A Real-Time Algorithm to Solve the Peer-to-Peer Ride-Matching

Problem in a Flexible Ridesharing System

Dec. 2017

Neda Masoud
Assistant professor
Department of Civil and Environmental Engineering
University of Michigan Ann Arbor, Ann Arbor, MI, 48109
[email protected]
Corresponding author

R. Jayakrishnan
Professor
Department of Civil and Environmental Engineering
University of California Irvine, Irvine, CA, USA, 92697
[email protected]

Masoud, N., and R. Jayakrishnan. A Real-Time Algorithm to Solve the Peer-to-Peer Ride-Matching Prob-
lem in a Flexible Ridesharing System. Transportation Research Part B: Methodological, V. 106, pp. 218-
236, 2017.

1
 Abstract
Real-time peer-to-peer ridesharing is a promising mode of transportation that has gained popularity dur-
ing the recent years thanks to the wide-spread use of smart phones, mobile application development
platforms, and online payment systems. An assignment of drivers to riders, known as the ride-matching
problem, is a central component of a peer-to-peer ridesharing system. In this paper, we discuss the fea-
tures of a flexible ridesharing system, and propose an algorithm to optimally solve the ride-matching
problem in a flexible ridesharing system in real-time. We generate random instances of the problem, and
perform sensitivity analysis over some of the important parameters in a ridesharing system. Furthermore,
we discuss two novel approaches to increase the performance of a ridesharing system.
Keywords: On-demand transportation, ridesharing, ride-matching, multi-modal transportation

1 Introduction
Congestion in urban transportation networks is one of the common problems faced by many countries around
the world. In addition to having a direct impact on travel time and fuel consumption, congestion imposes
indirect costs by increasing travel time uncertainty, as well as emission levels which adversely affect human
health and ecosystems. Managing congestion by expanding the infrastructure is costly and damaging to the
environment. An alternative way is to make more efficient use of the existing transportation infrastructure.
Public transportation is a conventional way of using the existing capacity on the roadway network more
efficiently.
Transit systems in urban networks mainly include buses and rail services. They typically carry multi-
ple passengers, and therefore can help reduce vehicle miles traveled (VMT) and ease congestion in urban
networks. One drawback of transit services is that they operate on fixed routes and schedules, which limits
their coverage of the network, both geographically and temporally. Urban transit systems are typically and
necessarily designed to satisfy peak period demands. Due to significant peaking behavior of demand, the
system capacity is often drastically under-utilized during off-peak periods, which causes significant cost
inefficiencies. Government regulations on fares exacerbate the cost concerns. Thus transit services usually
fail to act as financially independent entities, and are in need of subsidies.
Para-transit services were originally introduced to run as supplementary services alongside transit, and
as a means to increase the flexibility of public transportation. Para-transit services include all shuttle-like ser-
vices that serve customers such as airport travelers, employee or student commuters, or the elderly and dis-
abled. Since the passage of the Americans with Disabilities Act of 1990, however, the term para-transit has
been used more commonly to refer to the services provided to persons with disabilities, or to the elderly. In
this paper, we reserve the term para-transit to refer to such services. These services are demand-responsive,
and usually serve multiple passengers at a time, based on spatiotemporal proximity of the requests they
receive. Routes and schedules are fairly flexible and demand-dependent.
Although para-transit services can be beneficiary to demographics they target, these demographics are
fairly limited. In addition, most services that fall under this category are offered by non-profit organizations
that are supported by federal or state funding. Due to the limit on the amount of subsidies, it is not possible
to extend these services to the general public. These limitations alongside the desire for more comfortable
(and possibly quicker) travel options have resulted in a much higher demand for private sector alternatives,
such as taxis.
Taxis are a private form of demand-responsive transportation alternative. They provide door-to-door
transportation, but at a higher cost that not everyone can afford. Shared taxi/limousine/shuttle services
provide an opportunity for customers to share their trips, cutting the cost of the journey, and potentially
reducing the total miles traveled in the network. Over the years, different shared-use services have been de-
signed. Flexible route transit systems (Quadrifoglio et al. 2008; Li and Quadrifoglio 2010; Qiu et al. 2014),

2
and High-Coverage Point-to-Point Transit (HCPPT) (Cortés and Jayakrishnan, 2002) are a few examples.
HCPPT is perhaps the first conceptual system that envisaged the option of private drivers offering rides to
be shared, and entirely eliminated fixed route transit, though the somewhat misleading word “transit” was
used in the name.
In all alternative modes of transportation above, with the exception of HCPPT, drivers work as system
employees. Some of the discussed transportation alternatives are more flexible than others in terms of
routing and scheduling, and some have the potential to benefit the environment and customers by allowing
them to share (parts of) their trips. In the next set of transportation alternatives we discuss, drivers are not
employed by the agencies.
A different family of transportation alternatives which has attracted considerable attention during the
recent years is founded based on the principle of shared-use mobility (Shaheen and Chan, 2015). These
alternatives try to reduce the cost of flexible transportation by reducing (or completely eliminating) the
capital investments and operating costs.
Informal carpooling is one of the first and common forms of trip sharing in which a number of individuals
share a vehicle for their trips, which typically involve the same origin and/or destination. Examples include
parents who take turns in taking their children to school, colleagues who carpool to work, etc. In this form of
carpooling drivers own the vehicles and have personal interest in the trips, regardless of whether additional
passengers are present in the car. The incentives for informal carpooling may include saving time through
use of carpool lanes and not having to drive one’s own vehicle every single day, if participants take turns
in driving. This form of carpooling is usually pre-arranged, and happens among individuals who share
commonalities beyond the time and location of their trips.
Transportation Network Companies (TNC), such as Uber, Lyft, and Didi, are among the more recent
faces of shared-use mobility alternatives. TNCs use private vehicle owners and their personal vehicles to
provide flexible and on-demand transportation. These individuals collaborate with the company as inde-
pendent contractors, and not employees. This substantially reduces the cost of capital and human resources
for TNCs while generating revenue for both the company and the drivers. Essentially, the basic services
provided by TNCs act as a lower-cost alternative to taxi services, and hence not only do they not address the
problem of increasing travel demand and congestion, but they add to it as a result of the empty trips required
to pick up passengers. In terms of sustainability, TNCs impose the same cost to the environment as taxi
companies do. Recently, two of the more prominent TNCs, Uber and Lyft, have introduced sharing services
Uberpool and Lyft Line, respectively. Such services can certainly reduce the cost and environmental impacts
of the base services.
Peer-to-peer (P2P) ridesharing aims at capturing the benefits of TNCs while alleviating their adverse
impact on the environment. Ridesharing systems are founded on the principle of sharing economy. Sharing
economy, also known as collaborative consumption, is a fairly old concept that focuses on the benefits
obtained from sharing resources (products or services) that would otherwise go unused. This economic
model has gained more popularity in the recent years, giving birth to many P2P services in different fields
(for examples, refer to Böckmann, 2013). The rapid global spread of the internet during the past decade has
extended the domains of sharing economy to global populations, and has highlighted its benefits. Moreover,
new computer platforms allow easy and quick development of companion mobile applications that facilitate
sharing economy.
Similar to TNCs, drivers in a P2P ridesharing system use their personal vehicles to transport passengers,
and do not work as agency employees. Contrary to the TNC operations, drivers in a P2P ridesharing system
are making trips to perform activities of self interest (as in the case of informal carpooling), i.e., they do not
roam around the city only to pick up and drop off passengers. This setting can lead to services that are more
environmentally-friendly and cost-efficient compared to the sharing services offered by TNCs. Scoop is an
example of P2P ridesharing company which is currently operating in various cities throughout the US.
The overwhelming success and high acceptance rates of TNCs by the public suggest a bright future

3
 Table 1: Transportation alternatives differing in flexibility, cost, and environmental impacts

System Para-Transit Shared Informal P2P

Transit Taxi TNCs
components (disabled) Taxi/Van Carpooling Ridesharing
Employees ! ! ! !
Drivers
Peer to riders ! ! !
Pre-arranged ! ! ! ! ! !
Arrangement
On-demand ! ! ! ! !
Multiple passengers ! ! ! ! ! !
Financially self-sufficient ! ! ! ! ! ?

