Document [original]

This version is available at https://doi.org/10.14279/depositonce-8347

right to use is granted. This document is intended solely for

personal, non-commercial use.

other uses, in any current or future media, including reprinting/republishing this material for advertising

or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or

reuse of any copyrighted component of this work in other works.

Bischoff, J., Maciejewski, M., Schlenther, T., & Nagel, K. (2018). Autonomous Vehicles and Their Impact

on Parking Search. IEEE Intelligent Transportation Systems Magazine.

https://doi.org/10.1109/mits.2018.2876566

Bischoff, J., Maciejewski, M., Schlenther, T., & Nagel, K.

Autonomous vehicles and their impact on

parking search

Accepted manuscript (Postprint)Journal article |

IEEE INTELLIGENT TRANSPORTATION SYSTEMS 1

Autonomous vehicles and their impact on

parking search

Joschka Bischoff, Michal Maciejewski, Tilmann Schlenther, Kai Nagel

(Invited Paper)

Abstract—Parking is a major constraint for car users and

therefore an important factor in mode choice decisions. In this

paper we introduce a model to simulate parking search behavior

for cars within a multi-agent transport simulation, including

full simulation of all steps of parking search, such as walking

to and from the vehicle. This is combined with the capabilities

of privately owned autonomous vehicles (AVs), which may park

automatically, often in other locations than conventional cars,

once they are not in use. Three different strategies for AVs

to park are developed: (1) Conventional parking search, (2)

parking at a designated AV lot, and (3) empty cruising, where

vehicles do not use any parking space, but keep on driving.

We apply the simulation model to a residential neighborhood

in central Berlin, where parking pressure is generally high and

apply different shares of AV usage to the synthetic population

used. This allows a detailed evaluation of effects for both AV

and conventional vehicle owners. Results suggest that the usage

of designated parking lots may be the most beneficial solution

for most users, with both vehicle wait times and parking search

durations being the lowest.

Index Terms—parking search, autonomous vehicles, transport

simulation, MATSim

I. INTRODUCTION

Beyond well-known effects of traffic flow and congestion,

parking is a major issue of car usage in cities around the

globe. This is especially true in European cities, where on-

street, or curbside parking, is the predominating form of

vehicle parking.

In recent years, agent-based parking search models have

evolved and found usage in several cities. These models aim

at simulating parking search behavior for streets or quarters

of a city. This approach has proven suitable for modeling

additional traffic effects of parking and/or behavioral ques-

tions with regard to parking search behavior. At the same

time, agent-based transport simulations are a powerful tool

to simulate agents’ activities and travel patterns, as well

as the behavior related to it, on a large scale. There have

been several attempts to integrate parking search models

into transport simulations, however, to the knowledge of

the authors, there is currently only one working simulation

available offering both [1].

In the upcoming years, autonomous vehicles may change

the way people are using and parking their private vehicles

quite drastically, which will lead to a new interplay between

The authors are with the Department for Transport System Planning

and Transport Telematics at Technische Universität Berlin, Salzufer 17–

19, 10587 Berlin, Germany; e-mail: [email protected]. Michal Ma-

ciejewski is also with the Poznan University of Technology, Division of

Transport Systems.

Manuscript received September 15, 2017; revised December 18, 2017

users of conventional vehicles and AV users. Owners of AVs

may pick a totally different parking behavior as the parking

search process can be done without any human interaction.

This could possibly have an effect on users of conventional

vehicles.

In this paper, we introduce an approach to integrate

parking search behavior and privately owned autonomous

vehicles in an agent-based simulation to analyze possible

effects of interplay. The starting point is the parking search

simulation that has been introduced previously [1]. To

our best knowledge, there is currently no other research

that embeds parking search of AVs into microscopic urban

transport simulation that would allow to evaluate possible

effects of AV on the transport system as a whole.

II. STATE OF THE ART

Parking, parking search and parking choice have been

widely researched. On the behavioral side of parking search,

papers by Axhausen[2] and Polak and Axhausen[3] provide

a comprehensive overview of parking search behavior and

ways to model it. The effect of different parking strate-

gies and prices was also investigated by Shoup [4] , who

demonstrated why vehicle cruising occurs and explains the

influence of parking policy on the behavior of drivers.

