Evaluating SimPark as a parking policy analysis tool
SUBMITTED IN PARTIAL FULLFILLMENT FOR THE DEGREE OF MASTER
OF SCIENCE
Milou Bisseling
10427538
M
ASTER
I
NFORMATION
S
TUDIES
Data Science
F
ACULTY OF
S
CIENCE
U
NIVERSITY OF
A
MSTERDAM
25-06-2019
!
1
stExaminer
2
ndExaminer
dr. Elenna Dugundji
dr. Bas Schotten
Faculty of Science, Vrije Universiteit
Gemeente Amsterdam
Evaluating SimPark as a parking policy analysis tool
Milou Bisseling
miloubisseling@gmail.com University of Amsterdam
ABSTRACT
In this paper, the use of the agent-based parking model SimPark as a parking policy tool will be explored in a neighborhood with high parking demand. An existing case of removing on-street parking places will be examined with the model. �e e�ect of this policy on parking pressure and cruising for parking obtained from the model will be compared to the real-world data. �e neighborhood that is used investigated is the Frans Halsbuurt, which is located in Amsterdam, the Netherlands. In this neighborhood, all on-street parking places are moved to a newly built parking garage. �is transition from on-street to o�-street parking went through di�er-ent phases, starting in a situation with 587 on-street parking places. A�er the opening of the parking garage in May 2018, 600 o�-street parking places were built resulting in a total of 1167 parking places in the neighborhood. Finally, several months a�er the opening of the parking garage, most on-street parking places were removed. �ese di�erent situations will be simulated with an agent-based parking model named SimPark. SimPark is developed to simulate parking policy and to capture the e�ects on tra�c �ow and parking (behavior). �e simulation results will be compared with real data to examine whether SimPark is a useful policy tool for evaluating parking policy.
KEYWORDS
parking, parking policy, transport simulation, agent-based model, parking choice, parking behavior, Amsterdam, SimPark
1 INTRODUCTION
Over the past decades, urban parking policy has gained importance as car ownership increased rapidly and public space in the city got scarce [7]. According to Mingardo[9], who wrote an overview about parking policy in Europe, the evolution of parking policy can be categorized in di�erent phases. In the �rst phase, there is no parking regulation at all. �en in the second phase, a�er car ownership and congestion in cities rose, parking tari�s were introduced. Nowadays we are in between the second and the third phase: parking policy starts to become more pro-active, as policymakers try to manage parking demand in a strategic manner. Still, parking policy is mainly managed by parking managers, whose focus is operational and less based on data analysis. To get into this third phase, policymakers, strategic transport planners, parking managers, and academics should cooperate to produce a more evidence-based, cost-e�ective parking policy.
One of the main challenges featured by Mingardo [9] is the trend towards building o�-street parking facilities to improve accessibil-ity, air quality and the quality of the living environment in cities. As building o�-street parking facilities is expensive, the consequences
of replacing on-street parking places by o�-street parking places need to be examined carefully.
In line with the trend of moving parking places o�-street, the current city council of Amsterdam decided to remove 7000 to 10000 on-street parking places before 2025 [5]. A part of these removed parking places will be replaced by o�-street parking places in park-ing garages. �e city of Amsterdam has several projects where on-street parking places are moved o�-street. One of those projects is located in the Frans Halsbuurt, where it is decided by the munic-ipality together with the residents of the neighborhood to remove all on-street parking places. A�er the removal, people can park in the newly built parking garage. �e project had di�erent phases: in the �rst phase, only the on-street parking places were open. In the next phase, both o�-street and on-street were available. Finally, in the third phase, 85 percent of the on-street parking places were removed.
Despite the fact that parking policy will only become more im-portant over the upcoming years, up until recent years modeling has not played a big role in assessing urban parking policy in prac-tice. [8]. Several agent-based models have been developed whereby parking and parking behavior is simulated. According to Martens and Benenson[8], parking models can be divided into two di�erent categories. �e �rst category contains spatially implicit models wherein the choice of parking type or parking spot is predicted based on stated preferences. �e other category, spatially explicit models, simulate parking search and parking choice behavior. In the la�er, the whole process is simulated; from searching for a parking place to making the decision whether to continue searching or to stop searching and park the car [22]. In this manner, consequences for tra�c �ow, as for example congestion, can also be analyzed.
In this paper, a newly developed spatially explicit parking model SimPark by Vuurstaek et al.[22] is evaluated. According to Vu-urstaek, SimPark allows for more detailed parking simulations than the existing models. SimPark tries to simulate parking behavior as close to reality as possible by including the possibility to specify parking facilities and corresponding regulations. To simulate park-ing behavior the behavioral parkpark-ing choice model from Khaliq [6] is used which is based on discrete choice modeling. �e research is based on data from Belgian residents and shows that factors such as parking cost signi�cantly a�ect parking choice behavior.
�e goal of this research is to determine whether the agent-based parking simulation model SimPark can be used for analyzing the impact of parking policy on parking occupancy and cruising for parking which is the search time for a parking place. �e policy scenario will be assessed whereby on-street parking places will be moved o�-street in the Frans Halsbuurt in Amsterdam, the Netherlands. �e research will be divided into the following sub-questions:
(1) How can a population of agents be generated for SimPark from the Dutch travel survey data (Onderzoek Verplaatsin-gen in Nederland OViN)?
(2) Is the behavioral parking choice model from Khaliq suitable for the di�erent parking scenarios in the Frans Halsbuurt? (3) What are the consequences of moving on-street parking places o�-street in the Frans Halsbuurt for parking occu-pancy and cruising for parking?
(4) How do the results of SimPark regarding parking occu-pancy compare to the parking ticketing sales (NPR) data and to the parking garage data?
�e paper is structured in di�erent parts. In the related work, the literature relevant to the research questions is described while com-paring di�erent agent-based parking and parking choice models. Next, in the methodology, the used data and the methods used to answer the research questions are provided. A�er the methodology, the results of all research questions are discussed, to end with a conclusion and a discussion.
