Transportation Research Part C 132 (2021) 103362

Contents lists available at ScienceDirect

Transportation Research Part C

journal homepage: [Link]/locate/trc

Aggregate modeling and equilibrium analysis of the

crowdsourcing market for autonomous vehicles
Xiaoyan Wang a, Xi Lin a, Meng Li a, b, *
a
Department of Civil Engineering, Tsinghua University, Beijing 100084, PR China
b
Center for Intelligent Connected Vehicles and Transportation, Tsinghua University, Beijing 100084, PR China

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

Keywords: Autonomous vehicles (AVs) have the potential of reshaping the human mobility in a wide variety
Crowdsourcing of aspects. This paper focuses on a new possibility that the AV owners have the option of “renting”
Autonomous vehicles their AVs to a company, which can use these collected AVs to provide on-demand ride services
Mobility market
without any drivers. We call such a mobility market with AV renting options the “AV crowd­
Equilibrium
Pricing
sourcing market”. This paper establishes an aggregate equilibrium model with multiple transport
modes to analyze the AV crowdsourcing market. The modeling framework can capture the cus­
tomers’ mode choices and AV owners’ rental decisions with the presence of traffic congestion.
Then, we explore different scenarios that either maximize the crowdsourcing platform’s profit or
maximize social welfare. Gradient-based optimization algorithms are designed for solving the
problems. The results obtained by numerical examples reveal the welfare enhancement and the
strong profitability of the AV crowdsourcing service. However, when the crowdsourcing scale is
small, the crowdsourcing platform might not be profitable. A second-best pricing scheme is able
to avoid such undesirable cases. The insights generated from the analyses provide guidance for
regulators, service providers and citizens to make future decisions regarding the utilization of the
AV crowdsourcing markets for serving the good of the society.

1. Introduction

With the rapid development of artificial intelligence and communication technology in recent years (Hasan et al., 2020; Zhang
et al., 2019), autonomous vehicles (AVs) are undergoing rapid development. There has been an intense effort by researchers and
manufacturers to enable AVs to handle a wide range of driving conditions that can be encountered during the road travel. In 2018,
public road tests for AVs to provide on-demand ride service were performed, and in 2020, some major transportation network
companies (TNCs) such as Lyft, Waymo and Baidu have launched commercialized AV on-demand ride service programs in certain
areas around the globe. It is predicted that approximately 40% of vehicle travel could be autonomous in the 2040s (Litman, 2015), and
as suggested by Zmud et al. (2018), traveling by AV may become the predominant transport mode in the future. It is widely believed
that due to the increased mobility and safety, reduced congestion, emissions and travel costs (Litman, 2015), AVs have the potential to
reshape future human transport.
High-tech AVs can perform tasks that cannot be performed by manually driven vehicles (MVs). One of the most distinct differences
between AVs and MVs is that AVs owned by a person can be used for other tasks when not used by the owner such as traveling to other

* Corresponding author at: Department of Civil Engineering, Tsinghua University, Beijing 100084, PR China.
E-mail address: mengli@[Link] (M. Li).

[Link]
Received 8 February 2021; Received in revised form 16 August 2021; Accepted 18 August 2021
Available online 20 September 2021
0968-090X/© 2021 Elsevier Ltd. All rights reserved.
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Fig. 1. One day’s task chain of an autonomous vehicle.

places without drivers to complete certain tasks. One potential use is to serve other riders (Fig. 1 for illustration). This would provide
opportunities for people who do not own AVs to utilize the service.
When a large number of AV owners are able to share their AVs to provide ride services to others, a platform is required to aggregate
the information of those AVs with that of the customer demands, and act as a bridge connecting the AV owners with the customers.
Similar to the current TNCs such as Uber and Lyft, the platform identifies the dynamically-updated supply from AV owners, and then
matches the available AVs with the customers generated within the region in a real-time fashion. The platform needs to pay these AV
owners a certain amount of money to encourage them to share their AVs, and meanwhile the platform charges the customers a certain
fee for their trips; the difference between the platform gain from travelers and the platform payment to AV owners is the revenue that
should be equal to a considerable amount so that people have the incentive to establish and operate such a service.
The aforementioned new type of mobility market with AV renting options is called “the AV crowdsourcing market” in this study.
The concept of crowdsourcing, first introduced by Brabham (2008) and later perfected by other researchers using case studies (Estellés-
Arolas and González-Ladrón-De-Guevara, 2012), refers to the act of a company handing over a part of its tasks, particularly needs-
based problems (Poetz and Schreier, 2012), such as hatching ideas, designing algorithms and public supervision to the public
crowd who have the potential to finish the jobs better than employees (Brabham, 2008). Developed communication technologies
enable such business model to be realized, through which a wide range of the public including but not limited to social experts are
attracted and involved to produce a better outcome for the companies (Vander Schee, 2009). For the AV crowdsourcing market, our
primary interest is to answer two key questions: (1) can AV crowdsourcing benefit society? and (2) can the operating agency gain from
operating the crowdsourcing platform? The first question relates to the utilitarian objective of such business model that tells social
managers whether they should allow or regulate the market, and the second question is related to the spontaneity problem whether
people have the incentive to run such a business. To answer these questions, we establish an aggregate equilibrium modeling
framework in this study. The model considers three market subjects: travelers, AV owners and the crowdsourcing platform, where AV
owners and travelers may have some coincidence within a time period. The platform determines the payment to the AV owners and the
charge to the travelers, and the travelers and AV owners determine their mode choices and rental choices. Under a user’s utility-
maximization framework, we can capture the equilibrated state of the whole system. Then, based on the equilibrium state, the
platform can adjust the payment and pricing strategies to achieve different goals, including profit maximization (monopoly), social
welfare maximization (first-best), and constrained social welfare maximization (second-best). Numerical experiments and sensitivity
analyses are conducted to draw a variety of insights on the AV crowdsourcing market.
The rest of this paper is organized as follows. Section 2 reviews related literature on AVs, ride-providing services and crowd­
sourcing. In Section 3, we develop the static equilibrium model for the AV crowdsourcing market, and investigate some properties of
the model. Section 4 formulates the platform’s decision framework across different market scenarios with various transport modes, and
then propose a gradient-based algorithm to solve the problem. In Section 5, we use numerical examples to illustrate key points and
insights of our model, and then present some discussions. Finally, Section 6 concludes the paper.

2. Related literature

The body of literature on AVs has been continuously growing in recent years. Control of fully AVs may be the best investigated topic
(Wu et al., 2021), including route planning (Wang et al., 2020), longitudinal control, i.e. Cooperative Adaptive Cruise Control (Gong
and Du, 2018), and lateral control of autonomous vehicles (Yang et al., 2018). An et al. (2020) proposed a space–time routing
framework for control of AVs in dedicated AV zones. Chen et al. (2020) studied control of AVs using dedicated bus rapid transit (BRT)
lanes and Chen et al. (2020) studied control of AVs in mixed flow of AVs and manually driven vehicles (MVs). Control in intersection
has been studied with “signal-free” schemes (Xu et al., 2018) and “signalized” schemes (Yu et al., 2018) where the aim is to increase
intersection capacity and to reduce delay on the basis of driving safety. Recently, Chen et al. (2021) proposed a rhythmic control
scheme where CAVs pass through an intersection with a preset rhythm, and Lin et al. (2021), further expanded the rhythmic control to
a grid network. Meanwhile, studies found that AVs with communication technologies that can smooth oscillations and reduce crash
probability, will greatly improve the capacity of a highway (Bian et al., 2019; Michael et al., 1998) and reduce road congestion
(Fagnant and Kockelman, 2015; Sun et al., 2020). AVs are highly advantageous for providing high-level transport experience so that
they will change or even reshape future travel modes. One change will revolutionize the taxi industry. Literature on AV on-demand ride
service mainly focuses on its operation. For example, Vosooghi et al. (2019), considering multi-modal dynamic demand, investigated
shared autonomous vehicle operations including the fleet size, vehicle capacities, ridesharing and rebalancing strategies through
simulations. Autonomous aerial vehicle has also been widely explored (Lin et al., 2021; Wang et al., 2021) these days.

2
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Fig. 2. Conceptual design of an AV crowdsourcing market.