Several tools to simulate it are available. One of them

is PARKAGENT[5], a multi-agent, spatially explicit model

developed as an ArcGIS extension. It allows the simulation

of both streetside and garage parking lot locations in city

quarters. Agent simulation takes place only during parking

search, which is modeled in high detail. The biggest, and

to the knowledge of the authors, only simulation scenario

published about has been set up for parts of Tel Aviv.

On the contrary, the project SUSTAPARK [6] includes

a detailed traffic model of cities so that the influence of

parking search into a city’s overall traffic state can be

simulated. The model is applied for Leuven, Belgium. The

model is based on cellular automata. Parts of the software

are available under an open source license. A simulation of

behavioral change seems only possible in terms of parking,

but not in terms of other choice dimensions, such as

departure times or mode choice.

Several papers proposed at ETH Zürich have addressed

modeling and simulating parking choice in MATSim [7].

Their main focus is on parking choice modeling (e.g., the

choice between two differently priced garages) [8], without

modeling explicitly related physical activities, such as walk-

ing from or to parking lots, or the actual search for a space

in a lot. Instead, the length of the walking distance and the

IEEE INTELLIGENT TRANSPORTATION SYSTEMS 2

Activity

e.g. home

Activity

e.g. work

Leg

Mode: car

Activity

e.g. home

Activity

e.g. work

Leg

Mode: walk

Activity

Car interaction

Leg

Mode: car

Activity

Car interaction

Leg

Mode: walk

Activity

e.g. home

Activity

e.g. work

Leg (Departure)

Mode: av

Private AV

idle

Vehicle arrives

at pickup point

Agent Enters

Drive

Vehicle arrives

at destination

Agent leaves

Dispatch call

Empty ride

Private AV

idle

Free spot Garage Cruise

Parking strategy choice

Standard

car ride

Car ride

with parking

AV ride

with parking

Fig. 1. An agent’s car leg using MATSim’s standard approach (left side), with walk legs to and from parking in between (center) and using an AV (right)

parking costs reduce an agent’s daily score. This approach

is very fast from the computational perspective and hence

useful for simulating different pricing policies and the like,

including mode choice and adaption of agents towards new

modes, such as free-floating car sharing [9]. However, by

omitting explicit parking search, neither the time lost on

parking search is considered nor a possible increase in

congestion that affects other agents is assessed. Parking

search integration was also proposed [10], but apparently

never published.

AVs may have a disruptive effect on both private and

public transport, with several possible scenarios evolv-

ing from their introduction. On one hand, they offer an

enormous amount of opportunities as shared autonomous

vehicles (SAVs). These could offer taxi-like services in urban

areas and gradually help to reduce the number of vehicles

required to serve the demand. Recent simulation studies

suggest a replacement rate of up to 1:10, i.e., one SAV could

replace ten conventional cars [11]–[13]. Consequuently,

parking demand could be reduced significantly, both for

off- and on street parking [13], [14]. SAV user costs may be

on a similar level as car usage costs [15].

However, studies also suggest that SAV operations may

not be economically efficient for operators in areas of low

population density [16] and people will rather continue to

use their private vehicles instead, which may have full,

or at least some, autonomous functionalities, including

automated valet parking. Since the total number of vehicles

remains the overall same, the effect of private AVs on

parking is likely to be less beneficial than the use of shared

vehicles.

Also, there will be a phase where autonomous and

conventional vehicles will both be common in city traffic

and AVs. As to the question, how AVs, indifferent if shared

or private, may park once they are idle, most ideas are

speculative, including ideas of re-balancing fleet vehicles

[17]. Especially in areas with‘ high parking pressure and /

or cost, vehicle owners or fleet providers may chose to park

vehicles elsewhere or even let them cruise [18].

AVs, irrespective if shared or privately owned, may also

lead to a break through in car-to-car communication. This

may ultimately reduce parking cruise to a bare minimum,

as vehicles close to each other may share information about

free spaces in surrounding areas [19].

III. METHODOLOGY

In line with the parking choice approach described [10],

we also decided to use MATSim as a transport simulation

to integrate parking search behavior. The source code of

MATSim and its official extensions, including the park-

ing module used in this research, is open and available

at https://github.com/matsim-org/matsim. As an agent-

based, flexible and pluggable open-source software, its co-

evolutionary algorithms and customizable scoring functions

provide a very versatile base to extend an existing model

with parking search behavior [20]. Furthermore, it might

be possible to integrate some of the existing parking choice

scoring functionality described above with the approach, if

required.