2 RELATED WORK
2.1 Agent-based parking models
�e most agent-based parking models share some sort of a similar structure as shown in �gure 1. Firstly, they need a road network to know where and in which directions agents can drive (see last red block from above in �gure 1).
Secondly, models need travel diaries of people: trips made by agents (see �rst red block from above in �gure 1). An example input entry can look as the following: Agent x starts its �rst trip from home to work at 7 a.m.. It arrives at work a�er one hour of driving at 8 a.m. and works for 8 hours. A�er work, it goes back home again at 5 p.m. and arrives at home at 6 p.m. �e locations, arrival- and departure times and activities that agent x visits are all stored in a person’s x travel diary. �e agent’s travel diary consists of two entries, as agent x makes two trips: one from home to work, and the other one, back home again.
�irdly, in some of the models, the type of car is assigned to agent x. �is input can contain speci�cations about the car the agent is driving in as for example the dimensions, the capacity, or the fuel type of the car (see third red block from above in �gure 1). As agent x needs to know where it can park, parking facilities and regulations are added to the model (see second red block from above in �gure 1). Agent x will start its �rst trip from home, parked somewhere near its home. �en it will drive to work and will try to �nd the best parking place according to its preferences. As agent x drives an electric car, he prefers to park at an electric charging place, so he will search the parking place where his demands are met best. �e choices agent x makes will be determined by the route choice algorithm and the parking choice behavior algorithm that is given as input to the model (see orange blocks in �gure 1). As output, most models give a log �le from all actions taken by every agent in the model. For example, agent x chooses to skip the �rst ten parking places it sees, to park at the 11th parking place. Every action from agent x is logged, therefore, it is possible to obtain statistics such as the occupancy rate, and cruising for parking (the
amount of search time agent x needed to �nd a parking place) (see blue blocks in �gure 1).
Figure 1: System architecture SimPark
2.1.1 PARKAGENT.One of the �rst real-world frameworks of a spatially explicit model is called PARKAGENT and is developed by Martens and Benenson[8]. �ey analyzed parking search and choice behavior in Tel Aviv with an agent-based parking model. With this model, it is possible to analyze parking dynamics and to assessparking policies and evaluate them. �e model works with GIS data layers as input containing information about the network, vehicles, parking regulations, parking facilities, and destinations with the number of agents that visit the destination at a certain time.
2.1.2 SUSTAPARK.�e second model that is discussed here, is SUSTAPARK developed by Dieussaert et al.[13]. �is model is developed to evaluate the tra�c e�ects of parking search behavior in the city as a whole. �is model is also based on GIS layers for the spatial data, but the non-spatial data containing agents, activity schedules, and parking behavior is stored in an Access-database.
2.1.3 MATSIM extensions.�e last two models described are two MATSIM extensions, the �rst by Waraich and Axhausen [23] and the second by Bischo� and Nagel [3]. �e model by Waraich and Axhausen is focused on evaluating e�ects on parking policy regarding parking capacity and pricing to help design parking policy. �e MATSIM extension by Bischo� and Nagel focuses on adding constraints on parking regarding walking distance, parking supply, parking cost, and in addition, the possibility of car-sharing has been added in this model.
2.1.4 SimPark.SimPark is developed by Vuurstaek et al.[22] and aims to obtain a more realistic simulation compared to the existing models by adding more speci�cation options to the input data. �e goal of the extension of characteristics is to make the tool suitable for analyzing more complex parking policies.
In �gure 1 a system architecture of SimPark is given. �e model works with input �les in XML format, which are o�en intercon-nected. �ose input �les are added to the model together with algorithms for parking choice strategy and route choice behavior. �e former contains a set of rules about the parking search strategy of agents. �e la�er speci�es the kind of route the agent takes to the destination. Finally, the output that the model generates can
be set to di�erent kinds of statistics such as the occupancy rate per hour or the cruising for parking per car.
In the next paragraphs, the di�erences between the above-named agent-based parking models are explained regarding the di�erent sorts of input data.
2.2 Agents, trips, and vehicles
�e papers presenting the models discussed in section 2.1 do not specify how trips of agents are generated, neither do they give the source of the activity pa�erns. In Vuurstaek’s �rst design of SimPark[22], the trips of the agents in the model are predicted by FEATHERS: an activity-based transport demand model[1]. �ose predictions are made based on the Belgian travel survey OVG (On-derzoek VerplaatsingsGedrag). Instead of using FEATHERS, it is also possible to use other schedule generators as long as they pro-duce results in the right XML format.
All trips in the travel diaries of the agents need to have an origin and destination, and the time of departure and arrival. A�er the trip, an activity follows with the time, the location, and type of the activity. In addition, social demographics such as age, gender, employment, and other characteristics can be added to the agent.
�e vehicles used in most of the above-named agent-based park-ing models cannot be speci�ed with characteristics. In both PARK-AGENT and SUSTAPARK it is not possible to add car dimensions, the capacity of the car, or the type of fuel[2, 4, 22]. In the MATSIM extensions, it is possible to specify car dimensions, but the papers do not mention the use of those dimensions [3, 22, 23]. In SimPark it is possible to specify these kinds of characteristics, which are used as constraints at parking places. In a parking garage, it is for example possible that only cars with a maximum height of 2 meters can enter the parking garage. Furthermore, the speci�cation of motor type can, for example, be used for parking places for electric cars.
2.3 Parking regulations
As the main goal of most agent-based parking models is to provide a policy analysis tool, it is important to compare the possibilities of evaluating di�erent parking policies. In all agent-based parking models mentioned above, it is possible to evaluate parking poli-cies a�ecting parking infrastructure such as evaluating di�erent amounts of parking places [22]. When evaluating policies a�ecting (price) regulations however, the possibilities of most models are limited. Just as in the case of the model of Bischo� and Nagel where all parking places are free, evaluation of policies a�ecting prices is not possible.
