Research Article

Transportation Research Record

2022, Vol. 2676(4) 697–709
Ó National Academy of Sciences:
Analytical Model to Estimate Ride Transportation Research Board 2022

Pooling Traffic Impacts by Using the Article reuse guidelines:

sagepub.com/journals-permissions
Macroscopic Fundamental Diagram DOI: 10.1177/03611981211064892
journals.sagepub.com/home/trr

Aledia Bilali1 , Ulrich Fastenrath1, and Klaus Bogenberger2

Abstract
Ride pooling services are considered as a customer-centric mode of transportation, but, at the same time, an environmentally
friendly one, because of the expected positive impacts on traffic congestion. This paper presents an analytical model that can
estimate the traffic impacts of ride pooling on a city by using a previously developed shareability model, which captures the
percentage of shared trips in an area, and the existence of a macroscopic fundamental diagram for the network of consider-
ation. Moreover, the analytical model presented also investigates the impact that improving the average velocity of a city has
on further increasing the percentage of shared trips in an operation area. The model is validated by means of microscopic
traffic simulations for a ride pooling service operating in the city of Munich, Germany, where private vehicle trips are substi-
tuted with pooled vehicle trips for different penetration rates of the service. The results show that the average velocity in
the city can be increased by up to 20% for the scenario when all private vehicle trips are substituted with pooled vehicle trips;
however, the improvement is lower for smaller penetration rates of ride pooling. The operators and cities can use this study
to quickly estimate the traffic impacts of introducing a ride pooling service in a certain area and for a certain set of service
quality parameters.

Keywords
analytical modeling, on-demand mobility, ride pooling, shareability, traffic impact, macroscopic fundamental diagram, planning
and analysis, transportation network modeling

Our cities are experiencing growth in population every However, an effective ride pooling service depends
year, which contributes to increased traffic demand. The largely on the customers’ readiness to use it, which is
use of private vehicles—even though convenient—is not affected by individual choices and on the attributes of
sustainable, considering the large amount of parking the service. The service attributes which affect customers
space and street capacity that is required as a result of an the most are travel time, waiting time and service cost,
average occupancy of only 1.3 passengers per vehicle (1). and the lower they are, the higher is the attractiveness of
From the other side, traditional public transportation is the service (5).
typically an efficient and environmentally friendly mode From the operators’ perspective, the percentage of
of transportation; however, it may not be very attractive shared trips in an area, called shareability, influences the
for customers because of the lack of convenience and profitability of the service, and therefore plays an impor-
flexibility as a result of fixed line and schedule and a lim- tant role in deciding whether to offer a pooling service in
ited area of coverage.
The extended availability of smartphones and data
1
accessibility have made possible the emergence of ride BMW AG, Munich, Germany
2
pooling services. These services offer a user-centric and Traffic Engineering and Control, Technical University of Munich, Munich,
Germany
sustainable mobility option for the customer, which can
reduce the vehicle kilometers traveled in the system Corresponding Author:
because of sharing of trips with similar trajectories (2–4). Aledia Bilali, [email protected]
698 Transportation Research Record 2676(4)

an area or not. Santi et al. examined shareability in a city generalization for different city types is difficult. Albeit
via simulations by using the concept of shareability net- the traffic impacts of ride pooling are not directly investi-
works (6). To generalize the calculation of shareability gated in these studies (as, for instance, would be the case
for different cities, Tachet et al. established a mathemati- if the agent-based simulation were to be coupled with a
cal model, in which shareability depends on city para- microscopic traffic simulation), the computational time
meters (average speed and surface of the operating area) needed for these simulations is still very high and rises
and service attributes (detour time) and tested it for dif- with increasing problem size. Therefore, it is difficult to
ferent cities (7). Their model was extended by Bilali et al. simulate high-demand pooling states, and it is even more
to capture the additional influence of maximum waiting difficult and time demanding to investigate the direct
time, boarding time, and reservation time, and the impact traffic impact by integrating the agent-based simulation
that the modeling details have on shareability, showing and microscopic traffic simulation for the pooling case.
that, in particular, the choice of the optimization objec- To overcome the drawbacks of using agent-based
tive has a high effect on shareability (8–10). simulations and to be able to estimate quickly the impact
The before-mentioned studies derived shareability of ride pooling with only a little input data, this paper
without considering the traffic impact of ride pooling. presents a method to derive analytically the traffic
Therefore, average velocity—a commonly used measure impacts of ride pooling services. The main requirement is
of traffic efficiency—is assumed to be constant in space the existence of a macroscopic fundamental diagram
and time in these models. However, by introducing a ride (MFD) for a specific city. Additionally, the influence that
pooling service in a city, the number of vehicles on the the improvement of average velocity in the city has on
roads will decrease, thereby increasing the average velo- shareability is also modeled. The models presented in this
city in the network. As the velocity is in turn an input for paper are tested for the city of Munich using AIMSUN
the shareability model, the percentage of shared trips in as a microscopic simulation environment.
an area will increase even further. This effect is noted as
second-order effect of velocity on shareability and its
basic idea is illustrated in Bilali et al. for a synthetic grid Analytical Model
network (11). In this paper, the concept of this second- This section describes a model allowing the analysis of
order effect will be explored for a real city network and a traffic impacts of ride pooling. Firstly, an introduction to
more realistic analytical model capturing traffic effects of the shareability model is given, followed by a model for
ride pooling. the reduction of vehicle trips in the road network result-
The benefits of ride pooling, focusing on a particular ing from shared trips. Subsequently, the relation between
city, have been investigated a lot by researchers. Alonso average velocity and vehicle trip generation and the mod-
Mora et al. showed that 98% of taxi trips in New York ified shareability model are described. A detailed descrip-
city currently catered for by 13,000 taxis, can be substi- tion of the model parameters can be found in Table 1.
tuted by a fleet of only 3,000 pooled vehicles, reducing
the mean travel distance in the system (2). A study for
the city of Prague, Czech Republic, substituting private Shareability Model
vehicle trips with pooled trips, demonstrated that, when The benefits of ride pooling are reliant on the possibility
using ride pooling, vehicle kilometers will decrease to of sharing trips which have similar trajectories. The per-
60% of the current state (3). A similar study was per- centage of shared trips in an area is called shareability S
formed for the city of Munich, Germany, and the authors and differs based on city parameters, service quality para-
argue that the benefits of pooling are seen only after a meters, and the used optimization objective. The general
certain penetration rate of the service, for which the formula defining shareability S is given by the following
saved travel kilometers resulting from shared trips are Equation (7–10):
higher than the empty vehicle trips generated to pick-up
customers (4). 1  
Lon on 2Lon
All of the above studies investigate only the impact S=1  3
(1  e sq )(1  1 + 2L sq e
sq ) ð1Þ
on
2(Lsq )
from the pooled vehicle fleet and indirectly check the
traffic impacts by calculating the vehicle kilometers in the where the dimensionless quantity Lon sq (‘‘on’’ refers to
system, without examining the interaction with the other online or on-demand generation of requests, and ‘‘sq’’
vehicles that are present in the network. These studies are refers to the consideration of service quality parameters)
performed using agent-based simulations, which, even depends on the pooled passenger trip generation rate per
though providing a good estimation of the vehicle kilo- hour lp , city attributes (velocity v and surface area O),
meters in the system, are specific for a particular city and and service quality attributes (temporal deviation from
require a large amount of input data. Therefore, a the direct route or detour time D, maximum waiting time
Bilali et al 699

