inventory and Vehicle Miles Travelled (VMT). Most prior studies have already explored the impact of SAVs on vehicle ownership and VMT generation. Limited understanding has been gained regarding vehicle ownership reduction and unoccupied VMT generation potentials in the era of PAVs. Motivated by such research gap, this study develops models to examine how much vehicle ownership reduction can be achieved once private conventional vehicles are replaced by AVs and the spatial distribution of unoccupied VMT accompanied with the vehicle reduction. The models are implemented using travel survey and synthesized trip profile from Atlanta Metropolitan Area. The results show that more than 18% of the households can reduce vehicles, while maintaining the current travel patterns. This can be translated into a 9.5% reduction in private vehicles in the study region. Meanwhile, 29.8 unoccupied VMT will be induced per day per reduced vehicles. A majority of the unoccupied VMT will be loaded on interstate highways and expressways and the largest percentage inflation in VMT will occur on minor local roads. The results can provide implications for evolving trends in household vehicles uses and the location of dedicated AV lanes in the PAV dominated future.

Full text

The Impact of Private Autonomous Vehicles on Vehicle Ownership and Unoccupied VMT
1
Generation
2
3
4
Wenwen Zhang (Corresponding Author)
5
Assistant Professor
6
Virginia Tech, Department of Urban Affairs and Planning
7
Room 215, Architecture Annex
8
140 Otey Street NW, Blacksburg, VA, 24060
9
Tel: 540-232-8431; Email: wenwenz3@vt.edu
10
11
Subhrajit Guhathakurta
12
Professor, Director of Center for GIS
13
Georgia Institute of Technology, School of City and Regional Planning
14
760 Spring Street NW, Suite 217, Atlanta, GA, 30308
15
Tel: 404-385-1443; Email: subhro.guha@design.gatech.edu
16
17
Elias B. Khalil
18
Ph.D. Student
19
Georgia Institute of Technology, School of Computational Science & Engineering
20
Klaus Advanced Computing Building, 266 Ferst Drive, Atlanta, GA 30313
21
Tel: 404-429-3015; Email: elias.khalil@cc.gatech.edu
22
23
24

2
ABSTRACT
1
2
With 36 ventures testing autonomous vehicles (AVs) in the State of California, commercial
3
deployment of this disruptive technology is almost around the corner (California, 2017).
4
Different business models of AVs, including Shared AVs (SAVs) and Private AVs (PAVs), will
5
lead to significantly different changes in regional vehicle inventory and Vehicle Miles Travelled
6
(VMT). Most prior studies have already explored the impact of SAVs on vehicle ownership and
7
VMT generation. Limited understanding has been gained regarding vehicle ownership reduction
8
and unoccupied VMT generation potentials in the era of PAVs. Motivated by such research gap,
9
this study develops models to examine how much vehicle ownership reduction can be achieved
10
once private conventional vehicles are replaced by AVs and the spatial distribution of
11
unoccupied VMT accompanied with the vehicle reduction. The models are implemented using
12
travel survey and synthesized trip profile from Atlanta Metropolitan Area. The results show that
13
more than 18% of the households can reduce vehicles, while maintaining the current travel
14
patterns. This can be translated into a 9.5% reduction in private vehicles in the study region.
15
Meanwhile, 29.8 unoccupied VMT will be induced per day per reduced vehicles. A majority of
16
the unoccupied VMT will be loaded on interstate highways and expressways and the largest
17
percentage inflation in VMT will occur on minor local roads. The results can provide
18
implications for evolving trends in household vehicles uses and the location of dedicated AV
19
lanes in the PAV dominated future.
20
21
Keywords: Autonomous Vehicles, Vehicle Ownership, Unoccupied VMT
22
23

3
INTRODUCTION
1
Many vehicle manufacturers and IT companies have announced plans for deployment of
2
autonomous vehicles by the year 2020. As of June 27th, 2017, 36 ventures have received permits
3
to test prototypes of self-driving vehicles on road in California (California, 2017). This
4
revolutionary transportation technology will undoubtedly alter household vehicle ownership and
5
VMT generation patterns in cities (Fagnant & Kockelman, 2015a; Litman, 2014). The impact of
6
AV on vehicle ownership and VMT generation depends heavily on the business models of the
7
technology, including Shared AVs (SAVs) and Private AVs (PAVs). SAV is an envisioned self-
8
driving taxi system. The operation of the SAV system is centralized to optimize the performance
9
of the system. In the SAV model, consumers pay for mobility service rather than the fleet.
10
Alternatively, the PAV model echoes the current vehicle business model, but replacing
11
conventional vehicles with AVs.
12
Most of the existing studies focused on the impacts of SAVs, which are considered as
13
more environmentally sustainable compared with PAVs. For instance, agent-based simulation
14
models are developed to demonstrate the affordability and feasibility of the SAV system (Burns,
15
Jordan, & Scarborough, 2013; Spieser et al., 2014) and to explore the impacts of SAVs on
16
vehicle ownership, Greenhouse Gas (GHG) emissions, traffic flow, charging stations, and
17
parking demand (Chen, Kockelman, & Hanna, 2016; Fagnant & Kockelman, 2014; Greenblatt &
18
Saxena, 2015; Zhang & Guhathakurta, 2017; Zhang, Guhathakurta, Fang, & Zhang, 2015a,
19
2015b). Based on author’s best knowledge, to date, only one report has explored the impact of
20
PAVs on household vehicle ownership reduction potentials, using the 2009 National Household
21
Travel Survey (NHTS) data (Schoettle & Sivak, 2015). The study only considers time conflicts
22
in the household AV scheduling model, while other components such as the origins and
23
destinations of trips are not included. Additionally, the study does not provide implications for
24
unoccupied VMT generation, as the origins and destinations of trips are not provided in NHTS
25
data.
26
Despite SAVs being more heatedly discussed in the existing literature, the privately-
27
owned AVs (PAVs) may turn out to be more preferable to consumers, based on several recent
28
AV preference survey results. Bansal et al. (2016) conducted an opinion survey in Austin.
29
Among the 347 respondents, only 13% indicate they may be willing to give up personal vehicles
30
and rely exclusively on SAVs whose costs are $1/mile. Additionally, the most optimistic
31
scenario indicates over 35% of the respondents are unlikely to participate into the SAV program,
32
regardless the cost of the service. Another SAVs preference survey suggests that given various
33
trip characteristics profiles, more than 70% respondents choose not to use the SAV system
34
(Krueger, Rashidi, & Rose, 2016). Another stated preference survey reveals that only 5.4% of
35
the 1920 observations in North America are willing to rely exclusively on SAVs for commuting
36
purposes trips and only 40.63% are willing to participate into the SAV program (even at zero
37
membership cost) (Haboucha, Ishaq, & Shiftan, 2017). In sum, the majority of consumers may
38
still prefer to own a private AV in the near future. Therefore, it is critical to gain a more
39
comprehensive understanding regarding the impact of PAV on vehicle ownership and VMT
40
generation.
41
Motivated by the limited understanding of the impacts of PAV on household vehicle
42
ownership and unoccupied VMT generation, this study designs and implements a vehicle
43
scheduling algorithm to estimate the vehicle ownership reduction potentials and unoccupied
44
VMT generation in the era of PAV, using the 2011 travel survey data from Atlanta Metropolitan
45
Area. Statistical analyses are then conducted to identify critical factors (such as household travel
46