In PARKAGENT, it is only possible to evaluate price-related parking policies through �xed hourly rates [2, 22]. SUSTAPARK can evaluate parking policies related to prices or for people with parking permits. Other parking policies a�ecting parking regulations are not possible in this model [4, 22].
In SimPark, all input data can be speci�ed to a very detailed level. It is possible to assign parking regulations and other restrictions on the parking place level. A parking regulation can contain a maximum duration, allowed fuel types, properties of the driver (disabled, permit holder, or non-permit holder), time of the day et cetera. Furthermore, the parking fare can also be speci�ed to depend
on the region, time of the day, duration of parking, properties of the driver and vehicle. SimPark evaluates all rules that apply to a particular parking place and determines which regulation is valid for the particular agent. SimPark will always select the regulation with the lowest price, when di�erent regulations apply to the agent or to the parking place [22].
2.4 Parking search behavior
To get an accurate simulation from an agent-based parking model, agents need a realistic parking choice process. Every time the agent passes an empty parking space, the agent has to make a decision. It can choose to park or to continue driving. To simulate this behavior of agents, several parking place search strategies can be applied.
2.4.1 Take-the-first-empty-place methods.In the paper of Bischo� and Nagel[3], simple logic to parking behavior is applied. When the agent arrives within a certain range of the destination, and the occupancy rate in the neighborhood passes a threshold, the agent will choose the �rst empty parking place.
In PARKAGENT, when the agent approaches its destination, the agent starts to continuously re-estimate the occupancy rate on the road to the destination. When it estimates a high occupancy rate in the streets closer to the destination, the agent will decide to take the �rst empty parking place [2]. In case the agent has passed the destination, it decides to take the �rst empty place a�er passing the destination.
2.4.2 Utility-based methods.Another way to simulate parking behavior is by assigning a value to each parking place. For every moment in time, the parking place has a di�erent value for the agent, this value is called utility. Every agent will try to maximize its utility when deciding on which parking space to choose. A factor in�uencing utility can, for example, be the distance to the destination or the crowdedness in the street. �ose factors can also vary per agent: a resident might �nd it more important to park close by its destination than a tourist. In SUSTAPARK, the utility of a parking place is calculated continuously based on various factors (for example available parking spaces, price, distance, and search time)[4].
Waraich and Axhausen[23] also work with assigning utility to parking places but in a di�erent manner. When the agent starts driving, the model selects the empty parking place with the highest utility to the agent. A�er selection of this place, the agent will park there. In this way, the agent will always park at the place where it can get the highest utility. In reality, it is not likely that people will always drive to the place with the highest utility, but according to Waraich and Axhausen, it will still be possible to capture elements as parking capacity and pricing for evaluating parking policy.
To determine which factors a�ect the parking choice of an agent, discrete choice modeling can be used. Discrete choice models cal-culate the utility of alternative choices. An agent will choose the alternative with the highest possible utility. Khaliq[6] developed a model for parking choice behavior of agents for PARKAGENT using discrete choice modeling. �rough a stated choice question-naire taken in Hasselt (Belgium), Khaliq analyzes which factors in a street have a signi�cant e�ect on parking.
�e results of her model show that factors as parking cost, oc-cupancy, distance to destination, and security in the street a�ect parking choice behavior of agents. �e predicted values of the factors will be used as parameters in the utility function. When the agent starts to come close to the destination, the agent will start evaluating the above-named factors to calculate utility. Based on this utility the agent will decide to continue driving to park in that street. Khaliq’s model calculates the utility per street. So for every street, the utility is calculated. Based on the utility, an agent decides whether to park in this street or to continue driving.
2.5 Route choice behavior
In PARKAGENT two di�erent route choice algorithms can be ap-plied. �e �rst one is Dijkstra’s shortest path algorithm and the second one is a heuristic algorithm developed by Benenson himself [2]. �is second algorithm chooses the streetsegment which end is closest to the destination. Another way to simulate route choice is applied by Bischo� and Nagel, a random path is taken to the parking places before the agent starts driving [3].
In SimPark, when the agent is approaching its destination, the agent will start driving blocks near the destination. It will continue driving blocks until the agent is parked somewhere, except for when the search time will have exceeded a maximum. In that case, the agent will give up searching on-street parking and drive straight to the parking garage. In Khaliq’s paper this maximum amount of search time is set to 10 minutes.
3 METHODOLOGY
3.1 Description of the data
3.1.1 Travel diaries.�e data that is used for modeling car trips is obtained from the Dutch travel Survey OViN (Onderzoek Ver-plaatsingen in Nederland) executed on a yearly basis by the Dutch statistics bureau CBS (Centraal Bureau voor Statistiek). �is dataset contains information on the travel behavior of the Dutch popula-tion. For this research, respondents are asked to keep track of all the trips they make during one day of the year. All movements made by the respondent are described according to the place of origin and destination, the time, the type of transport, the motive for the move-ment (e.g. home, work, leisure, etc.), and multiple other variables. �e total sample size from the OViN data is around 0.25 percent of the total target population which contains approximately 16 million people [21]. To the right amount of agents for the simulation all data from 2010 to 2017 is combined[14–21], under the assumption that there were no major changes in the neighborhood or in the population over the past ten years. �e estimation of how many agents need to be in the simulation will be explained in section 4.1. 3.1.2 Parking facilities.For the location of parking facilities, data provided by the City of Amsterdam is used. �is data contains the latitude and longitude of all parking places for each part of the city. �e exact point of the latitude and longitude of the park-ing place is located in the middle of the parkpark-ing place, which is called the centroid. �e dataset is updated whenever the city gets a noti�cation of the fact that the parking situation has changed.
�is means that changes are not processed directly, but only when somebody reports the change to the city.