the passenger waits to be picked-up tmax , and boarding/

disembarking time tb ) (8).
The calculation of shareability is based on the notion
of the shareability shadow, which defines the geometric
shape of where in space the origins and destinations of a
trip should be to be shareable with an already existing trip,
without violating the time constraints (defined by service
quality parameters) (7). Depending on the relation that
service quality parameters have with each other, there are
two different shapes of the shareability shadow specified
by Bilali et al. and, therefore, two forms of Lon
sq (8).
For tmax .D  tb :
 2 
n lp 3
Lon
sq = (D  tb
O
0 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1
max
2  max 2 b
Figure 1. Additional effect of average velocity on shareability.
@ 2 1 t 1 t D  t
+  1+ sin1 max A
3p p D  tb p D  tb t
parameters is reflected in the Lon
sq quantity. For the case
ð2Þ when the optimization objective is to minimize the vehicle
kilometers traveled, this prediction model is tested in
and for tmax \D  tb
Bilali et al. with data from an agent-based simulation,
   max 3 ! and the results showed that the model provides a good
n2
l p  3 2 1 t
Lon
sq = (D  tb + : ð3Þ estimation of the shareability values for this optimization
O 3p 2 D  tb objective (10). Therefore, similar to Bilali et al., the share-
ability in this study for the case when the optimization
The above equations determine the value of shareabil- objective is to minimize the vehicle kilometers traveled is
ity when the optimization objective in the customer calculated by using Equation 4 (10).
matching problem is to maximize the percentage of trips
which will be shared (8). However, using the objective of kLon n
sq
maximizing the percentage of shared trips does not Sn, k (Lon b max
sq (v, D, t , t )) =   ð4Þ
1 + kLon n
necessarily mean that the distance traveled is minimized. sq
To maximize the percentage of shared trips, the pooling
algorithm might decide about sharing of the trips (when This shareability model returns the percentage of
the time constraints allow it) just for the sake of achiev- shared passenger trips in an area, while assuming that the
ing maximum percentage of shared trips, even though it average velocity in the city v = v0 remains constant.
might be more effective in relation to saved vehicle kilo- However, for a certain pooled passenger trip generation
meters to serve the customers one after the other (10). rate per hour lp , the number of total vehicle trips per
As the optimization objective to minimize vehicle kilo- hour in the system g is reduced based on the percentage
meters traveled is more favorable to improve the traffic of trips that are shareable S(v0 ). And, as for a lower num-
conditions, it is the one selected in this study. The share- ber of vehicles in the system, the average velocity in the
ability for this optimization objective is given by a predic- system gets higher and a change in average velocity v1 is
tion model defined by Bilali et al. and inspired by Santi encountered, which should be reflected in the shareability
et al. and shown in Equation 4 (6, 10). This form of equa- model and additionally increase the shareability value
tion also describes statistically the natural combinatorial S(v1 ). These interactions (recently considered in the liter-
effects of particle bonding processes in biochemistry. To ature by Bilali et al., and Lehe and Pandey) are described
use this prediction model, the results of the shareability by a loop diagram in Figure 1 (11, 12). Therefore, the
values derived by means of simulations for a base sce- analytical model developed in this study considers the
nario referring to a certain set of service quality para- synergetic effects that increased pooling rate and average
meters are needed. Therefore, by fitting the simulated network velocity have on each other. In the following
shareability data to the form of shareability equation sections, the reduced number of vehicle trips in the sys-
given by Equation 4, the parameters n and k are defined. tem as a result of ride pooling will be defined, followed
To calculate the shareability for another set of service by the description of an analytical model that relates
quality parameters, the parameters n and k are kept the average velocity with the vehicle trips generated per hour
same and the impact of the new service quality in the city.
 Table 1. Model Parameters

700
Constant parameters Math notation Description Value Unit
2
Surface area O Area of the operation 221 km
Network length L Total length of the network 2,450 km
Velocity at capacity vc Average velocity when the network is at capacity 39.2 km/h
Flow at capacity qc Average flow when the network is at capacity 457 vph
Average velocity for morning peak vo Initial average velocity for the morning peak for the base 39.2 km/h
scenario
Parabola’s parameter a Parameter of the parabolic function found by data fitting 0.062 na
Average vehicle trip length with origin and lod Average vehicle trip length of trips that have the origin and 5.16 km
destination in the area destination in the area of operation
Average vehicle trip length with only origin in the lo Average vehicle trip length of trips that have only the 17.5 km
area origin in the area of operation
Average vehicle trip length with only destination in ld Average vehicle trip length of trips that have only the 17.5 km
the area destination in the area of operation
% of lo within the area of operation po % of lo within the area of operation 51 %
% of lp within the area of operation pd % of lp within the area of operation 46 %
Detour time D Time delay caused by detour 5 minutes
Maximum waiting time tmax Maximum time waiting to be picked up 5 minutes
Boarding/disembarking time tb Time needed to board/disembark the vehicle 0 minutes
Vehicle occupancy F Maximum F passengers can share the trip at the same time 2 passengers

