1

CONGESTION PRICING IN A WORLD OF SELF-DRIVING VEHICLES: AN ANALYSIS OF 1

DIFFERENT STRATEGIES IN ALTERNATIVE FUTURE SCENARIOS 2

3

Michele D. Simonia, Kara M. Kockelmanb, Krishna M. Gurumurthyc, Joschka Bischoffd 4

5

a Dept. of Civil, Architectural, and Environmental Engineering, University of Texas at Austin, 6 m.simoni@utexas.edu 7

b Dept. of Civil, Architectural, and Environmental Engineering, University of Texas at Austin, 8 kkockelm@mail.utexas.edu 9

c Dept. of Civil, Architectural, and Environmental Engineering, University of Texas at Austin, 10 gkmurthy10@utexas.edu 11

d Department of Transport Systems Planning and Transport Telematics, Technical University of Berlin, 12 bischoff@vsp.tu-berlin.de 13

14

ABSTRACT 15

The introduction of autonomous (self-driving) and shared autonomous vehicles (AVs and SAVs) will affect 16

travel destinations and distances, mode choices, vehicle-miles traveled, and congestion. Although some 17

congestion reduction may be achieved (thanks to fewer crashes and tighter headways, long-term), car-trip 18

frequencies and VMT are likely to rise significantly in most settings, compromising the benefits of 19

driverless vehicles. Congestion pricing (CP) and road tolls are key tools for moderating demand and 20

incentivizing more socially and environmentally optimal travel choices. 21

This work develops multiple CP and tolling scenarios and investigates their effects on Austin, Texas 22

network conditions and traveler welfare, using the agent-based simulation model MATSim. Results suggest 23

that, while all pricing strategies reduce congestion, their social welfare impacts differ in meaningful ways. 24

More complex and advanced strategies may considerably improve traffic conditions, but do not necessarily 25

improve traveler welfare. The possibility to refund users by reinvesting toll revenues as traveler budgets, 26

plays a salient role in the overall efficiency of each CP strategy. 27

INTRODUCTION 28

Recent advances in autonomous vehicles (AVs) are generating much discussion across academia, industry, 29

and news media. Considerable progress has been in made in AV technologies, thanks to investments by 30

technology companies and auto manufacturers (Muoio, 2017), along with support of public institutions 31

(Kang, 2016). 32

Since driverless passenger vehicles represent a new travel option, many passenger-trips made via existing, 33

traditional modes, like cars, public transit, and bikes, will be replaced by trips in AVs and SAVs. While AV 34

and SAV benefits may accrue in road safety and energy consumption (Fagnant and Kockelman, 2015), 35

network congestion impacts are less well understood and may be quite problematic (Litman, 2017; Wadud 36

et al., 2016). 37

On one hand, automated technologies can ultimately road network performance by reducing traffic crashes 38

(and the delays those entail) and eventually increasing traffic throughput by ensuring tighter headways 39

between vehicles and by making better use of intersections. However, AVs and SAVs will likely increase 40

the number and the distance of motorized trips by making driving “easier” (by removing the driving burden) 41

and car-travel more accessible (to persons with disabilities and those not owning cars, for example). 42

Congestion may dramatically worsen, and demand management options will become even more valuable 43

along congested corridors and in urban regions. 44

2

Charging drivers for the delays or congestion they cause (on those behind them, for example) is a well-1

known concept among economists, traffic engineers and transport professionals. Various congestion pricing 2

policies now exist in cities like Singapore, London (UK), Stockholm (Sweden), Milan (Italy), Gothenburg 3

(Sweden).Most are limited to rather simplistic cordon or area-based tolls that do not vary by congestion 4

level. Smartphones and/or connected vehicles offer cities, states and nations an opportunity to implement 5

more economically efficient and behaviorally effective strategies, thanks to advanced communication and 6

location capabilities and fast information sharing. 7

This study investigates the effects of different congestion pricing strategies in future scenarios characterized 8

by strong market penetration of AVs and SAVs. Schemes include a joint (time-plus-distance-based) charge 9

that varies with the Austin region’s overall network condition, and a time-varying link-based charge that 10

reflects marginal delay costs on following travelers at the link level. These schemes’ traffic and social 11

welfare impacts are investigated and compared to those of two much simpler but rather classic strategies: a 12

distance-based (flat) toll and a facility-based toll (for the most congested 2-4% links of the road network). 13

To reflect the technology’s uncertain development costs, capabilities and adoption rates, this work estimates 14

two distinctive technology-adoption scenarios: one with relatively high private AV reliance and the other 15

with high SAV uptake. 16

Use of congestion pricing in AV and SAV scenarios is relatively unexplored, with the exception of a few 17

theoretical studies (as described below). Here, scenario simulations use the multi-agent travel-choice model 18

MATSim (www.matsim.org). MATSim enables simulation of tens of thousands of individuals and self-19

driving vehicles with flexibility in departure time, route, activity engagement and mode (and optionally 20

destination). 21

The remainder of this paper offers the following: a brief discussion of AVs’ mobility impacts and their 22

implications for congestion pricing, a description of the agent-based modeling framework used here, 23

discussion of several future mobility scenarios and alternative congestion pricing strategies, analysis of the 24

mobility and system welfare effects of such strategies, conclusions and policy recommendations. 25

MOBILITY IMPACTS OF SELF-DRIVING VEHICLES AND CONGESTION PRICING 26

Researchers interest on AVs’ impacts on travel behavior and traffic conditions is currently strong, with 27

different studies now having investigated issues ranging from travel behavior studies to advanced traffic 28

signal control applications. Milakis et al. (2017) provide a relatively comprehensive literature review, and 29

the research community’s major findings are summarized below, followed by a discussion of congestion 30

pricing. 31

Travel Costs and Traveler Preferences 32

AV travel costs will differ from those of conventional vehicles due to several factors. Fixed costs will be 33

somewhat higher due to expensive hardware and software required (Litman, 2017). However, thanks to 34

lower user effort and safer driving, travel time and operating costs fall. Parking costs can also be a key 35

consideration for many travelers (Litman and Doherty, 2011), and “parking search” time can represent a 36

significant portion of travel time in very busy settings (Shoup, 2006). 37

To date, there is little consensus in the scientific community about the impacts of AVs on travelers’ value 38

of travel time (VOTT). While some argue that autonomous travel might not substantially affect travelers’ 39

overall time-cost perceptions (Cyganski et al., 2015; Lenz et al., 2016; Yap et al., 2016) , most assume 40

and/or estimate (using stated-preference data) a lower VOTT (see, e.g., van den Berg and Verhoef, 2016; 41

Lamotte et al., 2017; Zhao and Kockelman, 2017; Bansal and Kockelman, 2017). 42

SAV costs are likely to be under $1/mile, thanks to no driver wages (Fagnant and Kockelamn 2016; Loeb 43

and Kockelman 2018; Chen et al. 2016; Bosch et al., 2017). Most of the previous literature on SAVs has 44

focused on the effects of replacing trips and on the required fleet sizes for full the replacement of 45

conventional trips (Burns et al., 2013; Spieser et al., 2014; Fagnant et al., 2015; Chen et al., 2016; Bischoff 46

3

and Maciejewski, 2016). Only a few studies have focused on investigating their adoption from a travel 1

demand modeling perspective (Krueger et al., 2016; Haboucha et al., 2017; Scheltes and Correia, 2017; 2

Martinez and Viegas, 2017). 3

Changes of traffic conditions 4

Autonomous driving will affect traffic conditions in several ways. Thanks to reduced reaction times and 5

shorter following distances, road and intersection capacity may eventually increase (Dresner and Stone, 6

2008), thereby reducing delays. Cooperative technologies like vehicle-to-vehicle (V2V) and vehicle-to-7

infrastructure (V2I) communication can improve network performance (Tientrakool et al., 2011; Shladover 8

et al., 2012; Hoogendoorn et al., 2014). Interestingly, Talebpour and Mahmassani (2016) use 9

microsimulations to show how automation can play a larger role than connectivity in terms of capacity 10

impacts. Most of these improvements will likely occur only for large adoption rates of self-driving and 11

connected vehicles. 12

However, reduced driving burdens, lower travel cost, and improved transport access, will probably increase 13

car use and travel distances in the near term, delivering more congested traffic conditions long before CAV 14

technologies can resolve many capacity issues (Gucwa, 2014; Fagnant and Kockelman, 2015; Milakis et 15

al., 2017). Empty SAVs driving in between trips will may compound this effect, with actual increases 16

depending on each region’s policies, mode options, parking costs, and trip-making patterns, for example. 17

For Berlin, Germany, SAVs (without dynamic ride-sharing among strangers) added 13% VMT (Bischoff 18

and Maciejewski, 2016). Re-balancing vehicles during off-peak times may help to minimize the effect of 19

those trips (Winter et al., 2017; Hörl et al., 2017). However, the possibility of pooling multiple travelers on 20