In addition to the latitude and longitude, the type of parking is also speci�ed. �ere are three di�erent types of parking places: diagonal, perpendicular and parallel parking. Hereby angle park-ing places are diagonal parkpark-ing places, and perpendicular parkpark-ing places have a 90-degree angle to the road. �ese di�erent types of parking places have di�erent minimum dimensions. An angle or perpendicular parking place has to be 4.5 x 2.4m at minimum, and a parallel parking place needs to be at least 6 x 2.5m [10].
3.1.3 Parking regulations.�e parking regulations data is also provided by the City of Amsterdam. All recent parking regulations can be found online, but as the regulations have changed in April 2019, parking regulations from before this date were used for the Frans Hals buurt case. On-street parking in the blue zone (see �gure 2) costs 4 euros per hour on weekdays and Saturdays, on Sundays parking is free. Starting from 14 April 2019 on-street parking in the blue zone will cost 6 euros per hour.
O�-street parking in the Albert Cuyp garage before April 2019 costs 0.5 euro per 7 minutes with a maximum of 47.50 euros. A�er April 2019 this will cost 0.17 euro per 2 minutes with a maximum of 51 euros. As all scenarios of the Frans Halsbuurt case date from before April 2019, the old tari�s are applied.
Figure 2: Parking tari�s in Amsterdam. �e Frans Halsbu-urt is located in the blue zone at the point with the red �ag with the white cross.
3.1.4 Vehicles.For the vehicle data in this case, it is decided to use one type of vehicle for simplicity. �is type of car will be assigned to every agent in the simulation. When choosing di�erent types of cars per agent, the fact that that cars have to �t in the parking places in the model has to be taken into account.
3.1.5 Network.�e road network is obtained through Open-StreetMap (OSM). OpenOpen-StreetMap is a voluntary collaborative project which provides free geographical information. �e Frans Halsbuurt area is exported from the OSM website and used as input for the model. As OSM �les are not always up-to-date it is advised to compare the �le with other maps or with the real-world situation.
3.1.6 Parking tickets.Parking ticket data is obtained through the Nationale Parkeer Rechten (NPR) database provided by the City of Amsterdam. �is dataset contains all of the parking ticket sales in Amsterdam. Parking tickets are sold at parking ticket machines, which are located at a maximum range of x meters from a parking place. For every sale information about the sale is stored in the database. A ticket sale contains the number of the ticketing machine, the date and time, and several other variables. For this case, the parking ticketing machines located in the Frans Halsbuurt are extracted from the database. As some parking ticket machine is located on the border of an area, not all cars that park in the area bought a ticket in the area. �erefore, it is assumed that this overlap is balanced
3.2 Methods
3.2.1 Generating agents.�e �rst question that has to be an-swered before generating a population for the simulation, is how many agents have to be generated for the model. To estimate the amount and the distribution of agents needed for the model, park-ing data provided by the municipality of Amsterdam is used. To �nd the number of permit holders that park in this neighborhood, the Albert Cuyp parking garage data is used. Assumed is that the amount of permit holders does not change a lot over time, as the amount of parking permits in the neighborhood is set. For the number of non-permit holders that visit the area, the parking ticket (NPR) data is used.
To generate a population of agents for SimPark the OViN data needs to be �ltered and processed on car trips made to or from the Frans Halsbuurt on weekdays. To start, the OViN travel survey data is �ltered on car trips and on postal code area of the Frans Halbuurt. A�er this stage, the dataset contains car trips which depart or arrive in the postal code area of the neighborhood. Secondly, weekends are �ltered out and the date of the trips is set to the day of the simulation. �e reason that weekends are not included is that Saturdays and Sundays are no business days, which will lead to people make di�erent kinds of trips.
In various cases in the dataset, the end time of an activity is not given. As the SimPark needs to have an activity start and end time, the end time will be calculated by subtracting the time of departure of the next trip (if it exists) from the start time of the activity. In this manner, these trips can also be used in the model.
As the Frans Halsbuurt is a very small area, only 30 agents could be collected from the OViN data. To obtain more agents, those of similar neighborhoods are added. To �nd neighborhoods with similar parking pa�erns, a clustering method developed by Schmidt[11] will be used. Schmidt used k-means clustering to categorize neighborhoods by occupancy rate pa�erns based on NPR data [11]. �e clustering looks at the normalized occupation rate per hour. �e results of this clustering method are given in �gure 3.
It has to be noted that the clustering done by Schmidt, is based on NPR data, which means that only the people that buy a parking ticket are in the occupancy rate. Permit holders will not be in this number as they do not buy parking tickets.
As the Frans Halsbuurt is in cluster 2, all trips from neighbor-hoods that are in cluster 2 are similar neighborneighbor-hoods concerning
Figure 3: Clustering based on NPR occupancy rate by Schmidt[11]. �e Frans Halsbuurt is located in cluster 2 (green). Other clusters are cluster 0 in blue and cluster 1 in orange.
occupancy from NPR data. �e problem is that the trips in the OViN data are speci�ed on the 4 digit postal code level. �erefore it is decided to only select postal code areas that contain neighbor-hoods mostly in the same cluster, this is analyzed visually in �gure 3. All postal code areas in this selection are added to a list. All trips containing a value of this list will be replaced with a randomly assigned address in the Frans Halsbuurt which matches with the activity and motive of the trip. For example, when a person went to one of the similar postal code areas to go shopping, he will get an address of a shop randomly assigned in the Frans Halsbuurt.
A�er assigning locations to the diaries of the agents, the type of agent is added. Permit holders are �ltered out by use of the following rules:
(1) when the goal of the trip is of the type home and the destination is in the Frans Halsbuurt
(2) when it is the �rst trip of the day and the place of origin is in the Frans Halsbuurt
(3) the neighborhood they live in is the Frans Halsbuurt (not assigned in most data)
�e rest of the agents will be assigned to the non-permit holder category.