Although there is scarcely any literature on AV on-demand ride service, studies on the traditional taxi market and on-demand ride
service market have been performed that provide great inspiration for our study. Studies on traditional taxi market include Yang and
Wong (1998) which used a network model to establish the equilibrium of the cruising taxi market and offered some policy-relevant
results such as average taxi utilization and average customer waiting time for decision making. Pricing is an effective approach to
deal with the congestion problem (Lin et al., 2018, 2020). Arnott (1996) and Yang et al. (2014) performed economic analysis of the taxi
market, with Yang et al. (2014) considered taxi service congestion externalities through imposing a delay pricing. Ride-sourcing
service becomes competitive with the traditional taxi service in recent years (Yu et al., 2017). The inefficient matching of
ride-sourcing service in high demand scenarios, i.e. wild goose chase, has been explored (Zhang et al., 2019) and suggested to be
avoided by adaptively adjusting the matching radius (Xu et al., 2020) or applying price or rationing discrimination (Castillo et al.,
2018; Urata et al., 2021). The matching process has been modeled as Cobb-Douglas function (Zha et al., 2016), urn-ball matching
function (Buchholz, 2021) or queue process (Xu et al., 2020). He et al. (2018) modeled the effect of the customers’ reservation
cancellation behaviors on ride-sourcing service. After ride-sourcing is popularized, ride-pooling becomes more welcomed nowadays
due to benefits such as lower prices, better utilization of vehicle resources and traffic alleviation, and is distinguished from
ride-sourcing mainly in the pairing rate and detour time (Ke et al., 2020). Daganzo and Ouyang (2019) presented a general analytic
framework for non-shared and shared taxi services to model the assignment, pick-up and drop-off process of on-demand rides. Ke et al.
(2020) investigated the impacts of ride-pooling service on road congestion by establishing the relationship between vehicle speed and
vehicle density as macroscopic fundamental diagram. Zhang and Nie (2019) and Vignon et al. (2021) modeled the on-demand ride
market with dual modes of solo and pooling services, with Zhang and Nie (2019) proposed the policy of charging congestion tax to
penalize single rides and Vignon et al. (2021) imposed congestion externality on the drivers’ income. Although traditional taxi,
ride-sourcing and ride-pooling services may be different in many aspects, they share the same economic rules that the platform will
probably suffer a deficit in the first-best scenario (Arnott, 1996; Zha et al., 2016; Vignon et al., 2021). In this paper, we will show in
some circumstances, the AV crowdsourcing platform could be profitable in the first-best scenario. In addition, to the best of our
knowledge, few works have conducted detailed analysis of the future transport modes in which AVs can drive autonomously and can
serve customers on their own (Narayanan et al., 2020; Stocker and Shaheen, 2018; Zmud et al., 2018). The potential AV crowdsourcing
market enabled by the driverless feature of AVs has been seldom investigated in the existing literature.
Crowdsourcing has boomed in various industries in recent years (Brabham, 2008). It can be applied to collect ideas to improve
products or services (Bayus, 2010; Poetz and Schreier, 2012; Bayus, 2013), city inspection (Glaeser et al., 2016; Kang et al., 2013) and
other applications. Platforms such as Threadless and IStockPhoto collect and share designer’s talents around the world; Kaggle and
TunedIT gather studies and solutions to difficult problems; other companies such as PeoplePerHour, Crowdspirit, Editzen and
MusikPitch are examples of crowdsourcing (Zhao and Zhu, 2014). A business model in the hotel industry similar to AV crowdsourcing
is shared lodging where one of the typical companies is Airbnb (Zervas et al., 2017). The AV crowdsourcing that we propose in this
paper has been referred to in Stocker and Shaheen (2018) and described as hybrid AV ownership with the same entity operating.
Stocker and Shaheen (2018) preliminarily listed potential shared autonomous vehicle business models but did not make a detailed
analysis of them; to bridge this gap, such models are investigated in the present paper.

3. Equilibrium model

This section introduces the equilibrium model of the AV crowdsourcing market. We consider a time period with duration h (e.g., 2
h) for which the market conditions are time invariant, and therefore we can treat it as a stationary state. In this hypothetical market,
there exists a crowdsourcing platform; the AV owners have the option to rent their AVs to this crowdsourcing platform, and the
crowdsourcing platform collects these AVs to provide on-demand ride service for other riders. In addition to the collected AVs, the
crowdsourcing platform can also buy some AVs from external markets to provide the service. On the other hand, travelers can choose
to use the on-demand ride service, to drive private (manual or autonomous) vehicles or to take public transit to fulfill their travel needs
(the average trip distances of taking different modes are identical); meanwhile, AV owners can choose to rent their cars to this
crowdsourcing platform. It is worth noting that our model is an aggregate model that ignores spatial effect, so all variables in the model
are actually in an “average” sense, including average trip time, average trip fare, etc. Traffic congestion induced by these trips is

3
X. Wang et al. Transportation Research Part C 132 (2021) 103362

considered in the framework. A schematic illustration is provided in Fig. 2.

In the following, we will introduce the details of the equilibrium model establishment. We first describe the equilibrium state of the
on-demand ride service; and then, we model the travelers’ mode choices as well as the AV owners’ rental choices based on the utilities
associated with these selections; and finally, we describe the equilibration of supply and demand in a steady-state mobility market with
AV renting options.

3.1. Ride services and traffic congestion

We separate one-day time into several periods and use piece-wise fixed level of demand to approximate the time-variant demand. In
this section, we focus on a single period where the platform set prices to equilibrate the supply and demand of on-demand rides. On-
demand ride services are composed of matching, pick-up and on-travel processes. The customer first makes a request online and
searches for an available AV, and then the successfully paired AV comes to the appointed place to pick up the customer, finally the AV
carries the customer to the destination. In the matching process, we assume the customers continuously send requests to available AVs
in the surrounded region (a partitioned matching block or the whole city) and the platform serves customers according to the nearest-
distance policy. Thus, for a single customer, the matching time wc should be monotonically decreasing with the number of available
AVs Nv which is in a time-averaged sense; from the perspective of stochastic process, more densely distributed idle AVs in the region
means a higher probability that a successful matching is made at one request, and thus a successful matching needs less matching time.
There should also be limNv →0 wc = ∞ and limNv →∞ wc = 0. Considering the number of idle AVs Nv = wt qo , where wt is the AVs’ average
searching time, and qo is the demand for on-demand rides which equals the number of matchings per hour at the stationary state, we
can construct the relation describing the matching time as:

(1)
− α1
wc = A(wt qo )

which satisfies the required properties. A is a scaling parameter capturing the level of matching technology, and α1 > 0 is a parameter
to depict the elasticity of wc with respect to Nv . Note that we respectively propose the matching time function and pick-up time
function, different from the existed literature (e.g., Zha et al., 2016) where the matching process includes the pick-up process.
Upon being matched, the AV drives itself to the customized location to pick up the customer. We assume that the matching al­
gorithm assigns to every request one of the closest wt qo available vehicles. Given the size of the city R, the average pick-up time
√̅̅̅̅
tp ≈ k(wt qo )− 1/2 R/v, where k is a parameter depending on the network topology and k ≈ 0.63 for networks that resemble rectangular
grids, and v is the cruising speed (Daganzo and Ouyang, 2019). The average trip time from picking up a customer to arriving at the
√̅̅̅̅
destination is tr = κ R/v, where κ is a constant reflecting the average travel distance. Then, we obtain the following equation between
the average pick-up time tp and the average trip time from picking up a customer to the arrival at the destination tr :

(2)
− 1
tp = θ(wt qo ) 2 tr

where θ = k/κ is a city-specific constant.

Next, we discuss the modeling of traffic congestion. Since we have constructed the relation between the average pick-up time tp and
the average trip time tr , in the following we focus on the effect of traffic congestion on the average trip time tr . The functional form is
given by the following equation:
⎡ ⎛ ⎞ ⎤
(3)
−1
tr = Ts ⎣qm + α(qa − qo ) + αqo ⎝1 + θ(wt qo ) 2 ⎠; R, v0 ⎦

In the above equation, Ts ( ⋅ ) is a continuously differentiable, monotonically increasing and convex function; qm , qa , qo are travel
demands per hour for manual driving, autonomous driving (including both private vehicles and crowdsourcing vehicles) and on-
demand ride service, respectively; α ∈ (0, 1) is a parameter representing the “relative occupation” of AVs compared to MVs
(because AV technology can reduce the vehicle headway in traffic flows). The equation states that the average trip time tr is a function
related to city size R and free-flow cruising speed v0 , and it is a monotonically increasing and convex function with regards to the term
( ( ))
qm + α(qa − qo ) + αqo 1 + θ(wt qo )− 1/2 . In this term, qm represents the MVs’ contribution to congestion, α(qa − qo ) represents
( )
privately-used AVs’ contribution to congestion, and αqo 1 + θ(wt qo )− 1/2 represents the crowdsourcing AVs’ contribution to traffic
( )
congestion; the latter has a factor of 1 + θ(wt qo )− 1/2 because it includes both the pick-up trips and delivery trips, where θ(wt qo )− 1/2
is associated with Eq. (2). The functional form of Ts ( ⋅ ) can be obtained from realistic data.
Despite their analytical formulation, there are some mathematical difficulties associated with Eqs. (1)–(3). The main difficulties lie
in the terms (wt qo )− α1 in Eq. (1) and (wt qo )− 1/2 in Eqs. (2) and (3), where there are negative powers associated with these variables; this
gives rise to discontinuity and possibly results in imaginary numbers in the process of solving the equilibrium. To address this issue, we
introduce some continuous counterparts to replace the above discontinuous functions. The idea is as follows (taking (wt qo )− α1 as an
example). Because the non-negativity of qo is ensured by Eq. (10), we only need to approximate (wt )− α1 in case that the discontinuity in
the function may occur with a negative wt . We first choose a very small positive number ∊. Then, when wt ⩾∊, the counterpart takes

4
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Fig. 3. Illustration of function approximation.

