Computers, Environment and Urban Systems 87 (2021) 101623

Contents lists available at ScienceDirect

Computers, Environment and Urban Systems

journal homepage: www.elsevier.com/locate/ceus

Optimization of carsharing fleet size to maximize the number of

clients served
Cristiano Martins Monteiro a, *, Cláudia A. Soares Machado b, Mariana de Oliveira Lage c,
Fernando Tobal Berssaneti d, Clodoveu A. Davis Jr a, José Alberto Quintanilha c
a
Department of Computer Science, Federal University of Minas Gerais – UFMG, at 2 Reitor Píres Albuquerque St, Belo Horizonte, MG 31270-901, Brazil
b
Department of Transportation Engineering, Polytechnic School, University of São Paulo – USP, 05508-070 São Paulo, Brazil
c
Institute of Energy and Environment, University of São Paulo – USP, 05508-010 São Paulo, Brazil
d
Department of Production Engineering, Polytechnic School, University of São Paulo – USP, 05508-010 São Paulo, Brazil

A R T I C L E I N F O A B S T R A C T

Keywords: Carsharing services aim at offering short-term car rentals. Those rentals can be offered in different modalities,
Carsharing such as round-trip and one-way. This work simulates the dynamics of possible carsharing clients searching for
Agent-based simulation resources availability and using the carsharing service. Clients can rent a vehicle if there is an available car at the
Simulation-based optimization
client’s origin station and an available parking slot at the client’s destination station. If there is no car or parking
slot available, the possible clients look for these resources in other stations nearby their origin and destination.
Together with this Agent-Based simulation, the fleet of vehicles is optimized to serve as many clients as possible
and to avoid wasting resources. This work proposes a Mixed-Integer Linear Programming Model to optimize the
fleet size of a carsharing service for the one-way and round-trip modes while simulating the clients interaction.
Different scenarios are analyzed using real parameters and spatial data from the city of São Paulo, Brazil. Those
scenarios are composed for round-trip and one-way, varying the number of possible clients, the maximum
number of available vehicles, different sets of carsharing stations and the corresponding number of parking slots.
We also explore the distance walked by clients to find an available vehicle to rent or a parking slot to deliver the
rented vehicle. The Agent-Based simulation showed that clients’ flexibility to walk was well aligned with a higher
allocation of vehicles in the same station, since it increased the possibility of sharing cars among nearby clients.
Also, results show that round-trip services can scale-up better than one-way services, and that distances walked
by the clients to be served are essential to make the most of the one-way mode’s vehicles and parking slots.

1. Introduction Although public transportation is often cheaper than private trans­

Mobility is essential for people living in urban areas. Regardless of
the purpose for moving, people spend effort, time and money by
commuting in their cities. Recent works aimed at improving that routine
understanding job inaccessibility inequalities (Tomasiello, Giannotti, &
Feitosa, 2020), by analyzing and predicting patterns of urban trips by
public transportation modes (Qi, Huang, Guan, & Fan, 2018), simulating
public transportation routes (Kieu, Ngoduy, Malleson, & Chung, 2019),
suggesting new pick-up and drop-off points (Monteiro, Martins, & Davis
Jr, 2018), reducing consumption costs of electric buses (Bartomiejczyk,
2018), and reducing travel time and costs by integrating different
transportation modes (Stiglic, Agatz, Savelsbergh, & Gradisar, 2018).
portation, issues such as inaccessibility, inflexibility of routes, low
supply at certain times, high travel duration and discomfort can demo­
tivate people to use public transportation modes, such as buses, sub­
ways, urban train, and so on. The pandemic of COVID-19 disease
required reducing crowding, with special concerns directed at the public
transportation system. In such scenarios, alternative transportation
modes emerge as viable commuting resources by avoiding many issues,
without considerably increasing the travel cost.
Some alternative transportation modes are currently based on the
concept of shared mobility. This concept, influenced by the shared
economy, consists in sharing or paying for limited use of vehicles as
needed, instead of owning them (Machado, Hue, Berssaneti, &

* Corresponding author.
E-mail addresses: cristianomartins@dcc.ufmg.br (C.M. Monteiro), claudia.machado@usp.br (C.A.S. Machado), mariana_lage@usp.br (M.O. Lage), fernando.
berssaneti@usp.br (F.T. Berssaneti), clodoveu@dcc.ufmg.br (C.A. Davis), jaquinta@usp.br (J.A. Quintanilha).

