Content uploaded by Morteza Taiebat
Author content
All content in this area was uploaded by Morteza Taiebat on Feb 25, 2019
Content may be subject to copyright.
Content uploaded by Miaojia Lu
Author content
All content in this area was uploaded by Miaojia Lu on Aug 16, 2018
Content may be subject to copyright.
Multiagent Spatial Simulation of
Autonomous Taxis for Urban Commute:
Travel Economics and Environmental Impacts
Miaojia Lu1; Morteza Taiebat2; Ming Xu, Ph.D., A.M.ASCE3; and Shu-Chien Hsu, Ph.D., A.M.ASCE4
Abstract: With the likelihood of autonomous vehicle technologies in public transport and taxi systems increasing, their impact on commut-
ing in real-world road networks is insufficiently studied. In this study, an agent-based model is developed to simulate how commuters travel
by autonomous taxis (aTaxis) in real-world road networks. The model evaluates the travel costs and environmental implications of substitut-
ing conventional personal vehicle travel with aTaxi travel. The proposed model is applied to the city of Ann Arbor, Michigan, to demonstrate
the effectiveness of aTaxis. The results indicate that to meet daily commute demand with wait times less than 3 min, the optimized
autonomous taxi fleet size is only 20% of the conventional solo-commuting personal car fleet. Commuting cost decreases by 38%, and
daily vehicle utilization increases from 14 to 92 min When using internal combustion engine aTaxis, energy consumption, greenhouse
gas (GHG) emissions, and SO2emissions are respectively 16, 25, and 10% higher than conventional solo commuting, mainly because
of unoccupied repositioning between trips. Given the emission intensity of the local electricity grid, the environmental impacts of electric
aTaxis do not show significant improvement over conventional vehicles. DOI: 10.1061/(ASCE)UP.1943-5444.0000469.© 2018 American
Society of Civil Engineers.
Author keywords: Autonomous vehicle; Commute travel; Multiagent simulation; Environmental impact.
Introduction
Since 1969, commuters in the US have primarily traveled to work
in personally owned vehicles; this has been the case for 90% of
all commuters during the past two decades (Santos et al. 2011).
Consequently, heavy traffic congestion easily occurs during peak
commute hours, which generates hefty travel costs and consider-
able environmental impacts. For example, Los Angeles currently
experiences the most severe traffic congestion in the US, with a
typical half-hour commute taking 60% longer during the morning
and 81% longer during the evening (Chew 2016). Light-duty ve-
hicles, including passenger cars and light-duty trucks, are respon-
sible for 61% of transportation greenhouse gas (GHG) emissions
in the US (EPA 2016). Every year, over 2,200 premature deaths
and at least $18 billion in health care costs in 83 of the largest
urban areas in the US can be partly attributed to air pollution from
traffic (Copeland 2011). Meanwhile, personal cars are unused for
approximately 95% of the day (OECD 2015). The 2009 National
Household Travel Survey (NHTS) data show that the average ve-
hicle ownership per licensed driver is 0.99 (Santos et al. 2011).
There are far more cars in the US than Americans need to reach
their desired destinations according to current travel patterns in
most locations (Fagnant and Kockelman 2014b).
Fully autonomous vehicles are expected to become a commercial
reality in the next decade. Given the higher capital cost of early adop-
tion, they are likely to be introduced first in public fleets and by
Transportation Network Compnies (TNCs), such as Lyft, Uber, and
Car2Go (Heard et al. 2018). Ridesharing and car-sharing companies
are teaming up with automakers to introduce fleets of driverless taxis,
which they envision becoming ubiquitous in urban areas. Autono-
mous taxis (aTaxis) may provide a solution to the problems presented
above. The trajectory of technological progress suggests that aTaxis
will eventually be able to travel anywhere a conventional vehicle can
go. The use of aTaxis in car-sharing services may compete with con-
ventional taxis or even shared taxi services because this new mode
can bypass the costs associated with drivers (Liang et al. 2016;
Zachariah et al. 2014). Specifically, aTaxi systems have the potential
to reduce the average wait time and enhance ridematching experien-
ces for passengers compared with conventional car-sharing programs
(such as Zipcar and Car2go) with fixed rental and return stations;
aTaxis may also reduce operating costs and provide more affordable
service for low-income populations in comparison with app-
based car-sharing programs (such as Uber) (Shen and Lopes
2015;Zhang et al. 2015a). Compared with personal vehicles, aTaxis
may transform transportation from an owned asset into a subscription
or pay-on-demand service, reducing vehicle ownership needs ac-
cordingly (Fagnant and Kockelman 2014b). Used in this way, aTaxis
may enable consumers to make more spontaneous trips, be more
productive, and/or have more time to relax during travel, in addition
to providing more predictable and shorter travel times and improving
rider safety (Burns et al. 2013).
This study analyzes the potential of using aTaxis as a transport
mode for commuting travel rather than as a full replacement of
1Ph.D. Candidate, Dept. of Civil and Environmental Engineering,
Hong Kong Polytechnic Univ., Hong Kong 99077, China.
2Ph.D. Candidate, School for Environment and Sustainability, Univ. of
Michigan, Ann Arbor, MI 48104; Ph.D. Candidate, Dept. of Civil and
Environmental Engineering, Univ. of Michigan, Ann Arbor, MI.
3Associate Professor, School for Environment and Sustainability,
Univ. of Michigan, Ann Arbor, MI 48104; Associate Professor, Dept. of
Civil and Environmental Engineering, Univ. of Michigan, Ann Arbor, MI.
4Assistant Professor, Dept. of Civil and Environmental Engineering,
Hong Kong Polytechnic Univ., Hong Kong 99077, China (corresponding
author). ORCID: https://orcid.org/0000-0002-7232-9839. Email: mark
.hsu@polyu.edu.hk
Note. This manuscript was submitted on August 31, 2017; approved on
April 20, 2018; published online on August 15, 2018. Discussion period
open until January 15, 2019; separate discussions must be submitted for
individual papers. This paper is part of the Journal of Urban Planning
and Development, © ASCE, ISSN 0733-9488.
© ASCE 04018033-1 J. Urban Plann. Dev.
J. Urban Plann. Dev., 2018, 144(4): 04018033
Downloaded from ascelibrary.org by University of Michigan on 08/23/18. Copyright ASCE. For personal use only; all rights reserved.
existing transportation networks. The objective of this study is to
optimize aTaxi fleet size to meet commuting demand, keeping wait
times below an acceptable threshold and minimizing the system
vehicle miles traveled (VMT). The corresponding environmental
performance and total travel cost of this system are evaluated using
an agent-based modeling (ABM) method. The commuting model
simulates heterogeneous travel patterns to anticipate aTaxi system
implications for various travelers who previously commuted in per-
sonal vehicles. This research contributes to the understanding of
the impact of autonomous vehicles in three areas. First, the simu-
lation is based on a real road network. Second, the hidden travel
costs related to the value of commuters’time are considered. Third,
the environmental impacts of internal combustion engine (ICE)
aTaxis and electric aTaxis are both evaluated.
The paper is organized as follows. First, the ABM literature on
autonomous vehicles is reviewed to inform the development of our
method for modeling the commute with aTaxis in an urban road
network. The method is shown and explained in detail in sub-
sequent sections. Next, the application of the model to Ann Arbor,
Michigan, in the US is presented, followed by the main results of
several scenarios. Conclusions are drawn from the simulation
results and potential directions for future research are offered.
Literature Review
Several modeling efforts have addressed the potential impacts of
autonomous vehicles on traffic networks. Fagnant and Kockelman
(2014b) designed an agent-based model for autonomous vehicle-
sharing throughout a grid-based urban area and concluded that
one shared autonomous vehicle (SAV) could replace approximately
11 privately owned vehicles, traveling 10% more distance than
comparable nonshared trips but also resulting in improved environ-
mental impact. Boesch and Ciari (2015) suggested that agent-based
transport models are suitable for modeling future transport scenar-
ios that incorporate autonomous vehicles. They discussed some
possible research questions related to autonomous vehicles, such
as potential future car fleet size, prospective demand patterns,
and possible interactions between public transport and autonomous
vehicles. Burns et al. (2013) applied a relatively simple analytical
model to the case of Ann Arbor, Michigan, and concluded that
autonomous vehicle-sharing could enhance mobility at consider-
ably lower cost than privately owned vehicles. Zellner et al.