Table 1
Citizen classification.
Class Car ownership Travel needs Number of citizens

1 No Yes dn
2 MV Yes dr
3 AV Yes da
4 AV No da′
5 MV & AV Yes db
6 MV & AV No db′

Table 2
Generalized utilities of all possible choices.
Citizen class Choice Rental utility Travel utility

1 (P, -) 0 − Fn − βP tn
( )
(O, -) 0 − Fo − βA tr − γ wc +tp

2 (P, -) 0 − Fn − βP tn
(M, -) 0 − βM tr
( )
(O, -) 0 − Fo − βA tr − γ wc +tp

3 (A, N) 0 − βA t r
(P, N) 0 − Fn − βP tn
( )
(O, N) 0 − Fo − βA tr − γ wc +tp
(P, R) pn0 − m − Fn − βP tn
( )
(O, R) pn0 − m − Fo − βA tr − γ wc +tp

4 (-, N) 0 0
(-, R) pn0 − m 0

5 (A, N) 0 − βA t r
(P, N) 0 − Fn − βP tn
(M, N) 0 − βM tr
( )
(O, N) 0 − Fo − βA tr − γ wc +tp
(P, R) pn0 − m − Fn − βP tn
(M, R) pn0 − m − βM tr
( )
(O, R) pn0 − m − Fo − βA tr − γ wc +tp

6 (-, N) 0 0
(-, R) pn0 − m 0

5
X. Wang et al. Transportation Research Part C 132 (2021) 103362

exactly the same form as (wt )− α1

; and when wt < ∊, the counterpart is a linear function such that the whole function is continuous and
smooth at ∊. The same approximation is used for (wt qo )− 1/2 . This approximation is reasonable because wt only have physical meaning
when they are positive, and a sufficiently small ∊ can ensure that the approximations are almost the same as the original values on the
positive horizon. In the following, we denote the two approximations as ξ1 (wt ) and ξ2 (wt ). The graphical illustration of the approx­
imation is shown in Fig. 3.
With the above-mentioned function approximation scheme, we can express the approximated counterparts of Eqs. (1)–(3) as
described below. These expressions will be used in the model analysis and algorithmic development.

wc = W c (wt , qo ; A, α1 ) (4)

tp = Tp (wt , qo , tr ; θ) (5)

tr = Tr (qm , qa , qo , wt ; R, v0 , θ) (6)

3.2. Choice modeling

In this subsection, we discuss the citizens’ choices, including travelers’ model choices and AV owners’ rental choices. According to
the car ownership and travel needs, we divide the citizens into 6 classes as follows: (1) have travel needs in this period and own no
private cars; (2) have travel needs in this period and own a private MV; (3) have travel needs in this period and own a private AV; (4)
have no travel needs in this period and own a private AV; (5) have travel needs in this period and own a private MV as well as a private
AV; (6) have no travel needs in this period and own a private MV as well as a private AV. We denote by dn , dr , da , da′ , db , db′ the
populations corresponding to the six classes’ citizens for one unit time with duration h, respectively (Table 1). We assume each traveler
with travel needs consumes exactly one trip in a unit time with duration h. A person’s choice can be written as (i, j) ∈ H x (Table 2): the
first element in the parentheses i is relevent to travel mode choices, and the second element j is relevent to AV rental choices, and H x is
the set of all possible choices for class x citizens, x∈{1,2,...,6}. In details, i∈{A,P,M,O} stand for traveling by private AV, public transit,
private MV or on-demand ride, respectively; and j∈{-,N,R} stand for no AV ownership, not renting AVs with AV ownership, or renting
AVs to the crowdsourcing platform, respectively. It should be mentioned that the two indices i and j must be non-conflicting, e.g.,
people cannot rent out their own AVs and meanwhile take private AVs to travel.
AV owners have the incentive to rent their cars to the crowdsourcing platform only if this action can earn profit. Suppose that the
rental revenue depends on the serving rates, i.e., for unit time with duration h, the rental revenue for the AV owner is given by pn0 ,
where p is the payment per ride and n0 is the unit-time expected number of on-demand rides served by an AV. The rental revenue can
be viewed as the “commission fee” for the vehicle providers. Then, the unit-time utility of a citizen renting his/her private AV to the
crowdsourcing platform is given by (pn0 − m), where m is a constant indicating the expected additional cost of sharing the private
vehicle for public use, such as energy consumption, vehicle insurance, vehicle depreciation, psychological costs and other costs not
included in our model. The AV owners who choose not to rent their cars will receive zero rental utility.
Next, we discuss the mode choices for citizens with travel needs in duration h. The travel (dis) utility includes time costs and
monetary costs. We respectively specify the perception of on-board time for AVs, public transit and MVs as βA , βP and βM , and the
relation among them is βA < βP < βM ; this is because driving a car costs the most human effort, and taking public transit should be
generally less comfortable than taking an AV. The perception of waiting for a crowdsourcing AV is specified as γ (The waiting time is
composed of matching and pick-up time). On the other hand, it is assumed that the average travel time tn and fare Fn by public transit
are constants, i.e., unaffected by traffic congestion (e.g., BRT vehicles or subway). Given the average fare Fo , the average trip time tr
and waiting time (wc+tp) for a crowdsourcing trip, the utility of taking on-demand ride service can be stated as
( ( ))
− Fo − βA tr − γ wc +tp . The utilities of other transport modes can be similarly expressed. The utilities for all choices are shown in
Table 2. Let V (i,j)
x indicates the (dis)utility for choice (i,j) by class x citizens, and then V (i,j)
x can be calculated by adding up the rental
utility and travel utility in each choice (i, j) for the citizens of class x.
In the AV crowdsourcing market, we assume that the sensitivities to travel utility and rental utility for one class of people are
identical; and the random variables are identically distributed with a Gumbel density function (Ben-Akiva and Lerman, 1985). Then,
the probability for the choice (i, j) by people of class x can be captured by the following logit model:
(i,j)
eμx Vx
Π(i,j)
x = ∑ (i,j)
∀(i, j) ∈ H x , x ∈ {1, 2, …, 6} (7)
eμx Vx
(i,j)∈H x

where μx is a nonnegative parameter representing the degree of uncertainty for the choices from the perspective of people of class x,
and H x is the set of all possible choices for people of class x. A larger parameter μx means that people have more knowledge about the
AV crowdsourcing market and thus can be more certain about their utilities, i.e., an increasing number of class x citizens would give up
the choice (i, j) when V (i,j)
x decreases by a unit. In modeling customer choices, however, it is currently unknown whether the sensitivity
coefficients associated with travel and rental choices are similar. If they are not identical, we need a nested logit model (Ben-Akiva and
Lerman, 1985) to capture the choice probabilities. The detailed discussions of the nested logit models are presented in Appendix B.

6
X. Wang et al. Transportation Research Part C 132 (2021) 103362

3.3. Supply and demand

This section establishes the equilibration between supply and demand. For crowdsourcing AVs, the supply equation can be written
as:
N = Ns + Nr (8)

where N is the total number of usable AVs for the crowdsourcing platform in unit time with duration h; Ns is the number of pre-
purchased AVs by the crowdsourcing platform; and Nr is the number of AVs collected from crowdsourcing in unit time with dura­
tion h. The number of AVs crowdsourced from the society can be expressed by the following equation:
( ) ( )
Nr = da ΠP,R
a + Πa
O,R
+ da′ Π−a′ ,R + db ΠP,R M,R
b + Πb + ΠO,R
b + db′ Π−b′ ,R (9)

where the first to last terms of the RHS represent the number of AVs from citizens belonging to classes 3–6, respectively.
Similarly, on the demand side, the equilibration can be expressed as:
( ) ( )
qo h = dn ΠO,−
n + dr ΠO,−
r + da ΠO,N
a + ΠO,R
a + db ΠO,N
b + ΠO,R
b (10)

qa h = qo h + da ΠA,N A,N
a + db Πb (11)
( )
qm h = dr ΠM,−
r + db ΠM,N
b + ΠM,R
b (12)
( ) ( )
qp h = dn ΠP,− P,−
n + dr Πr + da ΠP,N
a + ΠP,R
a + db ΠP,N
b + ΠP,R
b (13)

In Eq. (10), the LHS is the total number of customers using the on-demand ride services in unit time with duration h, and the RHS is
the sum of the travelers choosing the associated mode. Eqs. (11)–(13) express similar equilibria for the AV riders (including privately
used ones and crowdsourcing ones), for manual driving, and for taking public transit, respectively.
According to the stability requirement, the number of trips of the on-demand rides should be equal to the total number of
crowdsourcing AV riders during unit time with duration h:
Nn0 = qo h (14)

where N is the total number of crowdsourcing AVs and n0 is the expected number of the on-demand rides served by an AV per unit time
with duration h.
Lastly, the sum of searching time, pick-up time and on-board time for all AVs put into use should be equal to their total service time.
Therefore, the service time constraint in view of one hour is modeled as follows:
( )
N = qo ⋅ wt + tp + tr (15)
( )
where qo equals the number of crowdsourcing trips per hour and wt + tp + tr is the service time for a single trip by a crowdsourcing
AV.
Aggregating Eqs. (4)–(15) yields the equilibrium model of the AV crowdsourcing market. The following subsection will specify
some basic properties of the model.