https://doi.org/10.1016/j.compenvurbsys.2021.101623
Received 30 September 2020; Received in revised form 26 November 2020; Accepted 1 March 2021
Available online 16 March 2021
0198-9715/© 2021 Elsevier Ltd. All rights reserved.
C.M. Monteiro et al.
Quintanilha, 2018; Shaheen, Chan, Bansal, & Cohen, 2015). People can,
for example, rent a vehicle for a short term or get a ride with someone
driving by a similar path to reduce mutual costs or to diminish the
number of vehicles on the roads. Sharing underutilized vehicles can
improve commuting with more sustainable solutions (Machado et al.,
2018), reducing the urban emissions of polluting gases (Martin & Sha­
heen, 2011), parking needs (Tchervenkov, Balac, Hörl, Becker, &
Axhausen, 2018), and traffic jams (Luan, Cheng, Zhou, & Tang, 2018).
There are different alternatives of shared mobility offered as a ser­
vice (Machado et al., 2018; McKenzie, 2020; Zhu, Zhang, Kondor, Santi,
& Ratti, 2020). Carsharing is one of them (Machado et al., 2018). In
carsharing, the client rents a vehicle in a “as-needed” basis, in which the
rentals can be as short as for only some minutes. Carsharing rental
modalities can be free-floating or station-based. Free-floating rentals
allow the client to start and end the use of the vehicle anywhere within a
service operating area. Therefore, the free-floating modality does not
define stations where the vehicles must be parked before and after the
rental. On the other hand, in station-based rentals such places are pre­
defined by the carsharing company. Station-based rentals can be offered
in the round-trip or one-way modes. Round-trip rentals require clients to
deliver the rented car at the same station where the rental has started.
One-way rentals allow clients to return the vehicle to a different station
(Jorge, Barnhart, & Correia, 2015; Machado et al., 2018).
Unlike taxi or Transportation Network Companies (TNCs), such as
Uber and Lyft, the carsharing client is also a costless driver. Therefore,
carsharing fares are expected to be lower than the ones for ridesourcing
services (Monteiro et al., 2019). In order to offer carsharing services in
attractive prices for one-way and round-trip modes, carsharing fleet-
sizing and the location of stations must be adequate (Correia &
Antunes, 2012; Lage, Machado, Monteiro, Berssaneti, & Quintanilha,
2019; Genikomsakis, Gutierrez, Thomas, & Ioakimidis, 2017; Jorge,
Barnhart, & Correia, 2015; Jorge, Correia, & Barnhart, 2014; Monteiro
et al., 2019). Simulating the dynamics of those modes is important to
allow carsharing planning. With appropriate analysis, carsharing com­
panies can optimize their vehicle fleet to satisfy as many clients as
possible, without wasting resources. That includes not to allow high
discrepancies through the fleet sizes of each station, balancing the fleet
to serve upcoming clients from as many stations as possible.
Urban challenges like carsharing fleet-sizing can be dealt with
microscale simulations in order to understand the dynamics and more
precisely predict the system performance limits. These simulations can
be modeled by Agent-Based models where clients are represented by
agents. Agents interact among themselves according with the system
rules and spatial characteristics, emulating realistic scenarios which
arise urban macroscale patterns (Smolak, Rohm, Knop, & Siła-Nowicka,
2020; Tomasiello et al., 2020).
This work simulates the microscale carsharing clients interaction
using a proposed Mixed-Integer Linear Programming (MILP) model to
optimize the fleet size of a carsharing service for the one-way and round-
trip modes. The proposed model aims to maximize the number of served
clients and to minimize the number of needed vehicles to be allocated in
each carsharing station, giving preference to balancing the fleet size
among the stations.
The adopted model was applied in realistic scenarios from São Paulo,
Brazil, simulating the effect of essential parameters to assess the car­
sharing performance. The analyses presented here can be useful for
carsharing decision-makers, supporting simulations for avoiding service
bottlenecks. Next section discusses related work. Section III introduces
the model’s formulation, Section IV explains the data and parameters
used, and Section V presents and discusses numerical results. Finally,
Section VI concludes the paper.
2. Literature review
Complex transportation problems have recently been solved using
Simulation-Based Optimization (SBO) approaches (Genikomsakis et al.,
2
Computers, Environment and Urban Systems 87 (2021) 101623
2017; Jorge, Correia, & Barnhart, 2014; Jorge, Barnhart, & Correia,
2015; Monteiro et al., 2019). SBO approaches allow the evaluation of
parameter changes, being useful to support “what-if” analyses for the
decision-making process (Monteiro et al., 2019).
In this work, the SBO approach focuses in the carsharing company’s
strategic decisions regarding the optimal number of vehicles to be
allocated in each station, for the round-trip and one-way modes, through
different client demands. Microscale simulations were performed
considering the carsharing clients as agents which share vehicles and
compete for rentals or parking slots. The following works made contri­
butions related to those presented in this article.
A MILP model is proposed in (Correia & Antunes, 2012) to maximize
the profits of a carsharing company considering all the revenues and
costs involved. Since the one-way mode allows the clients to deliver the
vehicle in another station, the number of vehicles in each station will
change along the day due to their different demands. The work aims to
optimize the locations for carsharing stations, avoiding fleet unbalanc­
ing among the stations on the one-way mode. The authors showed, in a
case study in Lisbon, Portugal, the impact of the location of stations and
of different scenarios of client behavior.
That work was extended (Correia, Jorge, & Antunes, 2014) to eval­
uate the impact of one-way clients being flexible to walk to the closest or
second closest station to get an available vehicle or parking slot. The
authors simulated the walking time between stations rather than the
walking distance. Results show that flexible clients knowing which
stations have available resources would have to walk about eight mi­
nutes to get to those stations. However, scenarios with flexible clients
can serve twice as many clients than scenarios with inflexible clients.
Jorge et al. (Jorge, Correia, & Barnhart, 2014) compare two methods
for the relocation of carsharing vehicles to fix the unbalancing among
stations on the one-way mode, in Lisbon, Portugal. The model achieved
better profits, mainly while considering vehicle relocation.
Even though the one-way mode can support more clients, many
carsharing companies do not offer one-way mode, since variations
throughout the demand of clients can lead to unbalancing among sta­
tions. Jorge et al. (Jorge, Barnhart, & Correia, 2015) propose a MILP
model to optimize the design of a carsharing service based on round-trip
to also offer one-way rentals in Boston, USA. The results show that
adding an optimized one-way mode to the service could increase the
earned profit.
A Multi-Objective MILP model is proposed (Boyac, Zografos, &
Geroliminis, 2015) to maximize the carsharing company’s net revenue
and the users’ benefit. Their work was extended (Boyac, Zografos, &
Geroliminis, 2017) to focus in the relocation problem when the car­
sharing service allows reservations. Three mathematical programming
models were presented for clustering stations, optimizing operations
and the flow of workers during relocations.
Bruglieri et al. (Bruglieri, Pezzella, & Pisacane, 2018) proposed a
Multi-Objective MILP model for the one-way mode in Milan, Italy. The
proposed model has three objectives: minimizing the number of workers
needed to relocate vehicles; maximizing the number of vehicle re­
locations to increase the number of served clients; and minimizing the
lengthiest relocation route. Approximate heuristic solutions were
compared with exact MILP ones observing the computational time spent
for both methods. Results show the benefits of using the approximate
heuristic methods instead of waiting for the exact optimal solution.
Lu et al. (Lu, Chen, & Shen, 2017) proposed a stochastic MILP based
on Benders decomposition to analyze the number of necessary vehicles,
fleet used percentage, relocation costs of vehicles among stations, and
metrics for quality of service on carsharing from the Boston-Cambridge
area in Massachusetts, USA. The results showed that companies offering
both round-trip and one-way modes can have their profits for one-way
decreased if the demand of clients be generated by pricing and stra­
tegic customer behavior instead of natural market penetration and user
adoption. That happens because if the one-way demand be generated by
pricing, for example, round-trip clients could prefer making two one-
C.M. Monteiro et al.
way rentals since it could be cheaper. Therefore, instead of increasing
the demand, pricing could rather reduce the company’s profit margin
since one-way relocations can be costful.
Vehicle relocations are common for one-way carsharing, since they
address two issues:
1) not being able to serve clients in some station because earlier clients
rented all vehicles from that station and delivered them to other
stations;
2) clients unable to deliver the vehicle in the desired station because all
parking slots are occupied.
Issue (2) can be even more restrictive to the one-way dynamics,
because it can happen on the first rentals of the day. In addition, that
issue can happen even when the fleet is balanced, with vehicles
distributed to all the stations. If one-way stations have a low number of
parking slots, and the stations are far apart, relocation operations can be
expensive and unable to solve issue (2) quickly enough. That happens
because stations with a low number of parking slots will more often be
subject to issue (2), and due to the fact that the vehicle cannot be used
while being relocated. Even with relocations, a necessary approach to
avoid issue (2) consists in the carsharing company holding more parking
slots than vehicles to rent.
Although relocation is important for balancing the fleet, it may not
significantly increase the carsharing company’s profit. As simulated in
(Santos & Correia, 2019) for a demand of 25% of all potential clients in
the city of Lisbon, Portugal, and considering that all relocations would
result in new clients served, relocations would increase the company’s
profit by only 3.2%.
However, if the carsharing investment be complemented by more
vehicles and parking slots, carsharing companies can also avoid both
issues (1,2). Furthermore, the management of one-way rentals can be as
simple as the round-trip ones, operation costs can be reduced and the
rental prices can be cheaper. Lower rental prices can attract more cli­
ents, and also motivate them to walk to another station to get an
available vehicle or parking slot.
Similar collaborations of clients were already evaluated in the
literature for one-way carsharing with electric vehicles (Boyac & Zog­
rafos, 2019; Cocca, Giordano, Mellia, & Vassio, 2019). The authors from
(Cocca et al., 2019) presented benefits derived from clients driving free-
floating electric vehicles up to nearby charging stations, instead of
parking them in any other place. In (Boyac & Zografos, 2019), the au­
thors evaluated the collaboration of clients regarding their spatial flex­
ibility (being available to walk to other station) and temporal flexibility
(being available to start their rental earlier or after the expected).
Among other results, authors found that spatial flexibility could increase
the number of served clients in up to 25%, having a stronger effect on the
system performance than temporal flexibility, which increased the
number of served clients in less than 5%.
This paper differs from related works by proposing a MILP model to
optimize the fleet size in each carsharing station, in order to maximize
the number of served clients given issues (1, 2), but considering the
clients’ tolerance for walking between nearby stations. To the best of our
knowledge, this is the first article comparing how round-trip and one-
way dynamics benefit from client flexibility to walk different distances
to another station to get an available vehicle or to drive further to find a
parking slot.
Experimental results presented in Section IV seek answers to these
research questions:
• How does increasing the vehicles number for both station-based
modalities impact the number of clients served?
• How much does the saturation of parking slots affect serving one-
way clients?
• How many parking slots per vehicle are needed to support the one-
way mode?
3
Computers, Environment and Urban Systems 87 (2021) 101623
• How does the distance walked by clients between stations influence
the total of clients served per day?
• What share of clients would have to walk to be served?
Next section presents the proposed MILP.
3. MILP formulation
This section presents the parameters, variables, objective functions
and constraints defined for the model. The proposed formulation can be
used for both round-trip and one-way carsharing modes and, following
(Mehdi, Sirui, Sina, & Roorda Matthew, 2015), does not split the oper­
ating day in time intervals. By doing so, carsharing dynamics are
simulated through a continuous time span, representing the flow of
clients and vehicles more realistically. Therefore, there are no explicit
time dimension discretizing the agents interaction. Instead, agents
perform their tasks exactly according with the constraints defined in the
MILP formulation.
Simulated scenarios and parameters were drawn from real carshar­
ing company and vehicle dealerships both from the city of São Paulo,
Brazil. Since those companies neither work with electric vehicles nor
with free-floating carsharing, constraints regarding charging vehicle
batteries and street parking availability are not included in this formu­
lation. Besides, it was assumed that the vehicle dealership stores will
work as stations for this evaluated carsharing service and it will work
inside the dealerships opening hours, i.e. from 8 a.m. to 6 p.m. Vehicle
maintenance tasks are expected to happen after the dealerships opening
hours, when the carsharing service is not working anymore at that day.
Therefore, the maintenance tasks were not included in the formulation.
Further explanations about the data and evaluated scenarios are pre­
sented in the Section IV.
Table 1 presents the variables and Table 2 presents the constants
used in the MILP formulation.
The MILP is formulated using the weighted sum method as in (Boyac
et al., 2015; Boyac et al., 2017) and the epsilon-constraint programming
as in (Bruglieri et al., 2018). The weighted sum was used to unify the
first objective f1 to the second objective f2 in just one objective (Marler &
Arora, 2010), so that a common MILP solver can also find the optimal
solution. The epsilon-constraint programming was used to define a
threshold ε which limits the total number of allocated vehicles for each
simulation run. Therefore, every scenario was simulated up to V times
(maximum total number of vehicles), varying the threshold ε from 1 to
V, and forming the curve with the optimal solutions (Pareto Front)
(Mavrotas, 2009).
Eqs. 1 and 2 present, respectively, the first and second objective
functions. The first objective f1 aims to maximize the number of served
clients, where xsr ∈ X represents a client interested in renting a vehicle
from station s and deliver it to station r. The client variables X are bi­
nary, assuming the value 1 when the client could be served and
0 otherwise. The second objective f2 minimizes the highest number of
vehicles to be allocated to a station in order to avoid fleet size discrep­
ancies among stations. Eq. 3 presents the optimized objective function,
built using the weighted sum method.
Table 1
Model variables.
Variable Description
ns Number of vehicles allocated to station s ∈ S
γ Greatest ns value among all station s ∈ S
xsr ∈ X Client interested in renting a vehicle in station s to go from station s to
station r
esl ∈ X Earlier client who rented a vehicle from station s to l
uts ∈ X Earlier client who delivered a vehicle from station t to s
ypq ∈ Yxrs Replica of xsr flexible to rent from station p to station q
C.M. Monteiro et al.
Table 2
Model constants.
Constant Description
V Maximum total number of vehicles
W Weight for objective f2
ε Limits the number of vehicles through the Pareto Front
Ps Number of parking slots in station s
Txsrstart Rental start time for client xsr
Txsrend Rental end time for client xsr
Ms “Big–M” with the max. clients the station s can serve
Mxrs “Big–M” with the max. Ms among all stations of Yxrs
∑ parking slots. Inequality 9 limits the total number of vehicles to the
max f1 = max xsr (1)
xsr ∈X
min f2 = min γ (2)
∑ ns ≤ Ps
max fo = max f1 − Wf 2 = max xsr − Wγ (3)
xsr ∈X
The f2 subtraction in the optimized function works as a penalty for
stations with many more vehicles than others. Due that penalty, if f1 has
more than one optimal solution (i.e., there is more than one way to serve
the maximum number of clients), the optimized function will select the
solution with lower γ. In other words, among the solutions serving the
highest possible number of clients, the optimized function will select the
solution with the lowest number of vehicles allocated to the station with
the highest number of vehicles.
In order to prioritize the optimal solution for f1 (maximum number of
clients served), the subtracted penalty value must be smaller than one.
Otherwise, the optimized function can avoid serving some clients not to
being penalized by f2. To solve that, the subtracted penalty is adjusted to
always be smaller than one due to the weight W multiplying f2. Eq. 5
defines the W value based on the optimization parameters. The rationale
of Eq. 5 is based on the upper bound for γ, defined by the Inequality 4.
Since W multiplied by γ must yield a value lower than one, the
greater the γ, the lower the value of W. A naive approach would set W as
the lowest fraction allowed by the computer. However, that can lead to
numerical errors during the optimization computations. Therefore, a
safer way to set a small enough W consists in first defining an upper
bound to γ. As γ stands for the greatest number of allocated vehicles in
the same station, its value is limited to the highest number of parking
slots (Ps) among all stations, as shown in Inequality 4.
max(Ps ) ≥ γ
∀s∈S
(4)
Wmax(Ps ) ≥ Wγ
∀s∈S
Therefore, if W times the highest number of parking slots is equalled
to a value smaller than one, for example 0.1, it is possible to define a
constant value for W that works as a good multiplier to ensure that the
multiplication Wγ will always be lower than one. Eq. 5 shows that
rationale.
Wmax(Ps ) = 0.1
∀s∈S
(5)
0.1
W=
max(Ps )
∀s∈S
Inequalities 6 to 14 compose the MILP’s constraints. As shown in Eq.
1, clients are represented by binary variables. Replicas of client variables
were created to simulate scenarios where clients are flexible to walk to
get an available vehicle or to drive further to find an available parking
slot. Each replica represents the original client going to a different sta­
tion. Replicas are also binary variables. The replicas only differ from
their client’s original variable by the starting station, or destination
station, or both. Replicas from client xsr are represented as ypq ∈ Yxsr .
Since the replicas are created to simulate the possible stations in
4
Computers, Environment and Urban Systems 87 (2021) 101623
which a client will be served, at most one variable among all replicas and
the original variable can be served. In other words, if a client will be
served in a simulation, only his/her original variable or one replica can
be equal to one. All the other variables from that client must be zero.
That constraint is presented by Inequality 6.
∑
xsr + ypq ≤ 1 (6)
p
yq ∈Yxs
r
Inequality 7, together with the objective function f2, ensures that γ
will hold the greatest number of vehicles allocated in the same station.
Inequality 8 limits the number of vehicles in a station to its number of
current ε value, which varies to generate the Pareto Front solutions.
ns ≤ γ ∀s ∈ S (7)
∀s ∈ S (8)
∑
ns ≤ ε (9)
s∈S
Inequalities 10 to 14 organize the flow of vehicles and clients
through the stations. Inequality 10 limits client servicing if no vehicles
are available in his/her starting station at the time when he/she looked
for the service. As shown by Inequality 10, and according to the car­
sharing common rationale, there is an available vehicle in a station if the
difference between the number of clients who earlier delivered a vehicle
to that station, and the number of clients who earlier rented a vehicle
from that station yields a result greater than zero. Since a station s can
start the day with some vehicles already allocated, the variable ns is also
added to Inequality 10 to represent the first served clients of the day,
before any other rental delivers a vehicle to that station. Therefore, if
even after adding the variable ns to that calculation the result is zero,
client xsr, in his/her original variable, cannot be served.
∑ ∑
xsr ≤ ns + uts − esl (10)
∀xsr ∈X esl ∈X:T start start
uts ∈X:T end start
t <T s s <T s
us xr e
l
xr
However, Inequality 10 does not guarantee the cases in which there
are available vehicles. Even if that calculation results in a value greater
than zero, the optimization software used could prefer not to serve client
xsr in order to serve the next client, who would return the rented vehicle
in less time, for example. Aiming to make the simulation more realistic
by giving priority to clients who have arrived first and found an avail­
able vehicle, the Inequality 11 uses the Big–M method to ensure
servicing xsr without disturbing the computation of available vehicles.
The Big–M method consists in including into the formulation a constant
M with a positive big enough value to assure that some rule will be
satisfied (Bazaraa Mokhtar, Jarvis John, & Sherali Hanif, 2011).
∑ ∑
Ms xsr ≥ ns + uts − esl (11)
∀xrs ∈X uts ∈X:T end
us
start
t <T s
xr
esl ∈X:T start
e
l
start
s <T sxr
By applying the Big–M in the Inequality 11, if there is one available
vehicle in starting station s, the variable xsr will necessarily go up to one
due to the ≥ inequality. And if there is more than one available vehicle,
the binary variable xsr will not limit the constraint’s right side, since xsr
multiplied by the Big–M constant Ms will yield a result always greater
the constraint’s right side. Therefore, the Inequality 11 assures the pri­
ority to clients who arrived first, without preventing other clients to be
served since their station has available vehicle and parking slot.
In order to avoid numerical errors due to the multiplication between
xsr and Ms, the Ms constant value should not be greater than needed for its
constraint (Bazaraa Mokhtar et al., 2011). Therefore, as Eq. 5 defines,
the value of Ms was defined as an upper bound for the inequality where it
acts. Since the positive terms on the right side of Inequality 11 are the
number of allocated vehicles and the number of earlier clients who
delivered a vehicle to that station, a reasonable value for Ms is the
C.M. Monteiro et al. Computers, Environment and Urban Systems 87 (2021) 101623