4
pattern, socio-economic, demographic, and built environment characteristics) that are associated
1
with the vehicle ownership reduction potentials. Additionally, the study also examines the
2
temporal and spatial distributions of unoccupied VMT using the synthesized trip profiles
3
generated by the Atlanta Activity Based Travel model.
4
The remainder of the article is organized as follows. The subsequent section provides a
5
brief overview regarding the existing studies regarding the impact of AVs on vehicle ownership
6
and VMT generation. Section Three describes the data sources and methodology used to
7
examine vehicle ownership reduction and unoccupied VMT generation potentials under PAV
8
business model. Section Four presents and analyzes the model results. Conclusions and future
9
research directions are discussed in Section Five.
10
11
BACKGROUND
12
With autonomous vehicles technology almost around the corner, the literature regarding the
13
impact of AVs is proliferating. Many studies show this disruptive technology will improve travel
14
experience by reducing crashes (Harper, Hendrickson, & Samaras, 2016), improve fuel
15
efficiency (Fagnant & Kockelman, 2015a; Mersky & Samaras, 2016), and provide more reliable
16
travel time, at a cost that is significantly more affordable than current private sedans (Burns et al.,
17
2013; Litman, 2014). However, AVs, if owned privately, instead of shared among consumers,
18
are also expected to generate several negative externalities, such as excessive VMT generation
19
(Zhang et al., 2015b), Greenhouse Gas (GHG) emissions, and more transportation energy
20
consumptions (Greenblatt & Saxena, 2015), stemming primarily from changes in travel behavior.
21
The following sections summarize the existing studies regarding how AVs (either SAV or PAV)
22
may influence vehicle ownership and VMT generation.
23
Most literature has focused on how SAVs would reduce vehicle ownership, using agent-
24
based simulation models. Results show that one SAV can replace approximately 11-14 private
25
vehicles (i.e. approximately 90% of reduction rate), assuming consumers are willing to give up
26
personal vehicles and rely exclusively on SAVs (Bischoff & Maciejewski, 2016; Boesch, Ciari,
27
& Axhausen, 2016; Fagnant & Kockelman, 2014, 2015b; Martinez & Crist, 2015; Rigole, 2014;
28
Zhang et al., 2015b). The replacement rates vary slightly based on the population and
29
employment density in the studied region. To authors’ best knowledge, only one study, to date,
30
explored how PAVs will influence household vehicle ownership. Schoettle and Sivak (2015)
31
found that average household vehicle ownership can be reduced by 43% from 2.1 to 1.2, once
32
households replace conventional vehicles with AVs, using weighted National Household Travel
33
Survey (NHTS) data. However, in this study, the minimum required vehicle is estimated based
34
on the trip starting and ending time. The location of origin and destination is not accounted for in
35
their analyses, as such information is not provided in the NHTS data. While, in some cases, one
36
AV may not be sufficient to serve two non-overlapping trips if the relocation time is too long.
37
Therefore, Schoettle and Sivak’s pioneering work only provides an optimistic upper bound for
38
potential vehicle ownership reduction rate. Almost no other study, to date, has developed a
39
model to understand vehicle reduction potential while incorporating the spatial distributions of
40
origins and locations into the model. Additionally, little understanding has been gained regarding
41
what type of household (i.e. socio-demographic, economic, and travel behavior characteristics)
42
are more likely to reduce household vehicle ownership in the era of AVs.
43
The vehicle automation technology will undoubtedly change Vehicle Miles Travelled
44
(VMT) for various reasons. First, several studies suggest that VMT would increase by 10-14%,
45
once the AVs start to serve underserved population, especially those driving capability are
46