3.4. Existence of equilibrium solution

For the platform, on the demand side, the trip fare of on-demand ride service Fo is treated as the decision variable; on the supply
side, the commission fee per ride p is treated as the decision variable. When fixing these two decision variables, the utility equations
along with Eqs. (4)–(15) together constitute a system of nonlinear equations that can be used to obtain the equilibrium state of the AV
crowdsourcing market, i.e., the intermediate variables including choice splits of different travel modes, on-demand ride service rate
and total travel costs.
We examine the existence of the equilibrium solution in this subsection. The statement is presented as follows.
Proposition 1. The system of nonlinear Eqs. (4)–(15) has at least one solution when fixing Fo and p.
Proof. We apply the fixed point theorem to prove the existence. The proof is based on Schauder’s fixed-point theorem (Zou, 2008): let
𝒞 be a closed convex subset of the Banach space and suppose f : 𝒞 ↤ 𝒞 and f are compact (i.e., bounded sets in 𝒞 are mapped into
relatively compact sets), and then f has a fixed point in 𝒞.
Let the constant parameters dmv and dav respectively denote the total number of MVs and AVs owned by the citizens. We focus on
( ) ( ) ( )
four variables tr , n0 , tp , wc , and the system (4)–(15) can be treated as a mapping Φ such that ̂t r , ̂ ̂ c = Φ tr , n0 , tp , wc , and the
n 0 , ̂t p , w
( c) ( )
equilibrium requires ̂t r , ̂ ̂ = tr , n0 , tp , wc . The mapping works as follows: it first computes all choice splits with the logit
n 0 , ̂t p , w
model Eq. (7), and obtain n 0 by Eq. (14), obtain wt by Eq. (15); and then with wt we obtain ̂t r , ̂t p and w
̂ ̂ c by Eqs. (4)–(6). Clearly, the
mapping Φ is continuous. To proceed, we first prove the following result.
̂ c are upper bounded by constants.
Claim: When max(tr , tp ) → +∞, ̂t r , ̂t p , w

7
X. Wang et al. Transportation Research Part C 132 (2021) 103362

To prove this claim, we note a simple fact that when tr or tp approaches infinity, qo will approach zero with an exponential rate due
( )
to the presence of constant-utility public transit, so that qo tr +tp will also approach zero. Therefore, by Eq. (15), qo wt ≈ N. Mean­
while, N is lower bounded by a positive constant due to the presence of citizens with types 4 and 6 and the fact that n0 ⩾0. Thus, qo wt is
lower bounded by a positive constant, from which and Eq. (4) we know that w ̂ c is upper bounded by a constant. Meanwhile, qm and qa
are upper bounded by dmv and dav , respectively. Combining the above facts, based on Eqs. (5)-(6), we know that ̂t r , ̂t p are also upper
bounded by constants. The claim is proved.
On the other hand, ̂ n 0 is always upper bounded by a constant because N is lower bounded by a positive constant, and with Eq. (14),
we can easily identify the bounded nature of ̂ n0 .
Finally, we define the feasible ranges of tr , tp , wc and n0 to be all within [0, M1 ], and this region is denoted as 𝒟. When M1 is suf­
ficiently large, by the above reasoning we know that ̂ n 0 are upper bounded by constants smaller than M1 , and based on the claim stated
above as well as the continuity of Φ, we know that w ̂ c , ̂t r , ̂t p are also upper bounded by constants smaller than M1 . Therefore,
( c
)
Φ tr , n0 , tp , w ∈ 𝒟, and by Schauder’s fixed-point theorem the existence is guaranteed. The proof is completed.

4. Scenario analysis

Using the equilibrium model proposed in the previous section, we can then proceed to analyze several scenarios with different
objectives for decisions regarding the platform payment p and the platform charge Fo . The scenarios of interest include a monopoly
scenario, a first-best scenario and a second-best scenario. The first intends to maximize the crowdsourcing platform’s profit; the second
intends to maximize social welfare; and the third intends to maximize social welfare under the constraints that the crowdsourcing
platform’s profit is guaranteed to be no less than a preset threshold. For practical consideration, in this section we assume that the time-
of-day can be divided into several heterogeneous periods according to the levels of demands. In the following, we use ℘ to denote the
set of all heterogeneous periods, and Hk represents the duration of the kth period, ∀k ∈ ℘. Below, we first present the optimization
models in the above-mentioned three scenarios, and then develop the algorithms for solving these models.
In this section, we use a tuple τ to collectively denote all of the equilibrium variables defined in the last section; these are treated as
intermediate variables in the established optimization models.

4.1. Monopoly scenario

In the monopoly scenario, the crowdsourcing platform decides some key variables to maximize its daily profit. The decision
variables include the period-specific average fare for the on-demand ride service Fo,k and the period-specific payment to AV owners per
customer pk . The total number of purchased AVs, i.e., Ns , is a parameter that is not considered as a decision variable. Let qo,k denote the
period-specific customer demand per hour for the on-demand ride service; Nr,k denote the period-specific number of rented AVs; and
n0,k denote the period-specific number of customers served by a crowdsourcing AV per unit of time with duration h. The optimization
model for the profit maximization problem is then stated as:
{ }
∑ Hk ( )
max Fo,k qo,k h − Nr,k pk n0,k − Ns (g + z) − Cf (16)
Fo ,p,τ
k∈℘
h

s.t. (4) − (15) ∀k ∈ ℘

(17)
Fo,k , pk ⩾0 ∀k ∈ ℘

where g and z are the unit-day amortized purchase cost and maintenance cost of an AV respectively; and Cf represents the basic
operational cost of the crowdsourcing platform per day.
In the objective function (16), Hk Fo,k qo,k represents the revenue of the platform in the kth period, and Hhk Nr,k pk n0,k is the total
payment to the AV owners in the kth period. The formula (17) states that the fares and payments should be nonnegative. It should be
noted that the constraints (4)–(15) are period-specific; i.e., for each period k, there is a corresponding constraint set stating the
equilibration of the AV crowdsourcing market.

4.2. First-best scenario

Now we consider the case that a government agency is operating the crowdsourcing platform, and its goal is to maximize social
welfare. This scenario is called the first-best scenario. By deciding the period-specific fare Fo,k and payment pk , the optimization model
for the operating agency can be written as:
∑ Hk
max Sk (18)
Fo ,p,τ
k∈℘
h

s.t. (4) − (15), (17) ∀k ∈ ℘

where Sk represents the unit-time social welfare in the kth period, and the length of such time duration is h. The constraints are the same

8
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Fig. 4. Solution algorithms.

as in the monopoly scenario. The social welfare is composed of consumer surplus, provider surplus and the platform profit. The total
(unit-time) consumer surplus and provider surplus is given by:
( )
∑ 1 ∑ μ V (i,j)
dx,k ln e x x,k
x∈{1,2,…,6} x
μ (i,j)∈H x

the above formula is because that both the consumers and the providers are from the citizen side, and the formula captures their total
surplus under the logit-based discrete choice setting (Ben-Akiva and Lerman, 1985). Note that the consumer payment to the platform
and the provider gain from the platform are included in the disutility terms V x,k . Meanwhile, the platform’s profit is given by:
(i,j)

h [ ]
Fo,k qo,k h − Nr,k pk n0,k − ∑ Ns (g + z) + Cf
Hj
j∈℘

∑
where j∈℘ Hj represents the total time duration of one day, Fo,k qo,k h − Nr,k pk n0,k is the income of the platform from the consumers after
[ ]
paying the providers, and ∑ h H
Ns (g + z) + Cf is the unit-time amortized cost. By combining the above terms, we give the expression
j∈℘ j

for the unit-time social welfare in the kth period as:

( )
∑ 1 ∑ μ V (i,j) ( ) h [ ]
Sk = dx,k ln e x x,k + Fo,k qo,k h − Nr,k pk n0,k − ∑ Ns (g + z) + Cf (19)
μ
x∈{1,2,…,6} x (i,j)∈H x
Hj
j∈℘

where k ∈ ℘.

4.3. Second-best scenario

From the platform’s perspective, the on-demand ride service can produce a deficit under the first-best case, and a similar phe­
nomenon is identified in the traditional taxi and ride-sourcing markets (Arnott, 1996; Zha et al., 2016). Specifically, in the early stage
of AV adoption, it is likely that only a few citizens own AVs, and there will be a shortage of AVs that can be rented. In this case, if the
crowdsourcing platform lacks start-up fund to purchase AVs, the on-demand ride service provided by the crowdsourcing AVs will be of
little attraction since it requires excessively long pick-up distances, cutting down the platform’s profit. To guarantee a certain level of
profitability for the service provider, in this subsection we propose a second-best operational strategy that maximizes social welfare
under the constraint of a lower bound on the platform profit. The model formulation is illustrated as follows:
∑ Hk
max Sk (20)
Fo ,p,τ
k∈℘
h

s.t. (4) − (15), (17) ∀k ∈ ℘

∑ Hk ( ) [ ] (21)
Fo,k qo,k h − Nr,k pk n0,k − Ns (g + z) − Cf ⩾ρ Ns (g + z) + Cf ∀k ∈ ℘
k∈℘
h

9
X. Wang et al. Transportation Research Part C 132 (2021) 103362

where ρ⩾0 represents the basic rate for profit accumulation. The second-best formulation shows some differences with the previous
two models in that it contains an additional nonlinear constraint with inequalities, i.e., the formula (21). The next subsection will show
how to deal with this constraint.