TNC
Taxi
P2P ridesharing

Para-transit
Flexibility

Shared taxi
Carpool

Transit

Cost
Figure 1: Comparison of shared-use mobility services in terms of cost and flexibility

for the more environmentally-friendly ridesharing systems. Since founded in 2009, Uber has managed
to expand its operations in 75 countries and more than 450 cities worldwide. According to BUSINESS
INSIDER, in 2014 Uber had more than 160,000 registered driver-partners in the US, and the number of
new drivers subscribing to Uber each month increased exponentially, approaching 40,000 (Business Insider,
2015). The company reportedly earned $1.5 billion in revenue in 2015 (Business Insider, 2015).
Table 1 and Figure 1 summarize the characteristics of the transportation alternatives discussed in this
section. They suggest that ridesharing systems have the potential to outperform other alternatives in terms of
cost, flexibility, and impact on congestion and the environment, if they can manage to operate as financially
independent systems. This paper introduces the characteristics of a flexible P2P ridesharing system, and
proposes a methodology to match riders and drivers in such a system.

2 Literature Review
The ride-matching problem can be viewed as the engine of a ridesharing system. This problem can be
considered as a generalization of the classic Dial-a-Ride problem (DARP) in the transportation literature.

4
Dial-a-ride problem optimizes the pick-up and delivery of passengers in special settings that typically in-
volve door-to-door transportation and is commonly used in para-transit systems or shuttle-like services. In
such systems, passengers in need of rides contact service providers regarding their transportation needs, and
the service providers assign vehicles to their requests. In the basic form of DARP, vehicles and riders are
assumed to be homogeneous. All vehicles start from and return to the same depot (Savelsbergh and Sol,
1995; Cordeau and Laporte, 2007; Cordeau and Laporte, 2007). Over the years, researchers have studied
more practical forms of DARP, by assuming time windows for requests (Jaw et al., 1986; Psaraftis, 1983),
multiple depots (Cordeau and Laporte, 2007), heterogeneity among vehicles and passengers (Carnes et al.,
2013; Braekers et al., 2014), and the possibility of transfers for passengers, i.e., multi-hop property (Masson
et al., 2014; Stein, 1978; Liaw et al., 1996).
The ride-matching problem is very similar to the more advanced versions of DARP in definition and
therefore in mathematical modeling. The main characteristic of a P2P ride-matching problem that differ-
entiates it from DARP is the fact that drivers in a ridesharing system are also considered as customers.
Additionally, the set of vehicles (and their drivers) is not deterministic, i.e., their availability is not generally
known in advance. Once they register their trips, drivers are available only in narrow sprites in space and
time. These characteristics play an important role when it comes to developing efficient algorithms to find
optimal matchings. Since riders and drivers have very limited spatiotemporal proximity compared to DARP
where drivers are much more flexible, the standard formulation of DARP is not the most efficient way to
model a P2P ridesharing system. In a ridesharing system, it is possible to take advantage of the narrow
time windows within which riders and drivers are available to perform value reduction on the domains of
decision variables, rather than adding time window constraints to the optimization problem, hence reducing
the complexity of the problem (Masoud and Jayakrishnan, 2017).
If we were to base the mathematical modeling of a P2P ridesharing problem on the formulation of
DARP, we would have to add multiple sets of constraints (and decision variables depending on the properties
of the defined ridesharing system) based on the characteristics of drivers’ trips (e.g., time windows and
pre-determined origins and destinations of trips) and their preferences (e.g., non-smoking passengers or
passengers within a certain age group). These additional constraint sets increase the complexity of the
corresponding optimization problems. In this paper rather than introducing new constraint sets, we use
these additional sets of constraints to reduce the solution space of the problem, and develop an efficient
solution algorithm for the multi-hop ride-matching problem.
In practice, ridesharing systems usually match a rider with a driver if they share the same origin and
destination. From a computational point of view, this task can be easily accomplished by making a series of
comparisons between origins and destinations of riders and drivers in polynomial time. More sophisticated
ridesharing systems allow drivers to specify their routes, and assign a driver to a riders whose origin and
destination lie along the driver’s pre-specified route. Computationally, this is a simple form of matching
and is sufficient if the goal of the system is to match each driver with a single rider. Matching in this
one-to-one setting can be easily accomplished by modeling the system as a bipartite graph, with riders and
drivers forming two mutually exclusive sets of vertices. Although this type of matching can be done in
real-time, it does not meet the requirements of a fully flexible system. Schaub et al. (2014) show that even a
slight deviation from this simple model can render the matching problem computationally intractable. For
example, they show that in a system where a vehicle can carry multiple passengers, relaxing the vehicle’s
path (as opposed to assuming a pre-specified and fixed path for the vehicle) can make the problem NP-
complete, under specific objective functions.
The first feature of a fully flexible ridesharing system is its ability to propose multi-hop itineraries to
riders. This feature allows riders to transfer between vehicles, if necessary. In addition, incorporating
this feature in the model enables a ridesharing system to provide multi-modal itineraries by extending the
concept of vehicles in a ridesharing system to include different modes of transportation.
The general belief among transportation planners that transfers in transport systems are intolerable while

5
understandable, is not entirely accurate. The current unfavorable perception of transfers is mainly due to
evaluating transfers in the context of transit systems. Since transit operations are typically sparse in time
and space, making transfers from one transit line to another typically involves long wait times, as well
as (sometimes considerable) deviations from one’s shortest path. The idea of having passengers transfer
between vehicles/modes of transportation, however, can considerably improve connectivity in transportation
networks-the principle that motivated incorporating them in transit systems in the first place.
It is possible to take advantage of the benefits transfers have to offer while minimizing their potential
inconveniences by properly planning them in time and space. Transfers are in fact an inseparable part
of a high-quality multi-modal transport system (Vukan, 1981). Designing multi-hop ridesharing systems
can facilitate integrating these systems with public transportation, providing a seamless transport platform
where ridesharing systems can serve as feeders to public transit. Masoud and Jayakrishnan (2017) show that
considering transfers can cosiderably increase the number of matched riders. Agatz et al. (2009), Herbawi
and Weber (2011), and Ghoseiri (2013) are some other works in the literature that incorporate the concept
of transfers in their models.
Another desirable feature in a ridesharing system is its ability to optimally route drivers to place them
in spatiotemporal proximity of riders. Next to multi-hop routes, this feature plays an important role in
increasing the number of served riders (Masoud and Jayakrishnan, 2017). Multiple works in the literature
incorporate this feature in their models of ridesharing systems (Masoud and Jayakrishnan 2017; Agatz et al.
2009, 2011; Teodorovic and DellOrco 2005; Baldacci et al. 2004; Di Febbraro et al. 2013; Herbawi and
Weber 2012.)
A flexible ridesharing system should be equipped with a ride-matching problem that is capable to pro-
viding optimal (or high-quality, near-optimal) solutions in a short period of time. There are several studies
in the literature that present a mathematical model of the ride-matching problem, but do not discuss a solu-
tion methodology or use the general branch-and-bound method devised for mixed integer-linear programs
in their case studies (Agatz et al. 2009; Di Febbraro et al. 2013; Cangialosi et al. 2016). Others propose
heuristic solutions without proving/showcasing the accuracy of their proposed algorithms. Masoud and
Jayakrishnan (2017) present a solution methodology to solve the matching problem more efficiently (i.e.,
using less computational resources and faster, compared to direct us of commercial optimization software),
and to optimality. The final solution times, however, are not appropriate for real-time implementations.
Similarly, Regue et al. (2016) propose an aggregation/disaggregation algorithm that is capable of providing
high quality solutions for recurrent commuter trips with a set of shared vehicles, under the assumption that
trips are registered in advance. Their solution methodology, however, is not scalable to large systems.
Finally, a flexible ridesharing system should be equipped with a ride-matching algorithm that can find
an optimal match for a set of riders and drivers simultaneously. Ridesharing systems that ignore this aspect
of matching typically serve riders in a first-come, first-served (FCFS) order; that is, once a rider is matched,
the routes of the drivers involved in the rider’s itinerary are fixed. These drivers can still be considered in
the ride-matching problems solved for future riders, but with (partially) fixed routes. Fixing drivers’ routes
comes at the opportunity cost of fewer options for serving the riders that join the system in the future. This
is a missing component from most of the existing studies in the literature.
Table 2 lists the features of a flexible ridesharing system as introduced in this paper, and indicates the
extent to which current studies in the literature cover these features. This table indicates that there are a few
studies that include all aspects of a fully-flexible ridesharing system; however, the solution times reported in
these studies are too high for large-scale real-time implementations. In this paper, we propose a methodology
that can find the optimal solution to a semi-flexible ride-matching problem in which riders are served on a
FCFS basis and drivers are routed optimally by the system, in a matter of seconds. Next, we introduce a
mechanism that can help shift the solution of this semi-flexible system towards that of a fully-flexible one,
where the FCFS order of serving passengers is disturbed in order to improve system throughput.