SAVs with dynamic ride-sharing could reduce the overall VMT (Fagnant and Kockelman, 2018). 21

Congestion pricing for AVs 22

The idea of charging individual travelers for the marginal external congestion cost of their trips (associated 23

with the delays’ caused to other users) has a long tradition in the transportation economics and engineering 24

field (Pigou, 1924; Walters, 1961; Vickrey, 1963, 1969; Beckman, 1965). During the past fifty years, 25

several studies have investigated the topic of congestion pricing to identify optimal strategies and to add 26

more realistic features to their models based on travelers’ behavior and infrastructure performance (Small 27

and Verhoef, 2007; de Palma and Lindsey, 2011). The adoption of AVs and the emergence of mixed-flow 28

scenarios bring new theoretical and practical challenges for pricing. 29

From a theoretical perspective, AVs will affect the demand curve (volumes of road users). As discussed in 30

Section 2.1, the possibility to perform other activities while driving and the different level of comfort, will 31

likely affect AV users’ travel costs and perceived VOTT. As van den Berg and Verhoef (2016) highlight, 32

a decrease of VOTT would correspond to a reduction in queuing costs for AV users, ultimately increasing 33

their acceptance of increased travel time. Hence, from a marginal cost perspective, this phenomenon could 34

hurt conventional users, whose travel costs have not changed. 35

In addition to that, AVs will affect the supply curve (response of the transportation system to the demand). 36

As discussed in Section 2.2, AV implementation will likely yield improvements in network performance. 37

Lamotte et al. (2016) considers some of these aspects by investigating the problem of optimal allocation of 38

road infrastructure (between conventional and automated cars) and tolls with a bottleneck model. In case 39

of shared road infrastructure, congestion phenomena becomes more complicated. From a traffic flow 40

perspective, increasing levels of AVs would delay the onset of congestion (shift of Critical Density in the 41

Fundamental Diagram1), allowing additional demand without compromising traffic conditions. However, 42

larger portions of AVs will also accelerate the deterioration of traffic conditions in the congested traffic 43

1 The Fundamental Diagram relates a roadway’s traffic flow and density values (Lighthill and Whitman, 1956;

Richards, 1956).

4

regime (increase of congested wave speed). Hence, from a marginal cost perspective, higher shares of AVs 1

would be beneficial to all travelers up to a certain level of traffic demand (capacity). 2

From a practical standpoint, the high level of information and communication characterizing autonomous 3

driving might favor the introduction of more advanced tolling strategies. Ideally, tolls should reflect 4

changes in travel costs depending on category of road user, time of the day, real-time traffic conditions, trip 5

purpose, and presence of transportation alternatives like public transit (Vickrey, 1997; Arnott, 1998). 6

However, current road and congestion pricing strategies typically include: facility-based tolls (on bridges, 7

tunnel, highways) specific roads, tunnels and bridges; cordon-based tolls (charge applied when entering an 8

area, like in Stockholm); area-based tolls (charge applied when driving inside the area, like in London); and 9

distance-based tolls (like in heavy goods vehicles charges in Germany). 10

In the past, research in the field has been characterized by a clear distinction between “first-best pricing” 11

solutions, with strong analytical frameworks and the absence of constraints, and “second-best pricing” 12

solutions, characterized by sub-optimal, but more feasible mechanisms (Verhoef, 2002; de Palma et al., 13

2004; Zhang and Yang, 2004; May et al., 2008; Lawphongpanich and Yin, 2010). 14

The possibility of exchanging traffic and charges information in real-time to all the connected vehicles 15

would allow for more advanced pricing strategies that vary in time, space and fare more dynamically. In 16

this sense, AV technologies could almost bring second-best pricing systems to first-best pricing ones. 17

Driverless technologies can also facilitate congestion pricing in other ways. First, tolling systems could 18

become more feasible thanks to advanced communication technologies (wireless, GPS), cheaper than the 19

current tolling systems based on dedicated short-range communications (DSRC) and automated license 20

plate recognition since they do not need for additional road infrastructure. Second, the fact that “smart” 21

AVs can compute and communicate fares and routing options to travelers would help keeping pricing 22

schemes understandable and transparent. This might eventually increase public acceptability of congestion 23

pricing (Gu et al., 2018). 24

In Section 5.2, we present two congestion pricing schemes that leverage on the advanced communication 25

and computation capabilities of AV-SAVs to derive schemes closer to the concept of first-best pricing. To 26

the best of our knowledge, congestion pricing schemes involving both privately owned AVs and SAVs are 27

still relatively unexplored except for a few studies in specific simulation environments (agents’ route choice 28

under dynamic link tolls) (Sharon et al., 2017) or involving broader external costs for some specific modes 29

(limited to SAVs and conventional cars) (Kaddoura et al., 2018). In this study, we provide a comprehensive 30

analysis of different potential future scenarios (accounting for different market penetrations of AVs and 31

SAVs) and alternative congestion pricing schemes. 32

33

MODELING AVs AND SAVs WITH AN AGENT-BASED MODEL 34

In this section after providing a brief overview of the agent-based model MATSim, we present a description 35

of our modeling framework. We then focus on the modeling of AVs and SAVs. 36

General Framework of MATSim 37

MATSim simulates an entire daily plan of every single user and it considers endogenous mode choice, 38

departure time choice and route choice into a fully dynamic model. As opposed to models that use single 39

trips, this model allows for predictions on reactions to demand management strategies, such as tolls during 40

the span of a day, accounting for a higher level of realism. In fact, trips are typically linked to each other as 41

a part of a daily plan and not that meaningful just as stand-alone trips (Balmer et al., 2006). Activities often 42

have higher importance in the daily schedule than trips that simply represent connections among them. 43

Since MATSim represents traffic behavior at a highly disaggregated level by modeling individual agents 44

(with different socio-demographic characteristics), it is possible to investigate the effects of transport 45

5

policies on travel behavior and traffic more in depth than in traditional 4-steps models (Kickhöfer et al., 1

2011). The overall process (Figure 1) can be summarized in the following stages: 2

• Each agent independently develops a plan that expresses his/her preferences in terms of activities, trips 3

and their schedules during the day (Initial demand). 4

• The agents simultaneously perform all the plans in the mobility simulation’s (Mobsim’s) physical system. 5

Congestion phenomena are modeled using a queue model, which takes both the physical storage capacity 6

and the actual throughput (flow capacity) of a link into account. 7

• To compare different plans’ performances, each scored using a utility-type function (Scoring). 8

• Agents are able to remember their plans and improve them during the simulation by means of a learning 9

algorithm (Replanning). During the implementation the system iterates between plan generation and 10

traffic flow simulation. 11

• The cycle continues until the system has reached an equilibrium where no agent can improve anymore his 12

score (Analyses). 13

14

Figure 1: MATSim cycle (source: Horni et al., 2016) 15

The choice model generally adopted in MATSim is equivalent to the standard multinomial logit model. 16

Since the amount of plans in the memory of agents is limited, the worst performing one is replaced by a 17

new one at each iteration. Thanks to this feedback mechanism agents are able to improve their plans over 18

several iterations until the system reaches the “relaxed” state when agents cannot significantly improve 19

their plans and the outcome of the system becomes stable. This state is also referred as agent-based 20

stochastic user equilibrium (Nagel and Flotterod, 2009). 21

For further information about the simulation framework MATSim, see Horni et al. (2016). 22

Choice dimensions and parameters 23

Plans can be improved by changing the time of departure, varying the route and choosing different transport 24

mode through modules. Agents’ travel choices are modeled in MATSim through an iterative learning 25

mechanism based on a quantitative score, referred to as utility. For each iteration agents choose from an 26

existing set of daily plans according to a multinomial logit model. 27

Every daily plan is associated with a utility, accounting for a trip-related disutility and a performing activity 28

utility: 29

𝑉𝑝𝑙𝑎𝑛 = ∑(𝑉𝑎𝑐𝑡,𝑖 + 𝑉𝑡𝑟𝑖𝑝,𝑖)

𝑛

𝑖=1

(1)

where Vplan is the total utility of a daily plan; n is the total number of activities or trips; Vact,i is the utility for 30

performing activity i; and Vtrip,i is the utility of the trip to activity i. The first and the last activity are wrapped 31

around the day, and handled as one activity. Thus, the number of activities and trips is the same. The trip-32

related utility for a each mode is calculated as follows: 33

𝑉𝑞,𝑖 = 𝛽0,𝑞 + 𝛽𝑡,𝑞 ∙ 𝑡𝑖,𝑞 + 𝛽𝑐 ∙ 𝑐𝑖,𝑞 (2)