3.2.2 Parking choice behavior.A�er the right amount of agents is generated, the parking choice behavior of the agents has to be analyzed. As di�erent parking choice strategies can be applied to agents, a decision has to be made on which parking choice model to use. One option was to let the agent take the �rst empty parking space in a street when a certain threshold is reached, for example when the occupancy rate is high or when the agent is 100m from the destination.
But to get closer to reality than taking the �rst empty park-ing place, the behavioral model based on utility maximization of Khaliq[6] is used to simulate the parking choice behavior of the agents. �is model is estimated with data from Belgian residents, it is assumed that the model also holds for Amsterdam. �is behav-ioral model is based on utility maximization per street as the model
is developed for PARKAGENT which uses the utility per street. In SimPark agents decide whether to park or continue driving, per parking place instead of per street. Factors that a�ect the utility of a street may not be exactly the same as factors that a�ect the utility of a parking place. �is has to be kept in mind when using the parameters of Khaliq’s model on the parking place level. Next, to this point, other facts have to be noted which will be explained in the following paragraphs.
Firstly, the parameters that are estimated in Khaliq’s model con-tain non-signi�cant values which are quite high (see table in ap-pendix A). As it is not possible to re-estimate the model for this research, the model will be tested in the following manner: the situation as it is with non-signi�cant values will be compared with the situation where all non-signi�cant values will be set to zero. �is will be tested for the most common cases from the simulation logs. If the di�erence between the two cases is very small, the non-signi�cant values will be set to zero. An example of a case is a parking place with cost x and which has a distance of 100 meters from the destination. When the di�erences in the outcomes are big, we need to keep the non-signi�cant parameters as it is ’proven’ they cannot be set to zero
Secondly, according to Khaliq’s behavioral model the following factors signi�cantly in�uence on-street parking choice: on-street parking cost, expected parking duration, speed limit, and occupancy rate. O�-street parking choice, on the other hand, is signi�cantly a�ected by o�- and on-street parking costs, and by the expected parking duration. For each street the agent passes, the utility will be calculated for parking on- or o�-street by summing the parameters estimated by the model. A�er the parameters are summed, the ex-pected value is taken. Lastly, the normalized probability is assigned to di�erent choices. In SimPark, the utility will be calculated on a di�erent level, namely on the parking place level. It is assumed that the parameters that a�ect parking choice for a street are the same as the parameters for parking choice of a parking place.
�irdly, Khaliq’s model is estimated in a situation where there are on- and o�-street parking facilities. In the May scenario, all 567 parking places are on-street. Instead of three di�erent choices, the agent has two di�erent choices now: continue driving or park on-street. In Khaliq’s model, there is also a third choice: park o�-street. �erefore, the park on-street and park o�-street probability in Khaliq’s model is summed for the February scenario.
Lastly, as Khaliq’s model is based on a city instead of a neighbor-hood, in this paper the time a�er which the agent changes parking strategy is di�erent. In Khaliq’s model, the agent decides to give up searching for an on-street parking place a�er a maximum of 10 minutes. �e maximum on-street search time probably in the case of a neighborhood, as a neighborhood forms only a small part of a city. �erefore it is assumed that the maximum on-street search time in the neighborhood is 5 minutes, instead of 10 minutes as in Khaliq’s model.
3.3 Comparing four scenarios
To compare the di�erent situations, one day in three di�erent time periods will be simulated. Scenario 1 takes place before the opening of the Albert Cuyp garage on the 15th of May 2018. Scenario 2 is simulated on the 1st of October and from this day onwards, the
removal of on-street parking places will start. Finally, in scenario 3 on the 1st of February 2019, only a small amount of on-street parking places is le� (see table 1). As an extra scenario, the October scenario is also simulated with more agents. �is is done under the assumption that with a higher parking supply, more people will travel by car [12].
Scenario May Oct Oct 2 Feb Agents 654 654 762 654 On-street parking 567 567 567 81 O�-street parking 0 600 600 600 Total parking 567 1167 1167 681
Table 1: Number of agents and on- and o�-street parking places per scenario.
�e scenarios will be compared on cruising for parking i.e. time that an agent spends on searching for a parking place and on the average occupancy in the neighborhood. �e occupancy is calcu-lated by dividing the number of occupied parking places by the total number of parking places. Both cruising for parking and the occupancy rate will be distributed across time (average per hour).
3.4 Comparing scenarios to the real world
To examine the performance of the model, the simulation results will be compared to the real-world data. First, the occupancy rates from the simulation will be compared to the parking ticket data. It has to be noted that the parking ticket data only contains non-permit holders because non-permit holders do not buy parking tickets. Despite the fact that the permit holders are not in the parking ticket data, it might still be interesting to compare both pa�erns.
�e second comparison will be between the simulation parking garage occupancy and the real-world parking garage occupancy. To prepare the real-world data for calculating occupancy the following steps will be executed for both comparisons. First, the average cumulative mean for the entries and exits per hour is calculated. Next, to calculate the occupancy the exits are subtracted from the entries.
4 RESULTS
4.1 Generating agents
A�er �ltering, and preprocessing the Dutch travel survey data (OViN), the agents are assigned to either the permit holder category or the non-permit holder category. From these categories, the right amount of permit holders and non-permit holders is drawn. �e size of both groups is decided as follows: based on averages from the parking garage data provided by the City of Amsterdam and the parking ticket sales data (NPR), it is estimated that there are 275 permit holders and 379 non-permit holders in the May, October, and February scenarios. �e October 2 scenario is based on the maximum amount of permit-holders in the parking garage en the maximum amount of non-permit holders from the NPR-data. �e scenario is simulated with 301 permit holders and 461 non-permit holders, as it is expected that both groups of people use the car more o�en when the parking supply is higher.