The base concept behind the MATSim simulation is the

evolution of agents’ scores over multiple iterations, origi-

nating from a synthetic population created, for instance,

from census data. The score of an agent is summed up

based on a daily plan of performed activities (usually

positive) and traveling (usually negative) [21]. After each

iteration, a certain share of agents modify their plans (“day-

to-day replanning”). Typical modifications are changes of

departure times, travel modes and routes. If the modified

plan scores well, it is kept, otherwise discarded again. This

process is repeated over several iterations until a stochastic

user equilibrium is reached.

In order to model parking search, as well as other

processes that cannot be planned ahead for the whole

day, MATSim has been extended with several implemen-

tations that allow agents to change their behavior within

IEEE INTELLIGENT TRANSPORTATION SYSTEMS 3

an iteration. This feature is often referred to as within-

day replanning [22], [23] and has extended application

of MATSim to a wide range of different use cases, from

evaluating evacuation scenarios [24] through modeling con-

gestion effects of city-wide SAV fleets [25] to benchmarking

dynamic vehicle routing algorithms [26].

A. Simulation of parking search behavior

1) Simulation extension and modification: For an in-

tegration of parking search algorithms, a combination of

day-to-day and within-day replanning (the latter provided

by the DVRP extension [23]) needs to be used. Before

each iteration agents may include car trips into their daily

plans. These trips, or legs, need to be adjusted during

simulation runtime due to non-determinism of parking

search. Depending on the location of an agent’s vehicle, its

route may differ substantially between iterations and thus

needs be calculated ad-hoc.

2) Data and computational requirements: Apart from the

typical input data required for the MATSim simulation, the

number of parking spots on each link in the study area

is necessary. For links without this information, a direct

on-street parking spot is assumed. Due to the necessity to

route each agent ad-hoc, the computational requirements

are higher than in standard MATSim simulation runs.

3) Agent logic: During a typical MATSim iteration, an

agent starts traveling by car right after performing an activ-

ity. The route it travels along is either set at the beginning

of the iteration or comes from previous iterations. Upon

reaching its destination, the vehicle is removed from traffic

and the agent’s next activity starts. Thus, an agent may

be either in traveling (LEG) or activity performing state

(ACTIVITY ). See Fig. 1 (left) for an overview of this scheme.

In order to simulate parking search, the agent state space

needs to be adjusted. Namely, each car leg needs to be

split up into several stages: (1) determining the vehicle

location and walking there, (2) unparking the vehicle, (3)

route calculation and traveling to destination, including

searching for parking, (4) parking the vehicle, (5) walking

to destination. This means, a single car leg is split into

three legs and two activities (cf. to Fig. 1, center). A similar

approach is also used for the simulation of schedule-based

public transport in MATSim [27].

4) Parking search behavior: A person’s parking search

behavior may depend on several factors, such as the lo-

cation, the pricing of parking, personal experiences, the

willingness to park illegally, and many more [28]. Therefore

it is advisable to allow the search behavior to be agent-

specific. To achieve this, the search behavior is kept behind

an Interface with every agent having possibly a custom

implementation. MATSim’s finest granularity in terms of

traffic flow is link-based, allowing parking search to be

explicit on a link-to-link base.

In this paper, a simple random search logic is used,

as depicted in Fig. 2. Initially, an agent drives to their

destination along a path that has been pre-calculated upon

departure (a). Upon reaching the destination, it traverses

along a randomly selected sequence of neighboring links

to search for parking (b). Once a link has a free parking

spot, the vehicle gets parked (c) and the agent walks to its

destination (d). This behavior may be appropriate in areas

where there is a certain chance to find parking next to the

activity location making it worthwhile to look for a spot

here first.

B. Simulation of autonomous vehicles

MATSim comes with a set of extensions to simulate

dynamic modes, including shared autonomous vehicles, by

allowing dynamic dispatch of vehicles during the simulation

runtime, which, just like parking search, is another form

of within-day replanning. A detailed overview of these

extensions, including a description how to apply and adapt

the code, is available in [26]. Developed use cases include

simulation of taxi services [29], pooled DRT services [30]

and SAVs [11], [17], [31]. The basic principle behind all these

simulations is that upon departure of an agent, a vehicle

is assigned for picking up and transporting the agent to its

destination. The assignment of vehicles can have certain

constraints, depending on the use case. Empty vehicles may

be re-allocated or positioned at ranks, if required.