6

where 𝛽0,𝑞 corresponds to the alternative specific constant of mode q; 𝑡𝑖,𝑞 corresponds to the travel time of 1

leg i traveled with mode q; 𝛽𝑡,𝑞 corresponds to the marginal utility of traveling by mode q; 𝑐𝑖,𝑞 corresponds 2

to the monetary cost of leg i traveled by mode q; and 𝛽𝑐 corresponds to the marginal utility of monetary 3

cost. 4

To calculate the positive utility gained by performing an activity, a logarithmic form is applied (Charypar 5

and Nagel, 2005; Kickhofer et al., 2011): 6

𝑉𝑎𝑐𝑡,𝑖(𝑡𝑎𝑐𝑡,𝑖) = 𝛽𝑎𝑐𝑡 ∙ 𝑡𝑖∗ ∙ ln (

𝑡𝑎𝑐𝑡,𝑖

𝑡0,𝑖) (3)

where tact is the actual duration of performing an activity (when the activity is open), ti* is an activity’s 7

‘typical’ duration, and βact is the marginal utility of performing an activity at its typical duration. In the 8

equilibrium, all activities at their typical duration are required to have the same marginal utility; therefore, 9

βact applies to all activities. t0,i is a scaling parameter linked to an activity’s priority and minimum duration. 10

In this study, t0,i is not relevant, since activities cannot be dropped from daily plans. 11

The value of travel time saving (VTTS) is derived as follows: 12

𝑉𝑇𝑇𝑆 =𝛽𝑎𝑐𝑡 − 𝛽𝑡,𝑞

𝛽𝑚 (4)

where 𝛽𝑚 corresponds to the marginal utility of money. 13

The travel options modeled in this study include: car, public transit, bike and walked (modeled jointly), 14

AV, and SAV. The behavioral parameters for car and public transit used in this study are based on Tirachini 15

et al. 2014 and Kaddoura et al. 2015 and have been adjusted to reflect more realistically current travel costs 16

in the US (2017). The parameters used for the simulation are summarized in Table 1. In order to account 17

for aspects such as parking and walking times of car users we have derived an alternative specific constant 18

𝛽0,𝑐𝑎𝑟 = −0.1. In addition to that, car users pay a monetary cost proportional to the distance traveled 19

corresponding to $0.20-$0.30 per mile, depending on the scenario. Since, waiting, egress and access times 20

are not modeled in these experiments, public transit (PT) has been recalibrated yielding an alternative 21

specific constant 𝛽0,𝑃𝑇 = −1.5. This value also accounts for the average ticket cost and for the particular 22

reluctance of American society in using public transit. In similar fashion, the alternative specific constant 23

for walking/biking has been set to 𝛽0,𝑎𝑐𝑡𝑖𝑣𝑒 = −0.2. Similar to Kaddoura et al. (2015), the marginal utility 24

of traveling by car is set to zero. Even if this value is set to zero, traveling by car will be implicitly punished 25

by the opportunity cost of time (Horni et al., 2016). In this study, the marginal utility of money 𝛽𝑚 is equal 26

to 0.79 such that the VTTS for car users corresponds to about $18 per hour. This value has been obtained 27

according to the recommendations from the US DOT (2011). 28

The parameters for AVs have been mainly derived based on (Kockelman et al., 2017). The monetary costs 29

are estimated to be around $0.20 per mile. The purchasing costs might be higher than conventional cars, 30

but the higher initial investment would be compensated by the increased efficiency and better insurance 31

premiums. We assume AVs to have a null alternative specific constant in order to account for parking and 32

walking time reductions. The marginal disutility of traveling equal to +0.48 to reflect a marginal cost of 33

traveling equal to 50% of those of car users (corresponding to a VTTS of about $9 per hour), in line with 34

Gucwa (2014) and Kim et al. (2015)2. 35

As for SAVs, we assume the same alternative specific constant and marginal cost of traveling of AVs. 36

Unlike AVs, SAVs are characterized by waiting times depending on the availability of vehicles. We assume 37

2 Note that, in MATSim, setting a positive marginal disutility of traveling does not imply a gain of score from the trip

since agents are

7

the monetary costs to be composed of a flat fee, a distance fare that change depending on the scenario (see 1

the following sections for further details). 2

3

Table 1: Mode choice parameters used 4

Travel Mode 𝜷𝟎 𝜷𝒕

Car -0.1 0

Public Transit -1.5 -0.36

Walk/Bike -0.2 0

AV 0 +0.48

SAV 0 +0.48

5

In addition to travel choices, agents can modify their activities’ scheduling decisions by shifting, extending 6

or shortening activities considering aspects like the optimal duration, and opening/closing times of the 7

facility (Table 2). Activities performed outside opening times do not yield any positive gain of utility. 8

Furthermore, agents are subject to schedule penalty costs for being early or late accordingly to well-known 9

Vickrey’s parameters: α, β, and γ (Arnott et al., 1990). 10

11

Table 2: Travelers’ out-of-home activity attributes 12

Activity Type Optimal duration Opening time Closing time

Home 14 undefined undefined

Education 5 08:00 22:00

Work 7 07:00 undefined

Shopping 1 09:00 01:00

Leisure 2 09:00 01:00

13

Simulation of Shared mobility services and road capacity increase 14

The simulation of SAVs is performed by means of an extension of MATSim that allows for a dynamic 15

vehicle routing problem (DVRP) (Maciejewski et al., 2017). The DVRP contribution reproduces 16

dynamically demand responsive modes such as conventional taxis and ride hailing services. As opposed to 17

the standard vehicle routing in MATSim, which is conducted before each iteration starts, the DVRP module 18

allows an online dispatch of vehicle fleets. Vehicle dispatch is generally started the moment an agent wishes 19

to depart using such a mode. 20

For the simulation of large fleets of SAVs, a straightforward, rule-based dispatch algorithm is used, that 21

has been applied in previous case studies with more than 100,000 vehicles (Bischoff and Maciejewski, 22

2016). The algorithm aims at reducing the required number of vehicles during peak hours and thus 23

minimizes the required fleet to serve all the received requests. This issue is addressed by a “demand-supply 24

balancing” vehicle dispatch strategy, in which the system is classified into two mutually-excluding 25

categories, namely oversupply, with at least one idle SAV and no open requests, and undersupply, with no 26

idle SAVs and at least one open request. Both states are being handled in different ways. In the first case, 27

when a new request is placed, the nearest vehicle is dispatched towards it. In the latter case, when a vehicle 28

8

becomes idle it is dispatched to the nearest open request. In times of oversupply, requests are served 1

immediately, whereas in times of undersupply, a vehicle will first be dispatched to requests waiting in close 2

proximity and thus may leave requests waiting for a longer time. This helps to maximize the throughput of 3

the system. Despite its simplicity, this strategy provides solutions which are close to those of more complex 4

methods, such as solving iteratively the taxi assignment problem (Maciejewski et al. 2016). 5

In order to account for the capacity increase resulting from reduced reaction times and shorter following 6

distances, a specific MATSim module is adopted that allows for traffic simulation of mixed 7

autonomous/conventional flows (Maciejewski and Bischoff, 2017). This is achieved by lowering the 8

capacity (maximum flow) required by AVs to travel on a link by a factor of 1.5. This means that a link 9

which may otherwise be passed by a maximum of 1000 conventional vehicles per hour could be passed by 10

1500 AVs per hour. If a mixed flow of AVs and conventional vehicles are passing the link, the maximum 11

flow lies between these values, depending on the actual vehicles’ mix (Figure 2). The results of the model 12

are in line with those in Levin and Boyles (2016), who proposed a multiclass cell transmission model for 13

shared human and AV roads. 14

15

16

Figure 2: Impact of AVs on traffic flow 17

SIMULATION SCENARIOS 18

The impacts of different pricing schemes are investigated for three different scenarios. The first one, which 19

is referred to in this work as the “Base Scenario” corresponds to a realistic simulation of the city of Austin 20

and surroundings (Figure 3), comprising a considerable portion of the Austin metropolitan area (Greater 21

Austin). The studied region includes a series of satellite cities such as Round Rock, Cedar Park, and 22

Pflugerville. The road network used in the simulation consists of a high-resolution navigation network 23

including about 211,000 road segments (links). The population and its plans have been obtained by 24

adjusting those from Liu et al. (2017) who used CAMPO’s household data for Austin 2020. Although the 25

plans have not been formally validated, resulting trip distances and durations are reasonably realistic. 26

Normally, each agent needs to travel at least once per day to execute his plans. Instead of simulating the 27

full population, a sample of 5% (equivalent to 45,000 agents) is used for the experiments of this study. Link 28

capacities are downsized to match these with the sample size. The available transportation modes are car, 29

9

public transit and walk/bike (modeled jointly). In order to reflect current trends in availability of car as a 1

travel option, we assume 90 percent of agents to have access to car (either as a driver or passenger). 2