5
constrained for various reasons (Harper, Hendrickson, Mangones, & Samaras, 2016). Second,
1
VMT may also change dramatically, given variations in travel behaviors due to reduced travel
2
time costs and parking costs (Childress, Nichols, Charlton, & Coe, 2015; Levin & Boyles, 2015).
3
The changes in VMT may vary significantly based on the assumptions made in the simulations,
4
ranging from -35% to 20%. Childress et al., (2015) suggest that VMT increase the most the
5
perceived travel time costs are reduced by over 50%. Alternatively, VMT may decrease if the per
6
mile based travel cost of AVs surplus the existing sedans. Finally, AVs can also introduce a
7
significant amount of unoccupied VMT, during the relocation process. In the SAV model, 11%-
8
20% of unoccupied VMT are generated, when SAVs relocate to serve clients or balance the
9
spatial distribution of vehicles in the system. The range varies significantly depending on the
10
level of willingness to share rides among consumers and the trip density (Fagnant & Kockelman,
11
2014; Zhang et al., 2015b). However, it remains unclear how much unoccupied VMT will be
12
induced after replacing conventional household vehicles with AVs under the PAV business
13
model. Additionally, it is also critical to understand the spatial and temporal distribution of
14
unoccupied VMT to provide implications for future travel demand and the allocation of
15
infrastructure resources correspondingly. However, few research has contributed to these topics.
16
To fill up the existing gaps, this study aims to examine potentials of vehicle ownership
17
reduction in the era of PAV by incorporating the spatial distribution of origins and destinations
18
into the model using weighted Atlanta travel survey. Additionally, this study also determines the
19
spatial and temporal distribution of unoccupied VMT as a result of reduced household vehicle
20
ownership, using synthesized regional travel profile output from Atlanta activity based model.
21
The results will provide implications for vehicle reduction potentials and spatial and temporal
22
distribution of unoccupied VMT in the region. The model outputs will shine lights on future
23
transportation facility demand in a PAV dominated future.
24
25
DATA AND METHODOLOGY
26
Data
27
Two data sets are used in this study, including (1) 2011 Atlanta Travel Survey and (2)
28
synthesized Atlanta trip profile from the Atlanta activity based travel model (ABM). Both
29
datasets are generously provided by Atlanta Regional Commission (ARC). The 2011 Atlanta
30
Travel Survey contains 10,278 households, 9,901of which (96% of the weighted sample) have at
31
least one privately owned vehicles (excluding three households living outside of the region and
32
eleven households with partial members filled out the survey). According to the survey data,
33
each household produces approximately 9.12 vehicle trips per day and owns 1.99 vehicles, on
34
average (ARC, 2011). The origin and destination of the trips have already been geocoded with
35
longitudes and latitudes by ARC.
36
The synthesized trip profile includes characteristics of simulated trips for each
37
synthesized household in the 20-county Atlanta metropolitan area. There are 2,115,034
38
households and 19,235,738 vehicle trips in this dataset. The data contain several trip features,
39
including origin, destination, departure time (in 30 minutes intervals), travel mode, etc. The
40
attributes of trips are simulated using their marginal distributions collected in the 2011 travel
41
survey. Therefore, the 2015 trip profile can be considered as an extrapolated version of 2011
42
Atlanta Travel Survey for the entire 20-county metro area. This dataset occupies 4.2 Gigabytes
43
of space on the disk and, therefore, is computationally challenging to process.
44
We use two datasets to examine different research questions, considering the
45
characteristics/strength of each dataset. The travel survey includes information, such as
46

6
longitudes and latitudes of trip origins and destinations and a wide range of socio-economic and
1
demographic characteristics at the household and individual level. In other words, the travel
2
survey data are more refined than the synthetic trip profile from ABM. Therefore, we use Atlanta
3
travel survey trip records to determine vehicle reduction potentials of households and identify
4
household features that are correlated with the vehicle reduction potentials. Meanwhile, although
5
ABM trip profile only contains TAZ level trip origin and destination information and a limited
6
number of household characteristics, the data do provide all Origin-Destination (OD) pairs in the
7
study area. Therefore, the synthetic trips from ABM are used to determine the spatial distribution
8
of unoccupied/relocation VMT in the transportation network.
9
10
Methodology
11
The research methodology is three-fold. In step one, a greedy algorithm is designed to determine
12
the minimum number of private AVs needed to fulfill a household’s current travel demand. In
13
step two, mixed-integer programming (MIP) problems are formulated for households that can
14
reduce vehicle ownership. The problems are then solved using IBM CPLEX software to obtain
15
optimized vehicle route (i.e. the route that can minimize daily VMT for each household) to
16
determine the origins and destinations of unoccupied trips. The above described two model
17
components are applied to 2011 Atlanta travel survey data to examine vehicle reduction and
18
unoccupied VMT generation potentials. Last, in step three, the models from step one and two are
19
applied to the synthesized 2015 trip profile for the entire region to generate new Origin-
20
Destination (OD) matrices. A trip assignment model is then implemented in CUBE to allocate
21
unoccupied AV trips to the transportation network. The details for each step are described in the
22
following sections.
23
In Step one, a greedy scheduling algorithm is designed to determine the minimum
24
number of autonomous vehicles needed to satisfy the travel demand of all household members in
25
each household. Frist, the vehicle trips generated in each household are sorted based on the trip
26
departure time and are analyzed sequentially. At the beginning of the day, the vehicle inventory
27
for the household is set as zero. For each incoming household trip, the algorithm will find all the
28
AVs that will be available by the departure time of the trip. An AV is considered as available
29
when two criteria are met: 1) AV is not serving other household member when the current trip
30
departs and 2) There is sufficient time for AV to relocate from its location to the origin of the
31
upcoming trip. The potential relocation time is obtained using Google Maps Distance Matrix
32
Application Programming Interface (API) service. The Distance Matrix API returns Google’s
33
estimate of travel time given the provided trip origin, destination, and departure time. Therefore,
34
the congestion factor on relocation is considered in this process. If no AV is available to serve
35
the incoming trip, a new AV will be added to the household vehicle inventory. The location of
36
the AVs will always be updated to the destination of the last served trip. Additionally, the status
37
of AVs will be marked as busy until the end of the last served trip. After scanning all trips made
38
by the household, the number of AV saved in the household vehicle inventory will be the
39
minimum required number of AV to serve the household. Vehicle reduction potential is
40
calculated by subtracting the existing number of operation vehicles (not the total number of
41
owned vehicles) by the number of required AVs. It is assumed that extra vehicles that are not
42
identified as daily operational vehicle in the survey are kept for purposes other than travel and
43
therefore will not be eliminated after the introduction of AVs.
44
The above described greedy algorithm, however, cannot determine the excessive VMT
45
generation of the household, as the vehicle service route is not optimized. Therefore, in step two
46