Variable parameters Math notation Description Relation Unit

Average velocity v Average velocity in the network at a certain time of day na km/h
Average flow q Average flow in the network at a certain time of day na vph
Average density k Average density in the network at a certain time of day na vehicles per km
Shareability S Probability of finding the matching trip in a network to Equations 1–4 %
share ride
Ride pooling penetration rate p Fraction of pooled passenger trips/total passenger trips 0, 5, 25, 50, 75, 100 %
(p = lp =l)
The original private vehicle matrix A0 The original private vehicle matrix na na
Lone-passenger trip generation la Lone-passenger trips generated in an hour and performed ð1  pÞA0 passenger trip per hour
by private vehicles
Pooled passenger trip generation lp Pooled passenger trips requests generated in an hour and pA0 passenger trip per hour
performed by pooling service
Lone-vehicle trip generation ga Lone-vehicle trips generated in an hour by private vehicles ga = l a vehicle trips per hour
Pooled vehicle trip generation gp Pooled vehicle trips generated in an hour by pooling gp à Equation 5 vehicle trips per hour
service
Total vehicle trip generation in the operation area god org Total vehicle trips generated in an hour in the service area g = ga + gp vehicle trips per hour
of pooling service.
Vehicle trips generation with only origin in the area go Vehicle trips generation that have only the origin in the na vehicle trips per hour
area of operation
Vehicle trips generation with only destination in the gd Vehicle trips generation that have only the destination in na vehicle trips per hour
area the area of operation

Note: vph = vehicles per hour; na = not applicable.

Bilali et al 701

Vehicle Trip Reduction Model city. This section will present an analytical model to cap-
A distinction is made between passenger trips generation ture the relation of average velocity and vehicle trip gen-
rate l, for trips requested by the customers, and vehicle eration by exploiting the benefits of an MFD. Therefore,
trips generation rate g, for vehicle trips occurring in the it will be possible to analytically derive the improvement
street, which can serve more than one customer simulta- in average velocity coming as a result of the reduction of
neously. For simplicity, it is assumed that private vehicle vehicle trips in urban areas because of ride pooling.
trips will be replaced by pooled vehicle trips. However,
as the implementation of ride pooling in a city will be Macroscopic Fundamental Diagram (MFD). MFD (or net-
gradual, there will be passengers who will still use their work fundamental diagram) defines the functional form
private vehicles to travel alone and there will also be pas- of the relation between average velocity, traffic flow q
sengers who will use the ride pooling service. Therefore, (vehicles per hour [vph]) and traffic density k (vehicles
the total passenger trips generated per hour in the area l per kilometer). The relation between these parameters
is divided into lone-passenger trips la and pooled pas- was firstly developed for motorway segments by using
senger trips lp . The number of lone-vehicle trips gener- roadside sensors (13). Researchers discovered that a simi-
ated per hour in the system ga will be the same as the lar relation exists also for urban networks. Godfrey pro-
hourly number of lone-passenger trips la (ga = la ), as posed an MFD for a town center road network and
there is only one person per vehicle. As the pooled pas- empirically calibrated it (14). A summary of the MFDs
senger trips lp can be shared with each other, the num- for different networks can be found in Cassidy et al.
ber of pooled vehicle trips generated per hour in the (15). A functional form of the MFD for urban areas,
system, gp , depends on the percentage of shared trips in relating travel production (vehicle-meters) with accumu-
the operating area (shareability) S and the occupancy of lation (vehicles), was defined by Daganzo, and validated
the vehicles F in the form given by Equation 5. This indi- by Geroliminis and Daganzo (16, 17). Since then, the
cates that, for a shareability value higher than zero, the MFD was developed for different urban areas by using
hourly pooled vehicle trips gp will be lower than the detector or floating car data or analytically (17–20).
hourly pooled passenger trips lp , as there can be more MFD has also been exploited in ride hailing studies,
than one passenger served simultaneously by vehicles. which do not consider ride pooling, as a background in
This equation gives a lower bound for the reduction of dynamic modeling of urban traffic modeling to develop
pooled vehicle kilometers traveled, as it does not consider a revenue maximization platform (21). It was also used
the empty pick-up trips or the reallocation trips and
as a means for dynamic modeling and control of a net-
assumes that all the passenger trips which are shared have
work taxi dispatch system (22).
the same origin and destination. Therefore, by using this
The functional form of MFD is also going to be
assumption, this model underestimates the total number
exploited in this study and used as a basis for defining
of vehicle trips generated per hour in the system and,
the relationship between average velocity and vehicle trip
thereby, the positive traffic impacts in reality might be
generation in a network. The MFD for this study is
lower than the ones predicted by the analytical model.
derived by means of simulations. For each time interval,
Slp I, the average velocity vie and flow qie for each edge in the
gp = lp ð1  S Þ + ð5Þ network is obtained, and the weighted average velocity
F
and flow in the network are defined using the below
Therefore, the total number of vehicle trips per hour equations, as proposed by Geroliminis and Daganzo,
in the system g includes the lone-vehicle trips ga and the where le is the length of each edge e (17):
pooled vehicle trips gp and is given by Equation 6, where P
shareability S is calculated using Equations 1–4. vi le
  v = Pe2E e
i
ð7Þ
Slp e2E le
g = ga + gp = la + lp ð1  S Þ + ð6Þ P
F qi le
qi = Pe2E e ð8Þ
This reduction in total vehicle trips in the system is e2E le
expected to improve traffic conditions in the city by
improving the average velocity. As the relation of flow and velocity resembles a para-
bola, to connect these two parameters analytically, a
parabolic function is defined in the form given by
Analytical Relation of Average Velocity
Equation 9, where the vertex of the parabola is V(qc , vc ),
and Vehicle Trip Generation vc and qc are velocity and flow when the network is at
As previously mentioned, reducing the number of vehi- capacity, and a is a parameter which will be defined by
cles in the system will affect the average velocity in the fitting the data points to this function. A parabolic form
702 Transportation Research Record 2676(4)