For the simulation of privately owned AVs, a dedicated

routing algorithm has been implemented. It assigns each

agent always its own private AV. Once an agent finishes

its activity and wants to travel by car, its private AV is

dispatched from the current location to pick the agent up.

Depending on the AV location, the agent has to wait for

some minutes for the car to arrive. After dropping off the

agent at the destination, the AV is idle and will proceed

according to one of the following options:

•Free spot – Remain where it is, should there be parking

available. In this case, the AV will not move from the

link where the customer was dropped off. Otherwise

find a nearby parking spot in the surroundings. In

this case, the vehicle will use the same parking search

algorithm (described in the previous section) as for

conventional vehicles.

•Garage – Proceed to the nearest not- fully occupied

parking lot or garage. Using this strategy, the AV will

continue to the closest designated AV garage.

•Cruise – In this case, the AV will cruise in circles around

the last drop off point until it is required again within

a defined radius around the point.

While the first option results in the parking search be-

havior typical to human drivers, garages are only used for

AVs in this paper. If these are designed to be used only

by AVs, they may be built in a less space-consuming way.

The possibility to let vehicles cruise instead of searching

for parking is an often-feared scenario, especially for areas

where parking charges are high.

Once an agent has finished its activity and requires the

vehicle again, it is dispatched from its current idle location

to pick the agent up. Depending on the vehicle location, the

agent has to wait for some minutes for the car to arrive.

All the software developed in this paper is available freely

under open-source licenses.

IEEE INTELLIGENT TRANSPORTATION SYSTEMS 4

Direct route

Parking search

Experienced path

Walk route

Pre-computed

path

Agent location

Vehicle location

Activity location Activity location

(a) (b)

Fig. 2. Parking search process

IV. SCENARIO ADAPTATION AND PARKING INTEGRATION

The parking search framework developed is applied to an

existing MATSim Berlin scenario [32]. Data about parking

spaces and their occupancy during nighttime is available

for a distinct area in the Charlottenburg district, around the

Klausenerplatz. The area is surrounded by a motorway to

the east and to major arterial roads in the north and south,

so car users tend to park their vehicle within these barriers.

Consequently, parking pressure in the area is known to be

high. Fig. 3 provides an overview of the study area. Roughly

4 000 curbside parking spaces exist in the area and are used

in the model. These were counted during a student project

in 2016, were also the overall parking room occupancy in

the area was analyzed.

For computational reasons, the synthetic population of

the original scenario with 6 million agents was reduced to

those agents who perform at least one activity in the area or

its immediate surroundings, leaving roughly 37 000 agents.

Travel times on links outside the area were assumed to be

the same as in the base case using dynamically changing

network attributes. In order to evaluate the influence of

AVs on the parking situation, the simulation was run with

several degrees of AV usage: 0, 10 and 20 % of the popula-

tion was equipped with an AV. The case where AVs are not

present is referred to as the base case.

In all runs, 50 iterations were simulated. Agents had

Study area

AV garage

location

Fig. 3. Study area and the designated AV garage location

the choice of modifying their departure times within a 15-

minutes interval. For parking locations, iterations were seen

as days. This means an agent picks up a car in the morning

where it was parked last in the preceding iteration. For

AVs, all of the parking strategies defined in the previous

section (III-B) where used. Furthermore, a simulation was

run where AVs randomly (uniform distribution) pick one

of the idle strategies. There is a single garage of unlimited

IEEE INTELLIGENT TRANSPORTATION SYSTEMS 5

capacity to park AVs at the eastern edge of the study area.

For the cruising strategy, AVs may circle with a radius of up

to 2 000 meters around the position where they dropped

off their owner at the allowed cruising speed.

V. SIMULATION RESULTS

The introduction of AVs with self-parking and cruising

capabilities may have several impacts. Firstly, there are

individual impacts for users of both conventional vehicles

and AVs: depending on the parking strategy AVs choose,

they may or may not compete for parking space with

conventional cars and may have an impact on the parking

search duration of drivers of ordinary vehicles. Furthermore,

the waiting time for owners of AVs may differ depending

on the location of the idle vehicle. A nearby parking spot

may be more favorable here. Finally, there are effects on

the transport network as a whole: AVs that cruise or go to

parking in a special garage produce additional idle mileage,

which may be partly compensated by conventional vehicles

that now have a shorter parking search duration. A com-

prehensive overview of the simulation results is provided in

Table I.