6
 Table 2: Literature on P2P ridesharing

Flexible Multi- Multiple Formulation/ Optimal

Authors
Paths Hop Riders Solution Algorithm Solution
Wolfler Calvo et al. (2004) ! ! Greedy heuristic
Baldacci et al. (2004) ! ! Optimization !
Teodorovic and DellOrco (2005) ! ! Bee colony optimization (meta-heuristic)
Agatz et al. (2009) ! ! Optimization !
Agatz et al. (2011) ! Optimization !
Herbawi and Weber (2011) ! Genetic/Evolutionary algorithms
Herbawi and Weber (2012) ! ! Exact formulation + Heuristic solution
Di Febbraro et al. (2013) ! ! Optimization !
Ghoseiri (2013) ! ! Exact formulation + Heuristic solution !
Schaub et al. (2014) ! Bipartite matching !
Regue et al. (2016) ! ! ! Aggregation/disaggregation algorithm
Cangialosi et al. (2016) ! ! ! Optimization !
Masoud and Jayakrishnan (2017) ! ! ! Decomposition algorithm !
Current study ! ! ? Dynamic Programming algorithm !
3 Ridesharing System Architecture
Envision an on-demand, real-time ridesharing system where individuals request transportation services when
they need them. Let P denote the set of participants in such a system. We divide the set of participants into a
set of riders, denoted by R, and a set of drivers, denoted by D. A rider r ∈ R joins the system hoping to find
a set of drivers who can collectively serve his/her ride request (multi-hop property). A driver, on the other
hand, is planning to travel to perform an activity of self-interest, and is willing to share the unused capacity
of his/her vehicle with those who are in need of rides. We start the modeling process assuming that sets R
and D are mutually exclusive, although we study the impact of dropping this assumption in later sections.
Let us identify a set of stations in the network, denoted by S, where participants can start and end their
trips, and riders can transfer between drivers/transportation modes. Stations should be located at areas with
high demand levels for ridesharing, such as high density residential/recreational areas, business districts,
and central transit stations. Furthermore, let us discretize the study time horizon (e.g., a morning/evening
peak hour) into a set of short time intervals of length δt, and index the ith time interval by ti . In this
study, the length of each time interval is set to one minute. Discretizing the study time horizon into short
time intervals allows for using time-dependent travel time matrices, and alongside discretizing the space
dimension through introducing stations provides the basis for the proposed ride-matching algorithm in this
paper.
In order to participate in the ridesharing system, riders and drivers need to register their trips. During
the registration process, a participant p ∈ P is required to provide information on their origin and destination
stations, denoted by O p and D p , respectively, their max ride time (i.e., the max time they are willing to
spend on the trip), denoted by TpT B , and a travel time window for their trip, denoted by [TpED , TpLA ], where
TpED is the earliest time interval participant p is willing to start their trip, and TPLA is the latest time interval
by which they need to have arrived at their destination. In addition to max ride time and travel time window,
in an attempt to provide a pleasant ridesharing experience for the system users, riders are asked to inform the
system on the maximum number of transfers they are willing to make (denoted by Vr , ∀r ∈ R), and drivers
are asked to set a limit on the number of passengers they are willing to have on board at any moment in time
(denoted by Cd , ∀d ∈ D). In addition, all participants can indicate their individual preferences, such as age,
gender, etc., on people with whom they travel.
Although drivers use their personal vehicles, they are encouraged to leave the routing of their vehicles

7
to the system. The system attempts to find itineraries for riders using the empty seats in the drivers’ vehicles
(and possibly using the available transit system), taking into consideration the travel time windows and max
ride time constraints of riders and rivers, max number of transfers specified by riders, vehicle capacities, and
the participants’ individual preferences.
An on-demand ridesharing system has to serve riders on a FCFS basis, i.e., the system should search
for an itinerary for a rider as soon as the rider puts in a trip request. Therefore, the appropriate matching
algorithm for such a system is a many-to-one algorithm, where a single rider can be routed through transfer-
ring between multiple vehicles. If a match is found for a rider, the itineraries of the involved drivers will be
(partially) fixed and announced to them. These drivers can still contribute to the itineraries of future riders
by being routed through the portions of their paths that do not concern the current rider. Furthermore, these
drivers can pick up more passengers in the fixed portions of their paths, if their capacity constraints allow.
The objective of the system is to find the best itinerary for a rider given the spatiotemporal and quality of
service constraints on both riders and drivers.

4 The Ellipsoid Spatiotemporal Accessibility Method (ESTAM)

In this section we introduce what we call the “Ellipsoid Spatiotemporal Accessibility Method (ESTAM)”.
ESTAM generates for each rider a “time-expanded feasible network” in which the rider’s optimal itinerary
could be searched. This network is constructed from the links on the original network that are reachable by
drivers. ESTAM first uses a simple geometric tool to identify the set of stations spatially reachable by each
participant (driver or rider). Next, ESTAM finds the time intervals during which each station can be reached,
forming the set of feasible links for each participant. Finally, the rider’s time-expanded feasible network is
generated using the set of feasible links reachable by drivers who are within spatiotemporal proximity of the
rider. The three steps of ESTAM are discussed in more detail in the following.
Throughout this paper, we use the following problem instance to illustrate the concepts and demonstrate
different steps of the proposed methodology. This example contains a network with 16 stations, one rider,
and four drivers. The network and the characteristics of the participants’ trips are displayed in Figure 2.
Each participant starts and ends his/her trip from one of the stations. For simplicity, It is assumed that
participants have no restrictions on the people with whom they travel.

4.1 Approximating the Set of Reachable Stations

Spatial proximity of participants can be determined based on their origin and destination stations and max
ride times. Masoud and Jayakrishnan (2017) show that the region in the network accessible to participant
p ∈ P is restricted to inside and on the circumference of an ellipse. The focal points of this ellipse are the
participant’s origin and destination stations, and its transverse diameter is an upper-bound on the distance
the participant can travel given their max ride time. Such an ellipse is displayed in Figure 3.
For an ellipse with focal points f1 and f2 , the total straight traveling distance between any point M
on the circumference of the ellipse and the foci (i.e., d( f1 , M) + d(M, f2 ), where the function d measures
the Euclidean distance between its two arguments) is equal to the transverse diameter of the ellipse. The
transverse diameter corresponds to the max ride time of the participant, assuming an underestimate of the
traveling speed in the region. Consequently, any station located outside of the ellipse cannot be reached
without violating the max ride time constraint of the participant. Still, note that not all stations located
inside a participant’s ellipse are necessarily reachable by them due to the fact that actual travel times on the
network are typically higher than straight-line distance travel times on the map. Masoud and Jayakrishnan
(2017) show that this procedure does not cut off the optimal solution, but reduces the feasible region based
on the spatiotemporal characteristics of the trip.