number of parking slots at station s added to the number of possible duration: varying from 45 min to 4 h per rental.
earlier clients who can deliver a vehicle to station s. Eq. 12 defines the Ms The location of stations are based on real vehicle dealerships in São
constant. Paulo, Brazil. The number of clients generated who start or end a rental
in each station is proportional to the total number of trips made using an
Ms = Ps + ∣U∣ ∀uts ∈ U⊂X, xsr ∈ X ∣ Tuend < Txstart (12)
t
s
s
r individual mean of transport (such as taxi or private vehicle) in the São
Inequality 13 guarantees the replica’s service when needed. That Paulo district where the station is located. A São Paulo district is the
constraint ensures that if the original client xsr was not served, and if smallest official spatial unit adopted by the local government, and was
there is an available vehicle at the starting station p of replica ypq, the the spatial unit used in the dataset from which the demand was drawn.
replica ypq will be served. This dataset used is the Metrô (Companhia do Metropolitano de São
∑ ∑ Paulo) origin-destination survey, same used by (Tomasiello et al., 2020)
Mp xsr + Mp ypq ≥ np + uts − esl (13) to generate agents to their simulation. Therefore, the demand simulated
was divided throughout the city’s districts, simulating more clients in
p
∀yq ∈Yxs uts ∈X:T end start
t <T p
esl ∈X:T start start
s <T p
r us yq e yq
l
regions with individual trips and vice versa.
Inequality 14 imposes the issue (2) mentioned in section II. In that The carsharing company which provided the round-trip duration
constraint, a client xsr cannot be served if the destination station r has no range reported that the usual carsharing client in São Paulo used the
available parking slot. Inequality 14 has no effect on the round-trip vehicle to go to more than one destination before finishing the rental.
mode, since every vehicle on the round-trip mode starts the day in its Probably it is related to the client preferring to use another trans­
parking slot at the starting station and, after each rental, the vehicle portation service instead of carsharing when he/she has only one
must be returned to the same station. destination. Therefore, the carsharing demand in São Paulo do not
∑ ∑ follow conventional patterns of commuting, being expected to neither
Pr ≥ nr + utr − erl (14)
∀xrs ∈X
utr ∈X:T end start erl ∈X:T start start cause a demand peak exactly when people are going to work nor cause
t <T s r <T s
only other demand peak exactly when people are coming back to home.
ur xr e xr
l