A. Parking search duration

In the base case, the average duration to find a parking

space is 8:50 minutes. This time varies during the day and

is particularly long during late morning hours, when people

commute into the area for work or other activities. During

this peak hour, the average search time is almost 12 minutes

(see Fig. 4).

When AVs are introduced, parking search time for con-

ventional cars generally reduces. As anticipated, search

times are usually lower in the 20 % AV scenarios than in

the 10 % cases.

In the cases where AVs share free parking slots with

conventional vehicles (marked orange in Fig. 4), search

times are roughly as long as in the base case.

Search times notably drop by more than 30 seconds in

the 10 % case and over 90 seconds in the 20 % case when

AVs are not parking but are sent to cruise around (marked

yellow in Fig. 4). Parking pressure is reduced significantly.

The usage of a dedicated garage facility for AVs can

further improve the search time (marked blue in Fig. 4),

which is then calculated to be 8:06 minutes for a 10 %

AV share or 7:05 minutes for the 20 % AV share. While

both demand and supply of parking spaces are obviously

similar in the garage and in the cruise case, the AVs

cause additional congestion in the cruise scenario, which

increases the overall search time for car users.

The random selection of a strategy (marked in gray)

scores somewhere in between the cruise and the parking

slot strategy.

B. Wait vs. walk times

For AV users, the time it takes to recover a vehicle from

its idle location can be used to some extent productively, as

opposed to a walk to the car, which only may be perceived

5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23

Duration (minutes)

Time

Parking search durations

Find free Spot 10% Random 10% Cruising 10% Garage 10% Base Case - all car

Find free Spot 20% Random 20% Cruising 20% Garage 20%

Fig. 4. Parking search durations in the base case and with different levels

of AV usage and AV parking strategies

as a positive thing, but is hard to measure in an economical

sense. Despite this, long waiting times are undesirable,

especially for spontaneous departures and the waiting time

should therefore be short.

Both in the 10 % and 20 % scenario, wait times are

roughly the same within each strategy. For the random,

garage and nearest parking spot strategies these are gen-

erally similar in the range of 3:00 to 3:30 minutes. The

waiting time is higher for the cruise strategy, with around

five minutes. Obviously, this value could be smaller if the

cruising radius was set to a lower value. However, all in

all, wait times are short in all scenarios and will most

likely not be of real significance for AV owners. From the

car drivers’ perspective, walking between the vehicle and

activity location is often as important as parking search,

especially because the walk is made both ways. As long as

AVs compete with conventional cars for curbside parking,

the walk times remain at the same levels as in the base

case. However, once AVs are ordered either to park at

the designated garage or cruise, the availability of parking

spaces increases and thus the time spent by car drivers on

both finding a free slot and walking to/from the vehicle

decreases. It should be pointed out that savings made on

the walk time ought to be counted twice (leaving the car

and getting back). In the 20% Garage scenario, for instance,

the total walk time per each activity is reduced by more

than a minute compared to the base scenario. However,

despite some improvements resulting from AVs being sent

to the designated garage, non-AV drivers still need to spend

approximately 20 minutes on searching and walking (per

activity), which is much more than around 3.5 minutes of

waiting for the AV to come.

C. Vehicle kilometers

Due to their capability of driving without a passenger

onboard, AVs will increase the driven distances. This implies

that the vehicle kilometers traveled will increase in all

AV scenarios, mainly due to pick up trips, where agents

otherwise would have walked. Compared to the base case,

the additional empty mileage is around 2 % in the free slot

IEEE INTELLIGENT TRANSPORTATION SYSTEMS 6

TABLE I

SIMULATION RESULTS FOR DIFFERENT SHARES OF AVS AND PARKING SEARCH STRATEGIES

AV Share AV Parking Strategy

Average parking

search duration

(AV and car)

Average wait

time for AV

Average walk

time to/from car VKT car VKT AV total VKT

[mm:ss] [mm:ss] [mm:ss] [km] [km] [km]