7
Mixed-Integer Programming problems are formulated and solved to determine the minimum
1
amount of unoccupied VMT generated during AV repositioning process for households that can
2
reduce vehicle ownership. The notation of the problems are as follows:
3
4
!" # $
: the set of
%$%
AVs;
5
&" # '
: the set of
%'%
trips made by household members;
6
(" # )
: the set of %
)
% potential relocations between household trips with
(*+,*-
indicating
7
relocation miles generated after serving
&."
first and then
&/
;
8
01,*+,*-# 2
0
,
1
3
: if AV
!"
relocates to serve trip
&.
and
&/
9
10
For each household, a weighted directed graph (or network),
"4 5 6',)7
, is generated.
11
The nodes (
'
) in the graph represent vehicle trips generated by the household. The directed
12
edges (
)
) indicate the amount of relocation miles incurred if an AV serves both starting and
13
ending trips/nodes sequentially. If there is no enough relocation time between the trips, then no
14
edge will be generated, i.e., the two trips cannot be served by one PAV. In other words, if there
15
is no sufficient time for AVs to relocation from the destination of prior trip to the origin of
16
current trip before the departure time of the current trip, then the two trip nodes will not be
17
connected by an edge. Similar to Step One, the relocation time is obtained using Google
18
Distance Matrix API. The direction of the edge indicates the time sequence of the service. The
19
objective of this optimization problem is to find
%!%
disjoint path(s) in this graph, such that the
20
sum of edge costs (relocation distance) is minimized, see the objective function below:
21
22
89: ; ;;01,*+,*-<(*+,*-
=
/>1
=
.>1
?
1">"1
(1)
23
Having defined the variables (
01,*+,*-
), the problem graph
4 5 6',)7
, and the set of AVs
24
$
, we are now ready to describe the constraints of our MIP problem. First, each trip should be
25
served by exactly one AV. This suggests that the sum of
0
variables related to the incoming
26
edge(s),
@(AB6&.7
of node
9
should be equal to one. Additionally, the sum of
0
variables related
27
to the outgoing edge(s),
@(AC6&/7
of node
D
should also be constrained to one.
28
29
;;01,*+,*-
."
?
1>1
51,E"&.#"@(AB6&/7
(2)
;;01,*+,*-
."
?
1>1
51,E"&/#"@(AC6&.7
(3)
30
Second, the route of each AV should be contiguous, i.e. AVs cannot teleport from one
31
location to another to serve trips. In other words, the incoming and outgoing edges of one
32
node/trip should be assigned to one AV.
33
34

8
; 01,*+,*-
*-"#"FGHI6*+75 ; 01,*+,*-
*-"#"FGHJ6*+7,E"&.# ',! # $
"
(4)
1
Third, the number of AV(s) should not be more than the minimum required number of
2
AV(s). To implement these constraints, hypothetical starting and ending nodes,
KL
and
KM
, are
3
added into the graph to control the number of AVs assigned to put into the network. The starting
4
node has outgoing edges to all the trip nodes and the weight are all assigned to be zero. Similarly,
5
the ending node has zero cost weighted incoming edges from all the trip nodes.
6
7
;;01,NO,*+
=
*>1
?
1>1
5"%!%
"
(5)
;;01,*+,NP
=
*>1
?
1>1
5"%!%
(6)
8
This optimization algorithm is then applied to households that can potentially reduce
9
vehicle ownership to determine their unoccupied VMT generation. The optimization is
10
implemented in Python 2.7 using IBM CPLEX’s Python Application Programming Interface
11
(API) (IBM, 2017). Descriptive statistics of the model outputs are calculated using weighted
12
2011 Atlanta travel survey to examine the overall vehicle reduction and unoccupied VMT
13
generation potentials.
14
Finally, in step three, models from previous steps are applied to the 2015 synthesized trip
15
profile to obtain the origins and destinations of all AV relocation trips on a typical weekday.
16
There are some minor changes in the methodology from Step One and Two, so that the model
17
can be applied to ABM data. ABM only contains TAZ level trip origin and destination
18
information. Therefore, instead of using Google Distance Matrix API, we obtained relocation
19
time using the SKIM matrix from ABM. After applying the revised model to ABM data, we
20
obtained New Origin-Destination (OD) matrix, containing empty relocation AV trips, by time of
21
the day. The vehicle trips (original trip and empty relocation trips) are then assigned to road
22
segments by applying the all or nothing trip assignment process in CUBE voyager. Potential
23
changes in the spatial distribution of traffic volume are then identified through cross-comparing
24
with the current ARC baseline network outputs.
25
26
Model Assumptions and Scenarios
27
The assumptions and simplifications of the developed models are summarized as follows:
28
1) No change in the travel behaviors, i.e., no induced travel demand and no variations in the
29
travel patterns (origin, destination, departure time);
30
2) The estimated VMT changes stems exclusively from the re-routing of unoccupied AVs
31
from prior trip destination generated by another household member to the existing trip
32
origin;
33
3) Vehicles are only shared among household members not among households;
34
4) It assumes a 100% market penetration rate and heterogeneity in the preferences for AVs
35
is not considered
36
37

9
We also examined some other scenarios in the scenario development section. We specifically
1
explored the impact of schedule flexibility, i.e., the tolerance on trip departure and arrival time.
2
We discuss the results of the scenarios analysis in the schedule flexibility scenarios development
3
section.
4
5
RESULTS
6
Vehicle Reduction Potentials
7
The results show that approximately 18.3% of the households in the weighted survey have the
8
potential to reduce vehicle ownership even if they maintain the current travel schedule.
9
Compared with the weighted vehicle inventory in the region, approximately 9.5% vehicle
10
ownership reduction can be achieved overall. For households that can reduce vehicle ownership,
11
on average, 1.1 vehicles can be eliminated. The majority of the households cannot achieve
12
vehicle reduction given their overlapping trip schedules, especially during peak hours. More
13
vehicle can be reduced if household members start to re-schedule daily trips to accommodate
14
AVs.
15
We developed a logistic regression model to understand the correlations between
16
socioeconomic and demographic characteristics of households (i.e., explanatory variables) and
17
vehicle ownership reduction potentials (i.e., the dependent variable). The results, as displayed in
18
Table 1, indicate that current vehicle dependency plays an important role in vehicle ownership
19
reduction potential. Households with more operating vehicles are more likely to benefit from the
20
vehicle automation technology. Additionally, the reduction potential is also correlated with the
21
trip generation pattern of the household. The vehicle ownership reduction potential is larger for
22
households with higher trip generation rates and shorter trips. This type of travel pattern leaves
23
more room for PAV relocation in the future. The results also suggest households with different
24
socioeconomic features are more likely to benefit from PAVs. Families with higher income and
25
home owners (rather than renters) have more potential to reduce vehicle ownership. Additionally,
26
households with more workers are also more likely downsize their existing vehicle ownership,
27
once vehicles can relocate from work locations to serve family members in other places in the
28
region. Finally, the results also indicate that built environment features are also correlated with
29
vehicle ownership reduction potentials. The vehicle reduction potential for suburban households
30
is larger, as the estimated coefficients for variables such as log transformed housing unit density
31
and four-way intersection density are negative and significant at 95% level, while the coefficient
32
for log transformed distance to the Central Business District (CBD) is positive and significant.
33
Atlanta is a monocentric city, the model results suggest that households in suburban areas, which
34
are further away from downtown and are less intensively developed, are more likely to be able to
35
reduce vehicle ownership in the future.
36
37
TABLE 1 Logistic Regression Results Summary
38
Variables
Coefficients
Std. Err.
Z
P > |z|
Operating Vehicle Count
2.213
0.069
31.960
0.000
Trip Generation Rate
0.099
0.016
6.242
0.000
Average Trip Distance (Mile)
-0.042
0.007
-6.286
0.000
Low Income Household (annual income < 30,000)
-0.466
0.130
-3.587
0.000
Number of Workers
0.383
0.043
8.952
0.000
Home Renter Dummy
-0.612
0.133
-4.605
0.000
Single Adult Household
-2.154
0.162
-13.333
0.000