Next section presents how the clients and replicas were generated for
this optimization model.
4. Data and parameters
This section presents the data and parameters used in the optimiza­
tion. When the optimization starts, all the client and replica variables are
already defined for the solver to optimize using the formulation pre­
sented in Section III. Table 3 shows the simulation’s parameters.
The simulations represent one day of carsharing and were performed
varying the demand of clients, stations, parking slots and maximum
distance walked by the clients to be served. As shown in Table 3, the
simulated number of clients were 100, 300 and 500. These parameters
are consistent with the used in (Boyac & Zografos, 2019), being
reasonable for analysis of a local carsharing service. After client gener­
ation, properties such as starting and ending stations were maintained
for every simulation. Using the same client setup allows for fairer
comparisons as the other parameters vary.
The range between minimum and maximum round-trip duration
(from 1.5 h to 8 h) were provided by a carsharing company which
operates the round-trip mode in São Paulo. Although the company does
not set those minimum and maximum durations for rentals, clients
commonly follow that duration range since shorter trips will probably be
cheaper or easier if made by on-demand ride services (such as from
TNC), and longer trips will probably be cheaper if made using a common
rental car for a full day.
As the carsharing company does not offer one-way rentals, the ex­
pected duration range for one-way in the simulations was based on the
round-trip duration range. Since the round-trip mode consists in the
client having to drive back to the origin station, the duration on round-
trip is expected to be longer than in the one-way mode. In this simula­
tion, one-way rental duration was simulated being half of the round-trip
Thus, rental start and end times were generated randomly between the
vehicle dealerships opening hours, and the start and end times follow a
uniform distribution.
The original parking slots mentioned in Table 3 are the same used in
(Monteiro et al., 2019), regarding the stations based on real vehicle
dealerships from São Paulo. The 27 original parking slots are distributed
throughout 11 stations, in which the number of parking slots varies from
one to four. Additional parking slots are available at two vehicle deal­
erships that are not used for carsharing. Each of those parking locations
would have 2 additional parking slots. Since the total number of parking
slots (27 + 4 = 31) is still low for a metropolis such as São Paulo, partner
parking slots were also included.
Partner parking slots are alternative stations for the carsharing ser­
vice, available at commercial establishments such as parking garages,
hotels, colleges and shopping malls, selected using a “Location-Alloca­
tion” model as proposed in (Lage, Machado, Monteiro, Berssaneti, &
Quintanilha, 2019). The addition of partner parking slots yielded a set of
22 stations. Simulations were performed varying from 1 to 3 parking
slots in each partner station (thus, from 22 to 66 parking slots).
Combining the original number of parking slots to the others
mentioned in Table 3, we used six settings of parking slots:
• Only the original parking slots: total of 27
• Original + additional parking slots: 27 + 4 = 31
• Original + partner stations with one parking slot available: 27 + 22
* 1 = 49
• Original + additional parking slots + partner stations with one
parking slot: 27 + 4 + 22 * 1 = 53
• Original + partner stations with three parking slots available: 27 +
22 * 3 = 93
• Original + additional parking slots + partner stations with three
parking slots: 27 + 4 + 22 * 3 = 97