(2016) used an agent-based approach to examine how interventions
such as using autonomous shuttles and making streetscape en-
hancements for pedestrians and cyclists may mitigate the first/last
mile problem of public transit; they also considered other factors
such as parking fees and fuel costs. They simulated four Chicago
neighborhoods with different densities and income levels; auto-
mated shuttle buses were assumed to have no capacity constraints.
They concluded that a dedicated automated shuttle service could
support significant mode shifts by increasing the utilization of
public transit. Liang et al. (2016) simulated the use of electric au-
tomated taxis for the first/last mile of train trips with the objective
of maximizing daily profits by optimizing service zone locations
and which reservations were accepted. However, the model only
considered trips occurring in the service zone, thus ignoring inter-
zonal trips. Additionally, the model assumed that all the origins and
destinations of passengers’requests were coming from or going to
the center of the service zone. The automated taxis were treated as
“flows”rather than as independent vehicles, which means that the
model did not represent the battery recharging needs of specific
vehicles.
Zhang et al. (2015a) used agent-based modeling to study
the effect of shared autonomous vehicles on urban parking de-
mand by varying fleet size and passenger wait time in a hypotheti-
cal city laid out in a grid network. Their simulation results
indicated that with a low market penetration rate of 2%, SAV users
reduced their parking demand by 90%. Fagnant and Kockelman
(2015a) used an agent- and network-based simulation to deliver
a benefit-cost analysis for fleet size optimization with dynamic
ridesharing based on a system of SAVs in Austin, Texas. The
authors concluded that dynamic ridesharing could reduce overall
vehicle miles traveled, thus avoiding new congestion problems.
Chen et al. (2016) simulated the operation of shared autonomous
electric vehicles (SAEVs) under various vehicle range and charg-
ing infrastructure scenarios in a gridded city modeled roughly
on Austin, Texas and predicted that with each SAEV replacing
5–9 privately owned vehicles, the unoccupied VMT could be
reduced by 3–4%, with average wait times between 2 and
4 min. Martínez et al. (2016) developed an agent-based model
to simulate a station-based one-way car-sharing system by divid-
ing the city of Lisbon into a homogeneous grid of 200 m by
200 m cells in which trips were generated between two grid cells
at each hour. Martínez et al. (2014) proposed an agent-based sim-
ulation model to assess the market performance of newly shared
taxi service in Lisbon. They identified set of rules for space and
time matching between the shared taxis and passengers, but
ignored the interactions between passengers and vehicles (such
as the waiting time limit of passengers). Levin et al. (2017) used
realistic flow models to make predictions about the benefits of
replacing personal cars with SAVs and found that, without dy-
namic ridesharing, the additional unoccupied repositioning trips
made by SAVs increased congestion and travel times. However,
this model was based on a downtown grid network, and intrazonal
trips were not considered. Zhang and Guhathakurta (2017) exam-
ined the influence of SAVs on urban parking demand based on a
real transportation network with calibrated link level travel speeds;
however, in this research the trips always started and ended at the
traffic analysis zone (TAZ) centroid and intrazonal travel time was
ignored.
Table 1summarizes previous studies related to shared autono-
mous vehicle modeling. Most of the research done so far on this
topic has been simulated on a highly developed grid or hypothetical
city and is constrained by several assumptions, such as a grid-based
transportation network, constant travel speeds across the network,
and passengers with uniform travel behavior. Furthermore, the
planning and operation of autonomous taxis on commuting travel
has received less attention. The present work seeks to fill these
knowledge gaps.
Proposed Multiagent Model
This study utilizes agent-based modeling to simulate the antici-
pated autonomous vehicles’effect on commute travel. Agent-based
models (ABMs) are well suited for modeling and studying the im-
pacts of traffic behavior (Lu and Hsu 2017). Du and Wang (2012)
suggested that an ABM approach can explore explanations, testify
regarding assumptions, and predict changes in or the emergence of
individual behaviors resulting from urban change. ABMs enable
the representation of highly heterogenous and behaviorally com-
plex populations of agents, and enable the modeling, both spatially
and temporally, of large-scale interactions between agents for the
study of dynamic but coherent system behaviors (Eppstein et al.
2011). One of the benefits of the agent-based computational pro-
cess approach is that no complicated mathematical algorithms are
© ASCE 04018033-2 J. Urban Plann. Dev.
J. Urban Plann. Dev., 2018, 144(4): 04018033
Downloaded from ascelibrary.org by University of Michigan on 08/23/18. Copyright ASCE. For personal use only; all rights reserved.
required. The agents are driven by rational behaviors, and irrelevant
aspects are ignored. These features of ABMs may explain their
increasing popularity in studies of transportation logistics and
traffic flow. Miller and Heard (2016) suggested that agent-based
models can help define reasonable scenarios of technology
deployment and evaluate designs that can lower transportation-
related emissions.
The model is implemented with GAMA 1.7.0, a software
platform for constructing spatially explicit agent-based simulations.
Integrating a geographic information system (GIS) and traffic
simulation leads to a more realistic representation of real-world
transportation activities (Cai et al. 2012). Fig. 1shows how our
research was conducted according to the following steps:
1. Collect commute and spatial data of the study city, including the
road network, the geographic distribution of office, commercial,
and residential buildings, commuting speed, and a number of
commuting trips by trip start time.
2. Use agent-based modeling to understand how a system of aTaxis
will perform in meeting the daily commute demand.
3. Optimize fleet size to (1) ensure that wait times are below an
acceptable threshold during peak hours and (2) minimize total
VMT.
4. Once the fleet size is known, evaluate the available travel cost
and environmental impacts of this commuting system.
5. Compare the travel cost and environmental performance of the
aTaxi scenario with the personal car scenario.
Simulation Environment and Agents
Commuting demand is concentrated in two peak periods: 6:00–
9:00 a.m. and 4:00–6:00 p.m. Given that for the first possible
commute, the trip begins at 12:00 a.m., and for the last return
commute, the trip begins at 11:59 p.m. (Santos et al. 2011),
0:00:00–23:59:59 was chosen as the service period of the aTaxi.
Twenty four hours of commute behaviors were simulated in 5 min
time steps, resulting in 288-time steps in the 24-h service period.
In the model, office and residential buildings are represented
as the origin and destination of trips, and real road networks
are followed during the commutes.
There are two types of agents in the model—commuter agents
and aTaxi agents. Commuters, who place a request to an aTaxi,
and individual aTaxis, which set their shortest route paths for trans-
porting commuters to their destinations, behave according to
the well-known Floyd–Warshall algorithm (Aini and Salehipour
2012), one of the most efficient algorithms for finding the shortest
Fig. 1. Research workflow.
Table 1. Previous studies related to shared autonomous vehicle modeling
Reference Objects Method Transportation network Findings
Fagnant and
Kockelman (2014b)
SAV ABM Grid city One SAV could replace eleven private cars with 10%
more VMT and improve environmental impacts
Burns et al. (2013) SAV Analytical model None SAVs have lower costs than private cars
Zellner et al. (2016) Autonomous
shuttles
ABM None Autonomous shuttles could enhance the use of public
transit
Liang et al. (2016) Electric
automated taxis
Mathematical
models
Nodelink network Electric automated taxis used for first/last mile of train
trips
Zhang et al. (2015a) SAV ABM Grid city SAV users reduced their parking demand by 90% with
a low market penetration rate of 2%
Fagnant and
Kockelman (2015a)
SAV ABM Nodelink network Dynamic ride sharing could reduce overall vehicle
miles traveled, thus avoiding new congestion problems
Chen et al. (2016) Shared autonomous
electric vehicles
ABM Grid city Each SEAV could replace five to nine privately owned
vehicles
Martínez et al. (2016) Car sharing ABM Grid city Car sharing performed worse than private cars both in
terms of time and cost
Martínez et al. (2014) Shared taxi ABM Nodelink network Shared taxis could lead to reductions in average
waiting time and average taxi system fares
Levin et al. (2017) SAV Realistic flow
models
Nodelink network SAVs could increase congestion and travel times
without dynamic ride sharing
Zhang and
Guhathakurta (2017)
SAV Discrete event
simulation
Nodelink network
with calibrated speed
Parking land use could be reduced by 5% once SAVs
serve 5% of the trips within the city of Atlanta
© ASCE 04018033-3 J. Urban Plann. Dev.