0% base case 08:50 n/a 7:10 521 686 0 521 686

10% Free Spot 08:45 03:22 7:10 466 556 67 234 533 790

10% Garage 08:06 03:35 6:56 465 424 67 987 533 411

10% Cruise 08:16 05:09 6:58 465 667 428 827 894 494

10% Random 08:25 03:15 7:03 465 865 166 102 631 967

20% Free Slot 08:23 03:28 7:02 412 370 133 494 545 864

20% Garage 07:05 03:39 6:37 410 375 134 673 545 048

20% Cruise 07:15 04:54 6:42 410 560 684 854 1 095 414

20% Random 07:39 03:23 6:49 411 220 295 078 706 298

and garage strategies in the 10 % AV scenarios and around

4 % in the 20 % scenarios. This additional empty mileage

can most likely be coped easily due to the expected road

capacity increase, as literature suggests [25]. For most links

in the study area, the increase in daily volumes per link

is less than 200 for a 20 % AV share, as Fig. 5 shows for

these strategies. In the cruise scenario, the mileage driven

by all is increased by almost 80 % in the 10 % scenario and

more than doubled in the 20 % case. This shows clearly the

possible impact should vehicle owners be allowed to use

such strategies. This increase in mileage spreads along all

links in the network. For the study area, arterial roads would

have to cope with more than 4 000 additional vehicles per

day, and more than 1 000 additional vehicles in residential

streets, which is an increase of two to three times compared

to the base case.

VI. CONCLUSION

With the integration of parking search into a multi-agent

transport simulation we were able to show the influence

Find free Spot

Parking Lot Random

Cruise

≥ 4000

2000

1000

500

200

≤ 0

Additional daily

vehicles per link

Fig. 5. Changes in daily traffic volumes in the 20 % AV case compared to the base case.

IEEE INTELLIGENT TRANSPORTATION SYSTEMS 7

autonomous vehicles may have on city life. Only some

possible, of all the enormous capabilites AVs have, are

discussed here. These show both the potential of AVs to

solve parking problems in cities, but also highlight new

problems arising from them. Of the strategies tested in this

paper, sending AVs to a designated parking infrastructure

seems to be most promising: The negative effects of parking

lots, namely long access ways, can be overcome when

using AVs, and parking search for conventional vehicles

can be decreased. With a higher share of AVs, the re-

moval of curbside parking locations may also be discussed.

These positive side effects can easily compensate additional

mileage created by empty driving AVs. Should AVs simply

start searching for parking spaces like conventional vehicles,

the parking search situation will remain as it is now. This

would be only beneficial for AV users, who do not have to

spend time on parking search.

The option most feared in literature, namely an endless

cruising of vehicles, does not seem to be beneficial for AV

users themselves, as waiting times to recover the vehicle

are higher due to congestion and the actual vehicle’s loca-

tion. Additionally, cruising vehicles create significant direct

operating costs, which could only be recovered if parking

charges are even higher than these. This, however, seems

not very probable, considering that vehicle owners may let

their vehicles use cheaper parking lots farther away.

This also leads to additional research questions. These

include the choice between AV parking policies if costs for

parking and externalities are included into the model and

passed on to their owners. Further research should also deal

with a bigger area and a city-wide model, which may lead to

more interactions between agents. Working on both points

will also allow to provide more explicit public parking policy

recommendations for AVs.

ACKNOWLEDGMENT

The authors would like to thank the BMW Group for co-

funding parts of this paper.

REFERENCES

[1] J. Bischoff and K. Nagel, “Integrating explicit parking search into

a transport simulation,” Procedia Computer Science, vol. 109, pp.

881–886, 2017. [Online]. Available: http://www.sciencedirect.com/

science/article/pii/S1877050917310906

[2] K. Axhausen, “Ortskenntnis und Parkplatzwahlverhalten, report to

the Deutsche Forschungsgemeinschaft,” Institut für Verkehrswesen,

Universität (TH) Karlsruhe, Karlsruhe, 1989.

[3] J. Polak and K. Axhausen, “Parking search behaviour: A review

of current research and future prospects,” Transport Studies Unit,

University of Oxford, Working Paper 540, 1990.

[4] D. C. Shoup, “Cruising for parking,” Transport Policy, vol. 13,

no. 6, pp. 479–486, 2006, parking. [Online]. Available: http:

//www.sciencedirect.com/science/article/pii/S0967070X06000448

[5] I. Benenson, K. Martens, and S. Birfir, “Parkagent: An agent-based

model of parking in the city,” Computers, Environment and Urban

Systems, vol. 32, no. 6, pp. 431–439, 2008, geoComputation: Modeling

with spatial agents. [Online]. Available: http://www.sciencedirect.

com/science/article/pii/S0198971508000689

[6] K. Spitaels, S. Maerivoet, G. De Ceuster, G. Nijs, V. Clette, P. Lannoy,

K. Dieussaert, K. Aerts, and T. Steenberghen, “Optimising price

and location of parking in cities under a sustainability constraint

(sustapark),” Belgian Science Policy, vol. Final Report, 2009.