The Fig. 1 illustrates the distribution of simulated stations on the

Table 3 involved districts of São Paulo. Original, additional and partner stations
Simulation’s parameters. are represented, respectively, by red, orange and yellow points. The ten
regions indicate the São Paulo districts which have at least one simu­
Parameter Value
lated station.
Clients 100, 300 and 500 per day Simulations also considered the possibility that clients would be
Max. walking distance 0 m, 250 m, 500 m, 750 m and 1 km
Min. and Max one-way trip duration 45 min to 4 h
flexible to walk for some distance to reach a viable station. The
Min. and Max round-trip duration 1.5 h to 8 h maximum walking distances used in simulations were 0 m (where the
Number of original parking slots 27 clients are not flexible to walk), 250 m, 500 m, 750 m and 1 km. As the
Number of additional parking slots 4 number of clients, the parameters for walking distance are consistent
Number of partner parking slots 22 or 66
with the used in (Boyac & Zografos, 2019). The walking distance is

5
C.M. Monteiro et al. Computers, Environment and Urban Systems 87 (2021) 101623

Fig. 2. Round-trip scenarios with no walking.

Fig. 1. Simulated carsharing stations.

measured as the shortest distance on the São Paulo road network from
the desired station to another station with an available vehicle or
parking slot to deliver the shared vehicle. The shortest distance did not
consider driving right-of-way constraints, since clients would walk be­
tween the stations instead of driving. Next section presents experimental
results.

5. Experimental results

The simulations were performed on a Mac mini Server (Late 2012)

with operational system macOS Mojave 10.14.6, processor Intel Core i7
2.3 GHz, and RAM of 16 GB. The models were implemented using Py­
thon 3.7, with the wrapper PuLP1 version 1.6.0 and the solver CBC2
version 2.10.0. No swap operations between main and secondary
memories were needed.
Fig. 3. Round-trip scenarios with 500 m walking.
The time limit set for each CBC solver run was 20 min. All other
parameters follow CBC’s defaults. Therefore, the linear objective and
constraints were handled using the Simplex algorithm and the integrity
of integer variables was dealt using the Branch-and-Cut algorithm. The
constraint represented by Eq. 9 was updated before every CBC run to
increase the constant ε and to build the Pareto Front. That update was
made directly in the PuLP model from the previous solution of the Pareto
Front. By doing so, it was possible to avoid reloading the simulated data,
and also to avoid rebuilding the constraints for each ε value.
Subsection V.A discusses the optimized solutions, and subsection V⋅B
extends the analyses.

5.1. Simulation results

Figs. 2 to 7 present results. Each graph shows the variation of the

number of served clients in relation to the number of needed vehicles.
The number of parking slots is indicated by the color bar.
Fig. 2 shows round-trip scenarios considering that clients would not
walk to another station to rent a vehicle. It shows the number of served
clients considering a demand of 100, 300 or 500 potential clients, i.e.,
clients who seek the service, but will not necessarily be served. In this
Fig. 4. One-way scenarios with no issue 2) constraints and no walking.