J. Urban Plann. Dev., 2018, 144(4): 04018033
Downloaded from ascelibrary.org by University of Michigan on 08/23/18. Copyright ASCE. For personal use only; all rights reserved.
path between any two nodes in a given network (Floyd 1962;
Warshall 1962).
Commuters
Every commuter has two spatial parameters: home (a residential
building) and workplace (an office building). Population density is
based on the spatial distribution of commuters’home locations at
the beginning of the simulation. People commute between home
and workplace every weekday, with most starting their commute
between 6:00 and 9:00 a.m. and beginning their journeys home
around 4:00–6:00 p.m. Commuters’departure times from home
and workplace obey the normal distribution. The 20,000 commut-
ers have their choice of transportation, personal car or aTaxi.
Krueger et al. (2016) showed that travel cost, travel time, and
waiting time might be decisive factors influencing the adoption
of SAVs and the acceptance of dynamic ridesharing. In our model,
commuters have different hourly incomes that obey a lognormal
distribution. Commuters’waiting time limits are uniformly
distributed and vary from 1 to 5 min. Commuters can decide
whether or not to share vehicles with others. Commuters that
choose not to share bear a higher travel cost. Zhang et al.
(2015b) showed that the average hourly income for ridesharing
commuters is 13% lower than the national average. Hence, com-
muters’willingness to share is negatively correlated to their hourly
income in the model.
Autonomous Taxis
Based on commuters’willingness to share, there are two types of
aTaxis, one that can be shared by multiple passengers, and one that
can pick up and drop off a single passenger. The second condition
occurs when (1) a passenger is not willing to share an aTaxi with
others or (2) an aTaxi does not show up before reaching the waiting
time limit of the potential second passenger. Idle aTaxis are ran-
domly distributed in the city at the beginning of the simulation.
During the simulation, aTaxis park directly at the last passenger’s
destination if not assigned to the next trip. They picks commuters
up from their homes and bring them to their workplaces or
pick them up from their workplaces and bring them home. The
maximum capacity of an aTaxis is set as four. Only passengers that
have the same trip starting hour have the potential to share a ve-
hicle. The vehicles used in the model operate at different travel
speeds based on the time of day. To realistically simulate traffic
congestion during peak hours, vehicle travel speed depends on
the number of vehicles on the road and the road capacity [Eqs. (1)
and (2)]. In Eq. (2), the free-flow speed is a theoretical distance per
time unit that a vehicle can travel without the presence of other
vehicles (Jeerangsuwan and Kandil 2014); this value is set at
53.1 km/h (33.0mi=h) (Zhang et al. 2015a). The aTaxi optimizes
its route to deliver all onboard commuters to their respective desti-
nations. An optimized route is the shortest distance between the
highest αv(speed coefficient) to deliver all the commuters to their
destinations. The aTaxis’scheduled routes are first-come, first-
served for commuters willing to share rides, as will be explained
in detail in the next section
αv¼e
−Nroad
RC
αv∈½0.10;1.00#ð1Þ
v¼αv×vff ð2Þ
where Nroad = number of vehicles on the road; RC = road capacity;
v= vehicle speed; and vff = vehicle’s free flow speed.
Interactions among Agents
Ride Sharing
Ride sharing appears to be essential for sustainable adoption
of autonomous vehicle use to mitigate congestion and environ-
mental consequences (Taiebat et al. forthcoming). Fagnant and
Kockelman (2015a) showed that VMT may rise by over 8% if
no ride sharing is allowed in satisfying travel demand with autono-
mous taxis. Zhang et al. (2015b) found that autonomous vehicle
ride sharing offers superior service to a non-ride-sharing autono-
mous vehicle system, through shorter trips, lower trip costs, lower
VMT, and, in the long run, better environmental outcomes. In this
study, commuters can choose to participate in ridesharing if they
are willing.
There are four operational parameters in the model: waiting time
limit, occupancy, added distance, and in-vehicle time. Waiting time
limit is the maximum time a passenger will wait between when the
passenger requests a vehicle and when the vehicle arrives for
pickup. If passengers cannot get an aTaxi within the waiting time
limit, they use their personal car as usual. Occupancy is the number
of passengers in the aTaxi, which varies from 0 to 4. Ride sharing
occurs when occupancy is greater than 1. According to Zachariah
et al. (2014), to share a ride, an additional occupant cannot increase
the distance of any direct trip by more than 20%. Thus, added
distance should be 20% less than the random original distances
between passengers’homes and workplaces. For example, consider
two potential passengers who want to travel from their workplaces
to home, Passenger A and Passenger B. Passenger A’s home loca-
tion and workplace location are Ahand Aw, respectively, and
Passenger B’s home location and workplace location are Bhand
Bw, respectively. The following equations need to be satisfied
for ridesharing to occur. The aTaxi location when Passenger B asks
to share a ride is denoted by Brequest. The added distance algorithm
is defined in Eqs. (3)–(5) as
dBrequest−Bw ≤tB×vð3Þ
dAw−Bw−Ah−Bh≤1.2×dAw−Ahð4Þ
dAw−Bw−Ah−Bh≤1.2×dBw−Bhð5Þ
where d= distance; and t= waiting time limit.
The aTaxi first takes Passenger A home because of the first-
come, first-served rule. The aTaxi then stops to board additional
passengers if the maximum capacity has not been reached. This
study only considers ridesharing in SAV scenarios and assumes
all commuters drive individually with their personal vehicles in
the business as usual (BAU) scenario. In the SAV scenario, there
are two mode choices, aTaxi and personal car (PC). Passengers
choose different transport modes based on their waiting time limit
and the waiting time for the closest aTaxi. In the BAU scenario,
occupancy and added distance are set to 1 and 0, respectively,
and passengers’wait time is 0. In-vehicle time represents the time
spent in the traveling vehicle, which is converted into cost in
economic evaluations.
Travel Costs
Travel cost is the primary concern for people choosing among differ-
ent transport modes. One of the objectives of this study is to min-
imize the total travel cost in the commuting system, based on the
passenger perspective. Some studies have used detailed cost catego-
ries to estimate the total cost for the operation of an SAV system,
including vehicle costs (capital, running, and maintenance costs),
infrastructure costs, and fleet management service costs based on
© ASCE 04018033-4 J. Urban Plann. Dev.
J. Urban Plann. Dev., 2018, 144(4): 04018033
Downloaded from ascelibrary.org by University of Michigan on 08/23/18. Copyright ASCE. For personal use only; all rights reserved.
various operational scenarios (Bösch et al. 2017;Chen and
Kockelman 2016). This research only considers service cost for
commuters. Operational costs undoubtedly account for a large pro-
portion of a system’s costs for transportation network companies,
but from a consumer point of view, economics is the primary influ-
ence on the decision to adopt and utilize the system. In this study, the
explicit financial costs of the service for commuters are considered,
as well as the hidden costs associated with the time invested in vari-
ous mobility-related activities. This analysis has received less atten-
tion in the literature compared to the operational cost of systems.
Explicit Costs
The regular fare for UberX (nonsurge periods) consists of a base
fare of $1 and a $1.65 booking fee, plus $1.30 per mile plus $0.26
per minute. As aTaxis do not need drivers, their operating costs are
lower (Liang et al. 2016). In consideration of these cost reductions
and other factors, Fagnant and Kockelman (2015a) set their simu-
lated nonshared trip price to $1.00 per mile (less than a third of the
average taxi cab rates in Austin, Texas). Simulation results from
Burns et al. (2013) showed that the costs per trip-mile of personal
cars and SAVs were $0.75 and $0.41, respectively, without consid-
ering decreased parking costs and the value of time. Bauer et al.