8
 15 7 16
13 10 8

14
9 12 12
7
9 6
10 11 7
11 12

12 8
13 8
8 6 12 8
7 7
5

9 6 12
18 3

10 15
1 16 4
2

Participant 𝑂𝑝 𝐷𝑝 𝑇𝑝𝐸𝐷 𝑇𝑝𝐿𝐴 𝑇𝑝𝑇𝐵

𝑟1 5 12 2 39 37
𝑑1 7 12 23 40 17
𝑑2 5 14 3 28 25
𝑑3 6 7 11 25 14
𝑑4 1 2 10 29 19

Figure 2: Problem instance used throughout the paper to illustrate the concepts

This method can substantially reduce the number of stations a participant can visit, simply by evaluating
the coordinates of a set of candidate (and not all) stations in the equation of an ellipse. Storing stations in
an appropriate data structure based on their geographical location makes it easy to find the set of candidate
stations . We call the part of the network confined within and on the circumference of participant p’s ellipse
the “reduced network” the participant, and denote it by G p .
Figure 4 displays the reduced networks of the participants in our example. Rider 1 is planning to travel
from station 5 to station 12. This rider’s reduced network, Gr , is highlighted in Figure 4 as the section of
the network confined within the blue ellipse. The reduced networks for the four drivers in our example
are also demonstrated in this figure, using green ellipses. The shaded area shows the intersection of the
rider’s reduced network with those of the drivers. We refer to the set of stations within the shaded area as
the “feasible network” for rider r, denoted by GFr . These stations are highlighted in Figure 4. The feasible
network is defined for a rider, and contains only the section of the network that the rider can reach due to
the space proximity provided by drivers in set D.
Defining reduced networks for participants serves three purposes. First, by studying the reduced net-
works of all participants, it is possible to make an early diagnosis whether a rider can be served. For a rider
to be served, he/she needs to have both time and space proximity with at least one driver along a path that
connects the rider’s origin and destination stations. Therefore, the minimum requirement for a rider to be
served is to have space-proximity with at least one driver at both his/her origin and destination stations. In
Figure 4, for example, the rider has spatial proximity with driver 2 (traveling from station 5 to station 14) at
his/her origin station and with driver 1 (traveling from station 7 to 12) at his/her destination station. In the
absence of any of these two drivers, the rider could not be served.
In addition, studying the reduced networks makes it possible to shrink the driver set for each rider. For a
given rider, a driver who does not contribute to the rider’s feasible network (i.e., the driver’s ellipse does not
overlap with the rider’s ellipse) can be filtered out from the ride-matching problem. As an example, driver
4 in Figure 4 can be filtered out from the problem, reducing set D = {1, 2, 3, 4} to D = {1, 2, 3}. Driver 3
does not have proximity to the rider’s origin or destination, but can contribute to portions of the trip in case

9
 15 7 16
13 10 8

14
9 𝑀 12 12
7
9 6
10 11
11 7 𝑓2
12 8
13 8
8 6 12 8
𝑓1 7 7

9 6 12
18 3

10 15
1 16 4
2

Figure 3: The reduced network Gr on the original network G. Stations inside the reduced network are
highlighted.

of a transfer. In fact, in this particular example, the only way the rider can travel from station 5 to station
12 is by being picked up by driver 2, transferring to driver 3, and finally transferring to driver 1. The exact
transfer locations and whether this route is feasible at all under the participants’ temporal constraints will be
examined later. For instance, it might be the case that by the time the rider arrives at station 6 or 10, driver 2
needs to have traveled beyond those stations to get to his/her destination on time (thereby, a spatiotemporally
feasible path for the rider would not exist).
The third benefit of forming reduced networks is that the network-dependent computations can be done
much faster on G p than G. Different types of geometric containers have been proposed in the literature for
more efficient computation of shortest paths in large scale networks (Wagner and Willhalm, 2003; Wagner
et al., 2005). One example of the computational savings obtained by reducing G to G p manifests itself in the
second step of ESTAM, when time-proximity of stations is examined. To determine time-proximity, we need
to search for multiple shortest travel time paths between the origin and destination stations of participants.
Shortest travel time matrices can be computed more efficiently on G p . To demonstrate the computational
savings that can be obtained by replacing the original network with a set of reduced networks, we find the
shortest travel time paths between the origin and destination stations of participants for problem instances of
different sizes (i.e., various number of participants, and various network sizes), using Dijkstra’s algorithm,
both on G and G p . Results are displayed in Figure 5. Dijkstra’s algorithm is used here because it is the
shortest path algorithm we have used for the experiments in this paper; however, similar computational
savings are expected for any other shortest path algorithm. Figure 5 showcases the computational savings
that can be obtained by replacing the original network with the reduced networks.

4.2 Constructing Time-Expanded Reduced Networks

In the previous step, we constructed the participants’ reduced networks, and used them to form the feasible
reduced network for the rider. However, not all stations in the rider’s feasible reduced network are necessar-
ily reachable by the rider, since we have yet to study the time proximity of stations. To study time-proximity,
we have to find the set of time intervals during which station i ∈ G p can be reached by participant p.
To study time-proximity of participant p at each station in G p , we identify all paths in G p with travel
times less than or equal to TpT B . To avoid enumerating the paths, we find the first k-shortest paths in G p
such that the travel time on the kth shortest path is less than or equal to TpT B , and the travel time on the

10
 15 7 16
13 10 8

14
9 12 12
7
9 6
10 11 7
11 12

12 8
13 8
8 6 12 8
7 7
5

9 6 12
18 3

10 15
1 16 4
2

Figure 4: Reduced networks of all the participant in the example problem instance. The shaded area is the
rider’s feasible network, GFr . The stations inside GFr are highlighted.

Figure 5: The surface on the top/bottom shows Dijkstra’s running time on G/G p . Results are averaged over
10 randomly generated networks for each problem

11
 Figure 6: Time-expanded reduced network for driver 2 in our example

(k + 1)th shortest path is higher than TpT B . Recall that computing the shortest paths on reduced networks is
far more computationally efficient than the original network. After finding the jth shortest path ( j ≤ k) on
G p , we follow in order the stations on each path and use time-dependent travel time matrices to find the time
intervals during which each station s ∈ G p can be visited, denoted by Tsp . We can use this information to
construct the “time-expanded reduced network”, G p (t), for each participant p ∈ P. Each node in G p (t) takes
the form of a tuple (ti , si ), representing a visit to station si during time interval ti . Consequently, a link in the
form of (ti , si ,t j , s j ) represents a trip that starts at station si during time interval ti and ends at station s j at
time interval t j . Links in the form of (ti , si ,ti + 1, si ) represent “waiting” links, where an individual is waiting
at station si for one time interval, starting at time interval ti . Figure 6 displays the time-expanded reduced
network for driver 2 in our example. The time-expanded reduced networks for the rest of the participants in
our example are presented in Appendix A.
Note that it is possible to have more than one origin and/or destination node on a time-expanded reduced
network. Nodes in the form of (t, O p ), ∀t ∈ TOpp are all origin nodes, and nodes in form of (t, D p ), ∀t ∈ TDpp
are all destination nodes. We keep these nodes in origin and destination sets, denoted as OS(p) and DS(p),
respectively. In Figure 6, OS(d2 ) = {(3, 5), (4, 5), (5, 5)}, and DS(d2 ) = {(26, 14), (27, 14), (28, 14)}.

4.3 Constructing a Rider’s Time-Expanded Feasible Network

In section 4.1, we generated the feasible network for rider r, GFr , by scanning through the stations in the
rider’s reduced network, Gr , and checking whether these stations exist in any of the drivers’ reduced net-
works. We can repeat a similar procedure on the links in the time-expanded reduced network of the rider,
Gr (t). The result is the rider’s “time-expanded feasible network”, GFr (t). More specifically, in this step we
scan through the drivers’ time-expanded feasible networks, looking for links that they have in common with
the rider. For each link in the riders’ time-expanded reduced network, the drivers that also have that link in
their networks can potentially carry the rider on that link. Similarly, if a link in the rider’s time-expanded
reduced network does not appear in any of the drivers’ networks, that link is in fact not feasible for the rider,
since there are no drivers to transport the rider on that link. In other words, links in set GFr (t) are links that

12
can be traversed by the rider, and at least one driver. At this stage, we can filter out the driver set D even
further by eliminating drivers who do not contribute to the rider’s time-expanded feasible network. Note
that in the example used in this paper, no additional drivers are filtered out at this stage, since all remaining
drivers contribute to the rider’s time-expanded feasible network (see Figure 7(b)).

5 Searching for the Optimal Itineraries

The ESTAM algorithm constructs the time-expanded feasible network for the rider in our problem, on which
the rider’s optimal itinerary can be searched. Even though this network is substantially smaller than the
original network, the number of feasible routes for the rider can still be very large. Different combinations
of links on which to travel and drivers with whom to ride can make enumerating the paths computationally
prohibitive. In this section, we introduce a dynamic programming algorithm to search for a rider’s optimal
path on their time-expanded feasible network.