3

Figure 3: Simulation Network (source: Google Maps) 4

The two additional scenarios correspond to possible future scenarios characterized by the presence of AVs 5

and SAVs. Currently, it is not clear whether AVs will mainly replace privately owned vehicles or if they 6

are going to be adopted as shared taxis. On one hand, the auto industry is moving quickly to provide the 7

first “partially autonomous” models (Level 3) by 2020 and full autonomous models by 2030 (Level 4 and 8

Level 5) (Kockelman et al., 2017). Conversely, ride-sharing companies (Uber, Lyft, Didi) are already 9

running tests (Kang, 2016; Hawkins, 2017), making considerable investments (Buhr, 2017), and developing 10

important partnerships (Russell, 2017) to put driverless fleets on the road within a few years. Hence, an 11

“AV-oriented” Scenario and a “SAV-oriented” Scenario are included representing these two opposite 12

trends. In the AV-oriented scenario, it is assumed that a large portion of the population will switch from car 13

to AV (90% of agents having accessibility to car in the Base Scenario). In this scenario, the cost of AVs is 14

equal to the car cost ($0.20 per mile). SAVs are available too, but the fleet size is relatively small (one 15

vehicle every 30 agents) and they are characterized by lower prices than the current shared mobility services 16

($0.5 flat charge, 0.4 $/mile distance charge and 0.1 $/min time charge). For example, a trip of 5 miles, 17

from the northern suburbs to downtown, would vary approximately between $3.70 and $5.20 depending on 18

traffic conditions. In the SAV-oriented scenario, SAVs are largely available (one vehicle every 10 agents), 19

whereas most of the population is still car-dependent (only 10% has access to privately owned AVs whose 20

costs is increased to $0.30 per mile). Furthermore, we assume a decrease of availability of privately owned 21

vehicles to 60% in order to reflect a decrease of ownership (Litman, 2014). In this scenario, SAVs are 22

characterized by lower prices than in the AV-oriented scenario (a 50% reduction), assuming that main ride-23

sharing companies and local authorities would stipulate agreements on prices concerning the provision of 24

shared autonomous services. In this case, the same type of trip described above would cost approximately 25

between $1.80 and $2.60 (slightly higher than a public transit pass). 26

Results of MATSim simulations in terms of modal shift are reported in Figure 4. In the Base Scenario, a 27

car clearly appears as the dominant travel option, in line with the current situation. In the AV-Oriented 28

Scenario and SAV-Oriented Scenario, the introduction of two additional travel options (SAVs and AVs) 29

10

generate significant changes. PT trips decrease in the AV-Oriented Scenario (to 5% of the mode share), 1

whereas they increase SAV-Oriented Scenario (to 15% of the mode share). This result is partly due to the 2

lower ownership of private vehicles, which forces a considerable portion of commuters to travel with either 3

PT or SAVs. “Active trips” decrease to 4% and 7% respectively in the AV-Oriented Scenario and SAV-4

Oriented Scenario. As a result of this shift, congestion measured as daily total vehicle-miles traveled (VMT) 5

and daily total travel delay increase in both the scenarios (Table 3). The increased capacity due to 6

autonomous driving is offset by the increased trips particularly in the SAV-oriented Scenario by the empty 7

SAV trips that account for 6.6% of the total VMT. 8

9

Figure 4: Modal Share for the three different scenarios 10

11

Table 3: Traffic conditions of the three different scenarios 12

Base Scenario AV-oriented Scenario SAV-oriented Scenario

Total Daily VMT 2,845,406 3,106,669 3,032,629

VMT by Empty SAVs 2,845,406 4,741 201,828

Total Travel Delay

(veh-hours per weekday) 437,887 459,781 523,594

13

CONGESTION PRICING SCHEMES 14

In this study, we investigate the performance of four different congestion pricing schemes. A facility-based 15

and distance-based scheme are considered “traditional schemes” (Section 5.1), as they are well known in 16

the academia and real world. A link-based marginal cost pricing scheme and a joint distance-time scheme 17

are considered “advanced schemes” (Section 5.2) because they are more complex and they require cutting-18

edge technologies (such as those of connected-automated vehicle) for implementation. 19

20

Traditional congestion pricing strategies 21

Facility-based tolls are probably the most common form of congestion pricing since they do not require 22

particularly advanced technologies for implementation. In the past, this type of scheme has been 23

implemented mainly on tunnels, bridges and highway facilities that represent major bottlenecks. In this 24

study, a “Link-based Scheme” is applied to the one thousand most congested links during the morning 25

peak hours (7-9AM) and evening peak hours (5-7 PM). The tolled links are selected based on the 26

84%

7%

9%

Base Scenario

car PT walk/bike

9%4%

3%

84%

0%

AV-oriented Scenario

car PT walk/bike AV SAV

50%

15%

7%

5%

23%

SAV-oriented Scenario

car PT walk/bike AV SAV

11

volume/capacity (V/C) ratio calculated on hourly basis and aggregated for the peak hour periods. A 1

minimum threshold V/C ratio of 0.9 is chosen to identify the most congested links, resulting in the selection 2

of about 2-4% of the road network (3,911 links in the Base Scenario, 5,100 links in the AV-Oriented 3

Scenario, and 3,820 links in the SAV-Oriented Scenario). As it possible to see in Figure 4, the tolled links 4

include the most important segments of road infrastructure of the region such as Interstate 35 and the Texas 5

State Highway Loop 1. A flat toll rate of $0.20 is set to all the selected links regardless of the amount of 6

congestion and the characteristics of the link. 7

Here, the Distance-based Scheme toll varies simply with distance traveled, at a rate of $0.20 per mile 8

traveled by motorized vehicle between 7AM and 8PM. One could make it more time or location dependent, 9

requiring on board GPS to keep track of their position (and tally the owed charges before reporting back to 10

a fixed roadside or gas-pump-side device, for example). Of course, many nations, states and regions are 11

interested in distance-based tolls or VMT fees, especially when more fuel-efficient and electric vehicles 12

pay relatively few gas taxes. Clements and Kockelman (2018) discuss such tolling options, and the strengths 13

and weaknesses of various tolling technologies. 14

15

16

Figure 5: selected links in the Link-based Scheme for the Base Scenario (source VIA:Senozon) 17

Advanced congestion pricing strategies 18

The first advanced congestion pricing strategy investigated consists of a dynamic Marginal Cost Pricing 19

(MCP) scheme at link level. MCP in the context of road usage, consists in charging users for the extra cost 20

(shadow cost) that their trip cause to other users (Walters, 1961) due to congestion. According to MCP 21

models, an optimal static link-based toll 𝜏 can be derived for each link such that: 22

23

𝜏 = 𝑉 ∙𝜕𝑐

𝜕𝑉 (5)

24

12

Where 𝑉 corresponds to the traffic volume on the link and 𝑐 corresponds to the congestion costs that can 1

be related to 𝑉 by means of several functions. However, MCP presents some theoretical and practical 2

limitations, such as the necessity of dynamic features and difficulty of setting optimal link tolls for large 3

networks (De Palma and Lindsey, 2011). 4

Communication and automation technologies installed in AV/SAVs offer the opportunity to apply different 5

tolls on each link of a network such that vary dynamically according to traffic conditions. In the “MCP-6

based scheme” proposed, the cost of congestion on each link is derived by means of the Fundamental 7

Diagram (FD), that expresses a relation between traffic throughput 𝑞 (veh/h) and density 𝑘 (veh/km) 8

(Greenshields, 1935). According to the FD, the traffic throughput increases with density until reaching the 9

critical density corresponding to the link’s capacity. For values of density above the critical ones, the 10

throughput and (average) speed on the link decreases until zero. 11

Based on this concept, it is possible to derive for each link, its average speed 𝑢(𝑘, 𝑞) as function of density 12

and outflow, as follows: 13

14

𝑢(𝑘, 𝑞) =𝑞

𝑘 (6)

15

Then, for each link, the total delay accumulated during the time interval [𝑡, 𝑡 + ∆𝑡] corresponds: 16

17

𝑑 = [(𝑙

𝑢𝑡+∆𝑡−

𝑙

𝑢𝑡) ∙ 𝑛] (7)

18

where the first term corresponds to the marginal delay per time interval, which is given by the difference 19

of travel time on link of length 𝑙 at the average speed 𝑢 and at free-flow speed 𝑣. The second term 𝑛 20

corresponds to the additional or fewer number of link users over the time interval ∆𝑡, that can be expressed 21

as: 22

23

𝑛 = (𝑘𝑡+∆𝑡 − 𝑘𝑡) ∙ 𝑙 (8)