(2018) estimated that the lowest cost of service provided by shared
automated electric vehicles fleet could be $0.29–$0.61 per revenue
mile. Spieser et al. (2014) concluded that a mobility system featur-
ing autonomous vehicles could be almost half as expensive as a
system based on conventional human-driven cars. An average $1
per trip-mile fare for nonshared aTaxis was assumed, and personal
car costs were assumed to be $1.4 per trip-mile based on the afore-
mentioned price ratio of aTaxis and personal cars. In ride sharing
situations, the explicit cost after picking up the next passenger is
shared by all the passengers based on their trip distances.
Hidden Costs
Value of time (VOT) here is defined as “the monetary valuation
of the total time invested in mobility-related activities”(Ellram
2002;Spieser et al. 2014). The time spent requesting, waiting, enter-
ing, and traveling is monetized with passengers’VOT based on level
of comfort. Less comfortable trips incur a higher cost (Spieser
et al. 2014). For example, personal trips on local roads during
free-flowing traffic are priced at 50% of the median wage
(Manpower-Research 2015) but the cost of traveling during heavy
traffic is represented at 150% of the median wage (Victoria
Transport Policy Institute 2013). Commuters experience a higher
level of comfort in aTaxis, because they can use their travel time
to perform other activities (reading, eating, talking, texting, sending
email, or watching a movie). Zhang et al. (2015a) and Wadud (2017)
contended that the personal valuation of travel time may decline as
passengers reap productivity gains due to time free from driving. In
contrast, Yap et al. (2016) showed that in-vehicle time in autonomous
vehicles is experienced more negatively than in-vehicle time in man-
ually driven cars because of the influence of travelers’negative atti-
tudes regarding the trustworthiness and sustainability of autonomous
vehicles. After considering these research results, the personal trip
time in aTaxis and personal cars was priced at 20 and 67% of
personal wages, respectively (Spieser et al. 2014). For example,
when the wage is $28.40=h (the median wage in Ann Arbor), the
corresponding VOT in aTaxis is approximately $5.68=h, one-third
of the VOT in personal cars, which is $19.03=h. Table 2summarizes
the parameters for the total travel cost evaluation.
Environmental Impacts
According to Fagnant and Kockelman (2014a), even gasoline-
powered SAVs could substantially reduce negative environmental
impacts by consuming approximately 16% less energy and
generating 48% less volatile organic compound emissions per
person-trip compared to conventional vehicles. However, Miller
and Heard (2016) argued that the GHG emissions of autonomous
vehicles could decrease on a functional unit basis (i.e., per-
passenger-mile); overall transport-related GHG emissions increase
as VMT increases (Brown et al. 2014;Morrow et al. 2014). Added
VMT may also amplify drawbacks associated with high automobile
use, such as increased gasoline consumption and oil dependence
and higher obesity rates (Fagnant and Kockelman 2015b). Zhang
et al. (2015b) indicated that although SAV systems tend to generate
more VMT, the vehicle life cycle GHG and air pollutant emissions
and energy consumption could still be reduced due to fewer
cold starts and reductions in parking infrastructure requirements.
Fagnant and Kockelman (2014b) acknowledged that relative to
personal cars, the reduced parking needs of aTaxis could reduce
emissions as well as traffic congestion.
GHG and pollutant emissions from conventional vehicles
could be further ameliorated through the use of low-emission
and energy-efficient drivetrain technologies (Taiebat et al.
forthcoming). Fully electrically-powered fleets could eliminate
all tank-to-wheel emissions from car travel (OECD 2015). Chen
et al. (2016) showed that SAVs and electric vehicle technology
have natural synergies. Thus, electric aTaxis have been integrated
into this commuting system. Hawkins et al. (2013) found that elec-
tric vehicles (EVs) powered by the present European electricity
mix could decrease global warming potential (GWP) 10–24% com-
pared to conventional diesel or gasoline vehicles, assuming lifetimes
of 150,000 km. The specific energy requirements for operat-
ing light-duty vehicles are approximately 0.30–0.46 kW · h=mi
(Kintner-Meyer et al. 2007), and the average emission rates of the
DTE energy system serving Michigan electric customers are about
1.4 kg=MW · h (3.1lbs=MW · h) for SO2and 884.5 kg=MW · h
(1,950 lbs=MW · h) for CO2(Parks et al. 2007), making the SO2
and GHG emissions of electric aTaxis straightforward to estimate.
The vehicle life cycle inventories from Chester and Horvath
(2008,2009) are used, which include parking infrastructure. In
our model, it is assumed that personal cars and aTaxis are all con-
ventional gasoline sedans. Following Fagnant and Kockelman
(2015a), aTaxis are assumed to have a 250,000-mi service life; this
aligns with the expected 7-year service life of Canadian taxis, which
typically log more than 248,000 mi over their lifetimes (Stevens
and Marans 2009), although SAVs may actually offer longer service
due to their smoother automated driving profile. Life-cycle environ-
mental impacts of autonomous vehicles and light-duty vehicles
(Fagnant and Kockelman 2014b;Zhang et al. 2015b) were the basis
for the environmental impacts of aTaxis and personal cars shown
in Table 3. Only energy consumption,GHG emissions, and SO2
emissions are considered.
Case Study of the City of Ann Arbor
Model Experiment Settings and Initialization
In this section, a detailed view of a city’s existing commuting
patterns, topology, and other characteristics used to build a
Table 2. Components of total travel cost
Travel cost Personal car aTaxi
Explicit cost $1.40 per trip-mile for
nonshared trip
$1.00 per trip-mile for
nonshared trip
Hidden cost $19.03=h with median
wage level
$5.68=h with median
wage level
© ASCE 04018033-5 J. Urban Plann. Dev.
J. Urban Plann. Dev., 2018, 144(4): 04018033
Downloaded from ascelibrary.org by University of Michigan on 08/23/18. Copyright ASCE. For personal use only; all rights reserved.
transportation model are presented. Recently passed legislation in
Michigan allows self-driving vehicles to operate on any Michigan
roadway, which widens opportunities for autonomous vehicle devel-
opment (Burden 2016). Ann Arbor is representative of small- to
medium-sized cities in the United States, based on the data from
the 2009 NHTS. The city covers an area of 44 mi2and has a pop-
ulation of 117,770 (City-Data 2013). Among the 39,095 peoplewho
live and work in Ann Arbor, 50% (around 20,000) drive single-
passenger vehicles to work, 20% walk to work, 11% take the bus,
and 5% bike to work, according to the Washtenaw Area Transpor-
tation Study’s most recent transit profile, conducted in 2009
(Biolchini 2013). Our analyses focus on the 20,000 people that drive
alone in their commuter travels (the BAU scenario in this study).
The model is based on an area of 6.97 ×6.29 mi that contains
the city of Ann Arbor. Taking advantage of Ann Arbor Open Data,
the spatial information for buildings, roads, and city boundaries
are incorporated into the model (City-Services 2017). In Fig. 2,
the residential and office buildings are represented by different col-
ors (grey for residential and purple for office/commercial), which
serve as the origins and destinations of commuter travels within
Ann Arbor. The population density in the model is based on the
spatial distribution of residential buildings. The vehicles are shown
as red squares. For people shown as circles, different colors depict
the different objectives, with blue denoting “working”people trav-
eling from home to work, and yellow depicting “resting”people
traveling from work to home. The median income of Ann Arbor
residents is $56,835 per year, which translates into $28.4=h
(40 h=week, 50 weeks=year). Table 4shows the basic parameters
used in the Ann Arbor case study.
Model Validation
Using real-world data to calibrate and validate the behavior model
increases the credibility of and trust in this agent-based model and
its results. Three components were used to validate the commuting
Fig. 2. Ann Arbor commute model.