of the MFD is also noticed in Geroliminis and Daganzo, go povlo + gd pdvld

and Ramezani and Nourinejad (17, 22). kb = ð11Þ
L

(v  vc )2 = 4a(q  qc ) ð9Þ where

go (gd ) is the average number of vehicles generated
which have only the origin (destination) inside the area,
Analytical Vehicle Trips in a Network Based on MFD. Firstly, a and
simple method to define the traffic density of a network lo (ld ) is the average vehicle trip length for this vehicle
type.
analytically will be described. It is assumed that a ride
As the vehicle trips of this type are only partly inside
pooling service will be operated in an area defined by a
in the area, po and pd give the percentage of the vehicle
boundary, as illustrated in Figure 2. Within this area,
trip length types (2) and (3), respectively, that is, within
there are three different types of trip to be considered:
the area of operation.
(1) the ones that have the origin and destination within
The overall network’s traffic density is the sum of the
the area, (2) the ones that have the origin in the area and
traffic density of vehicle trips type (1), (2), and (3), as illu-
the destination outside, (3) the ones that have the desti-
strated in Equation 12.
nation in the area and the origin outside, and (4) the
trips that have both origin and destination outside the
god lodv + go povlo + gd pdvld
operation area but pass through it. The first type of trip k = kod + kb = ð12Þ
includes the lone-vehicle trips and the potential pooled L
vehicle trips. The second and third types of trip are the From the MFD relation, q = vk is substituted, and
ones that comprise what will be called here ‘‘background traffic flow q and, consequently, also the relation of flow
traffic’’ in the network as they are going to be there and vehicle trips generated per hour inside the network
regardless of the impact of the ride pooling service in (god ), are derived analytically by the below formulation:
vehicle trip reduction, as the impact that might come
from parking in the boundary of the area of service and god lod + go po lo + gd pd ld
q= ð13Þ
continuing the trip with ride pooling service is neglected L
in this study. To define the traffic density within the
From Equation 13 it is possible to derive the gener-
boundaries of this area, only the contribution of vehicle
ated number of vehicle trips of type (1), which have both
trip types (1), (2), and (3) are considered, and the contri-
the origin and destination in the operated area god (as
bution from type (4) is left out, as most of the cities have
these are the ones of interest for the ride pooling service)
a highway belt to reduce transit traffic within the city.
by Equation 14:
Traffic density is defined as the number of vehicles per
lane kilometer in the system. According to Little’s law qL  (go po lo + gd pd ld )
the average number of vehicles in the system is equal to god = ð14Þ
lod
the average time the vehicles spend in the system multi-
plied by the average number of vehicles generated (23). If By substituting q in Equation 9 of the MFD, the rela-
only the vehicles of type (1), which have both the origin tion of average velocity and vehicle trips generated per
and the destination within the area of operation, are con- hour inside the network (god ) given by Equation 15 is
sidered, the traffic density kod is given by Equation 10: obtained, from which two values are obtained, depending
on which state the network is: free flow state (Equation
god lodv 16) or congested state (Equation 17).
kod = ð10Þ
L  
god lod + go po lo + gd pd ld
where ðv  vc Þ2 = 4a  qc
L
god is the average number of vehicle trips per hour 0   1
Sl
generated inside the area (denoted by g in the Vehicle (la + lp ð1  S Þ + Fp )lod + go po lo + gd pd ld
Trip Reduction Model section), = 4a@  qc A
L
lod is the average vehicle trip length,
v is the average velocity in the network, and ð15Þ
L is the total network length.
For vehicle trips of type (2) and (3), which have only If v.vc :
their origin or destination within the area, the back- sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 ffi
ground density kb is given by Equation 11: god lod + go po lo + gd pd ld
v = vc + 4a  qc ð16Þ
L
Bilali et al 703

 2   max 3 !
(n(S)) l p  3 2 1 t
Lon
sqmod = (D  tb + :
O 3p 2 D  tb
ð19Þ

Substituting Lonsqmod in Equation 4, which gives the

shareability for the optimization objective of minimizing
the vehicle kilometers traveled in the system, a non-linear
equation of shareability is obtained. The Newton-
Raphsod iterative method is used to solve the fixed point
problem presented in Equation (25) to find the new val-
ues of the average velocity and, thereby, also the new
values of modified shareability.
Figure 2. Illustration of an operating area with position of n
origin–destination (OD) trips.
kLon
sq
F ðS Þ = S   mod
 =0 ð20Þ
Note: D = destination; O = origin. n
1 + kLon
sq mod

And if v\vc :
The model described in this section can capture analy-
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 ffi tically the traffic impact that ride pooling has on average
god lod + go po lo + gd pd ld
v = vc  4a  qc ð17Þ velocity and the improvement that it may additionally
L cause to the urban environment because of the addi-
tional increase of shareable trips. This implies that traffic
In this way, it is possible to derive the new average velo- improvement resulting from ride pooling will also be
city in the network, only by having knowledge of the MFD beneficial for operators to increase the chances of finding
and the reduced number of vehicle trips per hour resulting shareable trips as a result of further distances reached
from pooling, calculated by Equation 6 when shareability within the allowed detour time because of higher
value is known. The parabolic shape of this relation is also velocity.
supported by a recent study from Ke et al. (24).

Modified Shareability Model Simulation Setup

By capturing the impact that ride pooling has on improv- Operating Area
ing average velocity in the network, this study also cap- To validate the developed model and investigate the traf-
tures the additional impact that the change in average fic impact of introducing a ride pooling service, a ride
velocity has on the shareability value. Therefore, this pooling service in the city of Munich is considered, where
part describes the modified shareability model where, dif- private vehicle trips are substituted with pooled vehicle
ferent from previous studies, velocity is considered as a trips for different penetration rates of the service. The
dynamic parameter. Munich network is built in the microscopic traffic simu-
To define the modified shareability value, the dynamic lation environment AIMSUN (26). The operating area
velocity formulation, which depends on shareability, considered (Figure 3) is located around Munich city cen-
given in Equation 15, is substituted into the shareability ter, similar to the one in Bilali et al. (10). Its surface is
Equations 2 and 3. Therefore, the modified Lon sq mod is 221 km2 and the network length L is 2450 km. The traffic
given by the following equations: demand in this area represents private vehicle trips for
For tmax .D  tb Munich (27).
  To construct the MFD for this network and extract
(n(S))2 lp 3 the relevant information, microscopic simulations are
Lon
sqmod = (D  tb
O run for the time period from 06:00 to 24:00. To push the
0 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1
max
2  max 2 b network to capacity, two other simulations are also run
@ 2 1 t 1 t D  t
+  1+ sin1 max A for scenarios where the private vehicle trip demand is
3p p D  tb p D  tb t increased by 10% and 20%, respectively. One (network
ð18Þ average) flow-velocity data point is extracted every
10 min based on Equations 7 and 8, and it is plotted in
and for tmax \D  tb the MFD graph.
704 Transportation Research Record 2676(4)