4.4. Solution algorithms

The optimization models presented above are all nonlinear programs with complicated nonlinear constraints. To efficiently solve
these models, we treat them as bi-level forms containing an equilibrium problem in the lower level, and then incorporate them into a
gradient projection algorithmic framework. In detail, we first obtain the equilibrium given the fare Fo,k and payment pk ; and then we
calculate the gradient of the objective functions to Fo,k and pk in the current equilibrium state, according to which we update the
decision variables and project them to the feasible domain when they are to cross the boundary. Thus, the process is carried out
iteratively until the termination criterion is met (Fig. 4).
We use the implicit function theorem (Gray, 2015) to compute the gradient of equilibrium state with regard to fare Fo and payment
p, and then the gradient of objective functions with regard to Fo and p can be obtained using derivative rule for complex functions. The
gradient of equilibrium state is given by
( )− 1
∂τ ∂f ∂f
=− ⋅ (22)
∂(Fo , p) ∂τ ∂(Fo , p)

Here, f u are the choice utility functions; f π are the choice probability functions for which the original forms are given by Eq. (7); f n are
the supply/demand functions for which the original forms are given by Eqs. (8)–(15); and f t are the traffic-related functions for which
⎛ ⎞
fu
⎜ fπ ⎟
the original forms are given by Eqs. (4)–(6). Then, the equilibrium can be compactly expressed as f = ⎝ ⎟ ⎜ = 0. For ∂f/∂τ , we
fn ⎠
ft
observe that f u and f n are linearly related with τ . With regard to probability functions, f π , we have
∑ (i,j) (m,n)
eμx Vx − eμx Vx
∂fπ(m,n) μx Vx(m,n)
(23)
,x (i,j)∈H x
= μx e ( )2
∂Vx(m,n) ∑ (i,j)
eμx Vx
(i,j)∈H x

(w,z)
∂fπ(m,n) eμx Vx
(24)
,x (m,n)
= − μx eμx Vx ( )2
∂Vx(w,z) ∑ (i,j)
eμx Vx
(i,j)∈H x

where (w, z) ∈ H x , (m, n) ∈ H x and (w, z) ∕ = (m, n). The traffic-related functions f t are piece-wise functions consisting of a linear
function and a power function. ∂f t /∂τ can be obtained using the derivative rule for complex functions. For ∂f/∂(Fo , p), with the
exception of f u that is linearly related with Fo and p, all other functions in f have zero partial derivatives with regard to Fo and p.
In the second-best scenario, there exists a nonlinear constraint (21) that can make the gradient projection method proposed above
ineffective. To handle this issue, we introduce a Lagrange multiplier λ⩾0 to transfer the constraint into a term in the objective function.
Specifically, the altered objective function can be written as:
{[ ] }
∑ Hk ∑ Hk ( ) [ ]
max Sk + λ Fo,k qo,k h − Nr,k pk n0,k − (1 + ρ) Ns (g + z) + Cf (25)
Fo ,p,τ
k∈℘
h k∈℘
h

Since the Lagrange multiplier λ is unknown prior to solving the problem, in the solution process we must adjust its value accordingly in
order to obtain a maximum objective function value while ensuring the feasibility of the constraint. In most cases, the optimal solution
∑ ( ) [ ]
is identified at the points where k∈℘ Hhk Fo,k qo,k h − Nr,k pk n0,k = (1 +ρ) Ns (g + z) + Cf .

5. Numerical examples

We report the equilibrium results under different scenarios in this section.

5.1. Test settings

We consider a city with the area of approximately R = 400 km2 . The population density of this city is 5, 000 persons per square
kilometer, and 30% of the residents of the city own MVs or AVs. We set the AV ownership as approximately 3%; i.e., the number of AVs
owned by citizens is equal to 3% of the population. The time of a day can be divided into two periods according to travel demand: peak
and off-peak hours that last 4 h and 20 h, respectively. In the peak-hours, the trip generation per hour equals 15% of total population;

10
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Table 3
Population in different periods for all six classes.
Population dn dr da da′ db db′
Peak hours 210,000 81,000 4,500 25,500 4,500 25,500
Off-peak hours 70,000 27,000 1,500 28,500 1,500 28,500

Fig. 5. Approximation of customers’ matching time function.

Table 4
Optimal pricing and social impact.
Monopoly Second-besta First-bestb

Peak hours F*o $29.4/trip $27.4/trip $26.7/trip (− 9%)

p* $13.3/trip $16.5/trip $18.3/trip (+38%)
Off-peak hours F*o $24.3/trip $19.3/trip $16.2/trip (− 33%)
p* $9.4/trip $11.2/trip $12.0/trip (+28%)
Crowdsourcing platform’s profit $1.53 × 107 /day $1.20 × 107 /day $7.87 × 106 /day (− 49%)
( ) ( ) ( )
Social welfare $ − 6.54 × 107 /day $ − 6.29 × 107 /day $ − 6.26 × 107 /day (+4%)

a
In the second-best scenario, the basic rate for accumulating revenue is set as ρ = 20, and we obtain λ* = 0.16 at the optimal.
b
The percentage in the parentheses represents the comparison between the first-best scenario and the monopoly scenario.

in the off-peak hours, the trip generation per hour equals 5% of total population. The model input data for the population of six types of
citizens is shown in Table 3.
AV technological maturity index α is set to 0.7. The decision period is h = 1 h. An AV purchased by the crowdsourcing platform
costs $100, 000 with 10 years’ service life and its maintenance fee is $5, 000 per year. The pre-purchased number of AVs by the
crowdsourcing platform is Ns = 0; i.e., all supplied AVs are crowdsourced from the AV owners. The cost incurred by the crowdsourcing
platform is set to $6 × 105 per day. The additional cost of AV owners is set as m = $20. Public transit is captured by the average travel
time tn = 1 h and the average fare Fn = $6. For simplicity, we consider μx = μ = 0.2 for all classes of citizens. Traveler’s VOTs are βA =
$20/h, βP = $30/h, βM = $40/h, γ = $30/h (Vignon et al., 2021).
Setting κ = 1, the average trip time with free-flow speed by private vehicle is 30 min. Without loss of generality, the trip time
( )
function considering congestion Ts ( ⋅ ) is set as a quadratic function, i.e., Ts ( ⋅ ) = a + b[qm + α(qa − qo ) + αqo 1 + θ(wt qo )− 1/2 ]2 , and
the approximation takes similar form. The basic trip time term is a = 0.5 h, and the congestion related parameter is b = 2.67 ×
10− 12 h, indicating that traffic congestion can lead to a delay of at most 1 h.
Specifically, we provide the actual form of the matching function Eq. (1). Assume the platform conducts a match every period with
duration δ, and in each match the distribution of idle AVs will be refreshed; and in each match the idle AVs are generated following a
spatial Poisson’s process with intensity λ = wt qo /R, where wt qo is the total number of idle AVs and R is the area of the region; then the
customers’ average searching time is:
δ δ
wc = = πλ
Prob {Idle AVs exist in the surrounding unit-time circular zone} 1 − e−

and by applying Taylor’s approximation we can yield the following approximated matching function (the comparison between the

11
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Fig. 6. Trip number per day of different modes.

Fig. 7. Influence of people’s sensitivity to utilities. (a) Average trip fare. (b) Payment per ride. (c) Crowdsourcing platform’s profits per day. (d)
Social welfare per day.

12
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Fig. 8. Influence of people’s sensitivity to utilities on mode split. (a) Private autonomous driving. (b) Private manual driving. (c) Crowdsourcing
services. (d) Public transit.

curves before and after approximation is shown in Fig. 5):

δR δR 1
wc ≈ +δ≈ (wt qo )−
πwt qo π

where the term +δ is ignored because δ is generally with a duration of only a few seconds. Comparing it with Eq. (1), we know that the
parameters are chosen as A = δR/π and α1 = 1. The above formula will be used throughout the numerical experiments. We assume the
duration of each refreshing period δ = 0.0014 h, i.e. 5 sec.

5.2. Results

The first-best scenario reduces people’s travel cost and increases the AV owners’ revenue significantly (Table 4), illustrating that
the first-best scenario encourages AV owners to share the private AVs with others and encourages citizens to travel by on-demand ride
service. Payment in the first-best scenario is high: considering an AV owner renting out their private AV in off-peak hours every day, it
requires approximately 1.15 years for the owner to recoup the purchase cost of the vehicle. This suggests that AV crowdsourcing may
be a worthwhile investment for some citizens; in turn, this will promote AV acceptance by the public. Different from the current
manual driving on-demand ride providers, we find the AV crowdsourcing platform can still greatly benefit from the first-best scenarios
(Table 4). However, the results show that the crowdsourcing platform must cut approximately 48.7% of its profits to achieve the first-
best scenario. To balance the crowdsourcing platform’s profits and the overall social welfare, we explore the second-best scenario and
find a trade-off pricing strategy in which the social welfare is close to that of the first-best scenario while the crowdsourcing platform
only loses approximately 21.6% of its monopoly profits. By applying this second-best pricing strategy, it is easier to obtain the

13
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Fig. 9. Influence of population density. (a) Average trip fare. (b) Payment per ride. (c) Crowdsourcing platform’s profits per day. (d) Social welfare
per day.

agreement of the crowdsourcing platform to achieve a relatively high social welfare.