5.1 Preparing the Graph

Figure 7(a) shows the time-expanded feasible network for the rider in our example. Let us define rider r’s
origin/destination node as the node with the smallest/largest time index in the rider’s origin/destination set.
Our goal is to find an optimal itinerary to take the rider from his/her origin node to his/her destination node.
In our example, the origin node is (2, 5), and the destination node is (39, 12). We define the optimal itinerary
as one that minimizes a linear combination of in-vehicle travel time, wait time, and number of transfers.
Before searching for the optimal itinerary, two preliminary operations need to be performed on the graph.
First, nodes that are not descendants of the origin node or predecessors of the destination node should be
removed from the graph, since they cannot be part of a path that connects the origin node to the destination
node. An example of this operation is demonstrated in Figure 7(a). This figure displays the time-expanded
feasible network for the rider in our example (who is traveling from station 5 to station 12). The sub-graph
{(5, 5), (12, 6), (19, 10), (20, 10), (21, 10)} (marked by red in Figure 7(a)) has to be eliminated from the
graph, since the nodes in this subgraph are not predecessors of the destination node (39, 12). This task can
be easily accomplished by scanning the adjacency matrix of the graph.
The second operation is to topologically sort the graph. A topological ordering for such a graph always
exists, since the graph is directed and acyclic. (The direction of the links can be determined based on the
time dimension. The graph is acyclic since nodes have a time component, and the link generation procedure
does not generate links that represent traveling back in time.) Figure 7(b) displays the resultant graph after
performing these two preliminary operations. The numbers next to nodes in this figure show their topological
orders. Links on this graph are annotated by drivers who can carry the rider on them.
Conducting depth first search (DFS) on the “revised time-expanded feasible network” in Figure 7(b)
could determine whether an itinerary for the rider exists at all. Starting from the first node in the topolog-
ically ordered graph, if DFS does not conquer a node that contains the rider’s destination station, then the
rider cannot be served in the system. Otherwise, if DFS finds a path (establishing that a matching is spatially
and temporally feasible), we will use the proposed dynamic programming (DP) algorithm in the following
section to find the best itinerary for the rider.

5.2 The Dynamic Programing (DP) Algorithm

Let us start by defining sets DIN OUT as the set of drivers who enter and exit node j, respectively, and
j and D j
set DN j to include tuples (i, d) such that there is a link for driver d from node i to node j in the revised
time-expanded feasible network. Furthermore, we denote by V ( j, d) the minimum cost of the optimal path

13
 9
10

𝑑1 𝑑1

7
𝑑2 𝑑1 𝑑1
5 8
𝑑3 𝑑3
3 6
𝑑2 𝑑2
1 4
2

(a) Time-expanded feasible network. The dashed lines marked (b) Revised time-expanded feasible network (after performing
in red will not be part of the optimal solution and can be deleted. the preliminary operations). Numbers next to nodes show their
order in the topologically-ordered graph. Links are annotated
with drivers who can carry the rider on them.

Figure 7: Time-expanded feasible network for the rider in our example before and after revision

from the start node (node 1 in the topologically ordered graph) to node j, such that the last link on the path
is traversed by driver d. The goal is to find Vr∗ = min j∈DS(r), d∈DINj {V ( j, d)}.
We initialize the N r × Dr matrix V (·) to infinity, where N r is the number of nodes in rider r’s revised
time-expanded network, and Dr is the number of drivers that contribute to rider r’s revised time-expanded
feasible network. To keep track of the optimal solution, we introduce two additional matrices of the same
size, n pred and d pred , and initialize them to zero. These matrices keep the predecessor node and the driver
that takes the rider to the predecessor node on the optimal path, respectively.
We compute the cost of a path as a linear combination of travel time, wait time, and penalty for transfer.
Let us define the function C(i, j) to return the cost of traveling on link (i, j) = (ti , si ,t j , s j ). This cost can be
calculated using equation (1). The cost of travel as defined in equation (1) is the travel time on the link in case
a trip takes place (i.e., si 6= s j ), or a penalty W if the rider waits at a station (i.e., si = s j ) for one time interval.
Note that if the rider stays in the same station for multiple time intervals, the algorithm will automatically
accumulate W the appropriate number of times. This is due to the fact that the same station at different
time intervals corresponds to separate nodes on the time-expanded feasible network. Furthermore, note that
C(i, j) in equation (1) can be easily extended to C(i, j, d), in case a more customized driver-dependent cost
is desired. Additionally, we denote by CT the penalty (cost) of a transfer, and use the expression CT (1d6=d 0 )
to capture the total penalty of transfers, where the indicator function (1d6=d 0 ) returns the value 1 if d 6= d 0
(i.e., if the rider makes a transfer), and the value 0 otherwise.

(
(t j − ti + 1)dt si 6= s j
C(i, j) = (1)
W si = s j

We set the initial condition as V (1, d) = 0, for all d ∈ DOUT1 . In the revised time-expanded feasible
network in Figure 7(b), the initial condition is V (1, d1 ) = 0. After setting the initial condition, we traverse
the nodes in the (topologically ordered) graph in ascending order. For each node j, and each driver d in the
set DIN 0
j , V ( j, d) is computed as in equation (2). The set ED(i, d ) in equation (2) includes all drivers on the
0
optimal path to node i, excluding the driver on the last link (d ). We will discuss the necessity of introducing
this set further down this section.

14
 Figure 8: Example of an infeasible solution

{V (i, d 0 ) +CT 1{d 0 6=d} } +C(i, j)


V ( j, d) = min min (2)
i:(i,d)∈DN j d 0 ∈DIN 0
i \ED(i,d )

After computing V ( j, d) using the recursive function in equation (2), we use sets n pred ( j, d) and d pred ( j, d)
to keep track of the predecessor node and driver in the optimal solution, respectively:

{V (i, d 0 ) +CT 1{d6=d 0 } } +C(i, j)


n pred ( j, d) = argmini:(i,d)∈DN j min
d 0 ∈DIN 0
i \ED(i,d )

0
d pred ( j, d) = argmind 0 ∈DIN 0 {V (i, d ) +CT 1{d6=d 0 } }
i \ED(i,d )

Once the min-cost itinerary for the rider is found, these matrices can be used to retrieve the optimal itinerary
(i,e., the order of following the nodes, and the drivers to carry the rider between the nodes.)
Recall that set ED(i, d 0 ) includes all drivers on the optimal path to node i, excluding the driver on the
last link (d 0 ). Drivers in this set are excluded from set DINi in equation (2) to ensure the feasibility of the
solution. Infeasibility can arise in a scenario where a driver covers multiple links on the rider’s itinerary, but
these links belong to different feasible but conflicting paths of the driver (i.e., paths that cannot co-exist).
In such a scenario, the resulting itinerary would not be feasible, since it would require a driver to be at two
different locations at the same time. To clarify, an example is shown in Figure 8. In this example, driver d
has two potential paths to make his/her trip, while driver d 0 has a single path. There is only one itinerary for
the rider, marked by dashed red arrows. This itinerary consists of three links. The first link is covered by
driver d via this driver’s first potential path, the second link is covered by d 0 , and the third link is covered by
d via this driver’s second potential path. Clearly the resulting itinerary is not feasible, since d cannot carry
the rider on both links. Excluding drivers in set ED from the list of potential drivers in equation (2) ensures
that such an itinerary is never generated.
It is easy to incorporate individuals’ personal preferences into this framework. When forming the list of
drivers for a rider, drivers who do not meet the list of requirements by the rider can be easily excluded from
this list. Similarly, a driver can ask to be considered as a potential match only for riders with specific char-
acteristics. Using the proposed DP algorithm, not only the optimal itinerary but all the feasible itineraries
for a rider are generated. Therefore, if the solution that is deemed optimal based on the objective function
defined by the system needs to be re-considered for any reason (e.g., in case a rider cancels a ride, or a driver

15
gets stuck in traffic and cannot make it to the pick up point on time), we can easily retrieve the next best
solution.
Once a rider’s itinerary is fixed, we can use a straight forward procedure to find the optimal itineraries of
drivers. If a driver is part of a rider’s itinerary, then the portion of the driver’s path where he/she is scheduled
to carry the rider becomes fixed. The problem of finding the driver’s optimal itinerary becomes reduced
to finding the shortest travel time paths between the fixed potions. We can use Dijkstra’s algorithm on the
driver’s time-expanded reduced network to efficiently search for the shortest paths from the driver’s origin
to the origin of his/her first fixed portion, between the fixed portions, and from the destination of the last
fixed portion to the driver’s ultimate destination. The union of all these shortest paths and the fixed portions
constitutes the drivers optimal itinerary.