4.2 Parking choice behavior
First, it is tested if the non-signi�cant values could be set to zero by comparing the probabilities with the case where the non-signi�cant values are kept as they are. �is is evaluated in the most common cases from the simulation logs. It appears that in the most common case the normalized probabilities for parking o�-, on-street or to continue driving, are very di�erent when the non-signi�cant values are set to zero. �erefore, it is decided to keep the non-signi�cant values estimated in the model.
Khaliq’s behavioral parking model seems to give reasonable results for the October and February scenarios, and for the May scenario (a�er combining park on- and o�-street probabilities). �e decisions made by the agents are compared in table 2. In this table, all decisions the agents make are given. Everytime an agent passes a parking place it makes one of the three decisions (park on-street, o�-street or continue driving. One agent can, for example, choose to continue driving a�er passing the �rst parking place, and then choose to park at the second parking place it passes. �is agent has made two decisions: continue driving and park on-street. Distributions of all decisions made by the agents in di�erent scenarios seem to be quite similar.
May Oct Oct2 Feb Total # decisions 1474 1125 1288 401 Park o�-street % 21 31 34 31 Park on-street % 66 55 56 59 Continue driving % 13 13 11 10 Total # agents 654 654 762 654 Total # parking places 567 1167 1167 681 Table 2: Total number of decisions made per scenario and distribution between decisions.
In approximately 30 percent of the events, the agent chooses to park o�-street, in 60 percent the agent chooses to park on-street and in 10 percent of the cases the agent chooses to continue driving. �e only di�erence that can be seen between the scenarios is that the number of decisions made in the February scenario is about 30 percent of the number of decisions made in the other scenarios. A reason for the low amount of decisions made is that people automatically go o�-street when the maximum on-street search time of 5 minutes is reached. �erefore, it can be concluded that the behavioral model from Khaliq is not suitable for cases with very few or no on-street parking places as in the February scenario.
4.3 Comparing four scenarios
4.3.1 Cruising for parking.In all comparisons, on-street scenar-ios are visualized by solid lines and o�-street scenarscenar-ios by dashed lines. Furthermore, similar colors are kept to compare between di�erent �gures.
In �gure 4 the average cruising for parking is shown in minutes starting at the hour. �e scenarios May, Oct and Oct2 show similar results with a search time between 0 and 2 minutes with small peaks between 7 and 9 and between 16 and 18. �e pa�ern looks quite reasonable, as the peaks are located at the beginning and end of a business day.
Figure 4: Simulation average parking search time per hour (search started in hour) on an average weekday
�e February scenario has the highest search time with approxi-mately 5 minutes on average which is also the maximum on-street search time. In this scenario, there were 81 parking on-street park-ing places and 600 o�-street parkpark-ing places. As there are very few on-street parking places, and people prefer to park on-street, the parking places will be almost full most of the time. �e people that cannot park on-street all search for 5 minutes before they decide to park o�-street. �erefore the average search time per hour in February is approximately 5 minutes.
�e two October scenarios give similar results although there are more agents in the second October scenario than in the �rst. �is is presumably due to the fact that there are still more than enough parking places in the second scenario.
4.3.2 Occupancy rate. �e occupancy rate is shown in �gure 5. A clear pa�ern can be seen in all di�erent scenarios with a maximum between hour 12 and 14. �en between hour 15 and 18 we see a local minimum in all scenario’s. �e occupancy rate keeps rising to approximately hour 12 and then starts slowing down to hour 17 to rise again to hour 19. As the February scenario is the only scenario with only a few on-street parking places, approximately 80 percent of the places are full from the start of the simulation.
In all scenarios, the on-street occupancy rate is higher. �is seems to be logical as on-street parking gives a higher utility as it is cheaper, and the agent can o�en park closer to the destination (see appendix A).
�e two October scenario’s di�er lightly, the scenario with more people has a slightly higher occupancy rate. In the �gure with the absolute numbers it is shown that the di�erence between on-street occupied places is approximately between 0 and 50 for both on-street and o�-on-street parking places. So 100 more agents result in a maximum of 50 more occupied places in the October scenario with 1167 parking places in total.
Figure 5: Simulation on- and o�-street occupancy rate per hour
4.4 Comparing scenarios to the real world
To compare the simulation results to the real world data the follow-ing results are compared. First, the parkfollow-ing ticket data is compared to the on-street occupancy from the simulation. �en the parking garage data is compared to the simulation parking garage occu-pancy. �e real data contains an arrival time and a departure time. To calculate occupancy the following steps are executed: �rst both arrival times and departure times is are �ltered by weekdays, then the data is grouped per hour. Lastly, the data is made cumu-lative and then the cumucumu-lative departures are subtracted from the cumulative arrivals.
In the on-street occupancy simulation results in �gure 5 and real-world data in �gure �, a similar peak from hour 12 to 14 is seen in the simulation results in �gure 5. Furthermore, the local minimum situated around hour 17 can be seen in both �gures.
Second, the parking garage simulation results are compared to the real-world parking garage data. In the real data peaks are shown between hour 10 and 11 in October and at hour 12 in February. �en a second peak is shown between hour 17 and 18. �e second peak in February is higher than the �rst peak in February, while in October both peaks are of similar height.
To obtain the absolute real-world occupancies, the occupancy is added to number of occupied places at hour 0. �e monthly average of occupied places at hour 0 for the parking garage is known: in October 160 and in February 432 parking places on average are occupied at hour 0 (see �gure 7).
As �gure 7 shows, the real occupancy is much higher than the occupancy in the simulation results. �is could be due to the fact that the amount of agents in the simulation is too low. �e October 2 scenario with more agents comes closer to the simulation result of October than the �rst October scenario.
Figure 6: Upper �gure: Real on-street delta occupancy from parking ticket sales data (only non-permit holders). Lower �gure: Real o�-street delta occupancy from Albert Cuyp-garage data.