10
Log (distance to CBD)
0.102
0.018
5.776
0.000
Log (Housing Units Density)
-0.124
0.031
-3.997
0.000
Four-way Intersection Density (per Mile2)
-0.010
0.003
-2.939
0.003
Sample Size (N)
9007
Pseudo R-square
0.45
Log-likelihood
3070.0
1
Excessive VMT generation
2
In return for vehicle ownership reduction, the households will generate more unoccupied VMT
3
during the vehicle relocation process. Households, on average, will produce 29.8 more VMT per
4
day per reduced vehicle. The distribution of excessive VMT generation per household suggest
5
that most households increase VMT by around 10-20 miles per day, see Figure 1. The median
6
increase in VMT is 26.5 mile per household. Less than 10% of the households will generate 67
7
more miles per day.
8
9
10
FIGURE 1 Histogram of excessive VMT generation per household (weighted)
11
12
The VMT generation for households that can reduce vehicle ownership will increase by
13
59.5%, on average, compared with current generation patterns (i.e. 50.1 VMT per household per
14
day). The total VMT generation in the metropolitan area will rise by 13.3%, due to empty private
15
AV relocation. Some other factors, which are not modelled in this study, may inflate the empty
16
VMT generation in the future, including, but not limited to, cruising for less expensive parking
17
spaces, activity rescheduling, and changes in travel patterns.
18
The temporal distribution of the excessive VMT indicate that the increase is the most
19
significant during peak hours and hours during daytime, see Figure 2. The absolute increase in
20
VMT is the highest during 4-6pm, with more than 1.1 million VMT added into the network per
21
hour. During daytime (i.e. from 11 AM to 4 PM) the VMT generation increases by over 20%,
22
which is the largest percentage increase throughout the day.
23
24
0
10000
20000
30000
40000
50000
60000
70000
0-10
10-20
20-30
30-40
40-50
50-60
60-70
70-80
80-90
90-100
100-110
110-120
120-130
Weighted Frequency
Excessive VMT Generation per Household

11
1
FIGURE 2 Temporal distribution of unoccupied VMT
2
3
Spatial Distribution of Excessive VMT
4
The spatial distributions of unoccupied VMT by time of day are estimated by applying the
5
vehicle reduction and AV route optimization algorithm to the synthesized trip profile from
6
ARC’s activity based model. OD matrices of all empty AV relocation trips are generated by time
7
periods: early morning (EA), morning peak (AM), midday (MD), evening peak (PM), and night
8
(EV). New OD matrices are then generated by combining the current OD matrices with the
9
relocation OD matrices. Trips assignments are implemented in CUBE, using the new OD
10
matrices and local network. The trip assignment results provide updated traffic volume for each
11
road segment, based on which the Volume-to-Capacity (V/C) Ratio are re-estimated. It is
12
assumed that road capacity will remain unchanged, as the objective of this study is not to explore
13
whether the roads will be more congested or not, but to obtain an understanding of the spatial
14
distribution of unoccupied VMT in the region.
15
The changes in average V/C Ratio before and after the introduction of private AVs by
16
location of road segments are shown in Table 2. In the early morning, the V/C Ratio only
17
increases slightly, due to small travel demand at the beginning of the day. During this time of the
18
day, excessive VMT tend to locate primarily in suburban, exurban, and rural residential
19
neighborhoods. This may be due to the fact that the majority of the urban residents live in car-
20
oriented suburban communities in Atlanta Metropolitan. These communities tend to generate
21
some morning errands that may lead to extra unoccupied VMT generation. During morning and
22
evening peak hours, the V/C Ratio, on average, increases significantly by around 7.99% and 8.44%
23
respectively. The suburban, exurban, and rural neighborhoods, as well as urban commercial
24
zones are more likely to experience the most dramatic increase in the V/C Ratio. This indicates
25
that a large amount of relocation VMT is generated between commercial zones and residential
26
zones after the adoption of PAVs. Therefore, the larger the mismatch between work and
27
residential locations, the larger the overall relocation VMT generation will be in the future.
28
Currently, a majority of commuters live in suburban residential zones and work in the urban core
29
0
1000000
2000000
3000000
4000000
5000000
6000000
7000000
8000000
9000000
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
VMT Generation
Hour
Current
Unoccupied