6 Numerical Experiments
In this section, we generate multiple random instances of the ridesharing problem in a network with 49
stations. All problem instances contain 1000 participants. We vary the ratio of drivers and riders in each
problem instance, and assess the performance of the proposed methodology as the ratio of riders in the
problem changes.
We use one-min time intervals, and consider a randomly generated maximum ride time within the inter-
val [TpST , κ TpST ] for participant p ∈ P, where TpST is the shortest path travel time between the participant’s
origin and destination stations, and κ is called the “max ride time factor”, and takes a value greater than or
equal to 1. The participants’ earliest arrival times are generated uniformly within a one hour time period.
The latest arrival time for a participant is computed as the sum of their earliest departure time and max ride
time. All drivers are assumed to have four empty seats, and riders are assumed to accept up to 3 transfers.
In the interest of simplicity, we assume that participants have no personal requirements on the people with
whom they travel. In addition to running experiments with the default parameter values reported above, we
perform sensitivity analysis over the number of participants in the system and their max ride time factors, in
order to gain insight on the performance of the proposed methodology under different scenarios.
In the next two sections, we use the proposed algorithm to find the optimal solutions for the randomly
generated problems through simulations. In section 6.1, the origin and destination locations of participants
will be selected based on a uniform distribution. This type of selection serves as the default setting in our
simulations. In section 6.2, we study a more practical scenario, where trip ends are determined randomly,
but based on a clustering of the network.

6.1 Uniform Random Selection of Trip End Locations

For the problem instances generated in this section, the origin and destination stations of participants are
selected from 49 stations based on a uniform random distribution. The results we report in this section are
averaged over 10 simulation runs for each problem instance. As mentioned before, riders are considered in
a FCFS basis in simulations. Once a rider registers in the system, a ride-matching problem is solved for that
rider. For each problem instance in this section, the highest solution time among riders, averaged over 10
simulation runs, is less than 0.5 sec.
Figure 9(a) displays the number of matched riders and drivers. This figure shows that the number of
served riders is maximized in problem instances where the number of riders and drivers is about the same.
This figure suggests that the number of matched participants is maximized in the same range where the
number of served riders is maximized. In fact, when riders constitute about 45% of the total number of
participants, the number of matched riders and drivers are both maximized. Notice that while it is expected
in a one-to-one system (where each driver is matched with exactly one rider) for the number of matched

16
 (a) Distribution of number of matched participants (b) Distribution of number of transfers

Figure 9: Algorithm performance for the randomly generated problem instances, averaged over 10 runs

riders and drivers to be equal, this is not an obvious conclusion in a multi-hop system where each driver can
carry multiple riders and each rider can transfer between multiple drivers. Furthermore, notice that these
results are only valid for the case where trip origins and destinations are randomly selected. We show in
the next section that in more practical settings where demand is not uniformly distributed (and therefore trip
ends have a higher degree of spatial proximity), the number of matched participants can increase drastically.
Figure 9(b) shows the distribution of number of rider transfers for each problem instance. This figure
suggests that the majority of served riders do not have to make any transfers. This figure also suggests that
the number of additional riders that can be served as a result of allowing transfers in the system is not trivial.

6.2 Targeted Random Selection of Trip Ends in Clusters

In the previous section, we made the assumption that trip origins and destinations are completely random.
In reality, commercial and business districts are usually denser within certain geographical regions, and are
distinct from residential areas. In this section, we identify two geographically distinct regions on the network
to represent residential and business districts, and generate trips with origins in one region and destinations
in the other. Similar to the previous section, we generate different problem instances, varying the percentage
of riders in the problems. For each problem instance, we conduct 10 simulation runs, and report the average
results. Results are summarized in Figure 10.
Figure 10(a) shows the clustering of a sample network. The area on the top is considered to be the
business district, and the area on the bottom is assumed to be the residential area. Each problem instance
represents the simulation of a morning peak period, in which participants travel from a randomly selected
station in the residential area to one in the business district.
Figure 10(b) shows that the number of served riders under this more realistic scenario follows the same
general trend as the one observed in the scenario with randomly generated trip ends; however, in this case
the number of served riders has more than doubled. Figure 10(c) suggests that in this more practical setting,
when the system has the optimum ratio of drivers and riders, more than 60% of the participants will be
matched. Finally, Figure 10(d) shows the distribution of number of transfers for riders. Similar to Figure
9(b), most trips can be completed with 0-1 number of transfers. This consistency between results from
simulations with very low (random) and very high (clustered) spatial proximity among trips suggests that
a multi-hop system is not likely to produce solutions with large numbers of transfers. This observation is

17
 (a) A sample clustering of the network (b) Comparison of number of served riders between the com-
pletely random scenario and the clustering scenario

(c) Distribution of number of matched participants (d) Distribution of number of transfers

Figure 10: Algorithm performance for the scenario with distinct business and residential areas, averaged
over 10 runs.

important due to the perceived discomfort associated with high numbers of transfers, and the consequent
concerns over the acceptance of multi-hop systems that follows. Note that even though the number of
transfers is not high, they allow for higher coverage of the network by the same set of vehicles and therefore
lead to higher matching rates, as depicted in Figures 9(b) and 10(d).
The random demand distributions in section 6.1 and the two-clustered demand distribution in section 6.2
are two extreme cases. In reality, the demand distribution, and hence the performance of the system, falls
somewhere between these two extremes, such that there are a multitude of residential areas and business
districts, as well as regions that host both types of activities. A real-world study using the proposed method-
ology was conducted in the Los Angeles County, California. In this case study the proposed methodology
was used to connect travelers to transit using P2P ridesharing, serving as a solution to the first/last mile
problem (Masoud, Nam, et al., 2017). This study showcases the scalability of the proposed methodology in
larger, realistic networks with higher levels of demand.

18
 Figure 11: Sensitivity analysis over the number of participants in the system

6.3 The Critical Mass

For a ridesharing system to be able to work independently, a critical mass of participants is required. In
the previous sections, we demonstrated the performance of a ridesharing system with 1000 participants. We
showed that in a more practical scenario with distinct business and residential areas, more than 60%, and
in the completely random case about 35% of participants can be successfully matched. It is intuitive to
expect the average number of successful matches to increase with the number of participants. However, for
a system to survive, a minimum number of participants should be able to use the system successfully. This
success will encourage users to continue participating in the system, and generates positive word of mouth.
In this section, we conduct sensitivity analysis over the number of participants in the system. This anal-
ysis can shed light on the relationship between system performance and the number and ratio of participants.
Figure 11 shows the percentage of matched participants as the total number of participants changes from 200
to 1000. Similar to the previous sections, for a given number of participants, we generate multiple random
problem instances by changing the ratio of riders. The trip ends are selected at random with uniform prob-
abilities. Figure 11 suggests that increasing the number of participants leads to higher system performance
levels in terms of the matching rate. The peak performance of the system in all cases occurs within the same
range, where the ratio of riders in the problem is around 0.4. This analysis can help system operators trying
to reach a critical mass make strategic decisions on system expansion and targeted marketing by (i) identi-
fying rider to driver ratios that lead to higher matching rates, and (ii), establishing a relationship between
the matching rate and number of participants under any rider to driver ratio.

6.4 Max ride time factor

In this section, we look into the scalability of the proposed methodology by expanding the size of the time-
expanded network on which the search for the optimal solution is conducted. There are multiple parameters
that can affect the size of the rider’s time-expanded feasible network. For instance, increasing the size of
number of drivers, the spatio-temporal proximity among trips, number of stations, and max ride times of
participants can increase the size of the network. In this section, we study the impact of increasing κ, the
max ride time factor, on solution times and matching rates. Higher κ values correspond to more flexibility
in participants’ departure times and detours from their shortest paths. Consequently, participants can stay
longer, and reach farther in the network, leading to larger link sets.
Figure 12(a) shows the change in the percentage of served riders as κ increases. This figure suggests
that more flexibility leads to higher matching rates, as expected. It is interesting to note that the impact

19
 (a) Percentage of served riders (b) Solution times

Figure 12: Algorithm performance under different max ride time factor (κ) values

of this higher degree of flexibility is more substantial when the ratio riders is low, and diminishes as the
ratio of riders increases. When the rider to driver ratio is low, there are many opportunities for riders to
be served. With increased flexibility introduced by increasing the participants’ max ride times, network
coverage increases substantially to the point where at κ = 1.5, almost all riders can be served. As the rider
to driver ratio increases, the impact of higher degree of flexibility diminishes as the number of drivers is
decreasing and the number of riders is increasing. Eventually, at very high rider to driver ratios, the higher
flexibility does not have a substantial impact on the percentage of served riders since the resources (i.e., the
number of drivers) is very limited to begin with.
Figure 12(b) shows the changes in solution time as κ, and hence the size of the rider’s time-expanded
feasible network, increases. For the same value of κ, this figure shows that the solution time decreases as
the ratio of riders increases. The reason for this observation is that the number of participants is fixed, and
hence a higher ratio of riders translates into a lower ratio of drivers, and therefore a smaller time-expanded
feasible network for every rider. On the other hand, for a given ratio of riders, the size of the time-expanded
feasible network for each rider increases with κ, leading to higher solution times. Figure 12(b) shows that
even with higher values of κ, the solution times remain acceptable for a dynamic system.