Figure 7: O�street absolute occupancy calculated as the aver-age monthly occupancy at hour 0 summed with occupancy
For the on-street occupancy, it is not known exactly how many places are occupied on-street. �erefore the average occupancy rate from the neighborhood in the evening is taken from 2012 calculated by Trajan commissioned by the city of Amsterdam. �e occupancy in this research is 90 percent, and is used to obtain �gure 7. Again
Figure 8: Onstreet absolute occupancy calculated as the aver-age monthly occupancy at hour 0 summed with occupancy
it is shown that the real occupancy is higher than the simulation results.
5 CONCLUSION
In this paper, parking policy is evaluated with the agent-based parking model SimPark in the Frans Halsbuurt in Amsterdam, the Netherlands. In this neighborhood, almost all on-street parking places are moved o�-street. �e transition of this parking policy is divided into three di�erent phases with di�erent numbers of on- and o�-street parking places. �ose di�erent scenarios are simulated and the e�ect on parking occupancy and cruising for parking is compared between the di�erent scenarios and with real-world data. It is shown that a population of agents can be generated for the simulation of a neighborhood from the Dutch travel survey data (OViN) by using trips made in neighborhoods that have a similar non-permit holder occupancy pa�ern.
Furthermore, the use of the behavioral parking choice model developed by Khaliq[6] seems to be suitable for evaluating policy in the Frans Halsbuurt. In this model a value was given to three di�erent decisions for the agent: continue driving, park on-street or park o�-street. To make the model suitable for the scenario where there were no o�-street parking places, the probability for choosing to park on-street or o�-street are summed. �e decisions made by the agents in the simulations look plausible. Only in the scenario where there were very few on-street parking places the agents in the simulation have made fewer decisions, than the agents in other scenarios. It can be concluded that the behavioral model from Khaliq is not suitable for situations without or few on-street parking places. �e model should be adapted, or another model should be used. Furthermore, it could be the case that the behavioral
model that is based on Belgian residents from a small city, is not suitable for a city like Amsterdam.
�e simulation results from the di�erent scenarios were com-pared in terms of occupancy and cruising for parking. �e average cruising time per hour in the February scenario with very few on-street parking places has a higher average search time per hour than the other scenarios. Small, but not very clear peaks in search time, can be seen in the morning and around 5 p.m., which correspond to the start and of the business day. In the occupancy rate comparison, a clear pa�ern is shown in all di�erent scenarios. Occupancy is rising to 1 p.m., then falling to 5 p.m., and rising again to 7 p.m., a lot of people seem to arrive and depart from the neighborhood around 5 p.m..
Finally, the scenarios were compared to real-world data. �e occupancy rates from the simulations are compared to the park-ing ticket sales data. Although the simulation occupancy contains permit and permit holders and the real-world data only non-permit holders, a similar pa�ern can be seen in the data. A�er the on-street occupancy rate comparison, the parking garage occu-pancy results are compared to the real parking garage data. In this case, again a lightly similar pa�ern is shown with peaks around 12 and 18.
Although the pa�erns from the simulation results and the world data are similar, when absolute numbers are compared real-world, occupancy is always higher than the occupancy in the sim-ulation results. �is could be due to the fact that there are not enough agents in the simulation.
Considering the fact that the simulation results show similar pa�erns as the real-world data, SimPark seems to be adequate parking policy evaluation tool in situations where on-street parking places are moved o�-street. But when look at the absolute numbers in this case the occupied places in the real-world data are always higher than in the simulation results. �is could be due to the fact that there are too less agents, or due to the fact that the agents from the OViN data, do not make enough trips. �e tool has potential to be used as a parking policy analysis tool but it �rst has to be tested further. When the results from further testing are positive, the tool could be used to predict the e�ect of moving on-street to o�-street parking on parking pressure and cruising for parking for future cases in other neighborhoods.
6 DISCUSSION
One limitation in this study is the use of the non-signi�cant param-eters for the calculation of the utility to park on-street, o�-street or to continue driving. As it was not possible for this research to re-estimate the model, the decisions that are made by the agents are a�ected by factors that do not have a signi�cant in�uence on parking choice. In the future, the behavioral model should be re-estimated to re-run the simulations with signi�cant values only. Furthermore, the parameters are estimated for Belgian residents from Hasselt, it could be the case that the parameters do not hold for Amsterdam residents.
�e second limitation in this paper is that the clustering of the neighborhoods to �nd neighborhoods with similar parking pa�erns to obtain more trips from the OViN data is based on non-permit
holder data only, this could have the e�ect that neighborhoods are chosen with a similar parking pa�ern for non-permit holders but a very di�erent pa�ern for permit holders. �is could in�uence the results of the model. �erefore, clustering on di�erent factors should be compared.
Lastly, for future work to make be�er comparisons between the simulation results and the real-world data, output of the Sim-Park model should be split into the type of the agent. In that way parking ticket data can be compared with non-permit holders only, instead of comparison with all agents which is actually not a equal comparison.
Furthermore, it would be very interesting to further research changes in parking policy regarding the rise in parking tari� in April 2019 in the Frans Halsbuurt. �e o�-street tari� became cheaper than the on-street tari�. Unfortunately, the data was not available yet during the writing of this thesis, this could result in an extra performance measure for the use of SimPark as a parking policy tool.
ACKNOWLEDGEMENTS
I would like to take the opportunity to thank Jan Vuurstaek and dr. ir. Luk Knapen from IMOB University of Hasselt for the use of the SimPark so�ware and for their time, useful insights, and the endless help I got by interpreting the simulation results. I really enjoyed all the discussions during our (almost daily) Skype meetings, I really started to like writing this thesis (which I did not expect). I am also grateful to dr. Bas Scho�en from the municipality of Amsterdam who kept an eye during the thesis, who motivated me, and who also gave some key insights about my thesis. Furthermore, I would like to thank my supervisor dr. Elenna Dugundji for helping me with the statistics, and for coming up with the idea to use the clustering method from Schmidt. But also for being an inspiration for me when leading and forming the parking research work group, which I also need to thank. Lastly, the City of Amsterdam for the data, Datalab, Iva Gornischka, Marco van Leeuwen and Francina van der Star from the parking department and my best friend Veerle and sister Pleun for giving feedback to my thesis.