To examine the mode split in the equilibrium, one notable finding is that in the first-best scenario, the number of AV crowdsourcing
services is increased by roughly 40.1% during peak hours and by 70.5% in the off-peak hours compared with the monopoly scenario
(Fig. 6) as a result of higher payment and lower fare. This illustrates that AV crowdsourcing service has an overall beneficial effect on
the society, outweighing the congestion externality caused by deadheading. Additionally, in the first-best scenario, roughly 91.3% and
61.8% of the AV owners in the peak and off-peak hours who have no travel needs choose to rent AVs to the crowdsourcing platform,
respectively; roughly 31.8% and 12.6%of the AV owners in peak and off-peak hours who have travel needs choose to rent AVs to the
crowdsourcing platform, respectively. We find the payment so attractive that many people choose to rent their AVs to the crowd­
sourcing platform rather than drive them themselves. Other findings include that AV crowdsourcing plays a more important role in the
off-peak hours, i.e., serving up to 66.2% of the total trips for the first-best scenario and 38.8% for the monopoly scenario; while in peak
hours, public transit is the most utilized transport mode. It is important to note that in the first-best scenario, people will shift to other
modes from public transit due to its high time cost.

5.3. Sensitivity analysis

In this section, we examine the influence of a series of model parameters on the equilibrium results, including people’s recognition
of utility uncertainties (i.e., the scale parameter in the logit model), population density, AV market penetration rate, AV technology
maturity and the additional cost for sharing one’s AVs with the society (i.e., m).

14
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Fig. 10. Influence of the autonomous vehicle market penetration rate. (a) Average trip fare. (b) Payment per ride. (c) Crowdsourcing platform’s
profits per day. (d) Social welfare per day.

5.3.1. Scale parameter in the logit model

Without loss of generality, we assume that all classes of citizens are characterized by the same degree of uncertainty with regard to
choices, i.e., μx = μ (∀x ∈ {1, 2, …, 6}). As shown in Figs. 7(a) and (b), when people are more sensitive to utilities, it will be harder for
the crowdsourcing platform to attract citizens, and thus the trip fare decreases, and the payment increases. It is observed from Figs. 7(c)
and (d) that the crowdsourcing platform’s profit and the social welfare decline overall. Trips by public transit decrease, implying that
the shortcoming of its high time cost is amplified with increasing μ, and trips by other three transport modes increase (Fig. 8). One
noticeable deviation is that the payment during the peak hours for the first-best scenarios decreases (Fig. 7(b)), leading to slightly
fewer crowdsourcing trips and slightly greater platform profit. This indicates that more crowdsourcing trips during the peak hours are
discouraged with increasing μ, most likely due to the amplification of the congestion externality. Another deviation is the minor in­
crease of the platform profits for the monopoly scenario for μ⩾0.3 (Fig. 7(c)) due to an increasing number of crowdsourcing trips with a
small amount of price changes. In short, these experiments show that parameter μ has great effect on price setting and the equilibrium
state, and thus it is important for the crowdsourcing platform to investigate this parameter prior to setting the optimal prices in order to
maximize its profits; μ also influences the urban planner’s decision whether to regulate the AV crowdsourcing market, because the first-
best scenarios present a more limited improvement with a higher μ, i.e., 11.9% for μ = 0.1 and only 1.9% for μ = 0.5; thus, regulation
will have a limited effect when people are highly sensitive to utilities. We assume μ = 0.2 in the following experiments, i.e.,
approximately 73.1% of people make a choice with $5 higher utility.

5.3.2. Population density of the city

Figs. 9(a) and (b) show that the population density generally has little impact on prices with exceptions for a few cases, namely the
payment during peak-hours declines for the first-best scenarios. From this and the observation that average trip fare increases slightly,

15
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Fig. 11. Influence of autonomous vehicle technology maturity. (a) Average trip fare. (b) Payment per ride. (c) Crowdsourcing platform’s profits per
day. (d) Social welfare per day.

we know that the first-best scenario discourages people to rent AVs or take crowdsourcing service and thus alleviate the congestion
externality that grows with the population density. These imply that main impact of the population density on the equilibrium is the
congestion externality. Fig. 9(c) shows that the crowdsourcing platform’s profits increase, indicating that the crowdsourcing platform
must expect to be launched in more densely populated areas, such as a metropolis. Fig. 9(d) shows that individual welfare first in­
creases slightly and then decreases, most likely because a low population density induces insufficient number of crowdsourcing AVs,
increasing the individual travel cost. The highest individual welfare is obtained when the population density is approximately 2500
persons per square kilometers.

5.3.3. AV market penetration rate

We denote the autonomous vehicle market penetration rate as the ratio of the population who own AVs to the population who own
AVs or MVs. A greater market penetration rate means a larger pool for AV crowdsourcing. As shown in Figs. 10(a) and (b), fare and
payment decline due to a more sufficient supply and the rate of decline decreases with growing AV market penetration rate. As
observed from Fig. 10(c), the crowdsourcing platform’s profits increase with AV market penetration rate overall, except when the
penetration rate is more than 0.1, the crowdsourcing platform’s profits would not increase any more for the first-best scenario. This
demonstrates that the crowdsourcing platform would support AV popularization in most cases, but under regulation for the first-best
scenario, the increase of market penetration rate beyond 0.1 would not benefit the crowdsourcing platform anymore. Fig. 10(d)
demonstrates that social welfare increases and this increasing tendency is significant when AV penetration rate is small, suggesting that
the urban planner should have a supportive attitude toward AV popularization especially at its very beginning of launching process.

16
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Fig. 12. Influence of the number of autonomous vehicle purchased. (a) Average trip fare. (b) Payment per ride. (c) Crowdsourcing platform’s profits
per day. (d) Social welfare per day.

5.3.4. AV technology maturity

The maturity of AV technology (1 − α) represents the degree of improvement on road capacity, where α represents the “relative
occupation” of AVs compared to MVs. As Figs. 11(a) and (b) show, AV technology maturity’s impact on prices is ignorable except that
the platform raises payment in peak hours for the first-best scenarios to encourage more crowdsourcing trips, which results from higher
road capacity. Fig. 11(c) shows the crowdsourcing platform’s profits are slightly improved in the monopoly scenario, but are dete­
riorated under regulation. Fig. 11(d) points out that the advances of AV technology help improve the social welfare. However, the
maturity of AV technology has limited impact on price setting and the equilibrium state.

5.3.5. Number of AVs pre-purchased by the crowdsourcing platform

As shown by Figs. 12(a) and (b), the optimal fare and payment remain almost unchanged. The results presented in Fig. 12(c)
indicate that the crowdsourcing platform should purchase as many AVs as possible to obtain a higher daily profit. An examination of
Fig. 12(d) shows that social welfare improves with greater number of purchased AVs. Nevertheless, the impact of number of vehicles
pre-purchased is much smaller than the impact of AV market penetration rate on the equilibrium state.

5.3.6. Owners’ additional costs for sharing AVs

Additional costs for sharing private AVs mainly include energy consumption, vehicle insurance, vehicle depreciation and psy­
chological costs. Figs. 13(a) and (b) present the rise of fare and payment with increasing additional cost; the crowdsourcing platform is
more difficult to collect AVs, and thus must provide greater compensation to the AV owners and its supply would decline. One
exception is that payment during peak hours for the first-best scenario first decreases and then increases; this is because when
additional cost is low, the payment in peak hours decreases to discourage crowdsourcing trips for less congestion; and when additional

17
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Fig. 13. Influence of the additional cost to rent out private autonomous vehicles. (a) Average trip fare. (b) Payment per ride. (c) Crowdsourcing
platform’s profits per day. (d) Social welfare per day. (e) Number of crowdsourcing vehicles.

18
X. Wang et al. Transportation Research Part C 132 (2021) 103362

cost is high, the payment increases to compensate the AV owners’ expenses. Figs. 13(c) and (d) show that as additional cost becomes
larger, there is less difference between the first-best scenario and the monopoly case for both social welfare and platform’s profits.
These tests suggest that it is important for the crowdsourcing platform to investigate the AV owners’ additional cost prior to setting
prices; it may also be important for the urban planner to determine people’s additional cost m when deciding whether to impose
regulations because less additional cost means that regulation has more beneficial effect on the society, as shown in Fig. 13(d) (Fig. 3).

6. Conclusions

This paper investigates the AV crowdsourcing market that may emerge in the future. We first proposed an equilibrium model with
multiple transport modes: AV on-demand ride service, private AV, private MV and public transit, and proved the existence of the
equilibrium solution. Then, several scenarios, namely, the monopoly scenario, the first-best case and the trade-off second-best case
were explored.
We use numerical examples to illustrate some insightful findings concerning the AV crowdsourcing market, including the com­
parison of optimal pricing for trip fare and payment during the peak and off-peak hours, the crowdsourcing platform’s profits, the
social welfare and transport modes in each scenario. Sensitivity analyses of the optimal pricing and the equilibrium state with regard to
different parameters in the model are presented in a thorough manner. The main findings from the numerical examples are sum­
marized below.