[7] A. Horni, K. Nagel, and K. W. Axhausen, Eds., The Multi-Agent

Transport Simulation MATSim. Ubiquity, London, 2016. [Online].

Available: http://matsim.org/the-book

[8] R. A. Waraich, “Parking,” in The Multi-Agent Transport Simulation

MATSim, A. Horni, K. Nagel, and K. W. Axhausen, Eds. Ubiquity,

London, 2016, ch. 13. [Online]. Available: http://matsim.org/the-book

[9] M. Balac, F. Ciari, and R. A. Waraich, “Modeling the impact of

parking price policy on free-floating carsharing: Case study for Zurich,

Switzerland,” in World Conference on Transport Research - WCTR 2016

Shanghai, 2016.

[10] R. A. Waraich, “Modelling parking search behaviour with an agent-

based approach,” International Conference on Travel Behaviour Re-

search (IABTR), 2012.

[11] J. Bischoff and M. Maciejewski, “Simulation of city-wide replacement

of private cars with autonomous taxis in Berlin,” Procedia Computer

Science, vol. 83, pp. 237–244, 2016. [Online]. Available: http:

//www.sciencedirect.com/science/article/pii/S1877050916301442

[12] D. J. Fagnant and K. M. Kockelman, “The travel and environmental

implications of shared autonomous vehicles, using agent-based

model scenarios,” Transportation Research Part C: Emerging

Technologies, vol. 40, pp. 1–13, 2014. [Online]. Available: http:

//www.sciencedirect.com/science/article/pii/S0968090X13002581

[13] L. Martinez, “Urban Mobility System Upgrade – How shared self-

driving cars could change city traffic,” International Transport Forum,

Tech. Rep., 2015.

[14] W. Zhang, S. Guhathakurta, J. Fang, and G. Zhang, “Exploring the

impact of shared autonomous vehicles on urban parking demand:

An agent-based simulation approach,” Sustainable Cities and Society,

vol. 19, no. Supplement C, pp. 34–45, 2015. [Online]. Available: http:

//www.sciencedirect.com/science/article/pii/S221067071530010X

[15] P. Boesch, F. Ciari, and K. Axhausen, “Required autonomous vehicle

fleet sizes to serve different levels of demand,” Transport Research

Board, Tech. Rep., 2016.

[16] J. Bischoff and M. Maciejewski, “Autonomous taxicabs in Berlin –

a spatiotemporal analysis of service performance,” Transportation

Research Procedia, vol. 19, pp. 176–186, 2016.

[17] K. Winter, O. Cats, B. van Arem, and K. Martens, “Impact of relocation

strategies for a fleet of shared automated vehicles on service effi-

ciency, effectiveness and externalities,” in 2017 5th IEEE International

Conference on Models and Technologies for Intelligent Transportation

Systems (MT-ITS), Jun. 2017, pp. 844–849.

[18] T. Litman, “Autonomous Vehicle Implementation Predictions,”

Victoria Transport Policy Institute, 2017. [Online]. Available: http:

//www.vtpi.org/avip.pdf

[19] M. Caliskan, A. Barthels, B. Scheuermann, and M. Mauve, “Predicting

parking lot occupancy in vehicular ad hoc networks,” in 2007 IEEE

65th Vehicular Technology Conference - VTC2007-Spring, Apr. 2007,

pp. 277–281.

[20] A. Horni, K. Nagel, and K. W. Axhausen, “Introducing MATSim,” in

The Multi-Agent Transport Simulation MATSim, A. Horni, K. Nagel,

and K. W. Axhausen, Eds. Ubiquity, London, 2016, ch. 1. [Online].

Available: http://matsim.org/the-book

[21] K. Nagel, B. Kickhöfer, A. Horni, and D. Charypar, “A closer look at

scoring,” in The Multi-Agent Transport Simulation MATSim, A. Horni,

K. Nagel, and K. W. Axhausen, Eds. Ubiquity, London, 2016, ch. 3.

[Online]. Available: http://matsim.org/the-book

[22] C. Dobler and K. Nagel, “Within-day replanning,” ch. 30.