Table 3. Potential environmental impacts of aTaxis and personal cars per
vehicle-mile traveled
Environmental impact
Personal
cars aTaxis
Electric
aTaxis
Energy consumption (MJ=VMT) 4.96 4.35 3.48
GHG emissions (kg CO2eq=VMT) 0.36 0.34 0.27
SO2emissions (g=VMT) 0.12 0.10 0.60
Table 4. Basic modeling parameters
Parameter Value
Service area 6.97 ×6.29 mi
Average speed 44.4 km/h (27.60 mi=h)
a.m. peak 6:00–9:00
p.m. peak 16:00–18:00
Free-flow speed 53.1 km/h (33.0mi=h)
Commute period 0:00:00–23:59:59
Commuters’average hourly income $28.4=h
Maximum aTaxi occupancy 4
© ASCE 04018033-6 J. Urban Plann. Dev.
J. Urban Plann. Dev., 2018, 144(4): 04018033
Downloaded from ascelibrary.org by University of Michigan on 08/23/18. Copyright ASCE. For personal use only; all rights reserved.
model based on the BAU scenario: commute speed, commute
time, and commute trips by time of day. The commute speed
and commute time were collected from an Ann Arbor commuting
survey (City-Data 2013). The survey data indicated an average
commute speed of 44.4 km/h (27.60 mi=h); the corresponding
simulation result was 44.3 km/h (27.52 mi=h). The average
surveyed commute time within Ann Arbor was 10 min, and
the commute time from the simulation results was 7.44 min. This
difference can be explained by the inclusion of boarding and
alighting time in the survey data; commute time from the simu-
lation results only considered driving time. Data from the 2009
National Household Travel Survey (NHTS) were used to validate
the commute trips by time of day (Fig. 3). These data contain
extensive information about each commuting trip made by indi-
viduals living and working in small- to medium-sized cities, in-
cluding start times of daily trips to work and return trips home. In
Fig. 3, the morning peak hours of commuting travel are from 6 to
9 a.m., and the evening peak hours are from 4 to 6 p.m. In the
simulation, the start times of trips to work and trips home both
follow a normal distribution. The simulation data in the figure
have the best fit with the NHTS data.
Scenario Simulation
Several scenarios were used for the evaluation of autonomous taxi
performance in commuting trips. The same random number was used
in the simulation runs for different scenarios to ensure that any
difference in outputs would not be caused by noise from the random
number seed that starts the simulation. All simulation results were
generated from 100-run Monte-Carlo simulations. These scenarios
were generated by varying three principal parameters in the simula-
tion: fleet size, vehicle types, and operation strategies.
•Fleet size: In the BAU scenario, the fleet size equaled the com-
muting population (commuters who drive alone to work). In the
SAV scenarios, the aTaxi fleet size was also related to the com-
muting population, but was varied from 10 to 90% of the BAU
commuting population in 10% steps.
•Vehicle types: The BAU scenario represented the current
situation—20,000 people commuting alone in their cars. In the
SAV scenarios, there were two kinds of scenarios simulated—
an all aTaxi scenario and a mode choice scenario. In the all
aTaxis scenario, all personal cars were replaced with aTaxis,
and people could choose whether to share aTaxis with others
or not. In the all aTaxi scenario, 50% of the people driving
alone to work in the BAU scenario could only choose aTaxis
as their commute mode, while the other 50% of the people kept
their previous commute modes, such as walking or cycling,
which were not covered in this study. In the mode choice sce-
nario, the 50% of people driving alone to work could choose
aTaxis or personal cars based on their waiting time limit and
the waiting time for the closest aTaxi. The electric aTaxi sys-
tem was also simulated, with the environmental impacts com-
pared to the personal car system. Full battery-electric vehicles
have a limited range compared to gasoline vehicles and thus
need time for recharging (OECD 2015). Nonetheless, Taiebat
et al. (forthcoming) indicated that it is easier to integrate elec-
tric propulsion vehicles into a dynamic ride-sharing system
than into a non-ride-sharing system, as the former has longer
and more frequent chargeable breaks during the daytime. Elec-
tric aTaxis were assumed to have a fast battery recharge time of
30 min (using Level III chargers) and a vehicle range of 110 mi
(Chen et al. 2016).
•Operation strategies: In the optimized fleet size scenario, several
vehicle operation strategies were tested for further performance
optimization. At the beginning of the simulation, idle aTaxis
were randomly distributed in the city (Zhang et al. 2015a),
or the empty aTaxis were spatially clustered according to popu-
lation or building density. During the simulation, the aTaxis
either parked directly at the last passenger’s destination if they
were not assigned to their next trip (OECD 2015) or they gravi-
tated toward high-demand areas based on population or building
density after sending their last passengers to their destinations
(Zhang and Guhathakurta 2017).
Fig. 4shows travel times for the SAV and BAU scenarios
(the average wait time for the BAU scenario was 0 min, because
people can drive their own cars anytime they like). In the SAV sce-
narios, when all the commute modes are aTaxis (the all aTaxis sce-
nario), waiting time was reduced from 2.88 to 0.70 min, because
the fleet size was larger. In the SAV scenarios when passengers
have mode choice, the waiting time for the aTaxi fleet was rela-
tively short, between 0.61 and 0.13 min, because passengers could
choose the most convenient mode.
Table 5shows the VMT of the SAV and BAU scenarios. Com-
pared with the BAU scenario, as fleet size increased in the SAV
scenarios, total VMT and unoccupied VMT increased. This was
Fig. 3. Commute trips within Ann Arbor by start time of trip on weekdays.
© ASCE 04018033-7 J. Urban Plann. Dev.
J. Urban Plann. Dev., 2018, 144(4): 04018033
Downloaded from ascelibrary.org by University of Michigan on 08/23/18. Copyright ASCE. For personal use only; all rights reserved.
a result of the cruise distances that aTaxis accumulated when com-
muters requested rides. The total cruise distances were be longer
when there were more aTaxis. But total VMT did not increase dras-
tically with larger fleet sizes, because the service aTaxis provide
overlaps with commuting activities already performed without
aTaxis.
In the SAV scenarios, the simulation results of the all aTaxis and
mode choice scenarios were compared. In the mode choice sce-
nario, unoccupied VMT were much smaller than in the all aTaxis
scenarios. The total VMT in the all aTaxis and mode choice sce-
narios were very close. However, a significantly larger fleet size
(more vehicles) was needed in the mode choice scenario. For ex-
ample, only 4,000 aTaxis were needed to serve 20,000 passengers
in the all aTaxis scenario, but in the mode choice scenario, 10,555
personal cars and 2,539 aTaxis were needed. This was because pas-
sengers with mode choices turn to personal cars as their commuting
mode when aTaxis could not arrive within their waiting time limit.
It can be concluded that the waiting time is still a big challenge for
aTaxis compared with personal cars.
Results and Discussion
The final ideal fleet size was determined by passengers’wait time,
in-vehicle time and total VMT. The optimized fleet size was deter-
mined when the average waiting time was less than 3 min, the aver-
age in-vehicle time was less than 15 min per trip, and the VMT was
minimized throughout the simulation day (Zhang et al. 2015a,b).
The optimized fleet size in this study was 4,000, 20% of the fleet
size in the BAU scenario. The average wait time was 2.74 min, and
the VMT increased by 33.6% because of the unoccupied vehicle
travel of the aTaxis. Because there is little difference in total
VMT for the all aTaxi and mode choice scenarios and many fewer
vehicles are needed in the all aTaxis scenario, the optimized
scenario uses 4,000 aTaxis in the all aTaxis scenario.
To further minimize the total VMT and average wait time, sev-
eral operation strategies were tested. Fig. 5shows the operation
algorithm for aTaxis. Some of the operation strategies were men-
tioned previously—location of initial parking and behavior after
serving the last passenger. High-demand areas refer to high pop-
ulation density areas or high building density areas. Ride sharing
Fig. 4. Travel time for SAV and BAU scenarios.