24

Hence, the marginal cost pricing charge for each link i, during the time interval [𝑡, 𝑡 + ∆𝑡] can be derived 25

as: 26

27

𝜏𝑖 =𝑑𝑖 ∙ 𝑉𝑇𝑇𝑆

𝑛𝑖 (9)

28

where VTTS corresponds to the average value of travel time. For reasons of understandability and 29

acceptability, each link’s charge varies over intervals of 15 minutes and it is derived by aggregating traffic 30

condition measurements across 5 minute intervals. The fare has a maximum threshold value of $1. For 31

practical reasons, given the size of the network, only a subset of 15,020 central links shown in Figure 6 is 32

analyzed. 33

34

13

1

Figure 6: analyzed links in the MCP Scheme (source VIA:Senozon) 2

The second advanced congestion pricing scheme consists of a joint “Distance-Time-based scheme.” The 3

main rationale behind this approach lies in the fact that traditional distance-based schemes do not really 4

capture traffic dynamics. In certain cases, they can even become detrimental as drivers might be 5

incentivized to take shorter (but more congested) routes (Liu et al., 2014). Including a toll component that 6

charges users for time traveled, depending on the traffic conditions of the network could obviate this 7

problem. In a similar fashion to the surge pricing mechanism, where prices are dynamically inflated based 8

on demand, in the Distance-Time-based scheme the fare is identified as follows: 9

10

𝜏 = 𝜑(𝑙) + 𝜎[𝑡,𝑡+∆𝑡](𝑡) (10)

11

where 𝜑(𝑙) corresponds to the distance dependent component and 𝜎[𝑡,𝑡+∆𝑡](𝑡) corresponds to the travel 12

time dependent component for any given time interval [𝑡, 𝑡 + ∆𝑡]. The distance component is set to $0.10 13

per mile between 7AM and 8PM, which is half of the Distance-based scheme charge. The component 14

𝜎[𝑡,𝑡+∆𝑡](𝑡) varies each hour, based on traffic conditions across all twelve 5-min intervals in that hour. In 15

order to reflect changes of overall marginal cost of congestion on the network, the travel-time-dependent 16

component is derived as follows: 17

18

𝜎[𝑡,𝑡+∆𝑡] =(∑ 𝑑𝑖

𝑁𝑖 ) ∙ 𝑉𝑇𝑇𝑆

𝑆 ∙ 𝑟 (11)

where links’ delay 𝑑𝑖 is calculated using Eq. 7, 𝑆 corresponds to the total number of departures over the 19

time period [𝑡, 𝑡 + ∆𝑡], and 𝑟 corresponds to the average trip duration on the network, which is derived as 20

follows: 21

14

1

𝑟 =𝐿

𝑈 (12)

where 𝐿 and 𝑈 correspond respectively to the average trip length and average free-flow speed over the 2

network. 3

Since both advanced schemes are intended to be consistent with traffic dynamics, which in turn depend on 4

agents’ mode, departure time and route choices, we adopt a simulation-based feedback iterative process to 5

derive the final toll values 𝜏̅3. The approach utilizes as stopping criterion the overall change of network 6

delay∆𝐷, which is the sum of all links’ delays. For each iteration j, the algorithm performs the following 7

steps: 8

9

Identify toll values 𝜏�̅� for each time interval [𝑡, 𝑡 + ∆𝑡] by means of Eq. 9 or Eq. 10. 10

Perform a MATSIM simulation until new stochastic user equilibrium is reached. 11

Derive total network delay by aggregating links measurement: 𝐷 = ∑ 𝑑𝑖𝑁𝑖 12

Check the change of total delay ∆𝐷 between the current iteration j and the previous (j-1). If ∆𝐷 is 13

less than 5%, stop. Otherwise, return to step 1. 14

15

The resulting fares for the MCP-scheme for the AV-Oriented and SAV-Oriented Scenario are determined 16

after 10 to 15 iterations. Among the 15,020 links analyzed, between 2,000 and 3,000 are tolled (across the 17

various 15-min intervals), with an average fare of $0.92. 18

Figure 7 illustrates fares for the Distance-Time-based scheme for both the AV-Oriented and SAV-Oriented 19

Scenarios. The two schemes show similar trends in the variation of the travel time charge component during 20

the peak hours, with higher fares during the evening peak. As expected, since the SAV-Oriented Scenario 21

is characterized by higher delays than in the AV-Oriented Scenario, the levels of charge are higher. 22

3 We indicate with �̅� the vector of tolls for all the links considered.

15

1

Figure 7: Resulting fares for the Distance-Time based scheme in the AV-Oriented and SAV-Oriented scenario 2

RESULTS AND IMPLICATIONS 3

The impacts derived from the different congestion pricing schemes in each scenario are discussed in this 4

section. The evaluation of the schemes is carried out by means of a set of commonly used performance 5

indicators such as mode shift, change of traffic delay and motorized trips. The analyses continue with a 6

comparison of system welfare effects, followed by a discussion about the policy implications of the 7

different schemes. 8

Mode choice 9

All the schemes succeed in reducing car, AV and SAV trips to a different extent. In all the scenarios PT 10

and slow modes, witness a considerable increase of mode share (Table 4-7). 11

Overall, the demand for SAVs and AVs seem more elastic than the demand for standard vehicles given the 12

higher modal shift achieved for all the CP strategies. Because of their higher initial costs, AV and SAV 13

travelers are more incentivized than car travelers to adopt PT or slow modes in presence of tolls. For this 14

reason, CP strategies seem to be more effective in AV-oriented and SAV-oriented scenarios. 15

Among the traditional schemes, the Distance-based scheme generates larger changes in travelers’ mode 16

choice than the link-based scheme in the Base Scenario. These results are in line with previous studies about 17

distance-based schemes (Litman, 1999). Instead, in the scenarios characterized by large presence of AVs 18

and SAVs the modal shifts are comparable. This is an interesting outcome, since the two schemes are 19

conceptually very different from each other and might have different effects in terms of economic gains, 20

distributional effects and public acceptability. 21

The MCP-based scheme determines travel behavior changes comparable to the ones achieved with the 22

traditional Link-based scheme, although the total number of tolled links (and the area involved), and the 23

levels of charge are lower. 24

16

The Distance-Time based scheme achieves the highest reduction of private trips in both scenarios. 1

Compared to the “traditional” distance-based toll, including a travel time-based component in the fare 2

allows for further reduction of car, AV and SAV trips, despite its lower distance-based component. 3

4

Table 4: Modal shift from the link-based scheme 5

AV oriented SAV oriented Base (no SAVs-AVs)

Change of car trips (%) -0.8 -7.92 -6.81

Change of PT trips (%) 28.07 16.20 5.89

Change of walk/bike trips (%) 10.73 7.50 0.91

Change of AV trips (%) -37.94 -2.30 0.00

Change of SAV trips (%) -0.05 -13.48 0.00

6

Table 5: Modal shift from the distance-based scheme 7

AV oriented SAV oriented Base (no SAVs-AVs)

Change of car trips (%) -3.97 -7.31 -14.10

Change of PT trips (%) 35.55 14.74 2.89

Change of walk/bike trips (%) 11.11 5.74 11.29

Change of AV trips (%) -42.62 -1.07 0.00

Change of SAV trips (%) -0.06 -12.10 0.00

8

Table 6: Modal shift from the MCP-based scheme 9

AV oriented SAV oriented Base (no SAVs-AVs)

Change of car trips (%) -0.6 -9.62 -

Change of PT trips (%) 26.36 16.26 -

Change of walk/bike trips (%) 11.45 9.11 -

Change of AV trips (%) -37.09 -2.63 -

Change of SAV trips (%) -0.16 -13.12 -

10

Table 7: Modal shift from the Distance-Time based scheme 11

AV oriented SAV oriented Base (no SAVs-AVs)

Change of car trips (%) -5.74 -33.1 -

Change of PT trips (%) 38.96 40.26 -

Change of walk/bike trips (%) 16.03 17.30 -

Change of AV trips (%) -49.06 -3.99 -

Change of SAV trips (%) -0.18 -20.00 -

17

Traffic performance of the network 1

Both traditional and advanced CP strategies determine a significant reduction of private trips traveled by 2

AVs, SAVs and cars (Figure 8). Schemes with a distance dependent fee component do not necessarily 3

achieve the highest reduction of VMT. For example, the link-based and MCP scheme, determine higher 4

VMT reductions than the traditional distance-based scheme in both the AV-Oriented and the SAV-Oriented 5

scenario. In general, the traditional schemes seem to yield much higher improvements in the AV-Oriented 6

and SAV-Oriented scenario than in the Base Scenario because of higher elasticity of autonomous travelers. 7

The MCP-based scheme produces a VMT reduction comparable or higher that achieved with the link-based 8

scheme. The Distance-Time based scheme determines the largest decrease of travel demand in both the 9