Scenario Setup For the ride pooling service selected in this study,
To test the impact of pooling for a more congested net- where the optimization objective used by the operator for
the matching algorithm is to minimize the vehicle kilo-
work, the traffic demand of the base scenario is selected
meters traveled in the system, shareability in the area is
to be 10% higher than the current demand from private
derived by using Equations 2–4. The area of the city ðOÞ
vehicle trips in Munich and it is assumed that the pooling
is known and the average velocity (vo ) is obtained from
service is offered during the morning peak time from
the simulation of the base scenario. The service quality
07:00 to 10:00. The simulation for the base scenario is
attributes of detour time D and maximum waiting time
run, the results are extracted every 10 min, and the aver-
tmax are selected to be both equal to 5 min, the boarding
age velocity in the network for each time interval is
time tb is set to 0 for simplicity, and it is assumed that a
obtained by using Equation 7. The average velocity for
maximum of two passengers can share the trip at the
the morning peak time vo is found to be 39.2 km/h.
same time (vehicle occupancy F = 2).
To investigate the traffic impacts of a ride pooling ser-
For each scenario, one simulation is run for the morn-
vice offered within the area of Figure 3, private vehicle
ing peak time 07:00 to 10:00 and the network statistics
trips type (1) are substituted by pooled vehicle trips for
are extracted every 10 min. Similar to the base scenario,
different scenarios, where the penetration rate of ride
the average velocity in the network is calculated for each
pooling service p = lp =l is 0%, 5%, 25%, 50%, 75%,
time interval by using Equation 7 and then it is averaged
and 100%. The base scenario is the one where p = 0%,
for the morning peak time. All the model parameters and
meaning that all the vehicle trips in the system are not
values are shown in Table 1.
shared and are performed by private vehicles. For
p = 100%, all the private vehicle trips in the system are
substituted by ride pooling vehicle trips. Results
To model these pooling scenarios in AIMSUN, a dis-
tinction is made between the traffic demand generated in Macroscopic Fundamental Diagram (MFD) for Munich
the network from different vehicle types, and three Network
origin–destination (OD) matrices are created: matrix (B) The MFD for the operating area in the city of Munich is
for the background traffic (vehicle trips type [2], [3] and shown in Figure 4. As specified in the Scenario Setup
[4]), matrix (A) for the private (lone) vehicle trips type section, the blue data points to construct this MFD were
(1) (ga ) and matrix (P) for the pooled vehicle trips type extracted from the results of three scenarios with differ-
(1) (gp ). The latter two are the ones that have both their ent demand levels. The virtual queues of the vehicles
origin and destination within the area of operation. waiting to get into the network are kept at minimum to
Matrix (B) representing vehicle trips which have only control the state of the network and make sure that net-
the origin or the destination within the area (type [2] and work gridlock, which might occur when the input flow
[3]) or only pass through the area (type [4]), does not exceeds the supply function, is not happening.
change regardless of the reduction of pooled vehicle The x-axis correspond to the value of average traffic
trips. Matrix (A) for private vehicle trips type (1) and flow and the y-axis corresponds to average velocity. It is
matrix (P) for pooled vehicle trips will be changed shown that the area of consideration from the Munich
depending on the selected ride pooling scenario. network is at the free flow state most of the time, while
It is assumed that private vehicle and ride pooling reaching the unstable state at the network’s capacity dur-
trips follow the same OD distributions and scale as the ing the peak times. This form of the MFD for the city of
demand matrices based on A0 , the original private vehicle Munich is similar to the one observed in Dandl et al.
matrix. For each scenario considered, the value of lone- (28). The high average velocity values come as a result of
vehicle trips ga for matrix (A) of private vehicle trips is the large area considered, which contains city highways
changed by using Equation 21. For matrix (P) of pooled and arterials, where the speed limit is high. A clockwise
vehicle trips gd , firstly, the pooled passenger demand lp hysteresis loop is observed for the investigated scenarios
is obtained by using Equation 22 and then the pooled caused when the demand starts decreasing after the peak
vehicle trips gp is calculated by using Equations 5 and 6, time, showing that the system does not return to the free
knowing the shareability value for the designed ride flow state immediately if the initial congestion level in
pooling service and the specific pooled passenger trip the network is high (16). Therefore, the hysteresis phe-
demand lp . nomenon in this study occurs when, for the same average
flow in the network, the average velocity is higher during
ga = la = ð1  pÞA0 ð21Þ the congestion onset compared with its values during the
congestion offset. Geroliminis and Sun show that one
lp = pA0 ð22Þ
reason for the occurrence of this phenomenon is the
Bilali et al 705