6.5 P2P Ride Exchange

Ridesharing systems are in general spatiotemporally sparse. The temporal sparsity stems from the rather
tight time windows of participants, and the spatial sparsity is due to the participants’ fixed origin and des-
tination locations. This spatiotemporal sparsity limits the number of satisfied requests. In fact, one of the
contributors to the initial failure of P2P ridesharing systems in the US in the 1990’s was the very small
number of served rider requests.
Today, advancements in the communication technology and prevalence of smartphones has led to a larger
pool of riders and drivers interested in ridesharing. However, as discussed in the previous sections, a critical
mass of participants is still an important requirement for ridesharing systems to enable their independent
and budget-balanced operations. Therefore, a ridesharing system needs to make the best use of its limited
supply resources, i.e., the drivers.
A ridesharing system that looks to match riders one at a time is in fact wasting its very limited and
valuable resources by fixing the itineraries of matched drivers. Figure 13 shows an example of how fixing

20
 1 1

1 1

2 2

2 2

Figure 13: P2P ride exchange

drivers’ itineraries can deteriorate the performance of the system. This example includes two riders and two
drivers. Assume rider 1 registers in the system first, and hence the system starts by finding a match for rider
1. Vehicles 1 and 2 are both candidates to be matched with this rider. Eventually the system matches rider
1 with driver 2, since driver 1 has to take a larger detour to carry the rider. Driver 2’s itinerary, marked in
solid blue, is then fixed. Next, rider 2 registers their trip. This rider could have been matched with driver
2 through the route marked with the blue dashed line, if driver 2’s itinerary had not been fixed. However,
under the current state of the system, rider 2 cannot be matched. The ridesharing system, therefore, can only
match one rider with one driver.
In the example in Figure 13, had we had the information on both trips in advance, we could have
increased the system performance by solving a many-to-many ride-matching problem that included both
riders, as opposed to considering them one at a time. The solution to such a problem would match rider 1
with driver 1, and rider 2 with driver 2, serving both riders, and involving all 4 participants.
There are two issues that may prevent us from solving a many-to-many matching problem. First, if rider
2 has not joined the system yet at the time when we have to notify rider 1 of the results, then formulating
the many-to-many problem would not be a possibility. Second, solving the many-to-many problem could be
too time-consuming and not appropriate for real-time implementations.
To demonstrate the difference between a many-to-many ride-matching problem (which includes multi-
ple riders and multiple drivers) and a many-to-one problem (which includes one rider and multiple drivers)
in terms of solution time and matching rate, we solve the problem instances in section 6.1 using both ap-
proaches. For the many-to-many problem, we use the decomposition algorithm proposed by Masoud and
Jayakrishnan (2017) on a rolling time-horizon basis with 2, 5, and 10 min re-optimization periods. In this
approach, instead of matching one rider at a time, we solve a problem that includes all the riders who register
their trips within the mentioned re-optimization period. Figure14 depicts the results of this comparison.
Figure 14 suggests that the solution time for a many-to-many ride-matching problem is too high for real-
time implementations. On the other hand, the number of matched riders improves when a many-to-many
problem is solved. In other words, we can use the proposed methodology in this paper to solve a many-to-
one ride-matching problem to optimality; however, the modeling of the system as a many-to-one problem
itself is not optimal, and is constrained by the unwritten FCFS rule. This modeling approach is considered
due to the real-time and dynamic nature of requests, and the lower computational complexity of the many-
to-one problem. To address this issue, in this section we introduce what we call “P2P ride exchange” as a
mechanism to shift the solution of a many-to-one ride-matching problem towards that of a many-to-many
problem, while still serving riders in real-time. P2P ride exchange can help us move from the sub-optimal

21
 (a) Number of served riders (b) Solution times

Figure 14: Number of served riders and solution times for randomly generated problem instances under
different re-optimization periods (i.e., 2 min, 5 min, and 10 min) as well as in a FCFS system where a
matching problem is solved as soon as a rider requests a trip.

solution obtained by matching one rider as a time, towards the optimal solution that can be obtained by
including all riders in the matching problem, without the unattractive side-effect of the increased solution
time.
In this approach, we still solve the matching problem for one rider at a time, and based on the FCFS rule.
However, we do not fix the itineraries of matched drivers. If a rider can be served only using a driver that
has been previously matched and is unable to take a detour as a result of his/her commitment to prior riders,
then the two riders (the one whose trip is currently at stake, and the rider who was previously matched with
the driver) can start a negotiation to make a ride exchange. Note that incorporating P2P ride-exchange does
not increase the computational burden on the system due to the properties of the proposed algorithm in this
paper. It is easy for the system to propose alternative itineraries to riders involved in the exchange, since all
the feasible itineraries for riders are generated using the DP algorithm.
In the example in Figure 13, if we solve the ride-matching problem for rider 2 without fixing the itinerary
of driver 2, the solution will have driver 2 routed on the dashed path and matched with rider 2. However,
driver 2 was previously routed differently, and assigned to rider 1. These two riders can now engage in a
negotiation. Since rider 1 can also be matched with driver 1, there is a good chance that rider 1 will accept
the proposed ride exchange by rider 2, in exchange for money or credit toward future trips. In case the
negotiation is successful, the final matching would be equal to the matching obtained by solving a many-to-
many ride-matching problem.
In addition to the benefits that P2P ride exchange can offer in terms of system performance, it can also
make riders more engaged with the system. Riders can earn money or credit by selling their itineraries and
settling for less efficient itineraries suggested by the system. However, for such a system to work properly,
a detailed mechanism needs to be designed to ensure incentive-compatibility, individual rationality, and
budget balancedness of the proposed trade. A simple form of such a trade is described in Masoud et al.
(2017).

6.6 Overlapping Sets of Drivers and Riders

Initially, we made the assumption that riders and drivers form two mutually exclusive sets. In this section,
we study the scenario in which participants who register in the system as riders are doing so because they
prefer traveling as riders, and not because they do not own or have access to a personal vehicle. If the system

22
Figure 15: Number of served riders under different percentage of riders (who can switch and become drivers)

is not able to serve these participants as riders, then they will drive their own vehicles and join the system
as drivers. An extreme case would be to study the scenario where all riders have access to vehicles and are
willing to join the system as drivers. This scenario demonstrates the highest benefits a ridesharing system
can offer.
In this section, we study the impact of this assumption on the problem instances in section 6.1. Figure 15
shows the maximum number of served ride requests under different percentage of riders who are willing and
able to act as drivers if not successfully served as riders. The number of served riders reported in this figure
is the number of participants whose last status is a rider. For example, if a participant who is originally
registered as a rider switches to being a driver, they are considered as a driver and not a served rider. In
the case of 0%, the problem becomes equivalent to the problem of a system with two mutually exclusive
sets of riders and drivers. As this percentage increases, the shape of the curve starts to change. The best
performance of the system can be obtained when 100% of riders have access to vehicles, and are willing to
join the system as drivers. It is interesting to note that in this best case scenario, the number of served riders
has an increasing trend as the percentage of riders in the problem instances increases (and the percentage of
drivers decreases).
In the basic scenario with two mutually exclusive sets of riders and drivers, the number of served riders
reaches its peak when riders constitute around 45% of the participants. When working with mutually ex-
clusive sets of rider and drivers, the number of served riders decreases as the ratio of riders in the problem
increases, because there are fewer drivers available. If riders who cannot be matched join the system as
drivers, this practically increases the number of drivers in the ranges where previously there was a shortage
of drivers, and this leads to higher number of riders being served. Figure 15 suggests that encouraging this
flexible form of participation in a ridesharing system can drastically improve its performance.