AV-Oriented and the SAV-Oriented scenario. 10

However, this is just a single perspective to evaluate the effects of the strategies, as the changes in terms of 11

network daily travel delay show (Figure 9). Interestingly, in the AV-Oriented scenario all the CP strategies 12

generate a considerable reduction of delays (above 50%). The results vary significantly according to the 13

scheme. For example, the advanced CP strategies outperform the traditional schemes. In the SAV-Oriented 14

scenario, all the schemes determine similar reductions of delays (around 25-30%), except for the Distance-15

Time based scheme. Despite changes in the mode shift of the traditional schemes in the Base Scenario are 16

below 10%, the decrease of delay is above 30%. In this case, users seem to be more willing to reroute and 17

reschedule their trips rather than switching to public transit. 18

19

20

Figure 8: Reduction of motorized trips for the different scenarios according to the congestion pricing scheme 21

-80

-70

-60

-50

-40

-30

-20

-10

0

AV oriented SAV oriented Base

VM

T ch

ange

(%

)

Distance-based Scheme

Link-based Scheme

MCP-based Scheme

Distance-Time based Scheme

18

1

Figure 9: Reduction of traffic delay for the different scenarios according to the congestion pricing scheme 2

3

Welfare changes 4

Maximization of social welfare is an important criterion to evaluate transportation policy decisions. The 5

social welfare change due to the introduction of pricing policies can be estimated by means of the well-6

known “rule of the half”, which is given by the contribution of the gains/losses of existing users and 7

gains/losses of new users. The main drawbacks of this approach include the possibility of considering 8

different impacts of tolls on travel (changes in travel time and schedules) and users’ heterogeneity (Arnott 9

et al., 1990). 10

Multi-agent simulations like MATSim partly overcome these issues by allowing a “non-conventional” 11

economic appraisal based on the agents’ utilities as an economic performance indicator of the system. 12

Although income differences across the population by means of the agents’ utility score are not considered 13

in this study, it is possible to account for the main effects of congestion pricing schemes such as changes in 14

travel costs and travel behavior, and revenues. 15

In this study, change of total welfare ∆𝜔 between the original scenario (with no toll) and any congestion 16

pricing scenario is calculated as the sum of public revenues (first term in Eq. 13) and consumer surplus 17

(second term in Eq. 13): 18

∆𝜔 = 𝜏 + ∑ 𝛽𝑚

∙ (𝑉𝑗 − 𝑉𝑗′ )

𝐽

𝑗

(13)

19

where 𝜏 is the sum of the collected toll revenues (in presence of tolls), 𝛽𝑚 is the negative marginal utility 20

of monetary cost, and 𝑉𝑗 and 𝑉𝑗′correspond to the total daily utility score for agent j in the original scenario 21

and in the congestion pricing scenario, respectively. 22

Table 8 summarizes the social welfare impacts of the different schemes for each scenario. The most 23

effective strategy in terms of total welfare gains seems to be the Distance-based scheme in all scenarios 24

(assuming that the toll revenues could be fully reinvested). The Link-based scheme determines an 25

improvement of social welfare to a minor extent (between 2% and 4%, depending on the scenario). 26

-120.00

-100.00

-80.00

-60.00

-40.00

-20.00

0.00

AV oriented SAV oriented Base

Tota

l Del

ay c

han

ge (

%)

Distance-based Scheme

Link-based Scheme

MCP-based Scheme

Distance-Time based Scheme

19

Interestingly, the advanced CP strategies do not seem to yield improvements compared to the traditional 1

ones. The high levels of congestion of the autonomous scenarios, and the lower attractiveness of PT and 2

active modes as compared to autonomous transport, yield to relatively high tolls in those CP schemes that 3

explicitly depend on traffic delays. The MCP scheme, compared to the link-based one, is penalized by the 4

more limited area of action, particularly given the mobility configuration of Austin where traffic congestion 5

is rather spread out over the network, and not just limited to the city center. 6

When the revenues are not considered, all of the CP strategies achieve a reduction in social welfare. In this 7

case, the highest performance is achieved by the MCP-based scheme that determines the lowest reduction 8

of consumer surplus in both the AV-Oriented and SAV-Oriented scenario. This is an important aspect to 9

consider, since the ability to reinvest and the fraction of expendable revenues would determine whether a 10

scheme is favorable. 11

Finally, the Distance-based, the Link-based, and the Time-Distance scheme perform similarly regardless of 12

the level of private and shared AVs, whereas the MCP-based scheme behaves differently according to the 13

scenario. In the SAV-Oriented scenario, the social welfare changes are higher than in the AV-Oriented 14

scenario (about 0% and -10%, respectively). 15

16

Table 8: Welfare changes for alternative CP schemes for each scenario 17

AV-oriented

Scenario

SAV-oriented

Scenario

Base

Scenario

Original Scenario-total welfare (Million $/day) 7.264 5.340 14.242

AV-oriented

Scenario

SAV-oriented

Scenario

Base

Scenario

Distance-Based Scheme: consumer surplus change

($ per capita per day) -2.99 -1.13 -1.14

Distance-Based Scheme: welfare change with

revenues ($ per capita per day) 2.44 1.48 1.83

Total welfare change (%) 30.44 25.09 11.65

AV-oriented

Scenario

SAV-oriented

Scenario

Base

Scenario

Link-Based Scheme: consumer surplus change ( $

per capita per day) -2.17 -0.80 -0.40

Link-Based Scheme: welfare change with revenues

($ per capita per day) 0.33 0.28 0.32

Total welfare change (%) 4.12 4.72 2.01

AV-oriented

Scenario

SAV-oriented

Scenario

Base

Scenario

MCP-Based Scheme: consumer surplus change ($

per capita per day) -1.23 -0.75 -

MCP-Based Scheme: welfare change with revenues

($ per capita per day) -0.58 -0.04 -

Total welfare change (%) -9.35 -0.61 -

20

AV-oriented

Scenario

SAV-oriented

Scenario

Base

Scenario

Distance-Time Based Scheme: consumer surplus

change ($ per capita per day) -3.67 -2.17 -

Distance-Time Based Scheme: welfare change with

revenues ($ per capita per day) -0.63 -0.53 -

Total welfare change (%) -7.81 -9.08 -

1

Policy implications 2

All the investigated CP schemes are efficient in the sense that they decrease traffic demand and make the 3

traffic flows more efficient by reducing traffic congestion. 4

From an economic perspective, some of the schemes do not yield to an overall increase of social welfare, 5

whereas others might be progressive depending on the compensation schemes. However, benefits from 6

reduced emissions, noise and road damage were not considered in the analyses. 7

The portion of revenues collected from the tolls that can be reinvested in the social welfare is crucial for 8

the success and equity implications of the scheme. In the theory of congestion pricing, investments into 9

public transit service are considered a progressive solution, especially for zones with public transport access 10

(Ecola and Light, 2010). Other options could be by means of per-capita credits or income-based discounts. 11

Under these circumstances, it becomes particularly important to properly quantify the investment and 12

maintenance costs for different schemes. For example, the employment of cheaper satellite and cellular 13

technologies (compared to traditional tolling infrastructure) could lower the investment costs for advanced 14

CP schemes and make them more attractive. 15

CONCLUSION 16

AVs and shared AVs will affect people’s mobility and traffic. It is not clear whether the benefits of more 17

efficient traffic flows would compensate for the costs of increased trips and distance traveled. Congestion 18

pricing schemes represent an opportunity to internalize the negative costs of traffic congestion. The novel 19

transportation landscape, characterized by higher automation and connectivity, could facilitate the 20

implementation of traditional and more advanced strategies. 21

In this study, we adopt an agent-based model to investigate the potential mobility, traffic and economic 22

effects of different congestion pricing schemes in alternative future scenarios for the metropolitan area of 23

Austin. From a traffic perspective, all the mobility schemes yield to considerable reductions of congestion. 24

While advanced CP schemes are more effective than traditional ones in affecting travel demand and traffic, 25

they do not bring economic gains. In all the scenarios, the Distance-based and Link-based schemes yield 26

social welfare improvements. 27

The analysis of mobility scenarios by means of an agent-based model like MATSim allows a high level of 28

realism since it is possible to explicitly model several factors concerning transportation demand and traffic. 29

In the specific context of AVs-SAVs, the coexistence of different autonomous modes and cars is considered 30

(in addition to public transit and walk/bike), as well as: the impacts of autonomous driving on increased 31

capacity; the changes in travel costs and preferences, and the demand responsive mechanism of SAV 32

services (with the phenomenon of empty trips). In future studies of AV-SAV scenarios, it would be 33

interesting to include the effects of automation on destination choice and parking. 34