14, namely trip lengths and percentage of the trip length

inside the operating area for different vehicle trip types,
for the base scenario.
To get the length of the vehicle trips type (1), (2), and
(3), the centroid statistics in AIMSUN are used. The cen-
troids within the area of operation are extracted, and the
total number of vehicle trips type (1), (2), and (3), and
the total kilometers traveled by each of these vehicle trip
types, are calculated. By dividing the total kilometers per-
formed by each vehicle trip type with the total number of
vehicles of the respective type, the average trip length for
the vehicle trips type (1), (2), and (3) is calculated to be
5.16 km, 17.5 km, and 17.5 km, correspondingly.
As vehicle trips type (2) and (3) are only partly within
the area of operation, it is necessary to find the percent-
age of their trip length that contributes to the road net-
Figure 3. Munich operating area.
work traffic within the area. Therefore, all possible paths
connecting each origin with each destination are
excerpted, and only the ones which start or end within
the operation area are filtered. For the path that is mostly
dissimilar spatial and temporal distribution of traffic used by the vehicles, the identification numbers and
congestion (29). The correlations of the loop size of the lengths of the sections are obtained, and whether these
MFD, congestion heterogeneity, and network perfor- sections are within or outside the area of operation is
mance are further examined in Hemdan et al. (30). At checked. The length of the trip performed within the area
the point where the traffic flow is at maximum at the will be similar to the sum of the length of the sections
MFD graph, the network is in its optimum state and the which are situated inside the area. Dividing the vehicle
network’s traffic flow at capacity qc is derived. The cor- trip length within the area by the total vehicle trip length
responding velocity is the network’s velocity at capacity returns the percentage of the vehicle trip length that is
vc : The values of traffic flow and velocity at capacity are inside the operation area. In this case, the values are 51%
equal to 39.2 km/h and 457 vph, respectively. In the con- and 46% for vehicle trips type (2) and (3), respectively.
gestion regime, where the velocity and traffic flow are
lower than the ones at capacity, the chaotic nature of
Average Velocity and Vehicle Trip Generation Relation. The rela-
traffic leads to network gridlocks, and even small distur-
tion between average velocity and vehicle trip generation
bances effect the system; that is why the data points in
per hour for the vehicle trips which have both their ori-
this regime are widely scattered. These results are in line
gin and destination within the area (god ) is illustrated in
with previous studies, which show that, in the congestion
Figure 5. The dark blue data points are coming from the
regime, bifurcation occurs and the MFD becomes multi-
simulation input of generated vehicle trips and the simu-
valued (31).
lated average velocity in the network. For the light blue
To get a functional form for this MFD, the equation
data points, the x-value showing the vehicle trips per
of parabola specified in Equation 15 is used. As the
hour is derived analytically using Equation 14, and the
majority of the data points belong to the free flow state
y-value is the simulated average velocity.
and the base scenario has data points only in this state,
For the simulated data points, it can be seen that
the data is fit to the parabolic function defined in
there are a few scattered data points, which belong to the
Equation 16 for v.vc , where the vertex of the parabola
unstable state of the network and correspond to the hys-
is defined by the point where the network is at capacity
teresis shown in MFD in Figure 4. Comparing the simu-
V (qc , vc ). The parameter a of the parabolic function is
lated data points with the analytically derived ones, it is
found to be 0.62. The fitted curve is shown by the black
possible to see quite a good correlation between them,
line in Figure 4.
denoting that the analytical formulation for defining the
generated vehicle trips in the network based on network
and trip information holds for the free flow state of the
Average Velocity and Vehicle Trip Generation Relation network. Further investigations are, however, needed for
Network Information. To express the relation of average the congested state of the network to check the validity
velocity and vehicle trip generation, it is necessary to of the model also for this state. As this is not the case for
extract the network information mentioned in Equation the network in this study, it is not considered.
706 Transportation Research Record 2676(4)

analytically, by having information of only trip informa-

tion and the MFD relating average velocity and average
flow, is quite important to define traffic impacts of ride
pooling by capturing the change in average velocity if
the vehicle trip generation in the network changes. As
the shareability value S for the defined ride pooling sys-
tem for a certain ride pooling passenger demand lp can
be obtained using (Equations 1–4), it is then possible to
derive the reduced number of vehicles in the street god
because of pooling, based on Equation 6. By using this
value as an input to this function, it is possible to calcu-
late the new average velocity in the network. This func-
tion will be used as an input to analytically derive the
Figure 4. Macroscopic fundamental diagram (MFD) for Munich.
additional improvement in the shareability values as a
result of improved average velocity, described in the next
section.

Traffic Impacts of Ride Pooling

To investigate the traffic impacts of ride pooling service,
various scenarios are designed, where the penetration
rate of ride pooling service p ranges from 5% to 100%,
and matrix (A) of private vehicle trips and matrix (P)
of pooled vehicle trips are changed as specified in the
Scenario Setup section.
Figure 6 depicts the average velocity every 10 min for
the morning peak hour for different penetration rates of
ride pooling. It is observed that, at the beginning of the
morning peak, when the network is not in congested
Figure 5. Relation of average velocity (v) and vehicle trip state even for the base scenario, there is no significant
generation (god ).
improvement of average velocity in the network for all
the scenarios tested. When the network starts to get con-
gested and the average velocity starts decreasing, the
benefits of pooling become noticeable. This emphasizes
the advantage of pooling in improving traffic condition,
especially during peak times, and shows that it is possible
to see higher benefits of pooling in cities with high levels
of congestion.
As expected, it is shown that, the higher the penetra-
tion rate of pooling, the higher the increase in average
velocity. This comes as a result of high demand for pool-
ing, which increases the chances of finding shareable trips
(shareability). Therefore, as more trips are shared, fewer
vehicles are present in the road network and higher velo-
cities are observed. Figure 7 illustrates the improvement
Figure 6. Average velocity for different pooling penetration of average velocity compared with the base scenario and
rates. shows that when the penetration rate of pooling is 100%
the velocity can increase by up to 20% compared with the
base scenario. For scenario P5, when the penetration is
The analytical function relating average velocity and 5%, the effect of ride pooling on average velocity is quite
vehicle trip generation (god ) using Equation 16 is illu- small, indicating that when this service is introduced the
strated by the light blue line in Figure 5. It can be noted effect on traffic is not expected to be seen immediately.
that the analytical function fits well with both the simu- However, with increasing market share, the positive
lated and analytical data. Deriving this relation impact of pooling will be more prominent. For instance,
Bilali et al 707

Figure 7. Difference in average velocity compared with the base Figure 8. Analytical (original and modified) and simulated
scenario. shareability depending on the pooled passenger trips (lp ).