Table 5. Vehicle miles traveled for SAV and BAU scenarios
SAV
fleet size
VMT—aTaxi (miles) VMT—PC (miles) Unoccupied VMT (miles) Total VMT (miles)
All aTaxis Mode choice All aTaxis Mode choice All aTaxis Mode choice All aTaxis Mode choice
2,000 160,394 123,047 0 32,799 3,247 746 160,394 155,846
4,000 170,246 113,118 0 55,822 8,686 1,253 170,246 168,940
6,000 171,652 111,735 0 59,839 9,691 1,315 171,652 171,574
8,000 171,457 111,174 0 60,289 9,643 1,264 171,457 171,463
10,000 171,419 111,650 0 59,693 9,666 1,306 171,419 171,343
12,000 171,334 111,900 0 59,455 9,624 1,302 171,334 171,355
14,000 171,193 112,481 0 58,736 9,602 1,308 171,193 171,217
16,000 171,463 112,111 0 59,353 9,671 1,292 171,463 171,464
18,000 171,450 111,735 0 59,775 9,670 1,267 171,450 171,510
BAU 0 127,462 0 127,462
Note: VMT—aTaxi is the VMT traveled by aTaxis; VMT—PC is the VMT traveled by personal cars (PC); and unoccupied VMT is the cruise distance,
between car location at time of request and pickup location, that aTaxis accumulate when commuters request a ride.
© ASCE 04018033-8 J. Urban Plann. Dev.
J. Urban Plann. Dev., 2018, 144(4): 04018033
Downloaded from ascelibrary.org by University of Michigan on 08/23/18. Copyright ASCE. For personal use only; all rights reserved.
only occurs when all the conditions for ride sharing are satisfied.
The low rate of ridesharing can be explained. Table 6shows some
representative simulation results. The first column shows the origin
condition—empty aTaxis are randomly distributed in the initial
stage and park at the location of the last passenger’s destination
before receiving a new request. The second column shows the best
simulation results, in which the total VMT are minimized and the
average wait time is less than 3 min. Although the fourth and fifth
columns show less wait time and higher ridesharing rates, the total
VMT is significantly larger. Thus, the operation algorithm in the
second column (in which empty vehicles park based on population
density at the beginning of the simulation and wait at the location of
the last passenger’s destination until receiving a new request) was
used for the following simulation.
In the optimized fleet size scenario, vehicle utilization for daily
commuting improved to 92 min as opposed to the BAU scenario, in
which privately owned vehicles were typically used for 14 min in
daily commute travel. The average occupancy was 1.3 in the opti-
mized fleet size scenario. This may reflect the low probability of
matching trips that satisfy the ridesharing algorithm, a phenomenon
in accord with the findings of Zhang et al. (2015a).
Total travel cost is composed of explicit costs and hidden costs,
which are highly sensitive to the level of VMT and VOT. The more
vehicle miles traveled, the greater the total travel cost. The VMT in
Fig. 5. Operation strategies of aTaxis.
Table 6. Simulation results of respective operation strategies
Results
Number
12345
Initial parking based on
Population density NYNYY
Building density NNYNN
Drive toward areas with high
Population density NNNYN
Building density NNNNY
Fleet size 4,000 4,000 4,000 4,000 4,000
Total VMT (miles) 170,246 168,233 168,293 290,331 290,680
Unoccupied VMT (miles) 8,686 8,635 8,681 8,246 8,389
In-vehicle time (min) 12.85 12.94 12.93 14.26 14.29
Wait time (min) 2.74 2.68 2.69 1.54 1.54
Total ridesharing 4,112 4,195 4,063 4,582 4,472
Note: Y = yes; and N = no.
© ASCE 04018033-9 J. Urban Plann. Dev.
J. Urban Plann. Dev., 2018, 144(4): 04018033
Downloaded from ascelibrary.org by University of Michigan on 08/23/18. Copyright ASCE. For personal use only; all rights reserved.
aTaxis is higher because of the distance that vehicles travel
while unoccupied as they drive to pick up passengers. The lower
the value of time, the lower the total travel cost. In aTaxis, passen-
gers are relieved from driving and can use their time as desired.
Their productivity can be improved by working while riding in
aTaxis. Therefore, the VOT for aTaxis is significantly lower. Over-
all, for ride-sharing trips in the optimized SAV scenario, the aver-
age total cost per mile was approximately $1.29 ($1.0 in explicit
costs and $0.29 in hidden costs), which was 38% lower than the
non-ride-sharing trips in the BAU scenario.
In contrast, the environmental performance of the aTaxis system
was not positive, because the environmental impact of the transpor-
tation system is highly related to VMT, and VMT was higher in
the SAV scenarios because of unoccupied vehicle travel. In the
optimized SAV scenario, system energy consumption, GHG emis-
sions, and SO2emissions were 16, 25, and 10% higher, respec-
tively, than in the BAU scenario. The environmental results were
consistent with Miller and Heard (2016): autonomous vehicles
may be more environmentally friendly on a functional unit basis
(i.e., per passenger-mile), but overall transport-related GHG emis-
sions increase as VMT increases. Environmental outcomes did not
improve in the electric aTaxi scenario when the fleet size was set to
4,000. Although the corresponding system energy consumption
and GHG emissions were 7 and 1% lower, respectively, than those
in the BAU scenario, total SO2emissions increased by 560% com-
pared to the BAU scenario. This was mainly due to the carbon
emission intensity of Michigan’s grid mix. Thus, environmental
performance did not improve as expected with the introduction
of autonomous vehicles for commuting in Michigan.
It is also found that, although aTaxis require far fewer vehicles
than are currently on the road, the total distance traveled was
greater because of unoccupied aTaxi travel as the vehicles accom-
modate the geographical distribution of demand. To explore road
conditions with the introduction of aTaxis, road occupancy was
studied (Fig. 6). Road occupancy represents the total number
of vehicles using a specific road on one weekday. In the opti-
mized SAV scenario, average road occupancy increased by 12%
compared with the BAU scenario; however, as suggested by
Zakharenko (2016), increased traffic does not necessarily cause an
increase in congestion, because the SAVs are expected to run effi-
ciently. Traffic congestion should be further investigated with more
factors, such as direction of travel. This unexpected traffic problem
was due to low rates of ride sharing and increased VMT in the SAV
scenarios. The results indicate that policymakers and planners
should not view vehicle automation through rose-colored glasses
as a solution to traffic jams and environmental implications.
In the simulations for Ann Arbor, aTaxis were only used for end-
to-end trips, because the city has no mass transit system. Using
aTaxis to connect the first/last mile of trips on transit systems will
be explored further in future work. Given the relatively small size of
Ann Arbor, the results from this work are not representative for
other cities, especially large metropolitan areas in which the aver-
age commute time is over 1 h per day. Future studies will develop
similar agent-based models for large metropolitan areas with long,
complex commute patterns. In addition, we only considered how
the income of commuters affects their willingness to share rides.
The role of social and racial factors, which are equally important
to ride sharing, will be further examined in the future. Meanwhile,
more realistic features can be added to this modeling framework,
such as the consideration of traffic signals and further validation of
the model through vehicle trips that cross main intersections.
Conclusions and Policy Recommendations
This study developed a simulation model to evaluate the travel
costs and environmental impacts of aTaxis for commuting. The
major contribution of the model described in this paper is to
simulate aTaxis traveling on a real road network in which all ve-
hicles start and end their trips and travel on the road. Moreover,
hidden travel costs related to commuters’value of time were con-
sidered, and the environmental impacts of aTaxis were estimated
to compare electric aTaxis, gasoline aTaxis, and conventional
gasoline cars.
Fig. 6. Road occupancy of the optimized: (a) SAV scenario (fleet size ¼4,000); and (b) BAU scenario (fleetsize ¼2,000).
© ASCE 04018033-10 J. Urban Plann. Dev.
J. Urban Plann. Dev., 2018, 144(4): 04018033
Downloaded from ascelibrary.org by University of Michigan on 08/23/18. Copyright ASCE. For personal use only; all rights reserved.