In the specific field of travel demand management, additional studies can be performed to investigate the 35

distributional effects of different CP schemes (considering income heterogeneity across the population) and 36

possible compensation measures. 37

21

ACKNOWLEDGEMENTS 1

The authors thank Michal Maciejewski and Amit Agarwal for fruitful discussions on the MATSim 2

simulation, and Felipe Dias for support in the analyses. The study was partly funded by the Texas 3

Department of Transportation under Project 0-6838, “Bringing Smart Transport to Texas”. 4

5

22

REFERENCES 1

Arnott, R., De Palma, A., & Lindsey, R. (1990). Economics of a bottleneck. Journal of Urban Economics, 2

27(1), 111-130 3

Arnott, R. (1998). William Vickrey: Contributions to Public Policy. International Tax and Public Finance, 4

5(1), 93-113. 5

Balmer, M., Axhausen, K., & Nagel, K. (2006). Agent-based demand-modeling framework for large-scale 6

microsimulations. Transportation Research Record No. 1985, 125-134. 7

Bansal, P., & Kockelman, K. M. (2017). Forecasting Americans’ Long-Term Adoption of Connected and 8

Autonomous Vehicle Technologies. Transportation Research Part A: Policy and Practice, 95, 49-63. 9

Bösch, P. M., Becker, F., Becker, H., & Axhausen, K. W. (2017). Cost-based analysis of autonomous 10

mobility services. Transport Policy. 11

M.J. Beckman. (1965). On optimal tolls for highways, tunnels and bridges. Vehicular Traffic Science, 12

American Elsevier, New York, pp. 331-341 13

Bischoff, J., & Maciejewski, M. (2016). Simulation of city-wide replacement of private cars with 14

autonomous taxis in Berlin. Procedia Computer Science, 83, 237-244. 15

Buhr, S. (2017). Lyft launches a new self-driving division and will develop its own autonomous ride-hailing 16

technology. Tech Crunch.com (July 21). Retrieved at: https://techcrunch.com/2017/07/21/lyft-launches-a-17

new-self-driving-division-called-level-5-will-develop-its-own-self-driving-system/ 18

Burns, L. D., Jordan, W. C., & Scarborough, B. A. (2013). Transforming personal mobility. The Earth 19

Institute, 431, 432. 20

Chen, T. D., Kockelman, K. M., & Hanna, J. P. (2016). Operations of a Shared, Autonomous, Electric 21

Vehicle Fleet: Implications of Vehicle & Charging Infrastructure Decisions. Transportation Research Part 22

A: Policy and Practice, 94, 243-254. 23

Clements, L., & Kockelman, K. (2018). Technologies for Congestion Pricing. Under review for publication 24

in Travel Behavior and Society, and available at 25

http://www.caee.utexas.edu/prof/kockelman/public_html/TRB18CBCPtech.pdf. 26

Correia, G. H., & van Arem, B. (2016). Solving the User Optimum Privately Owned Automated Vehicles 27

Assignment Problem (UO-POAVAP): A model to explore the impacts of self-driving vehicles on urban 28

mobility. Transportation Research Part B: Methodological, 87, 64-88. 29

Cyganski, R., Fraedrich, E., & Lenz, B. (2015). Travel time valuation for automated driving: a use-case-30

driven study. In 94th Annual Meeting of the Transportation Research Board. Washington, DC: 31

Transportation Research Board. 32

De Palma, A., Lindsey, R., & Quinet, E. (2004). Time-varying road pricing and choice of toll locations. 33

Road pricing: Theory and evidence. Research in Transportation Economics, 9, 107-131. 34

De Palma, A., & Lindsey, R. (2011). Traffic congestion pricing methodologies and technologies. 35

Transportation Research Part C: Emerging Technologies, 19(6), 1377-1399. 36

Ecola, L., & Light, T. (2010). Making congestion pricing equitable. Transportation Research Record: 37

Journal of the Transportation Research Board, (2187), 53-59. 38

Dresner, K., & Stone, P. (2008). A multiagent approach to autonomous intersection management. Journal 39

of Artificial Intelligence Research, 31, 591-656. 40

Fagnant, D. J., & Kockelman, K. (2015). Preparing a nation for autonomous vehicles: opportunities, barriers 41

and policy recommendations. Transportation Research Part A: Policy and Practice, 77, 167-181. 42

23

Fagnant, D. J., Kockelman, K. M., & Bansal, P. (2015). Operations of Shared Autonomous Vehicle Fleet 1

for Austin, Texas, Market. Transportation Research Record: Journal of the Transportation Research Board, 2

No. 2536, 98-106. 3

Fagnant, D. J., & Kockelman, K. M. (2018). Dynamic Ride-sharing and Fleet Sizing for a System of Shared 4

Autonomous Vehicles in Austin, Texas. Transportation, 45(1), 143-158. 5

Gu, Z., Liu, Z., Cheng, Q., & Saberi, M. (2018). Congestion Pricing Practices and Public Acceptance: A 6

Review of Evidence. Case Studies on Transport Policy, 6(1), 94-101 7

Gucwa, M. (2014). Mobility and energy impacts of automated cars. Proceedings of the Automated Vehicles 8

Symposium, San Francisco. 9

Gurumurthy, K. M., & Kockelman, K. M. (2018). Analyzing the Dynamic Ride-Sharing Potential for 10

Shared Autonomous Vehicle Fleets Using Cell Phone Data from Orlando, Florida, under review for 11

publication in Computers, Environment and Urban Systems . 12

Haboucha, C. J., Ishaq, R., & Shiftan, Y. (2017). User preferences regarding autonomous vehicles. 13

Transportation Research Part C: Emerging Technologies, 78, 37-49. 14

Hawkings, A.J. (2017). Uber’s self-driving cars are now picking up passengers in Arizona. The Verge 15

(Feb. 21). Retrieved at: https://www.theverge.com/2017/2/21/14687346/uber-self-driving-car-arizona-16

pilot-ducey-california 17

Hoogendoorn, R., van Arem, B., & Hoogendoorn, S. (2014). Automated driving, traffic flow efficiency and 18

human factors: A literature review. Transportation Research Record, 2422, 113–120. 19

Hörl, S., Ruch, C., Becker, F., Frazzoli, E. & Axhausen, K. (2017). Fleet control algorithms for automated 20

mobility: A simulation assessment for Zurich. ETH Zürich Research Collection. Retrieved at: 21

https://www.research-collection.ethz.ch/handle/20.500.11850/175260 22

Horni, A., Nagel, K., & Axhausen, K. W. (Eds.). (2016). The multi-agent transport simulation MATSim. 23

London: Ubiquity Press. 24

Kaddoura, I., Kickhöfer, B., Neumann, A., & Tirachini, A. (2015). Optimal public transport pricing: 25

Towards an agent-based marginal social cost approach. Journal of Transport Economics and Policy (JTEP), 26

49(2), 200-218. 27

Kaddoura, I., Bischoff, J. & Nagel, K. (2018). Towards welfare optimal operation of innovative mobility 28

concepts: External cost pricing in a world of shared autonomous vehicles. VSP Working paper, 18-01. 29

Retrieved at http://www.vsp.tu-berlin.de/publications/vspwp/ 30

Kang, C. (September, 2016). “No Driver? Bring It On. How Pittsburgh Became Uber’s Testing Ground” 31

on The New York Times. Retrieved at: https://www.nytimes.com/2016/09/11/technology/no-driver-bring-32

it-on-how-pittsburgh-became-ubers-testing-ground.html 33

Kang, C. (2016). Self-Driving Cars Gain Powerful Ally: The Government. The New York Times (Sept. 20). 34

Retrieved at: https://www.nytimes.com/2016/09/20/technology/self-driving-cars-guidelines.html 35

Kickhöfer, B., Grether, D., & Nagel, K. (2011). Income-contingent user preferences in policy evaluation: 36

application and discussion based on multi-agent transport simulations. Transportation, 38(6), 849. 37

Kim, K., Rousseau, G., Freedman, J., & Nicholson, J. (2015, May). The travel impact of autonomous 38

vehicles in metro Atlanta through activity-based modeling. In 15th TRB National Transportation Planning 39

Applications Conference, Atlantic City, May (Vol. 18). 40

Kockelman, K., Sharon, G., Simoni, M., Albert, M., Fritz, H., Hutchinson, R., ... & Pourrahmani, E. (2017). 41

An assessment of autonomous vehicles: traffic impacts and infrastructure needs (No. FHWA/TX-17/0-42

6847-1). University of Texas at Austin. Center for Transportation Research. 43

24

Krueger, R., Rashidi, T. H., & Rose, J. M. (2016). Preferences for shared autonomous vehicles. 1

Transportation Research Part C: Emerging Technologies, 69, 343-355. 2

Lamotte, R., De Palma, A., & Geroliminis, N. (2016). Sharing the road: the economics of autonomous 3

vehicles. Retrieved at: https://hal.archives-ouvertes.fr/hal-01281425/document 4

Lawphongpanich, S., & Yin, Y. (2010). Solving the Pareto-improving toll problem via manifold 5

suboptimization. Transportation Research Part C: Emerging Technologies, 18(2), 234-246. 6