[23] M. Maciejewski, “Dynamic transport services,” in The Multi-Agent

Transport Simulation MATSim, A. Horni, K. Nagel, and K. W.

Axhausen, Eds. Ubiquity, London, 2016, ch. 23. [Online]. Available:

http://matsim.org/the-book

[24] G. Lämmel, D. Grether, and K. Nagel, “The representation and imple-

mentation of time-dependent inundation in large-scale microscopic

evacuation simulations,” Transportation Research Part C: Emerging

Technologies, vol. 18, no. 1, pp. 84–98, 2010.

[25] M. Maciejewski and J. Bischoff, “Congestion effects of autonomous

taxi fleets,” Transport, vol. 0, no. 0, pp. 1–10, 2017. [Online].

Available: http://dx.doi.org/10.3846/16484142.2017.1347827

[26] M. Maciejewski, J. Bischoff, S. Hörl, and K. Nagel, “Towards a

testbed for dynamic vehicle routing algorithms,” in Highlights of

Practical Applications of Cyber-Physical Multi-Agent Systems: Interna-

tional Workshops of PAAMS 2017, Porto, Portugal, June 21-23, 2017,

Proceedings, J. Bajo, Z. Vale, K. Hallenborg, A. P. Rocha, P. Mathieu,

P. Pawlewski, E. Del Val, P. Novais, F. Lopes, N. D. Duque Méndez,

V. Julián, and J. Holmgren, Eds. Springer International Publishing,

2017, pp. 69–79.

IEEE INTELLIGENT TRANSPORTATION SYSTEMS 8

[27] M. Rieser, “Adding Transit to an Agent-Based Transportation Simula-

tion: Concepts and Implementation,” Ph.D. dissertation, TU Berlin,

Berlin, 2010.

[28] S. Belloche, “On-street parking search time modelling and validation

with survey-based data,” Transportation Research Procedia, vol. 6, pp.

313–324, 2015.

[29] M. Maciejewski, J. Bischoff, and K. Nagel, “An assignment-based

approach to efficient real-time city-scale taxi dispatching,” IEEE

Intelligent Systems, vol. 31, no. 1, pp. 68–77, Jan. 2016.

[30] J. Bischoff, M. Maciejewski, and K. Nagel, “City-wide Shared Taxis: A

Simulation Study in Berlin,” submitted to ITSC 2017, Tech. Rep., 2017,

available http://www.vsp.tu-berlin.de/publications/vspwp/ as WP 17-

11.

[31] S. Hörl, “Agent-based simulation of autonomous taxi services with

dynamic demand responses,” Procedia Computer Science, vol. 109,

pp. 899–904, 2017. [Online]. Available: http://www.sciencedirect.

com/science/article/pii/S1877050917310943

[32] A. Neumann, “Berlin i: Bvg scenario,” in The Multi-Agent

Transport Simulation MATSim, A. Horni, K. Nagel, and K. W.

Axhausen, Eds. Ubiquity, London, 2016, ch. 53. [Online]. Available:

http://matsim.org/the-book

Joschka Bischoff is a research associate and PhD

candidate working at the department for trans-

port systems planning and transport telematics

at TU Berlin. His main field of expertise is the

simulation of dynamic transport modes, including

autonomous vehicles and taxis. Previously, he has

studied transportation planning and operations.

Michal Maciejewski is an assistant professor in

Department of Transport Systems at Poznan Uni-

versity of Technology and a senior researcher in

the Department of Transport Systems Planning

and Transport Telematics at TU Berlin. His re-

search focuses on dynamic vehicle routing, on-

demand transport services, e-mobility and au-

tonomous vehicles.

Tilmann Schlenther Tilmann Schlenther is a grad-

uate student at TU Berlin and co-worker at the

department for transport systems planning and

transport telematics at TU Berlin. He has been

working on integrating parking data and the sim-

ulation of parking search behavior intensively.

Kai Nagel Kai Nagel is professor for transport

systems planning and transport telematics at TU

Berlin, specializing in modelling and large-scale

simulation of travel behavior and traffic flow. He

has a PhD in Computer Science from the Uni-

versity of Cologne; from 1995 to 1999 he was at

Los Alamos National Laboratory as part of the

TRANSIMS team. He is one of the creators of

MATSim.