The opt imized fleet size was obtained with minimized VMT and
reasonable average wait times for passengers; this study determined
the optimized fleet size to be 20% of the fleet size in the BAU
scenario. The results of the optimized fleet size scenario show a
38% reduction in total commute costs and an increase in the daily
vehicle utilization from 14 to 92 min; however, daily road occu-
pancy increased by 12%. The system’s energy consumption,
GHG emissions, and SO2emissions were 16, 25, and 10% higher,
respectively, compared to the BAU scenario. This was mainly due
to increased unoccupied VMT and less ride sharing. The unsatis-
factory environmental performance of aTaxis was not improved
when gasoline aTaxis were converted to electric aTaxis; the corre-
sponding energy consumption and GHG emissions were 7 and 1%
lower, respectively, than those in the BAU scenario, but SO2
emissions increased to 560% compared to the BAU scenario.
Our simulation results show that aTaxis do not exhibit signifi-
cant improvements in environmental performance compared to per-
sonal car use until more people are willing to share aTaxis rides. A
clear policy implication of this study is that aTaxi fleets do not nat-
urally lead to the higher environmental performance of transporta-
tion systems. Thus, tailored regulations must be in place before
deployment of this technology to ensure that the design and oper-
ation of aTaxi systems are environmentally compliant. Our model is
not designed as an accurate forecasting tool but rather as an initial
test of the potential application of aTaxis to commuting travel. The
model can be used to evaluate other prototypes in order to inform
policy discussions among planners and decision makers, as well as
to highlight gaps in existing methods that other model developers
can consider in order to improve future simulations.
Acknowledgments
The authors appreciate the financial support from the Hong Kong
Research Grants Council (Grant No. 25220615). M. T. would like
to thank Dr. Geoffrey M. Lewis (Center for Sustainable Systems,
University of Michigan, Ann Arbor) for invaluable input and
feedback.
Supplemental Data
The aTaxi simulation video is available online in the ASCE Library
(www.ascelibrary.org).
References
Aini, A., and A. Salehipour. 2012. “Speeding up the Floyd-Warshall
algorithm for the cycled shortest path problem.”Appl. Math. Lett.
25 (1): 1–5. https://doi.org/10.1016/j.aml.2011.06.008.
Bauer, G. S., J. B. Greenblatt, and B. F. Gerke. 2018. “Cost, energy, and
environmental impact of automated electric taxi fleets in Manhattan.”
Environ. Sci. Technol. 52 (8): 4920–4928. https://doi.org/10.1021/acs
.est.7b04732.
Biolchini, A. 2013. “Top 5 ways not driving to work can make a difference
in Ann Arbor.”The Ann Arbor News. Accessed December 13,
2016. http://www.annarbor.com/news/top-5-ways-not-driving-to-work
-can-make-a-difference-in-ann-arbor/.
Boesch, P. M., and F. Ciari. 2015. “Agent-based simulation of autonomous
cars.”In Proc., American Control Conf., 2588–2592. Piscataway, NJ:
IEEE.
Bösch, P. M., F. Becker, H. Becker, and K. W. Axhausen. 2017. “Cost-
based analysis of autonomous mobility services.”Transp. Policy
64: 76–91. https://doi.org/10.1016/j.tranpol.2017.09.005.
Brown, A., J. Gonder, and B. Repac. 2014. “An analysis of possible energy
impacts of automated vehicle.”In Road vehicle automation, 137–153.
New York: Springer.
Burden, M. 2016. “Snyder signs new Michigan self-driving vehicles
law.”The Detroit News. Accessed December 13, 2016. http://www
.detroitnews.com/story/business/autos/2016/12/09/autonomous-car-law
/95199544/.
Burns, L., J. William, and B. Scarborough. 2013. Transforming personal
mobility. New York: Earth Institute, Columbia Univ.
Cai, H., J.-S. Oh, and C. Y. David Yang. 2012. “Integrating GIS and micro-
scopic traffic simulation to analyze impacts of transportation infrastruc-
ture construction.”J. Comput. Civ. Eng. 26 (4): 478–487. https://doi.org
/10.1061/(ASCE)CP.1943-5487.0000136.
Chen, T. D., and K. M. Kockelman. 2016. “Management of a shared
autonomous electric vehicle fleet: Implications of pricing schemes.”
Transp. Res. Rec. 2572: 37–46. https://doi.org/10.3141/2572-05.
Chen, T. D., K. M. Kockelman, and J. P. Hanna. 2016. “Operations of a
shared, autonomous, electric vehicle fleet: Implications of vehicle &
charging infrastructure decisions.”Transp. Res. Part A: Policy Pract.
94: 243–254. https://doi.org/10.1016/j.tra.2016.08.020.
Chester, M., and A. Horvath. 2008. Environmental life-cycle assessment of
passenger transportation: A detailed methodology for energy, green-
house gas and criteria pollutant inventories of automobiles, buses, light
rail, heavy rail and air v. 2. Berkeley, CA: UC Berkeley Center for
Future Urban Transport: A Volvo Center of Excellence.
Chester, M., and A. Horvath. 2009. Life-cycle energy and emissions inven-
tories for motorcycles, diesel automobiles, school buses, electric buses,
Chicago rail, and New York City rail. Berkeley, CA: UC Berkeley
Center for Future Urban Transport: A Volvo Center of Excellence.
Chew, J. 2016. “This city has the worst traffic congestion in the US.”
Fortune. Accessed December 13, 2016. http://fortune.com/2016/03/22
/los-angeles-tomtom-traffic/.
City-Data. 2013. “Work and jobs in Ann Arbor, Michigan (MI) detailed
stats.”Accessed December 13, 2016. http://www.city-data.com/work
/work-Ann-Arbor-Michigan.html.
City-Services. 2017. “City of Ann Arbor’s data catalog.”Accessed August
19, 2017. http://www.a2gov.org/services/data/Pages/default.aspx.
Copeland, L. 2011. “Study blames 2, 200 deaths on traffic emissions.”USA
Today. Accessed December 13, 2016. http://usatoday30.usatoday.com
/news/nation/2011-05-25-traffic-pollution-premature-deaths-emissions
_n.htm.
Du, J., and Q. Wang. 2012. “Exploring reciprocal influence between
individual shopping travel and urban form: Agent-based modeling
approach.”J. Urban Plann. Dev. 137 (4): 390–401. https://doi.org/10
.1061/(ASCE)UP.1943-5444.0000084.
Ellram, L. M. 2002. “Total cost of ownership.”In Proc., Handbuch Indus-
trielles Beschaffungsmanagement, 659–671. New York: Springer.
EPA. 2016. US transportation sector greenhouse gas emissions 1990–
2014. Washington, DC: Office of Transportation and Air Quality.
Eppstein, M. J., D. K. Grover, J. S. Marshall, and D. M. Rizzo. 2011.
“An agent-based model to study market penetration of plug-in hybrid
electric vehicles.”Energy Policy 39 (6): 3789–3802. https://doi.org/10
.1016/j.enpol.2011.04.007.
Fagnant, D. J., and K. M. Kockelman. 2014a. The future of fully automated
vehicles: Opportunities for vehicle- and ridesharing, with cost and
emissions savings. Washington, DC: US Dept. of Transportation.
Fagnant, D. J., and K. M. Kockelman. 2014b. “The travel and environmen-
tal implications of shared autonomous vehicles, using agent-based
model scenarios.”Transp. Res. Part C: Emerging Technol. 40:
1–13. https://doi.org/10.1016/j.trc.2013.12.001.
Fagnant, D. J., and K. M. Kockelman. 2015a. “Dynamic ridesharing and
optimal fleet sizing for a system of shared autonomous vehicles.”
In Proc., Transportation Research Board 94th Annual Meeting.
Washington, DC: Transportation Research Board.
Fagnant, D. J., and K. M. Kockelman. 2015b. “Preparing a nation for
autonomous vehicles: Opportunities, barriers and policy recommenda-
tions.”Transp. Res. Part A: Policy Pract. 77: 167–181. https://doi.org
/10.1016/j.tra.2015.04.003.
Floyd, R. W. 1962. “Algorithm 97: Shortest path.”Commun. ACM 5 (6):
345. https://doi.org/10.1145/367766.368168.
© ASCE 04018033-11 J. Urban Plann. Dev.