12
area. Atlanta is a typical monocentric city, rendering a large amount of empty relocation trips
1
between commercial zones and suburban neighborhoods outside of the perimeter (I-285).
2
Additionally, the results also indicate that the V/C Ratio increases the most during the midday
3
period, when more PAV coordination will take place among household members. However, the
4
midday traffic condition will not be as congested as morning and evening peak hours, as the
5
overall V/C Ratio is still substantially lower during midday compared to peak hours. During
6
night time, the V/C Ratio in the region increases by approximately 6.79%. The traffic volume
7
inflates the most on road segments in suburban residential and exurban areas during night, due to
8
relocations among non-work-related household activities in the evening.
9
In sum, road segments located in suburban, exurban and rural areas will experience
10
higher percentage of increments in traffic volume after the coming of PAVs. This result is
11
consistent with the logistic regression results, indicating suburban households are more likely to
12
reduce vehicle ownership. However, roads in CBD and urban areas will remain more congested
13
than other areas in the region, given the higher V/C Ratios throughout the day. Cities with more
14
segregated land use may experience a higher percentage increase in the V/C Ratios in the future.
15
16
TABLE 2 Changes in V/C Ratios before and after AVs by Area Types
17
Time
Period
Scenarios
CBD
Urban
Commercial
Urban
Residential
Suburban
Commercial
Suburban
Residential
Exurban
Rural
Overall
EA
BAU
0.104
0.105
0.104
0.100
0.092
0.075
0.063
0.094
AV
0.105
0.106
0.105
0.101
0.093
0.076
0.064
0.095
Changes
0.73%
0.90%
0.86%
0.98%
1.25%
1.24%
1.28%
1.07%
AM
BAU
0.398
0.417
0.411
0.376
0.337
0.234
0.165
0.343
AV
0.415
0.446
0.439
0.406
0.367
0.253
0.177
0.371
Changes
4.29%
7.03%
6.87%
8.09%
8.98%
8.18%
7.17%
7.99%
MD
BAU
0.368
0.358
0.334
0.297
0.255
0.166
0.124
0.269
AV
0.391
0.392
0.362
0.326
0.283
0.183
0.135
0.295
Changes
6.28%
9.54%
8.35%
9.74%
10.71%
10.41%
9.51%
9.75%
PM
BAU
0.473
0.493
0.484
0.443
0.400
0.281
0.201
0.408
AV
0.503
0.540
0.523
0.486
0.441
0.307
0.217
0.447
Changes
6.51%
9.38%
8.07%
9.87%
10.14%
9.20%
8.10%
9.44%
EV
BAU
0.201
0.225
0.220
0.209
0.190
0.129
0.098
0.183
AV
0.211
0.240
0.233
0.223
0.204
0.139
0.104
0.195
Changes
4.77%
6.55%
5.76%
6.90%
7.45%
7.50%
6.04%
6.79%
* BAU: Business as Usual
18
19
The changes in V/C ratios by types of road segments are shown in Table 3. The results
20
from all time periods suggest that the majority of unoccupied traffic volumes are loaded on
21
minor arterial roads, where the V/C ratios surge dramatically regardless the time of the day.
22
During morning and evening peak hours, the V/C ratios inflate by 4.99% and 4.39% on
23
expressways correspondingly, second to minor arterials. While during off peak hours (except for
24
early morning hours), the V/C ratios rise more on principal arterials rather than expressways.
25
These suggest that the relocation trips during midday and night time are shorter local trips
26

13
compared with relocation trips incurred during peak hours. The average length of relocation trips
1
declined from 18.5 miles during peak hours to 15.6 miles during off peak hours.
2
3
TABLE 3 Changes in V/C Ratios before and after AVs by Road Types
4
Time
Period
Scenarios
Interstate /
Freeway
Expressway
Parkway
Principal
Arterial
Minor
Arterial
EA
BAU
0.297
0.171
0.169
0.124
0.098
AV
0.298
0.172
0.169
0.125
0.099
Changes
0.31%
0.68%
0.29%
0.50%
1.00%
AM
BAU
0.650
0.514
0.549
0.468
0.375
AV
0.669
0.539
0.564
0.486
0.403
Changes
3.02%
4.99%
2.76%
3.73%
7.53%
MD
BAU
0.503
0.404
0.367
0.353
0.280
AV
0.516
0.421
0.383
0.37
0.305
Changes
2.65%
4.13%
4.15%
4.93%
9.16%
PM
BAU
0.692
0.557
0.554
0.547
0.440
AV
0.709
0.581
0.576
0.573
0.479
Changes
2.40%
4.39%
4.05%
4.71%
8.68%
EV
BAU
0.441
0.298
0.321
0.239
0.197
AV
0.449
0.306
0.33
0.249
0.21
Changes
1.86%
2.66%
2.83%
4.00%
6.83%
* BAU: Business as Usual
5
6
The intra-zonal relocation trips (i.e. trips that start and end in the same traffic analysis
7
zones) are not loaded on the transportation network, as local roads are not included in the ARC’s
8
activity based model. In this study, we analyzed the impact of relocation trips on local roads by
9
examining the density of intra-zonal relocation trips at the TAZ level. Figure 3 illustrates the
10
spatial distribution of intra-zonal relocation trips density by time of the day. The number of intra-
11
zonal relocation trips peaks during midday at over 1.41 million trips. The amount of intra-zonal
12
relocation trips is 1.37 million during evening peak hours, which is slightly less than the amount
13
incurred during midday. However, the spatial distribution of intra-zonal trips varies significantly
14
during midday and evening peak hours. During midday, most of the intra-zonal trips are located
15
in commercial zones adjacent to expressways in the region, indicating that the local roads in the
16
commercial zones may experience a larger percent of increase in the R/C Ratio after the adoption
17
of PAVs. On the other hand, the majority of intra-zonal relocation trips sprawled into suburban
18
commercial and residential zones, which are further away from the expressways, during evening
19
peak hours, suggesting local roads in suburban residential and commercial zones may witness a
20
larger increase in R/C Ratio.
21
These results indicate that the generation of short (intra-zonal) relocation trips follows the
22
trip generation patterns and location of work and residential places in the region. During the
23
noon, more intra-zonal relocation trips are generated in the commercial zones in the region to
24
serve work-based trips among household members. While, during night, a significant amount of
25
relocation trips is likely to be generated to support household members to run evening errands.
26
The results also indicate that zones with more mixed or diversified land use tend to have larger
27
amount of intra zonal relocation trips. To alleviate future traffic pressure on local roads in these
28