1
https://pythonhosted.org/PuLP/
2
https://projects.coin-or.org/Cbc

6
C.M. Monteiro et al. Computers, Environment and Urban Systems 87 (2021) 101623

scenario, all results reached guaranteed optimality. In other words, it is

assured that the result shown is the best possible combination of vehicles
and parking slots to serve as many customers as possible. In this work,
the solutions with optimality guarantee are indicated on the legend by
“O. G.”.
The scenarios with higher client demand present lower saturation
while increasing the number of served clients. This indicates that
increasing the number of vehicles and the number of available parking
slots to more than 100 each can be beneficial. Consider the scenarios
with 100 and 500 potential clients, and with 97 available parking slots.
The scenario with 500 potential clients can serve 138% more clients
than the scenario with 100 clients, using only 27% more vehicles.
Fig. 3 shows round-trip scenarios where the clients are flexible to
walk up to 500 m between stations while looking for an available vehicle
to rent. Results follow the same patterns as in Fig. 2, and the number of
served clients does not increase significantly. The maximum number of
clients served increased from 228 to 235, a variation of only seven cli­
ents, equivalent to 3.1%. The summary of the variation for all walking
distances simulated is presented in Table 4.
Fig. 5. One-way scenarios with no issue 2) constraints and 500 m walking. As round-trip clients must deliver the vehicle at the same station
from which it was rented, every vehicle on round-trip can have a dedi­
cated parking slot. Therefore, the one-way issue (2) mentioned earlier
does not affect the round-trip performance. That pattern appears on
Figs. 2 and 3, where, in all simulated scenarios, there is no need to use
more parking slots than offered vehicles.
Fig. 4 presents the results for one-way rentals with no walking be­
tween stations and without constraints regarding issue (2). This scenario
is equivalent to the “partial-floating” case defined on (Repoux, Boyac, &

Table 4
Maximum clients served, vehicles and walking %.
Scenario Max. Raise Max. Raise Walking
Clients (%) Vehicles (%) Clients (%)

Round-trip no 228 – 97 0 –
walking
Round-trip 250 m 232 1.8 97 0 1
Round-trip 500 m 235 1.3 97 0 5
Round-trip 750 m 238 1.3 97 0 16
Round-trip 1 km 248 4.2 97 0 32
One-way no 285 – 89 – –
constraints no
walking
Fig. 6. One-way with all parking slots subject to issue (2) and no walking. One-way no 287 0.7 89 0 1
constraints 250 m
One-way no 289 0.7 87 − 2.2 5
constraints 500 m
One-way no 326 12.8 87 0 24
constraints 750 m
One-way no 378 16 88 1.1 44
constraints 1 km
One-way constraints 219 – 55 – –
on partners no
walking
One-way constraints 230 5 62 12.7 3
on partners 250 m
One-way constraints 278 20.9 79 27.4 8
on partners 500 m
One-way constraints 309 11.2 74 − 6.3 20
on partners 750 m
One-way constraints 357 15.5 69 − 6.8 35
on partners 1 km
One-way all 192 – 46 – –
constraints no
walking
One-way all 193 0.5 43 − 6.5 3
constraints 250 m
One-way all 252 30.6 51 18.6 4
constraints 500 m
Fig. 7. One-way with all parking slots subject to issue (2) and 500 m walking. One-way all 292 15.9 68 33.3 20
constraints 750 m
One-way all 329 12.7 69 1.5 35
constraints 1 km

7
C.M. Monteiro et al.
Geroliminis, 2014), where clients can park the vehicle outside the sta­
tion, but nearby it, when there are no available parking slots in the
station.
The number of clients served is greater than in the round-trip mode,
because one-way rentals tend to be shorter. However, having more cli­
ents served does not necessarily result in higher revenue or profit, nor
even in more idle vehicles. The increase in the number of served clients
shows signs of saturation even for the demand of 500 clients.
Fig. 5 shows the one-way results also without issue (2) constraints,
but considering that the clients are flexible to walk up to 500 m to find
an available vehicle. Some simulations using demand of 500 clients did
not obtain optimality guarantee due to the higher complexity of the one-
way optimization. Therefore, there is no guarantee that the results with
the “+” sign indicated in the legend are the solution with best possible
combination for that number of vehicles and parking slots. Each point
shown on the graph without optimality guarantee used up to three at­
tempts of 20 min processing to find an optimal solution. New simula­
tions using a longer run time can achieve the optimal solutions.
However, since the non-optimal solutions maintained the same pattern
as the optimal solutions, we expect that the optimal solutions are not
significantly different from the solutions shown in Fig. 5.
Even though the number of served clients increased only from 285 to
289 (four more clients, 1.4%), the number of vehicles required to serve
those 289 clients was reduced from 89 to 87 (two fewer vehicles, 2.2%).
Thus, the walking distance thresholds caused not only an increase in the
number of served clients, but also a reduction in the number of vehicles
needed to serve that many clients.
Figs. 8 and 9 present the effect of clients walking to be served, but in
scenarios with the one-way issue (2) applied only on the partner sta­
tions. In practical terms, it is as if the original and additional stations
(owned by the carsharing company) can always find enough space to
park the one-way delivered vehicles (i.e., parking on the street, or
elsewhere), but partner stations are limited to the preset parking slots.
Differently from Fig. 5, in which the clients are flexible to walk and
the issue (2) constraints are not applied, all solutions shown on the Fig. 8
have optimality guarantee. That pattern also occurs on the next figures,
indicating that, for the proposed MILP, adding the the client walking
distance thresholds makes the optimization problem harder to solve
than when adding issue (2) constraints.
Regarding the carsharing performance, adding issue (2) constraints
reduces the maximum number of served clients. This number was
reduced from 285 to 219 (66 fewer clients, 23.2%), and the number of
needed vehicles drops from 89 to 55 (34 fewer vehicles, 38.2%).
Therefore, although issue (2) reduces the number of clients and conse­
quently the revenue obtained, the investment on vehicles can also be
Fig. 8. One-way with partners subject to issue (2) and no walking.
8
Computers, Environment and Urban Systems 87 (2021) 101623
Fig. 9. One-way with partners subject to issue (2) and 500 m walking.
reduced.
Besides, Fig. 8 shows the scenario with 31 parking slots overcoming
even the one with 93 parking slots. That happens because solutions
using 27 and 31 parking slots are not affected by the one-way issue (2),
since those parking slots are not from partners. After applying issue (2)
constraints on the partners’ parking slots, the difference between the
maximum number of served clients using 31 and 97 parking slots
dropped from 80 to only 14 clients. This relative difference dropped
from 28.1% to only 6.4%.
Fig. 9 shows how clients walking up to 500 m can overcome issue (2)
on the partner stations. The maximum number of served clients
increased from 219 to 278 (up 59 clients, 26.9%) and the number of
required vehicles increased from 55 to 69 (up 14 vehicles, 25.5%).
Differently from the scenarios shown in Fig. 5, where the clients walked
and the parking slots were not subject to issue (2) constraints, the
number of served clients increased together with the number of needed
vehicles. This indicates that, as the clients are flexible to walk to other
stations seeking for available vehicles and parking slots, the drawbacks
from issue (2) are reduced.
Figs. 6 and 7 present the results of applying issue (2) constraints on
all parking slots. The performance of these scenarios with only original
and additional parking slots (27 and 31 slots respectively) were also
impaired. However, Figs. 6 and 7 show the greatest increase in the
number of clients served due to the flexibility to walk. The maximum
number of served clients increased from 192 to 252 (an increase of
31.3%), using only five more vehicles (a 12.12% increase).
Nevertheless, those performance benefits are only possible if a higher
number of clients are flexible to walk to another station. The following
subsection discusses that connection.
5.2. Relations among served clients, vehicles and parking slots
This subsection discusses the relationship between the number of
served clients, vehicles, and parking slots through the simulated pa­
rameters. Table 4 presents simulation results regarding walking dis­
tances, inclusion of issue (2) constraints, maximum number of served
clients, vehicles required, and clients who walked to be served. Each line
shows the percent increase in the maximum number of served clients,
and in the maximum number of needed vehicles if compared to the re­
sults from the line above it. The percentage of served clients who had to
walk is also shown.
In general, the greatest increasing percentage either of served clients
or needed vehicles happened on the scenarios with 500 or 750 m of
walking. The increasing percentages of one-way served clients were
higher than those obtained by (Boyac & Zografos, 2019), in which the
C.M. Monteiro et al. Computers, Environment and Urban Systems 87 (2021) 101623