Lenz, B., Fraedrich, E. M., & Cyganski, R. (2016). Riding in an autonomous car: What about the user 7

perspective? Retrieved at: 8

http://elib.dlr.de/108174/1/Lenz_Fraedrich_Cyganski_HKSTS_2016_RIDING%20IN%20AN%20AUTO9

NOMOUS%20CAR%20WHAT%20ABOUT%20THE%20USER%20PERSPECTIVE.pdf 10

Levin, M. W., & Boyles, S. D. (2016). A multiclass cell transmission model for shared human and 11

autonomous vehicle roads. Transportation Research Part C: Emerging Technologies, 62, 103-116. 12

Lighthill, M. J., & Whitham, G. B. (1955, May). On kinematic waves. II. A theory of traffic flow on long 13

crowded roads. In Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering 14

Sciences (Vol. 229, No. 1178, pp. 317-345). The Royal Society. 15

Litman, T. (1999). Distance-based charges; a practical strategy for more optimal vehicle pricing. Victoria 16

Transport Policy Institute. 17

Litman, T., & Doherty, E. (2011). Transportation Cost and Benefit Analysis II–Parking Costs. 18

Transportation Cost and Benefit Analysis Techniques, Estimates and Implications, 5-4. 19

Litman, T. (2017). Autonomous vehicle implementation predictions. Victoria Transport Policy Institute, 20

28. Retrieved at: http://www.vtpi.org/avip.pdf 21

Litman, T. (2016). Smart congestion relief: Comprehensive analysis of traffic congestion costs and 22

congestion reduction benefits. Retrieved at: http://vtpi.org/cong_relief.pdf 23

Liu, Z., Wang, S., & Meng, Q. (2014). Optimal joint distance and time toll for cordon-based congestion 24

pricing. Transportation Research Part B: Methodological, 69, 81-97. 25

Liu, J., Kockelman, K. M., Boesch, P. M., & Ciari, F. (2017). Tracking a system of shared autonomous 26

vehicles across the Austin, Texas network using agent-based simulation. Transportation, 1-18. 27

Maciejewski, M., Bischoff, J., & Nagel, K. (2016). An assignment-based approach to efficient real-time 28

city-scale taxi dispatching. IEEE Intelligent Systems, 31(1), 68-77. 29

Maciejewski, M., Bischoff, J., Hörl, S., & Nagel K. (2017). Towards a Testbed for Dynamic Vehicle 30

Routing Algorithms. Cham: Springer International Publishing, pp. 69–79. 31

May, A., Shepherd, S., Sumalee, A., & Koh, A. (2008). Design tools for road pricing cordons. Road 32

Congestion Pricing in Europe: Implications for the United States, Cheltenham, UK, 138-155. 33

Milakis, D., Van Arem, B., & Van Wee, B. (2017). Policy and society related implications of automated 34

driving: A review of literature and directions for future research. Journal of Intelligent Transportation 35

Systems, 1-25. 36

Muoio, D. (September, 2017). “RANKED: The 18 companies most likely to get self-driving cars on the 37

road first” on Business Insider. Retrieved at: http://www.businessinsider.com/the-companies-most-likely-38

to-get-driverless-cars-on-the-road-first-2017-4/#18-baidu-1 39

Nagel, K., & Flötteröd G. (2009). Agent-based traffic assignment: going from trips to behavioral travelers. 40

12th International Conference on Travel Behaviour Research (IATBR), Jaipur, India. 41

Pigou, A.C. (1924). The economics of welfare. Transaction Publishers. London, 1920, 876 42

25

Richards, P. I. (1956). Shock waves on the highway. Operations Research, 4(1), 42-51. 1

Russell, J. (March, 2017). “China’s Didi Chuxing opens U.S. lab to develop AI and self-driving car tech” 2

on Tech Crunch. Retrieved at: https://techcrunch.com/2017/03/08/didi-us-research-lab/ 3

Schakel, W. J., Van Arem, B., & Netten, B. D. (2010, September). Effects of cooperative adaptive cruise 4

control on traffic flow stability. In Intelligent Transportation Systems (ITSC), 13th International IEEE 5

Conference (pp. 759-764). 6

Scheltes, A., & de Almeida Correia, G. H. (2017). Exploring the use of automated vehicles as last mile 7

connection of train trips through an agent-based simulation model: An application to Delft, Netherlands. 8

International Journal of Transportation Science and Technology, 6(1), 28-41. 9

Sharon, G., Levin, M. W., Hanna, J. P., Rambha, T., Boyles, S. D., & Stone, P. (2017). Network-wide 10

adaptive tolling for connected and automated vehicles. Transportation Research Part C: Emerging 11

Technologies, 84, 142-157. 12

Shladover, S. E., Su, D., & Lu, X.-Y. (2012). Impacts of cooperative adaptive cruise control on freeway 13

traffic flow. Transportation Research Record No. 2324, 63–70. 14

Shoup, D. C. (2006). Cruising for parking. Transport Policy, 13(6), 479-486. 15

Small, K. A., Verhoef, E. T., & Lindsey, R. (2007). The Economics of Urban Transportation. Routledge. 16

Spieser, K., Treleaven, K., Zhang, R., Frazzoli, E., Morton, D., & Pavone, M. (2014). Toward a systematic 17

approach to the design and evaluation of automated mobility-on-demand systems: A case study in 18

Singapore. Road Vehicle Automation (pp. 229-245). Springer International Publishing. 19

Talebpour, A., & Mahmassani, H. S. (2016). Influence of connected and autonomous vehicles on traffic 20

flow stability and throughput. Transportation Research Part C: Emerging Technologies, 71, 143-163. 21

Tientrakool, P., Ho, Y. C., & Maxemchuk, N. F. (2011). Highway capacity benefits from using vehicle-to-22

vehicle communication and sensors for collision avoidance. In Vehicular Technology Conference (VTC 23

Fall), IEEE (pp. 1-5). 24

Tirachini, A., D. A. Hensher, & J. M. Rose (2014): ‘Multimodal pricing and optimal design of urban public 25

transport: the interplay between traffic congestion and bus crowding’, Transportation Research Part B, 61, 26

33–54 27

US Department of Transportation, 2011. The value of travel time savings: departmental guidance for 28

conducting economic evaluations revision 2. Retrieved at: 29

https://www.transportation.gov/sites/dot.dev/files/docs/vot_guidance_092811c.pdf 30

Yap, M. D., Correia, G., & Van Arem, B. (2016). Preferences of travellers for using automated vehicles as 31

last mile public transport of multimodal train trips. Transportation Research Part A: Policy and Practice, 32

94, 1-16. 33

van den Berg, V. A., & Verhoef, E. T. (2016). Autonomous cars and dynamic bottleneck congestion: The 34

effects on capacity, value of time and preference heterogeneity. Transportation Research Part B: 35

Methodological, 94, 43-60. 36

Verhoef, E. T. (2002). Second-best congestion pricing in general networks. Heuristic algorithms for finding 37

second-best optimal toll levels and toll points. Transportation Research Part B: Methodological, 36(8), 38

707-729. 39

Vickrey, W. S. (1963). Pricing in urban and suburban transport. The American Economic Review, 53(2), 40

452-465. Chicago 41

Vickrey, W. S. (1969). Congestion theory and transport investment. The American Economic Review, 59(2), 42

251-260. 43

26

Vickrey, W. (1997). Public economics: selected papers by William Vickrey. Cambridge University Press. 1

Wadud, Z., MacKenzie, D., & Leiby, P. (2016). Help or hindrance? The travel, energy and carbon impacts 2

of highly automated vehicles. Transportation Research Part A: Policy and Practice, 86, 1-18. 3

Winter, K., Cats, O., van Arem, B. & Martens, K. (2017). Impact of relocation strategies for a fleet of 4

shared automated vehicles on service efficiency, effectiveness and externalities. 2017 5th IEEE 5

International Conference on Models and Technologies for Intelligent Transportation Systems (MT-ITS), 6

Naples, 2017, pp. 844-849. doi: 10.1109/MTITS.2017.8005630 7

Walters, A.A. (1961). The theory and measurement of private and social cost of highway congestion. 8

Econometrica: Journal of the Econometric Society: 676-99. 9

Zhang, X., & Yang, H. (2004). The optimal cordon-based network congestion pricing problem. 10

Transportation Research Part B: Methodological, 38(6), 517-537. 11

Zhao, Y., & Kockelman, K. (2017). Anticipating the Regional Impacts of Connected and Automated 12

Vehicle Travel in Austin, Texas. International Journal of Sustainable Transportation. 13