for scenario P25, when the penetration rate of ride pool- consideration. These data points show how improvement
ing is 25%, the average velocity rises by up to 10% com- in average velocity resulting from ride pooling (given by
pared with the base scenario. This suggests that ride the simulation results) further increase the simulated
pooling services have to gain a considerable market share shareability values. A very good correlation with the
to profit from their positive impacts on traffic congestion. modified shareability model is shown for all of them,
validating in this way the result of the analytical model.
Even though the impact of an increased velocity on the
Modified Shareability Model modified shareability in the case study is small, it is nev-
The results of the previous section show that, when pri- ertheless an important finding, as it implies that, for cities
vate vehicle trips are substituted with ride pooling, traffic which are more congested than Munich, the likelihood of
congestion is expected to improve and, therefore, the finding shareable trips will further increase, because of
velocity will increase depending on pooled passenger the higher possibility for improvement in average
demand, which effects the shareability value. Until now, velocity.
velocity was considered as a constant parameter for the
shareability model. Using Equations 18–20 it is possible
to integrate a dynamic velocity into the shareability Conclusion
model and the result is illustrated in Figure 8.
Summary
The dotted gray curve shows the shareability curve for
the original shareability model using the constant average In this study, an analytical model to investigate the
velocity for the base scenario vo , and the solid gray curve impact of ride pooling on traffic efficiency was developed
shows the modified shareability curve for a dynamic velo- by using a shareability model and the MFD for a city. A
city. The difference between the two curves for the case model was developed that captures the relation of aver-
study is not very big, as the average velocity was already age velocity and vehicle trip generation in a network, to
rather high in the base scenario (vo = 39:2km=h) and its analytically check the change in average velocity when
improvement was limited; however, an improvement of pooling is introduced. Moreover, a modified shareability
the shareability values when the velocity increases can model was introduced, which derives further benefits on
still be seen, suggesting that this might increase addition- shareability from an improved average velocity resulting
ally the chances of finding shareable trips and, thus, the from pooling. Different scenarios were developed for a
average velocity in the city. The black data points in ride pooling service offered in the city of Munich, where
Figure 8 are the calculated shareability values consider- private vehicle trips are substituted by different levels of
ing the improvement in average velocity for the simulated ride pooling penetration rate ranging from 5%, to 100%
scenarios with p equal to 5%, 25%, 50%, 75%, and (when all the private vehicle trips are substituted by ride
100%. The leftmost point represents the shareability pooling vehicle trips). The results show that this analyti-
value when p equals 5% and the rightmost point repre- cal model provides a very good and fast estimation of the
sents the shareability when p equals 100%. For each of traffic impacts of ride pooling on the urban environment,
these scenarios, the simulated shareability data points requiring only a few input data and the existence of the
are calculated by using the new improved average velo- MFD for a city, which allows for a generalization of a
city for the whole period of the peak hours under model to other cities, even without the need of network
708 Transportation Research Record 2676(4)

simulations and calibrations, in cases when MFD is Funding

derived analytically or via empirical data (17–20). The The author(s) received no financial support for the research,
analytical model is beneficial for operators to quickly authorship, and/or publication of this article.
assess the traffic impact of pooling in a certain area of
service and for a certain sets of service quality attributes.
ORCID iDs
These insights could be used in discussions with cities to
allow or prioritize the operation of such a service. Aledia Bilali https://orcid.org/0000-0001-7116-7666
Klaus Bogenberger https://orcid.org/0000-0003-3868-9571

Future Work
References
Future work will include the validation of the developed 1. Mitchell, W. J., C. Borroni-Bird, and L. D. Burns. Rein-
model for other cities, especially cities with higher levels venting the Automobile. In Personal Urban Mobility for
of congestion in their network, where it could be possible the 21st Century (W. J. Mitchell, C. E. Borroni-Bird, and
to check if the general assumptions made for this model L. D. Burns, eds.), MIT Press, Cambridge, MA, 2010.
hold for the congestion regime as well. This would also 2. Alonso-Mora, J., S. Samaranayake, A. Wallar, E. Fraz-
allow further investigation into whether the assumption zoli, and D. Rus. On-Demand High-Capacity Ride-Shar-
used for a parabolic functional form of the MFD is valid ing via Dynamic Trip-Vehicle Assignment. Proceedings of
for the congestion regime, or if the functional form of the the National Academy of Sciences of the United States of
MFD needs to be adjusted to a skewed/asymmetric para- America, Vol. 114, No. 3, 2017, pp. 462–467.
3. Fiedler, D., M. Certicky, J. Alonso-Mora, and M. Cap.
bola as in Daganzo, and Laval and Castrillón (16, 20). In
The Impact of Ridesharing in Mobility-on-Demand Sys-
addition, the validity of the parameters of the prediction tems: Simulation Case Study in Prague. Proc., 21st Interna-
model of shareability (given by Equation 4 and found by tional Conference on Intelligent Transportation Systems
data fitting) for other cities is a topic requiring further (ITSC), Maui, HI, IEEE, 2018, pp. 1173–1178.
investigation. Furthermore, the simple model used to 4. Engelhardt, R., F. Dandl, A. Bilali, and K. Bogenberger.
derive the reduced number of vehicles resulting from ride Quantifying the Benefits of Autonomous on-Demand Ride-
pooling based on the shareability will be extended to cap- Pooling: A Simulation Study for Munich, Germany. Proc.,
ture analytically the impact of partially shared trips and IEEE Intelligent Transportation Systems Conference (ITSC),
how this depends on different optimization objectives Auckland, New Zealand, IEEE, 2019, pp. 2992–2997.
used for the matching algorithm. Moreover, the impact 5. Krueger, R., T. H. Rashidi, and J. M. Rose. Preferences
for Shared Autonomous Vehicles. Transportation Research
of empty vehicle trips generated for passenger pick-ups
Part C: Emerging Technologies, Vol. 69, 2016, pp. 343–355.
or re-allocation procedures could also be investigated.
6. Santi, P., G. Resta, M. Szell, S. Sobolevsky, S. H. Stro-
The impact of induced demand from other modes, gatz, and C. Ratti. Quantifying the Benefits of Vehicle
because of increased velocity as a result of ride pooling, is Pooling with Shareability Networks. Proceedings of the
also an interesting topic for further consideration. National Academy of Sciences of the United States of Amer-
ica, Vol. 111, No. 37, 2014, pp. 13290–13294.
Acknowledgments 7. Tachet, R., O. Sagarra, P. Santi, G. Resta, M. Szell, S. H.
The authors would like to thank the anonymous reviewers, Strogatz, and C. Ratti. Scaling Law of Urban Ride Shar-
whose comments helped in improving the quality of the paper; ing. Scientific Reports, Vol. 7, 2017, p. 42868.
Florian Dandl and Philipp Franeck for constructive discus- 8. Bilali, A., F. Dandl, U. Fastenrath, and K. Bogenberger.
sions; Gabriel Tilg for sharing his expertise about the MFD; Impact of Service Quality Factors on Ride Sharing in
and Majid Rostami for the support with AIMSUN and the Urban Areas. Proc., 6th International Conference on Mod-
calibration of the Munich network. els and Technologies for Intelligent Transportation Systems
(MT-ITS), Cracow, Poland, IEEE, 2019, pp. 1–8.
Author Contributions 9. Bilali, A., F. Dandl, U. Fastenrath, and K. Bogenberger.
An Analytical Model for on-Demand Ride Sharing to
The authors confirm contribution to the paper as follows: study
Evaluate the Impact of Reservation, Detour and Maxi-
conception and design: A. Bilali, U. Fastenrath, K.
mum Waiting Time. Proc., IEEE Intelligent Transportation
Bogenberger; data collection: A. Bilali; analysis and interpreta-
Systems Conference (ITSC), Auckland, New Zealand,
tion of results: A. Bilali, U. Fastenrath, K. Bogenberger; draft
IEEE, 2019, pp. 1715–1720.
manuscript preparation: A. Bilali. All authors reviewed the
10. Bilali, A., R. Engelhardt, F. Dandl, U. Fastenrath, and K.
results and approved the final version of the manuscript.
Bogenberger. Analytical and Agent-Based Model to Evalu-
ate Ride-Pooling Impact Factors. Transportation Research
Declaration of Conflicting Interests
Record: Journal of the Transportation Research Board,
The author(s) declared no potential conflicts of interest with 2020. 2674: 1–12.
respect to the research, authorship, and/or publication of this 11. Bilali, A., M. A. A. Rathore, U. Fastenrath, and K. Bogen-
article. berger. An Analytical Model to Evaluate Traffic Impacts
Bilali et al 709