7 Conclusion
In this paper, we define a flexible ridesharing system as a dynamic, real-time, and multi-hop system with
the ability to find itineraries for riders by means of optimally routing drivers. Real-time ride-matching is a
central component of a flexible P2P ridesharing system. It ensures that users who look for rides not long
before their desired departure times have a chance to be served in the system. We propose an optimal and
real-time ride-matching algorithm that maximizes the number of served riders in the system, while making
the trips as comfortable as possible by taking into consideration users’ preferences on whom to ride with,
and by minimizing the number of transfers and waiting times for riders. Numerical experiments suggest that

23
the proposed algorithm can solve matching problems in large-scale ridesharing systems in a fraction of a
second. Furthermore, results suggest that allowing transfers can have a considerable impact on the number
of served riders.
We start developing the framework under the realistic assumption that riders in a real-time system need to
be served on a first-come, first-served (FCFS) basis. We later introduce the concept of P2P ride exchange to
reverse the negative impacts of the FCFS rule on system performance, without increasing the complexity of
the problem. Furthermore, we investigate the impact of considering intersecting sets of riders and drivers on
system performance, and demonstrate using numerical experiments that under the scenario where a portion
of unmatched riders are willing to take the driver role, the matching rate improves significantly.

References
Agatz, N., A. Erera, M. Savelsbergh, and X. Wang (2009). Sustainable passenger transportation: Dynamic
ride-sharing. ERIM Report Series Reference No. ERS-2010-010-LIS.

Agatz, N. A., A. L. Erera, M. W. Savelsbergh, and X. Wang (2011). Dynamic ride-sharing: A simulation
study in metro atlanta. Transportation Research Part B: Methodological 45(9), 1450 – 1464.

Baldacci, R., V. Maniezzo, and A. Mingozzi (2004). An exact method for the car pooling problem based on
lagrangean column generation. Operations Research 52(3), pp. 422–439.

Böckmann, M. (2013). The shared economy: It is time to start caring about sharing; value creating factors
in the shared economy. University of Twente, Faculty of Management and Governance.

Braekers, K., A. Caris, and G. K. Janssens (2014). Exact and meta-heuristic approach for a general heteroge-
neous dial-a-ride problem with multiple depots. Transportation Research Part B: Methodological 67(0),
166 – 186.

Business Insider (2015). Uber Just Released Its First Report On Its Drivers Here Are The Numbers.
http://www.businessinsider.com/uber-driver-data-report-2015-1. Accessed: 2015-06-20.

Cangialosi, E., A. Di Febbraro, and N. Sacco (2016). Designing a multimodal generalised ride sharing
system. IET Intelligent Transport Systems 10(4), 227–236.

Carnes, T. A., S. G. Henderson, D. B. Shmoys, M. Ahghari, and R. D. MacDonald (2013). Mathematical

programming guides air-ambulance routing at ornge. Interfaces 43(3), 232–239.

Cordeau, J.-F. and G. Laporte (2007). The dial-a-ride problem: models and algorithms. Annals of Operations
Research 153(1), 29–46.

Cortés, C. and R. Jayakrishnan (2002). Design and operational concepts of high-coverage point-to-point
transit system. Transportation Research Record: Journal of the Transportation Research Board (1783),
178–187.

Di Febbraro, A., E. Gattorna, and N. Sacco (2013). Optimization of dynamic ridesharing systems. Trans-
portation Research Record: Journal of the Transportation Research Board (2359), 44–50.

Ghoseiri, K. (2013). Dynamic rideshare optimized matching problem. PhD Dissertation at University of
Maryland.

24
Herbawi, W. and M. Weber (2011, Nov). Ant colony vs. genetic multiobjective route planning in dynamic
multi-hop ridesharing. In 2011 IEEE 23rd International Conference on Tools with Artificial Intelligence,
pp. 282–288.

Herbawi, W. M. and M. Weber (2012). A genetic and insertion heuristic algorithm for solving the dynamic
ridematching problem with time windows. In Proceedings of the 14th Annual Conference on Genetic and
Evolutionary Computation, GECCO ’12, New York, NY, USA, pp. 385–392. ACM.

Jaw, J.-J., A. R. Odoni, H. N. Psaraftis, and N. H. Wilson (1986). A heuristic algorithm for the multi-vehicle
advance request dial-a-ride problem with time windows. Transportation Research Part B: Methodologi-
cal 20(3), 243 – 257.

Li, X. and L. Quadrifoglio (2010). Feeder transit services: choosing between fixed and demand responsive
policy. Transportation Research Part C: Emerging Technologies 18(5), 770–780.

Liaw, C.-F., C. C. White, and J. Bander (1996, Sep). A decision support system for the bimodal dial-a-ride
problem. IEEE Transactions on Systems, Man, and Cybernetics - Part A: Systems and Humans 26(5),
552–565.

Masoud, N. and R. Jayakrishnan (2017). A decomposition algorithm to solve the multi-hop peer-to-peer
ride-matching problem. Transportation Research Part B: Methodological 99, 1 – 29.

Masoud, N., R. Lloret-Batlle, and R. Jayakrishnan (2017). Using bilateral trading to increase ridership and
user permanence in ridesharing systems. Transportation Research Part E: Logistics and Transportation
Review 102, 60–77.

Masson, R., F. Lehuédé, and O. Péton (2014). The dial-a-ride problem with transfers. Computers & Oper-
ations Research 41(0), 12 – 23.

Psaraftis, H. N. (1983). An exact algorithm for the single vehicle many-to-many dial-a-ride problem with
time windows. Transportation Science 17(3), 351–357.

Qiu, F., W. Li, and J. Zhang (2014). A dynamic station strategy to improve the performance of flex-route
transit services. Transportation Research Part C: Emerging Technologies 48, 229–240.

Quadrifoglio, L., M. M. Dessouky, and F. Ordóñez (2008). Mobility allowance shuttle transit (mast) ser-
vices: Mip formulation and strengthening with logic constraints. European Journal of Operational Re-
search 185(2), 481–494.

Regue, R., N. Masoud, and W. Recker (2016). Car2work: A shared mobility concept to connect commuters
with workplaces. Transportation Research Record: Journal of the Transportation Research Board 2542,
102–110.

Savelsbergh, M. W. P. and M. Sol (1995). The general pickup and delivery problem. Transportation Sci-
ence 29(1), 17–29.

Schaub, T., G. Friedrich, and B. O’Sullivan (2014). ECAI 2014: 21st European Conference on Artificial
Intelligence, Volume 263. IOS Press.

Shaheen, S. and N. Chan (2015). Mobility and the sharing economy: Impacts Synopsis. Technical report,
TRANSPORTATION SUSTAINABILITY RESEARCH CENTER, University of California Berkeley.

Stein, D. M. (1978). Scheduling dial-a-ride transportation systems. Transportation Science 12(3), 232–249.

25
Teodorovic, D. and M. DellOrco (2005). Bee colony optimization–a cooperative learning approach to
complex transportation problems. In Advanced OR and AI Methods in Transportation: Proceedings of
16th Mini–EURO Conference and 10th Meeting of EWGT (13-16 September 2005).–Poznan: Publishing
House of the Polish Operational and System Research, pp. 51–60.

Vukan, V. (1981). Urban Public Transportation: Systems and Technology. Prentice-Hall, Englewood Cliffs,
NJ, EU.

Wagner, D. and T. Willhalm (2003). Geometric speed-up techniques for finding shortest paths in large sparse
graphs. In Esa, Volume 3, pp. 776–787. Springer.

Wagner, D., T. Willhalm, and C. Zaroliagis (2005). Geometric containers for efficient shortest-path compu-
tation. Journal of Experimental Algorithmics (JEA) 10, 1–3.

Wolfler Calvo, R., F. de Luigi, P. Haastrup, and V. Maniezzo (2004). A distributed geographic information
system for the daily car pooling problem. Computers & Operations Research 31(13), 2263 – 2278.

A Appendix
Figure 16 shows the time-expanded reduced networks of the rider, and drivers 1, 3, and 4.

26
 (a) Rider (b) Driver 1

(c) Driver 3 (d) Driver 4

Figure 16: Time expanded networks of participants in the example presented in Figure 2

27