cumulative increasing along the scenarios barely surpassed 25%. In this Table 5
simulation, the increase of served clients by walking up to 500 m instead Rates of one-way vehicles per parking slot.
of 250 m was up to 30.6% when considering all constraints. A possible Parking slots scenario 27 31 49 53 93 97
reason for the increasing percentages in this simulation be higher than
One-way no constraints no 1 1 1.09 1.13 1.48 1.09
the observed in (Boyac & Zografos, 2019) is due to different distances walking
between the simulated stations, since the stations in both works are not One-way no constraints 250 m 1 1 1.04 1.02 1.16 1.09
equally distanced between themselves. One-way no constraints 500 m 1.04 1.07 1.09 1.02 1.16 1.15
The number of clients who had to walk to be served always increased One-way no constraints 750 m 1 1 1 1.02 1.12 1.17
One-way no constraints 1 km 1 1 1.02 1 1.13 1.10
as the walking distance increases. Therefore, in order to maximize the One-way cstr. on partners no 1 1 1.63 1.77 2.38 1.76
number of served clients, the walking distance can certainly be used to walking
replace or strengthen the one-way relocation results. However, the One-way cstr. on partners 250 m 1 1.03 1.75 1.67 1.79 1.56
benefits of client walking starts to saturate for lengthier walking dis­ One-way cstr. on partners 500 m 1.04 1.11 1.75 2.04 1.39 1.41
One-way cstr. on partners 750 m 1 1 1.20 1.13 1.39 1.31
tances, depending on the distances among stations.
One-way cstr. on partners 1 km 1 1 1.09 1.08 1.37 1.23
Table 5 presents the rate of the number of parking slots divided by One-way all constraints no 2.08 1.94 3.77 3.53 2.33 2.11
the number of needed vehicles, considering the demand of 500 clients walking
per day. The greater the result, the more parking slots are needed per One-way all constraints 250 m 1.69 2.07 3.50 3.31 2.33 2.26
vehicle. Round-trip mode is not shown in Table 5 since their rates would One-way all constraints 500 m 1.17 1.30 3.50 3.12 1.98 1.90
One-way all constraints 750 m 1.17 1.30 1.48 1.47 1.50 1.42
always be 1. That happens for round-trip because all vehicles must be One-way all constraints 1 km 1.25 1.30 1.29 1.47 1.45 1.41
returned to their starting station after the rentals, and the starting sta­
tion has available parking slots for their allocated vehicles.
The highest rates occur on the scenarios with issue (2) constraints optimizations are independent between themselves, they can share
apply to all stations, with the shortest (or nonexistent) walking distance, similar optimal solutions. We also propose to generate different numbers
and using 49 and 53 parking slots. On those scenarios, there are 22 of potential clients and apply other statistical distributions. Another
partner stations with only one parking slot each, making issue (2) even future work consists in proposing methods for calculating price dis­
more restrictive. Thus, besides the walking distance, having more counts to be offered to clients who agree to walk, and analyzing the
parking slots in the same station helps to reduce the effects of issue (2). expected profit for the carsharing company after applying such
discounts.
6. Conclusions and final remarks
Declarations of Competing Interest
In this work, a MILP model was proposed for carsharing fleet-sizing
on the round-trip and one-way modalities. Instances were evaluated by None.
varying the number of clients interested in the service, settings of sta­
tions, number of vehicles, parking slots, and the maximum distance that Acknowledgements
clients would walk between stations to obtain an available vehicle or an
available parking slot to deliver the rented vehicle. The impact of The authors acknowledge the National Council for Scientific and
delivering a vehicle to a station with all parking slots occupied was also Technological Development (Conselho Nacional de Desenvolvimento
discussed. Científico e Tecnológico - CNPq) with grant number 304037/2015-0,
Those analyses were important to observe how the carsharing per­ CAOA Group, Zazcar - Smart Mobility for their help, and the Engineering
formance can saturate when all variables do not increase together and Technological Development Foundation (Fundação para o Desenvolvi­
proportionally. In addition, the analyses showed that one-way carshar­ mento Tecnológico da Engenharia - FDTE) and Coordination for the
ing services and without the relocation operation of vehicles should Improvement of Higher Education Personnel (Coordenação de Aperfei­
prioritize the number of available parking slots, mainly locating them in çoamento de Pessoal de Nível Superior - CAPES) for their help and
the same station or in nearby stations. Therefore, if a client could not financial support.
park the rented vehicle in the desired location, the client could find and
park it in another nearby parking slot. References
The distance walked by the clients was shown to contribute to reduce
Correia, G. H. A., & Antunes, A. P. (2012). Optimization approach to depot location and
the impact of missing parking slots and the need of relocation opera­
trip selection in one-way carsharing systems. Transportation Research Part E: Logistics
tions. However, as the maximum distance that clients are flexible to and Transportation Review (vol. 48,(1), 233–247.
walk increases, the percentage of clients who would need to walk to be Bartomiejczyk, M. (2018). Driving performance indicators of electric bus driving
served also increases. In real situations, it is likely that requiring clients technique: naturalistic driving data multicriterial analysis. In IEEE Transactions on
Intelligent Transportation Systems, no. 99 (pp. 1–10).
to walk lengthy distances would decrease the demand. Bazaraa Mokhtar, S., Jarvis John, J., & Sherali Hanif, D. (2011). Linear Programming and
Also in a real environment, generic unexpected events as car crashes Network Flows. 1em plus 0.5em minus 0.4em John Wiley & Sons.
or unscheduled vehicle maintenance can interfere on the vehicles Boyac, B., & Zografos, K. G. (2019). Investigating the effect of temporal and spatial
flexibility on the performance of one-way electric carsharing systems. Transportation
availability. Besides, external events such as traffic and rainfall can slow Research Part B: Methodological, 129, 244–272.
down the sharing dynamics, hampering the flow of clients between Boyac, B., Zografos, K. G., & Geroliminis, N. (2015). An optimization framework for the
stations. In these cases, increasing the number of vehicles and parking development of efficient one-way car-sharing systems. European Journal of
Operational Research, 240(3), 718–733.
slots would help the carsharing service to be more resilient and reliable An integrated optimization-simulation framework for vehicle and personnel relocations
for its clients. of electric carsharing systems with reservations. Transportation Research Part B:
In future works, those unexpected or external events can be modeled Methodological, 95(2017), 214–237.
Bruglieri, M., Pezzella, F., & Pisacane, O. (2018). A two-phase optimization method for a
as new constraints on the MILP formulation. By doing so, the carsharing multiobjective vehicle relocation problem in electric carsharing systems. Journal of
company can organize its fleet to reduce the impact of such possible Combinatorial Optimization, 1–32.
bottlenecks on the service. We intend to compare the performance of the Cocca, M., Giordano, D., Mellia, M., & Vassio, L. (2019). Free floating electric Car
sharing: A data driven approach for system design. IEEE Transactions on Intelligent
MILP model using different mathematical formulations. In addition, we
Transportation Systems, 20(12), 4691–4703.
should evaluate parallelism and distributed processing techniques dur­ Correia, G. H. A., Jorge, D. R., & Antunes, D. M. (2014). The added value of accounting
ing the optimization. These techniques can be applied mainly between for users’ flexibility and information on the potential of a station-based one-way car-
the optimizations for a varying number of vehicles. Although these