1. AV crowdsourcing service is beneficial for the society overall. This provides guidance for urban planners to have a supportive
attitude toward AV development and popularization.
2. Renting out private AVs to the crowdsourcing platform provides high return, encouraging people to rent their cars instead of
driving them themselves. This verifies a new opportunity for adventurous investors such as the wealthy to invest into AVs.
3. Regulation may be easier for the crowdsourcing platform to accept in some second-best scenario where the social welfare is close to
that of the first-best scenario. Regulation is particularly necessary when people are not sensitive to utilities, or the AV owners are
more inclined to rent out their AVs to gain profits.
4. It is important for the crowdsourcing platform to know about the people’s sensitivity to utilities and AV owners’ additional cost
prior to setting prices, because these two factors significantly affect the optimal pricing. Additionally, the crowdsourcing platform
should expect to be launched in areas with high population density and should purchase as many AVs in advance as possible to
obtain higher profits. Nevertheless, even without pre-purchased vehicles, the platform is highly likely to gain profits under the
premise that the market penetration rate of AVs is sufficiently high (above 5%).

This paper could be further advanced by considering the competition between MV and AV on-demand ride services in the mobility
market, as is closer to the reality before the society is fully accustomed to AV technologies. Additionally, as shared autonomous ve­
hicles will probably be operated together with public transit, the joint planning and operation of two modes to form an integrated
mobility-as-a-service (MaaS) system should be explored in the future. In terms of wider implications, our results suggest that with the
presence of AV crowdsourcing business, even when most AVs are owned privately, these AVs will be mostly used for serving other
customers in the society, which motivates the discussion of AV ownership in the future, including where it will converge and how it
will evolve. On the other hand, for these AV owners, AVs will not be purchased merely for personal usage, but also for investment
(buying vehicles to earn money from day to day), and the capital value of AVs merits careful investigation.

Acknowledgements

The research is supported in part by Tsinghua-Daimler Joint Research Center for Sustainable Transportation.

Appendix A. Nomenclature

Variables
Fo Average trip fare for AV on-demand ride service
p Payment per ride to an AV owner
N Total number of usable AVs for the crowdsourcing platform per hour
Nr Number of rented AVs per hour
qm Travel demand for manual driving per hour
qa Travel demand for autonomous driving (including privately used and crowdsourcing ones) per hour
qo Travel demand for on-demand ride services per hour
qp Travel demand for public transit per hour
tr Average trip time from picking up a customer to arriving at the destination by a car
̂t r The image of tr in the mapping Φ
n0
̂ The image of n0 in the mapping Φ
̂t p The image of tp in the mapping Φ
̂c
w The image of wc in the mapping Φ
(continued on next page)

19
X. Wang et al. Transportation Research Part C 132 (2021) 103362

(continued )
tp Average pick-up time for a crowdsourcing AV
wc On-demand ride customers’ average searching time for an available AV
wt AVs’ average searching time for on-demand ride customers
n0 Expected number of on-demand rides served by an AV per unit time with duration h
V(i,j)
x
(dis) Utility for class x citizens to choose (i, j)
Πx
(i,j) Probability for class x citizens to choose (i, j)
S Social welfare in unit time with duration h in a fixed-level demand period
λ Lagrange multiplier
Set
Hx All possible choices for class x citizens
𝒟 Region of a hypercube with edges of [0, M1 ]
𝒯x Choices of travel modes by class x citizens
ℛx Choices of whether to rent out AVs by class x citizens
𝒯 ix Choices of whether to rent out AVs after deciding travel mode i by class x citizens
ℛjx Choices of travel modes after making rental decision j by class x citizens
℘ All heterogeneous periods in one day with regard to demand level
τ Collective equilibrium variables
f Deformation of all equilibrium equations
fu Deformation of the utility equations
fπ Deformation of the probability equations
fn Deformation of the supply/demand equations
ft Deformation of the traffic-related equations
Functions
Ts ( ⋅ ) Trip time function considering traffic congestion
ξ1 (wt ) Approximate continuous function for (wt )− α1
ξ2 (wt ) Approximate continuous function for (wt )− 1/2
Wc ( ⋅ ) Approximate continuous function for representing wc
Tp ( ⋅ ) Approximate continuous function for representing tp
Tr ( ⋅ ) Approximate continuous function for representing tr
Mappings
( )
Φ A mapping derived from the system (4)-(15) with original images tr , n0 , tp , wc
Parameters
h Duration of a decision period, also known as unit time
H Duration of a period with a certain level of demand
δ Duration of a period when the platform refreshes search results
k A parameter depending on the network topology
κ A constant parameter depending on the average travel distance
θ A city-specific constant which is a ratio of k over κ
α A parameter representing the “relative occupation” of AVs compared to MVs
R City size (area)
v Cruising speed of cars
v0 Free-flow cruising speed of cars
dn Number of people with travel needs and own no private vehicles per unit time with duration h
dr Number of people with travel needs and own MVs per unit time with duration h
da Number of people with travel needs and own AVs per unit time with duration h
da′ Number of people with no travel needs and own AVs per unit time with duration h
db Number of people with travel needs and own MVs and AVs per unit time with duration h
db′ Number of people with no travel needs and own MVs and AVs per unit time with duration h
m A constant indicating the expected additional cost to rent out AVs for AV owners
βA Perception of on-board time for taking AVs
βP Perception of on-board time for taking public transit
βM Perception of on-board time for driving MVs
γ Perception of waiting time for taking crowdsourcing AVs
Fn Average fare of taking public transit
tn Average travel time of taking public transit
μx Scale parameter for class x citizens in logit model
μrx Scale parameter for class x citizens on rental choices in nested logit model
μtx Scale parameter for class x citizens on travel choices in nested logit model
Ns Number of pre-purchased AVs by the crowdsourcing platform
∊ A constant truncating the original function to obtain its approximation form
g Unit-day amortized purchase cost of an AV
z Unit-day amortized maintenance cost of an AV
Cf Unit-day operational cost of the crowdsourcing platform
ρ Basic rate for the AV crowdsourcing platform to accumulate profit
a Basic trip time parameter by car
b Congestion related parameter
dmv Total number of MVs owned by the citizens
dav Total number of AVs owned by the citizens
M1 A large positive number

20
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Appendix B. Discussion about Logit Model in Section 3.2

Let μrx denote the scale parameter of logit model with regard to rental choices; μtx denote the scale parameter of logit model with
regard to transport mode choices. If μrx > μtx , citizens first choose travel modes and then choose whether to rent their AVs to the
crowdsourcing platform. Let 𝒯 x denote the choice set of travel modes by class x citizens and 𝒯 ix denote the rental choice set after they
have chosen travel mode i. Therefore, the probability of class x citizens choosing (i, j) ∈ H x can be obtained using the nested logit
model:
(∑ (i,j)
)μtx /μrx
μrx Vx
j∈𝒯 i x e
r (i,j)
eμx Vx
Π(i,j)
x =∑ (∑ )μtx /μrx ∀(i, j) ∈ H x , x ∈ {1, 2, …, 6} (B-1)
μrx Vx(i,j) ∑
j∈𝒯 i x e μrx Vx(i,j)
i∈𝒯 x j∈𝒯 i x e

If μrx < μtx , citizens first choose whether to rent their AVs to the crowdsourcing platform and then choose travel modes. Let ℛx
denote the rental choice set by class x citizens and ℛjx denote the choice set of travel modes after they have made rental choice j.
Therefore, the probability of class x citizens choosing (i, j) ∈ H x can be obtained using the nested logit model:
(∑ )μrx /μtx
μtx Vx(i,j)
i∈ℛj x e
t (i,j)
eμx Vx
Π(i,j)
x =∑ (∑ )μrx /μtx ∀(i, j) ∈ H x , x ∈ {1, 2, …, 6} (B-2)
μtx Vx(i,j) ∑
i∈ℛj x e μtx Vx(i,j)
j∈ℛx i∈ℛj x e

If μrx = μtx = μx , people decide simultaneously whether to rent their AVs and the type of travel mode they take. Therefore, the
probability of class x citizens choosing (i, j) ∈ H x is reduced to the logit model:
(i,j)
eμx Vx
Π(i,j)
x =∑ ∀(i, j) ∈ H x , x ∈ {1, 2, …, 6} (B-3)
μx Vx(i,j)
(i,j)∈H x e