of on-Demand Ride Pooling. Proc., IEEE 23rd Interna- 23. Little, J. D. C., and S. C. Graves. Little’s Law. In Building
tional Conference on Intelligent Transportation Systems Intuition (F. S. Hillier, D. Chhajed, and T. J. Lowe, eds.),
(ITSC), Rhodes, Greece, IEEE, 2020, pp. 1–6. Springer, Boston, MA, 2008, pp. 81–100.
12. Lehe, L. J., and A. Pandey. Hyperdemand: A Static Traffic 24. Ke, J., H. Yang, and Z. Zheng. On Ride-Pooling and Traf-
Model with Backward-Bending Demand Curves. Econom- fic Congestion. Transportation Research Part B: Methodo-
ics of Transportation, Vol. 24, No. 3, 2020, p. 100182. logical, Vol. 142, No. 9, 2020, pp. 213–231.
13. Greenshields, B. D. A Study of Traffic Capacity. Highway 25. Venkateshan, S. P., and P. Swaminathan. Solution of Alge-
Research Board, Vol. 14, 1935, pp. 448–477. braic Equations. In Computational Methods in Engineering.
14. Godfrey, J. W. The Mechanism of a Road Network. Traf- Elsevier, Academic Press, Waltham, 2014, pp. 155–201.
fic Engineering & Control, Vol. 11, 1969, pp. 323–327. 26. Barceló, J., and J. Casas. Dynamic Network Simulation
15. Cassidy, M. J., K. Jang, and C. F. Daganzo. Macroscopic with AIMSUN. In Simulation Approaches in Transportation
Fundamental Diagrams for Freeway Networks: Theory Analysis (R. Kitamura, and M. Kuwahara, eds.), Springer-
and Observation. Transportation Research Record: Journal Verlag, New York, NY, 2005, pp. 57–98.
of the Transportation Research Board, 2011. 2260: 8–15. 27. Dandl, F., B. Bracher, and K. Bogenberger. Microsimula-
16. Daganzo, C. F. Urban Gridlock: Macroscopic Modeling tion of an Autonomous Taxi-System in Munich. Proc., 5th
and Mitigation Approaches. Transportation Research Part IEEE International Conference on Models and Technologies
B: Methodological, Vol. 41, No. 1, 2007, pp. 49–62. for Intelligent Transportation Systems (MT-ITS), Naples,
17. Geroliminis, N., and C. F. Daganzo. Existence of Italy, IEEE, 2017, pp. 833–838.
Urban-Scale Macroscopic Fundamental Diagrams: Some 28. Dandl, F., G. Tilg, M. Rostami-Shahrbabaki, and K.
Experimental Findings. Transportation Research Part B: Bogenberger. Network Fundamental Diagram Based Rout-
Methodological, Vol. 42, No. 9, 2008, pp. 759–770. ing of Vehicle Fleets in Dynamic Traffic Simulations. Proc.,
18. Geroliminis, N., and C. F. Daganzo. Macroscopic Model- IEEE 23rd International Conference on Intelligent Transporta-
ing of Traffic in Cities. Presented at 86th Annual Meeting tion Systems (ITSC), Rhodes, Greece, IEEE, 2020, pp. 1–8.
of the Transportation Research Board, Washington, D.C., 29. Geroliminis, N., and J. Sun. Hysteresis Phenomena of a
2007. Macroscopic Fundamental Diagram in Freeway Networks.
19. Daganzo, C. F., and N. Gerolim. An Analytical Approxi- Procedia - Social and Behavioral Sciences, Vol. 17, No. 1,
mation for the Macroscopic Fundamental Diagram of 2011, pp. 213–228.
Urban Traffic. Transportation Research Part B: Methodo- 30. Hemdan, S., A. M. Wahaballa, and F. Kurauchi. Quantifi-
logical, Vol. 42, No. 9, 2008, pp. 771–781. cation of the Hysteresis of Macroscopic Fundamental Dia-
20. Laval, J. A., and F. Castrillón. Stochastic Approximations grams and Its Relationship with the Congestion
for the Macroscopic Fundamental Diagram of Urban Net- Heterogeneity and Performance of a Multimodal Network.
works. Transportation Research Part B: Methodological, Journal of Transportation Technologies, Vol. 8, No. 1,
Vol. 81, 2015, pp. 904–916. 2018, pp. 44–64.
21. Luo, Q., and R. Saigal. Dynamic Pricing for on-Demand 31. Daganzo, C. F., V. V. Gayah, and E. J. Gonzales. Macro-
Ride-Sharing: A Continuous Approach. SSRN Electronic scopic Relations of Urban Traffic Variables: Bifurcations,
Journal, 2017. In SSRN Journal 352 (6289), p.1056. DOI: Multivaluedness and Instability. Transportation Research
10.2139/ssrn.3056498. Part B: Methodological, Vol. 45, No. 1, 2011, pp. 278–288.
22. Ramezani, M., and M. Nourinejad. Dynamic Modeling
and Control of Taxi Services in Large-Scale Urban Net-
The authors remain responsible for all findings and opinions pre-
works: A Macroscopic Approach. Transportation Research
sented in the paper.
Part C: Emerging Technologies, Vol. 94, 2018, pp. 203–219.