9
C.M. Monteiro et al.
sharing system: An application in Lisbon, Portugal. Journal of Intelligent
Transportation Systems, 18(3), 299–308.
Genikomsakis, K. N., Gutierrez, I. A., Thomas, D., & Ioakimidis, C. S. (2017). Simulation
and design of fast charging infrastructure for a university-based e-Carsharing system (pp.
1–10). IEEE Transactions on Intelligent Transportation Systems, no. 99.
Jorge, D., Barnhart, C., & Correia, G. H. A. (2015). Assessing the viability of enabling a
round-trip carsharing system to accept one-way trips: Application to Logan Airport
in Boston. Transportation Research Part C: Emerging Technologies, 56, 359–372.
Jorge, D., Correia, G. H. A., & Barnhart, C. (2014). Comparing optimal relocation
operations with simulated relocation policies in one-way carsharing systems. IEEE
Transactions on Intelligent Transportation Systems, 15(4), 1667–1675.
Kieu, L.-M., Ngoduy, D., Malleson, N., & Chung, E. (2019). A stochastic schedule-
following simulation model of bus routes. Transportmetrica B: Transport Dynamics
(vol. 7,(1), 1588–1610.
Lu, M., Chen, Z., & Shen, S. (2017). Optimizing the profitability and quality of service in
carshare systems under demand uncertainty. Manufacturing & Service Operations
Management, 20(2), 162–180.
Luan, X., Cheng, L., Zhou, Y., & Tang, F. (2018). Strategies of car-sharing promotion in
real market. In 2018 3rd IEEE International Conference on Intelligent Transportation
Engineering (ICITE). 1em plus 0.5em minus 0.4em IEEE (pp. 159–163).
Machado, C. A. S., Hue, N. P. M. S., Berssaneti, F. T., & Quintanilha, J. A. (2018). An
overview of shared mobility. Sustainability (vol. 10,(12), 4342.
Marler, R. T., & Arora, J. S. (2010). The weighted sum method for multi-objective
optimization: New insights. Structural and Multidisciplinary Optimization, 41(6),
853–862.
Martin, E. W., & Shaheen, S. A. (2011). Greenhouse Gas Emission Impacts of Carsharing
in North America. IEEE Transactions on Intelligent Transportation Systems, 12(4),
1074–1086.
Mavrotas, G. (2009). Effective implementation of the ε-constraint method in multi-
objective mathematical programming problems. Applied Mathematics and
Computation, 213(2), 455–465.
McKenzie, G. (2020). Urban mobility in the sharing economy: A spatiotemporal
comparison of shared mobility services. Computers, Environment and Urban Systems,
79, 101418.
Mehdi, N., Sirui, Z., Sina, B., & Roorda Matthew, J. (2015). Vehicle relocation and staff
rebalancing in one-way carsharing systems. Transportation Research Part E: Logistics
and Transportation Review (vol. 81,, 98–113.
10
Computers, Environment and Urban Systems 87 (2021) 101623
Monteiro, C. M., Machado, C. A. S., Lage, M. D. O., Berssaneti, F. T., Davis, C. A., Jr., &
Quintanilha, J. A. (2019). Maximizing carsharing profits: An optimization model to
support the carsharing planning. In 25th International Conference on Production
Research (ICPR 2019), Chicago, USA, Aug. 2019.
Monteiro, C. M., Martins, F. V. C., & Davis, C. A., Jr. (2018). Optimization of bus stops,
new pick-up and drop-off locations for public transportation. Journal of Information
& Data Management (vol. 9)(3).
Lage, M. O., Machado, C. A. S., Monteiro, C. M., Berssaneti, F. T., & Quintanilha, J. A.
(2019). Location suitable for the implementation of carsharing in the City of São
Paulo. In 25th International Conference on Production Research 2019. Chicago, USA:
ICPR 2019. Aug.
Qi, G., Huang, A., Guan, W., & Fan, L. (2018). Analysis and prediction of regional
mobility patterns of bus travellers using smart card data and points of interest data.
In IEEE Transactions on Intelligent Transportation Systems, no. 99 (pp. 1–18).
Repoux, M., Boyac, B., & Geroliminis, N. (2014). Simulation and optimization of one-way
car-sharing systems with variant relocation policies. In hEART 2014-3rd symposium
of the European Association for Research in transportation.
Santos, G. G. D., & Correia, G. H. A. (2019). Finding the relevance of staff-based vehicle
relocations in one-way carsharing systems through the use of a simulation-based
optimization tool. Journal of Intelligent Transportation Systems, 23(6), 583–604.
Shaheen, S., Chan, N., Bansal, A., & Cohen, A. (2015). Shared mobility: A sustainability &
technologies workshop: Definitions, industry developments, and early understanding,
Transportation Sustainability Research Center. University of California.
Smolak, K., Rohm, W., Knop, K., & Siła-Nowicka, K. (2020). Population mobility
modelling for mobility data simulation. Computers, Environment and Urban Systems,
84, 101526.
Stiglic, M., Agatz, N., Savelsbergh, M., & Gradisar, M. (2018). Enhancing urban mobility:
Integrating ride-sharing and public transit. Computers & Operations Research, 90,
12–21.
Tchervenkov, C., Balac, M., Hörl, S., Becker, H., & Axhausen, K. W. (2018). How much
parking space can carsharing save? Arbeitsberichte Verkehrs-und Raumplanung (vol.
1374).
Tomasiello, D. B., Giannotti, M., & Feitosa, F. F. (2020). ACCESS: An agent-based model
to explore job accessibility inequalities, Computers. Environment and Urban Systems,
81, 101462.
Zhu, R., Zhang, X., Kondor, D., Santi, P., & Ratti, C. (2020). Understanding spatio-
temporal heterogeneity of bike-sharing and scooter-sharing mobility. Computers,
Environment and Urban Systems, 81, 101483.