References

An, Y., Li, M., Lin, X., He, F., Yang, H., 2020. Space-time routing in dedicated automated vehicle zones. Transp. Res. Part C: Emerg. Technol. 120, 102777.
Arnott, R., 1996. Taxi travel should be subsidized. J. Urban Econ. 40 (3), 316–333.
Bayus, B., 2010. Crowdsourcing and individual creativity over time: The detrimental effects of past success. SSRN Electron. J.
Bayus, B.L., 2013. Crowdsourcing new product ideas over time: An analysis of the dell ideastorm community. Manage. Sci. 59 (1), 226–244.
Ben-Akiva, M.E., Lerman, S.R., 1985. Discrete Choice Analysis: Theory and Application to Travel Demand. ↱ The MIT Press.
Bian, Y., Zheng, Y., Ren, W., Li, S.E., Wang, J., Li, K., 2019. Reducing time headway for platooning of connected vehicles via v2v communication. Transp. Res. Part C:
Emerg. Technol. 102, 87–105.
Brabham, D.C., 2008. Crowdsourcing as a model for problem solving: An introduction and cases. Convergence 14 (1), 75–90.
Buchholz, N., 2021. Spatial equilibrium, search frictions and efficient regulation in the taxi industry. In: Young IO Economists Conference.
Castillo, J.C., Knoepfle, D.T., Weyl, E.G., 2018. Surge pricing solves the wild goose chase. Soc. Sci. Res. Netw.
Chen, X.-D., Li, M., Lin, X., Yin, Y., He, F., 2021. Rhythmic control of automated traffic – part i: Concept and properties at isolated intersections. Transp. Sci.
forthcoming.
Chen, X., Lin, X., He, F., Li, M., 2020. Modeling and control of automated vehicle access on dedicated bus rapid transit lanes. Transp. Res. Part C: Emerg. Technol. 120.
Chen, Z., Lin, X., Yin, Y., Li, M., 2020. Path controlling of automated vehicles for system optimum on transportation networks with heterogeneous traffic stream.
Transp. Res. Part C: Emerg. Technol. 110, 312–329.
Daganzo, C.F., Ouyang, Y., 2019. A general model of demand-responsive transportation services: From taxi to ridesharing to dial-a-ride. Transp. Res. Part B: Methodol.
126, 213–224.
Estellés-Arolas, E., González-Ladrón-De-Guevara, F., 2012. Towards an integrated crowdsourcing definition. J. Inf. Sci. 38 (2), 189–200.
Fagnant, D.J., Kockelman, K., 2015. Preparing a nation for autonomous vehicles: opportunities, barriers and policy recommendations. Transp. Res. Part A: Policy
Pract. 77, 167–181.
Glaeser, E.L., Hillis, A., Kominers, S.D., Luca, M., 2016. Crowdsourcing city government: Using tournaments to improve inspection accuracy. Am. Econ. Rev. 106 (5),
114–118.
Gong, S., Du, L., 2018. Cooperative platoon control for a mixed traffic flow including human drive vehicles and connected and autonomous vehicles. Transp. Res. Part
B: Methodol. 116, 25–61.
Gray, J., 2015. Implicit Functions. Springer International Publishing, Cham, pp. 259–269.
Hasan, M., Mohan, S., Shimizu, T., Lu, H., 2020. Securing Vehicle-to-Everything (V2X) communication platforms. IEEE Trans. Intell. Vehicles 5, 693–713.
He, F., Wang, X., Tang, X., 2018. Pricing and penalty/compensation strategies of a taxi-hailing platform. Transp. Res. Part C Emerg. Technol. 86, 263–279.
Kang, J., Kuznetsova, P., Luca, M., Yejin, C., 2013. Where not to eat? Improving public policy by predicting hygiene inspections using online reviews. In: Proceedings
of the Conference on Empirical Methods in Natural Language Processing.
Ke, J., Yang, H., Li, X., Wang, H., Ye, J., 2020. Pricing and equilibrium in on-demand ride-pooling markets. Transp. Res. Part B: Methodol. 139, 411–431.
Ke, J., Yang, H., Zheng, Z., 2020. On ride-pooling and traffic congestion. Transp. Res. Part B: Methodol. 142, 213–231.
Lin, X., He, F., Li, M., Liu, Y., 2020. On the existence of system-optimum–inducing tolling functions for transportation networks. Transp. Sci. 54, 1535–1554.
Lin, X., Li, M., He, F., 2018. Nonlinear pricing in linear cities with elastic demands. Transp. Res. Part C: Emerg. Technol. 95, 616–635.
Lin, X., Li, M., Shen, Z., Yin, Y., He, F., 2021. Rhythmic control of automated traffic – part ii: Grid network rhythm and online routing. Transp. Sci. forthcoming.
Lin, X., Wang, C., Wang, K., Li, M., Yu, X., 2021. Trajectory planning for unmanned aerial vehicles in complicated urban environments: A control network approach.
Transp. Res. Part C - Emerg. Technol. 128.
Litman, T., Transportation Research Board, 2015. Autonomous vehicle implementation predictions: Implications for transport planning. In: Transportation Research
Board 94th Annual Meeting.

21
X. Wang et al. Transportation Research Part C 132 (2021) 103362

Michael, J.B., Godbole, D.N., Lygeros, J., Sengupta, R., 1998. Capacity analysis of traffic flow over a single-lane automated highway system. J. Intell. Transp. Syst. 4
(1–2), 49–80.
Narayanan, S., Chaniotakis, E., Antoniou, C., 2020. Shared autonomous vehicle services: A comprehensive review. Transp. Res. Part C: Emerg. Technol. 111, 255–293.
Poetz, M.K., Schreier, M., 2012. The value of crowdsourcing: Can users really compete with professionals in generating new product ideas? J. Prod. Innov. Manage. 29
(2), 245–256.
Stocker, A., Shaheen, S., 2018. Shared automated mobility: Early exploration and potential impacts. In: Meyer, G., Beiker, S. (Eds.), Road Vehicle Automation 4.
Springer International Publishing, Cham, pp. 125–139.
Sun, J., Zheng, Z., Sun, J., 2020. The relationship between car following string instability and traffic oscillations in finite-sized platoons and its use in easing
congestion via connected and automated vehicles with idm based controller. Transp. Res. Part B: Methodol. 142, 58–83.
Urata, J., Xu, Z., Ke, J., Yin, Y., Wu, G., Yang, H., Ye, J., 2021. Learning ride-sourcing drivers’ customer-searching behavior: A dynamic discrete choice approach.
Transp. Res. Part C: Emerg. Technol. 130, 103293.
Vander Schee, B., 2009. Crowdsourcing: Why the power of the crowd is driving the future of business. J. Consum. Market. 26, 305–306.
Vignon, D.A., Yin, Y., Ke, J., 2021. Regulating ridesourcing services with product differentiation and congestion externality. Transp. Res. Part C: Emerg. Technol. 127,
103088.
Vosooghi, R., Puchinger, J., Jankovic, M., Vouillon, A., 2019. Shared autonomous vehicle simulation and service design. Transp. Res. Part C: Emerg. Technol. 107,
15–33.
Wang, C., Gong, S., Zhou, A., Li, T., Peeta, S., 2020. Cooperative adaptive cruise control for connected autonomous vehicles by factoring communication-related
constraints. Transp. Res. Part C: Emerg. Technol. 113, 124–145.
Wang, K., Song, M., Li, M., 2021. Cooperative multi-UAV conflict avoidance planning in a complex urban environment. Sustainability 13.
Wu, J., Ahn, S., Zhou, Y., Liu, P., Qu, X., 2021. The cooperative sorting strategy for connected and automated vehicle platoons. Transp. Res. Part C: Emerg. Technol.
123, 102986.
Xu, B., Li, S.E., Bian, Y., Li, S., Ban, X.J., Wang, J., Li, K., 2018. Distributed conflict-free cooperation for multiple connected vehicles at unsignalized intersections.
Transp. Res. Part C: Emerg. Technol. 93, 322–334.
Xu, Z., Yin, Y., Ye, J., 2020. On the supply curve of ride-hailing systems. Transp. Res. Part B: Methodol. 132, 29–43.
Yang, D., Zheng, S., Wen, C., Jin, P.J., Ran, B., 2018. A dynamic lane-changing trajectory planning model for automated vehicles. Transp. Res. Part C: Emerg. Technol.
95, 228–247.
Yang, H., Wong, S., 1998. A network model of urban taxi services. Transp. Res. Part B: Methodol. 32 (4), 235–246.
Yang, T., Yang, H., Wong, S.C., 2014. Taxi services with search frictions and congestion externalities. J. Adv. Transp. 48 (6), 575–587.
Yu, C., Feng, Y., Liu, H.X., Ma, W., Yang, X., 2018. Integrated optimization of traffic signals and vehicle trajectories at isolated urban intersections. Transp. Res. Part B:
Methodol. 112, 89–112.
Yu, J.J, Tang, C.S., Shen, Z.-J.M., Chen, X., 2017. Should on-demand ride services be regulated? An analysis of chinese government policies. SSRN Electron. J.
Zervas, G., Proserpio, D., Byers, J.W., 2017. The rise of the sharing economy: Estimating the impact of airbnb on the hotel industry. J. Mark. Res. 54 (5), 687–705.
Zha, L., Yin, Y., Yang, H., 2016. Economic analysis of ride-sourcing markets. Transp. Res. Part C: Emerg. Technol. 71, 249–266.
Zhang, K., Chen, H., Yao, S., Xu, L., Ge, J., Liu, X., Nie, M., 2019. An efficiency paradox of uberization. Soc. Sci. Res. Netw.
Zhang, Z., Li, M., Lin, X., Wang, Y., He, F., 2019. Multistep speed prediction on traffic networks: A deep learning approach considering spatio-temporal dependencies.
Transp. Res. Part C: Emerg. Technol. 105, 297–322.
Zhang, K., Nie, M., 2019. To pool or not to pool: Equilibrium, pricing and regulation. Soc. Sci. Res. Netw.
Zhao, Y., Zhu, Q., 2014. Evaluation on crowdsourcing research: Current status and future direction. Inform. Syst. Front. 16 (3), 417–434.
Zmud, J., Williams, T., Outwater, M., Bradley, M., Kalra, N., Row, S., 2018. Updating regional transportation planning and modeling tools to address impacts of
connected and automated vehicles, volume 1: Executive summary. Technical Report.
Zou, H., 2008. Chapter 7 - fixed point theory and elliptic boundary value problems. In: Chipot, M. (Ed.), Handbook of Differential Equations, volume 6 of Handbook of
Differential Equations: Stationary Partial Differential Equations. North-Holland, pp. 503–583.

22