14
areas, designated dropping off and picking up stations may be considered to promote walking
1
and reduce empty cruising.
2
3
4
FIGURE 3 Spatial distributions of intra-zonal trips by time of the day
5
6
SCHEDULE FLEXIBILITY SENARIOS DEVELOPMENT
7
The above experiments are conducted based on the assumption that individuals do not have
8
flexible activity schedules. In this section, we relaxed such assumption to determine how the
9
results may vary if household members collaborate closely to reduce vehicle ownership. In the
10
elasticity tests, we allow individuals to be dropped off 5, 10, and 15 minutes later than the
11
current arrival time and results are tabulated in Table 4. As expected, more households can
12
reduce vehicle ownership if delays are allowed. The percent of households that can reduce
13
vehicle ownership increases from 18.3% to 24.1% when the activity schedules are relaxed by 15
14
minutes. The overall vehicle reduction rates also inflate from 9.5% to 12.3%. Moreover,
15
marginal effects of schedule flexibility on vehicle reduction increases, as significantly more
16
households can reduce vehicle ownership when the delay tolerance increases from 10 to 15
17
minutes compared with 5 to 10 minutes. The average vehicle ownership reduction, however, is
18
quite stable across different tests. On average, households can only eliminate one vehicle
19
regardless of schedule flexibility. The total VMT generation will also increase significantly when
20
more households share PAVs among members. The results suggest that the schedule flexibility is
21
also associated with excessive VMT generation at the household level. When no delays are
22
tolerated the empty VMT per day per reduced vehicle is the lowest across all scenarios. The
23
empty vehicle relocation VMT increases slightly first when 5 minutes of delays are tolerated and
24

15
then declines when the schedules become more flexible. This is due to the fact that it is easier to
1
optimize the PAV daily routes to reduce relocation VMT when larger delays are allowed.
2
However, overall the relocation VMT still increases significantly due to more households are
3
able to achieve vehicle ownership reduction and generate empty VMT.
4
5
TABLE 4 Flexibility Scenarios Results
6
Trip Delay Tolerance
No Delay
5 minutes
10 minutes
15 minutes
% HH Can Reduce Vehicle Ownership
18.3%
20.0%
21.7%
24.1%
Total Vehicle Ownership Reduction
9.5%
10.0%
10.9%
12.3%
Avg. Vehicle Ownership Reduction
1.089
1.099
1.104
1.112
Total Empty VMT Generation
13.3%
14.6%
15.7%
17.3%
Empty VMT per Day per Reduced Vehicle
29.8
30.7
30.6
30.2
Median Relocation Length per HH (Miles)
26.5
27.1
27.0
26.5
7
8
CONCLUSIONS
9
In this study, we developed a greedy algorithm to examine vehicle ownership reduction
10
potentials after replacing private conventional vehicles by AVs. We also formulated MIP
11
problems to minimize the AV relocation VMT and optimize AV routes, while fulfilling all
12
households travel demand. After applying the models to the Atlanta metropolitan area, we found
13
that even if consumers do not change the existing travel pattern, approximately 18% of the
14
households can reduce vehicle ownership. If the schedule is relaxed by 15-minute time windows
15
(i.e. arriving at destination 15 minutes after the current arrival time is allowable) up to 24.1% of
16
the households are likely to at least eliminate one of the current private vehicles. The logistic
17
regression model results show that higher income families, who live in suburban neighborhoods
18
and generate more shorter trips, are more likely to be able to reduce vehicle ownership once
19
PAVs are adopted.
20
In return for vehicle ownership reduction, a significant amount of unoccupied VMT will
21
be generated in the region. For households who can reduce vehicle ownership, approximately
22
29.8 unoccupied VMT are generated per day per reduced vehicle. In the region, total VMT will
23
increase by at least 13%. Such increase only includes unoccupied VMT generated during the
24
vehicle relocation process. Other excessive VMT, such as cheaper parking lots cruising VMT,
25
changes in travel behaviors (destination selection), will inevitably inflate the estimation. The
26
majority of the occupied VMT occurs during evening peak hours. The spatial distribution
27
patterns of the excessive VMT indicates that regions with more disaggregated land use patterns,
28
especially larger mismatch between work and residential zones may experience larger VMT
29
increases in the future in the PAV dominated future. Finally, most short intra-zonal repositioning
30
trips take place in midday, which leads to larger percentage increase in the V/C Ratios during
31
midday.
32
The designed and implemented models can be used as pioneering tools to analyze the
33
vehicle ownership reduction and unoccupied VMT generation potentials in the era of PAV. The
34
results of this study can inform policy makers regarding the challenges of PAVs, if widely
35
adopted in the region, on existing transportation infrastructure, so that adaptation policies can be
36
drafted to prepare for the coming of AVs. Such policies may include travel demand management
37

16
tools, such as unoccupied VMT fees during peak hours to alleviate pressures on existing
1
infrastructures and the design of dedicated AVs lanes to improve road capacity on expressways,
2
where most of the unoccupied VMT are loaded.
3
While our results offer new understanding regarding regional light duty vehicle inventory
4
and unoccupied VMT generation after the coming of PAVs, there are several aspects that merit
5
future research efforts. Our models are developed based on the assumptions that travel behaviors,
6
such as the trip generation rates, the choice of destinations, and the travel schedules of household
7
members, will not vary significantly in the future. To gain more understanding regarding VMT
8
generation, future efforts may employ stated preferences survey to examine evolving trends in
9
travel behaviors. There has already been a wealth of literature regarding how different business
10
models of AVs, especially the Shared AVs (SAVs) will influence regional vehicle inventory and
11
VMT generation. It is critical to synthesizing current understandings to draw a comprehensive
12
picture regarding how different market penetration of various business models of AVs will
13
influence travel demand and consequently travel energy consumptions in the future. Last but not
14
least, more regional attitude surveys should be conducted to characterize early adopters’ socio-
15
demographics and economic features to understand potential AV adoption trajectories, especially
16
the adoption rates of PAVs, SAVs, and Transit Complementary AVs in the future. This
17
information can provide critical guidance to plan for AVs during the transition period, which is
18
not examined in this study.
19
20
ACKNOWLEDGE
21
The authors are grateful for ARC’s generous provision of the geocoded 2011 travel survey and
22
2015 Activity Based Model outputs.
23