J. Urban Plann. Dev., 2018, 144(4): 04018033
Downloaded from ascelibrary.org by University of Michigan on 08/23/18. Copyright ASCE. For personal use only; all rights reserved.
Hawkins, T. R., B. Singh, G. Majeau-Bettez, and A. H. Strømman. 2013.
“Comparative environmental life cycle assessment of conventional and
electric vehicles.”J. Ind. Ecol. 17 (1): 53–64. https://doi.org/10.1111/j
.1530-9290.2012.00532.x.
Heard, B. R., M. Taiebat, M. Xu, and S. A. Miller. 2018. “Sustainability
implications of connected and autonomous vehicles for the food supply
chain.”Resour. Conserv. Recycl. 128: 22–24. https://doi.org/10.1016/j
.resconrec.2017.09.021.
Jeerangsuwan, T., and A. Kandil. 2014. “Agent-based model architecture
for mesoscopic traffic simulations.”In Proc., 2014 Int. Conf. on
Computing in Civil and Building Engineering, 1246–1253. Reston,
VA: ASCE.
Kintner-Meyer, M., K. Schneider, and R. Pratt. 2007. Impacts assessment
of plug-in hybrid vehicles on electric utilities and regional US power
grids. Part 1: Technical analysis,1–20. Richland, WA: Pacific
Northwest National Laboratory.
Krueger, R., T. H. Rashidi, and J. M. Rose. 2016. “Preferences for shared
autonomous vehicles.”Transp. Res. Part C: Emerging Technol.
69: 343–355. https://doi.org/10.1016/j.trc.2016.06.015.
Levin, M. W., K. M. Kockelman, S. D. Boyles, and T. Li. 2017. “A general
framework for modeling shared autonomous vehicles with dynamic
network-loading and dynamic ridesharing application.”Comput. Envi-
ron. Urban Syst. 64: 373–383. https://doi.org/10.1016/j.compenvurbsys
.2017.04.006.
Liang, X., G. H. de Almeida Correia, and B. van Arem. 2016. “Optimizing
the service area and trip selection of an electric automated taxi system
used for the last mile of train trips.”Transp. Res. Part E: Logist. Transp.
Rev. 93: 115–129. https://doi.org/10.1016/j.tre.2016.05.006.
Lu, M., and S. C. Hsu. 2017. “Spatial agent-based model for environmental
assessment of passenger transportation.”J. Urban Plann. Dev. 143 (4):
04017016. https://doi.org/10.1061/(ASCE)UP.1943-5444.0000403.
Manpower-Research. 2015. Singapore yearbook of manpower statistics.
Singapore: Ministry of Manpower.
Martínez, L. M., G. Correia, and J. M. Viegas. 2014. “An agent-based
simulation model to assess the impacts of introducing a shared-taxi
system.”J. Adv. Transp. 49 (3): 475–495. https://doi.org/10.1002/atr
.1283.
Martínez, L. M., G. H. de Almeida Correia, F. Moura, and M. Mendes Lopes.
2016. “Insights into carsharing demand dynamics: Outputs of an
agent-based model application to Lisbon, Portugal.”Int. J. Sustainable
Transp. 11 (2): 148–159. https://doi.org/10.1080/15568318.2016
.1226997.
Miller, S. A., and B. R. Heard. 2016. “The environmental impact of autono-
mous vehicles depends on adoption patterns.”Environ. Sci. Technol.
50 (12): 6119–6121. https://doi.org/10.1021/acs.est.6b02490.
Morrow, W. R. III, J. B. Greenblatt, A. Sturges, S. Saxena, A. Gopal, D.
Millstein, N. Shah, and E. A. Gilmore. 2014. “Key factors influencing
autonomous vehicles’energy and environmental outcome.”In Road
vehicle automation, 127–135. New York: Springer.
OECD (Organisation for Economic Co-Operation and Development).
2015. “Urban mobility system upgrade: How shared self-driving cars
could change city traffic.”In International transport forum policy
papers, 6. Paris: OECD.
Parks, K., P. Denholm, and A. J. Markel. 2007. Costs and emissions asso-
ciated with plug-in hybrid electric vehicle charging in the Xcel Energy
Colorado service territory. Golden, CO: National Renewable Energy
Laboratory.
Santos, A., N. McGuckin, H. Y. Nakamoto, D. Gray, and S. Liss. 2011.
Summary of travel trends: 2009 national household travel survey.
Washington, DC: US Dept. of Transportation.
Shen, W., and C. Lopes. 2015. “Managing autonomous mobility on
demand systems for better passenger experience.”In Proc., PRIMA
2015: Principles and Practice of Multiagent Systems Engineering,
20–35. Berlin, Germany: Springer.
Spieser, K., K. Treleaven, R. Zhang, E. Frazzoli, D. Morton, and
M. Pavone. 2014. “Toward a systematic approach to the design and
evaluation of automated mobility-on-demand systems: A case study in
Singapore.”In Road vehicle automation, 229–245. New York: Springer.
Stevens, M., and B. Marans. 2009. Toronto hybrid taxi pilot. Toronto:
Toronto Atmospheric Fund.
Taiebat, M., A. Brown, H. Safford, S. Qu, and M. Xu. Forthcoming. “A
review on energy, environmental, and sustainability implications of
connected and automated vehicles.”Environ. Sci. Technol., in press.
Victoria Transport Policy Institute. 2013. Transportation cost and benefit
analysis—Travel time costs. Victoria, BC, Canada: Victoria Transport
Policy Institute.
Wadud, Z. 2017. “Fully automated vehicles: A cost of ownership analysis
to inform early adoption.”Transp. Res. Part A: Policy Pract. 101:
163–176. https://doi.org/10.1016/j.tra.2017.05.005.
Warshall, S. 1962. “A theorem on boolean matrices.”J. ACM 9 (1): 11–12.
https://doi.org/10.1145/321105.321107.
Yap, M. D., G. Correia, and B. van Arem. 2016. “Preferences of travellers
for using automated vehicles as last mile public transport of multimodal
train trips.”Transp. Res. Part A: Policy Pract. 94: 1–16. https://doi.org
/10.1016/j.tra.2016.09.003.
Zachariah, J., J. Gao, A. Kornhauser, and T. Mufti. 2014. “Uncongested
mobility for all: A proposal for an area wide autonomous taxi system
in New Jersey.”In Proc., Transportation Research Board 93rd Annual
Meeting. Washington, DC: Transportation Research Board.
Zakharenko, R. 2016. “Self-driving cars will change cities.”Reg. Sci.
Urban Econ. 61: 26–37. https://doi.org/10.1016/j.regsciurbeco.2016
.09.003.
Zellner, M., D. Massey, Y. Shiftan, J. Levine, and M. Arquero. 2016.
“Overcoming the last-mile problem with transportation and land-use
improvements: An agent-based approach.”Int. J. Transp. 4 (1):
1–26. https://doi.org/10.14257/ijt.2016.4.1.01.
Zhang, W., and S. Guhathakurta. 2017. “Parking Spaces in the Age of
Shared Autonomous Vehicle.”Transp. Res. Rec.: J. Transp. Res. Board
2651: 80–91. https://doi.org/10.3141/2651-09.
Zhang, W., S. Guhathakurta, J. Fang, and G. Zhang. 2015a. “Exploring the
impact of shared autonomous vehicles on urban parking demand: An
agent-based simulation approach.”Sustainable Cities Soc. 19: 34–45.
https://doi.org/10.1016/j.scs.2015.07.006.
Zhang, W., S. Guhathakurta, J. Fang, and G. Zhang. 2015b. “The perfor-
mance and benefits of a shared autonomous vehicles based dynamic
ridesharing system: An agent-based simulation approach.”In Proc.,
Transportation Research Board 94th Annual Meeting. Washington,
DC: Transportation Research Board.
© ASCE 04018033-12 J. Urban Plann. Dev.
J. Urban Plann. Dev., 2018, 144(4): 04018033
Downloaded from ascelibrary.org by University of Michigan on 08/23/18. Copyright ASCE. For personal use only; all rights reserved.