17
REFERENCES
1
ARC. (2011). ARC Regional Travel Survey. Retrieved from http://atlantaregional.org/wp-
2
content/uploads/tp-2011regionaltravelsurvey-030712.pdf
3
Bansal, P., Kockelman, K. M., & Singh, A. (2016). Assessing public opinions of and interest in new
4
vehicle technologies: An Austin perspective. Transportation Research Part C: Emerging
5
Technologies, 67, 1–14.
6
Bischoff, J., & Maciejewski, M. (2016). Simulation of City-wide Replacement of Private Cars with
7
Autonomous Taxis in Berlin. Procedia Computer Science, 83, 237–244.
8
Boesch, P. M., Ciari, F., & Axhausen, K. W. (2016). Autonomous Vehicle Fleet Sizes Required to Serve
9
Different Levels of Demand. Transportation Research Record: Journal of the Transportation
10
Research Board, 2542, 111–119.
11
Burns, L. D., Jordan, W. C., & Scarborough, B. A. (2013). Transforming personal mobility. The Earth
12
Institute, 431, 432.
13
California, D. (2017, June 27). Testing of Autonomous Vehicles. Retrieved July 28, 2017, from
14
https://www.dmv.ca.gov/portal/dmv/detail/vr/autonomous/testing
15
Chen, T. D., Kockelman, K. M., & Hanna, J. P. (2016). Operations of a shared, autonomous, electric
16
vehicle fleet: Implications of vehicle & charging infrastructure decisions. Transportation
17
Research Part A: Policy and Practice, 94, 243–254.
18
Childress, S., Nichols, B., Charlton, B., & Coe, S. (2015). Using an Activity-Based Model to Explore the
19
Potential Impacts of Automated Vehicles. Transportation Research Record: Journal of the
20
Transportation Research Board, 2493, 99–106.
21
Fagnant, D. J., & Kockelman, K. (2015a). Preparing a nation for autonomous vehicles: opportunities,
22
barriers and policy recommendations. Transportation Research Part A: Policy and Practice, 77,
23
167–181.
24
Fagnant, D. J., & Kockelman, K. M. (2014). The travel and environmental implications of shared
25
autonomous vehicles, using agent-based model scenarios. Transportation Research Part C:
26
Emerging Technologies, 40, 1–13.
27
Fagnant, D. J., & Kockelman, K. M. (2015b). Dynamic ride-sharing and optimal fleet sizing for a system
28
of shared autonomous vehicles. In Transportation Research Board 94th Annual Meeting.
29
Retrieved from
30
https://pdfs.semanticscholar.org/e960/62a0e83885f41d2d83090989c71aec44f2a5.pdf
31
Greenblatt, J. B., & Saxena, S. (2015). Autonomous taxis could greatly reduce greenhouse-gas emissions
32
of US light-duty vehicles. Nature Climate Change, 5(9), 860–863.
33
Haboucha, C. J., Ishaq, R., & Shiftan, Y. (2017). User preferences regarding autonomous vehicles.
34
Transportation Research Part C: Emerging Technologies, 78, 37–49.
35
Harper, C. D., Hendrickson, C. T., Mangones, S., & Samaras, C. (2016). Estimating potential increases in
36
travel with autonomous vehicles for the non-driving, elderly and people with travel-restrictive
37
medical conditions. Transportation Research Part C: Emerging Technologies, 72, 1–9.
38
Harper, C. D., Hendrickson, C. T., & Samaras, C. (2016). Cost and benefit estimates of partially-
39
automated vehicle collision avoidance technologies. Accident; Analysis and Prevention, 95(Pt A),
40
104–115.
41
IBM. (2017). IBM ILOG CPLEX Optimization Studio CPLEX User’s Manual. Retrieved from
42
https://www.ibm.com/support/knowledgecenter/SSSA5P_12.7.0/ilog.odms.studio.help/pdf/usrcpl
43
ex.pdf
44
Krueger, R., Rashidi, T. H., & Rose, J. M. (2016). Preferences for shared autonomous vehicles.
45
Transportation Research Part C: Emerging Technologies, 69, 343–355.
46
Levin, M. W., & Boyles, S. D. (2015). Effects of Autonomous Vehicle Ownership on Trip, Mode, and
47
Route Choice. Transportation Research Record: Journal of the Transportation Research Board,
48
2493, 29–38.
49

18
Litman, T. (2014). Autonomous vehicle implementation predictions. Victoria Transport Policy Institute,
1
28. Retrieved from http://leempo.com/wp-content/uploads/2017/03/M09.pdf
2
Martinez, L., & Crist, P. (2015). Urban Mobility System Upgrade--How shared self-driving cars could
3
change city traffic. In International Transport Forum, Paris.
4
Mersky, A. C., & Samaras, C. (2016). Fuel economy testing of autonomous vehicles. Transportation
5
Research Part C: Emerging Technologies, 65, 31–48.
6
Rigole, P.-J. (2014). Study of a Shared Autonomous Vehicles Based Mobility Solution in Stockholm. diva-
7
portal.org. Retrieved from http://www.diva-portal.org/smash/record.jsf?pid=diva2:746893
8
Schoettle, B., & Sivak, M. (2015). Potential impact of self-driving vehicles on household vehicle demand
9
and usage. Retrieved from https://deepblue.lib.umich.edu/handle/2027.42/110789
10
Spieser, K., Treleaven, K., Zhang, R., Frazzoli, E., Morton, D., & Pavone, M. (2014). Toward a
11
Systematic Approach to the Design and Evaluation of Automated Mobility-on-Demand Systems:
12
A Case Study in Singapore. In Road Vehicle Automation (pp. 229–245). Springer, Cham.
13
Zhang, W., & Guhathakurta, S. (2017). Parking spaces in the age of shared autonomous vehicles: How
14
much parking will we need and where? Transportation Research Record: Journal of the
15
Transportation Research Board.
16
Zhang, W., Guhathakurta, S., Fang, J., & Zhang, G. (2015a). Exploring the impact of shared autonomous
17
vehicles on urban parking demand: An agent-based simulation approach. Sustainable Cities and
18
Society, 19, 34–45.
19
Zhang, W., Guhathakurta, S., Fang, J., & Zhang, G. (2015b). The Performance and Benefits of a Shared
20
Autonomous Vehicles Based Dynamic Ridesharing System: An Agent-Based Simulation
21
Approach. In Transportation Research Board 94th Annual Meeting.
22