REFERENCES
[1] Tom Bellemans, Bruno Kochan, Davy Janssens, Geert Wets, �eo Arentze, and Harry Timmermans. 2010. Implementation framework and development trajec-tory of FEATHERS activity-based simulation platform. Transportation Research Record 2175, 1 (2010), 111–119.
[2] Itzhak Benenson, Karel Martens, and Slava Bir�r. 2008. PARKAGENT: An agent-based model of parking in the city. Computers, Environment and Urban Systems 32, 6 (2008), 431–439.
[3] Joschka Bischo� and Kai Nagel. 2017. Integrating explicit parking search into a transport simulation. Procedia Computer Science 109 (2017), 881–886. [4] Karel Dieussaert, Koen Aerts, �´er`ese Steenberghen, Sven Maerivoet, and Karel
Spitaels. 2009. SUSTAPARK: an agent-based model for simulating parking search. In AGILE International Conference on Geographic Information Science, Hannover. 1–11.
[5] PvdA Groenlinks, D66 and SP. 2018. Coalitieakkoord Amsterdam 2018. (May 2018). h�ps://assets.amsterdam.nl/publish/pages/887342/coalitieakkoord amsterdam 2018.pdf
[6] Annum Khaliq, Peter JHJ Van Der Waerden, and Davy Janssens. 2017. A discrete choice approach to de�ne individual parking choice behaviour for parkagent model. WIT Press.
[7] Michael Kodransky and Gabrielle Hermann. 2010. Europe’s parking u-turn: from accomodation to regulation. Institute for Transportation and Development Policy. [8] Karel Martens and Itzhak Benenson. 2008. Evaluating urban parking policies
with agent-based model of driver parking behavior. Transportation Research Record 2046, 1 (2008), 37–44.
[9] Giuliano Mingardo, Bert van Wee, and Tom Rye. 2015. Urban parking policy in Europe: A conceptualization of past and possible future trends. Transportation Research Part A: Policy and Practice 74 (2015), 268–281.
[10] Nederlandse Normalisatie Instituut (NNI). 2013. NEN 2443 Parkeren en stallen van personenauto’s op terreinen en in garage. h�ps://www.nen.nl/pdfpreview/ preview 175451.pdf. (March 2013).
[11] Jeroen Schmidt. 2019. Imputing Parking Usage on Sparsely Monitored Areas Within Amsterdam �rough the Application ofMachine Learning. Master’s thesis. University of Amsterdam.
[12] Donald C Shoup. 1999. �e trouble with minimum parking requirements. Trans-portation Research Part A: Policy and Practice 33, 7-8 (1999), 549–574. [13] �´er`ese Steenberghen, Karel Dieussaert, Sven Maerivoet, and Karel Spitaels. 2012.
SUSTAPARK: An Agent-based Model for Simulating Parking Search. Journal of the Urban & Regional Information Systems Association 24, 1 (2012).
[14] Centraal Bureau voor de Statistiek (CBS); Rijkswaterstaat (RWS). 2014. On-derzoek Verplaatsingen in Nederland 2010 - OViN 2010 versie 2.0. h�ps://doi. org/10.17026/dans-zhs-ghwg. (3 September 2014). DOI:h�p://dx.doi.org/h�ps: //doi.org/10.17026/dans-zhs-ghwg
[15] Centraal Bureau voor de Statistiek (CBS); Rijkswaterstaat (RWS). 2014. On-derzoek Verplaatsingen in Nederland 2011 - OViN 2011 versie 2.0. h�ps: //doi.org/10.17026/dans-xv2-hapb. (1 July 2014). DOI:h�p://dx.doi.org/h�ps: //doi.org/10.17026/dans-xv2-hapb
[16] Centraal Bureau voor de Statistiek (CBS); Rijkswaterstaat (RWS). 2014. On-derzoek Verplaatsingen in Nederland 2012 - OViN 2012 versie 2.0. h�ps: //doi.org/10.17026/dans-2bs-q7u2. (1 July 2014). DOI:h�p://dx.doi.org/h�ps: //doi.org/10.17026/dans-2bs-q7u2
[17] Centraal Bureau voor de Statistiek (CBS); Rijkswaterstaat (RWS). 2014. Onder-zoek Verplaatsingen in Nederland 2013 - OViN 2013. h�ps://doi.org/10.17026/ dans-x9h-dsdg. (1 July 2014). DOI:h�p://dx.doi.org/h�ps://doi.org/10.17026/ dans-x9h-dsdg
[18] Centraal Bureau voor de Statistiek (CBS); Rijkswaterstaat (RWS). 2015. Onder-zoek Verplaatsingen in Nederland 2014 - OViN 2014. h�ps://doi.org/10.17026/ dans-x95-5p7y. (31 March 2015). DOI:h�p://dx.doi.org/h�ps://doi.org/10.17026/ dans-x95-5p7y
[19] Centraal Bureau voor de Statistiek (CBS); Rijkswaterstaat (RWS). 2017. On-derzoek Verplaatsingen in Nederland 2015 - OViN 2015 versie 2.0. h�ps: //doi.org/10.17026/dans-z2v-c39p. (27 July 2017). DOI:h�p://dx.doi.org/h�ps: //doi.org/10.17026/dans-z2v-c39p
[20] Centraal Bureau voor de Statistiek (CBS); Rijkswaterstaat (RWS). 2017. Onder-zoek Verplaatsingen in Nederland 2016 - OViN 2016. h�ps://doi.org/10.17026/ dans-293-wvf7. (27 July 2017). DOI:h�p://dx.doi.org/h�ps://doi.org/10.17026/ dans-293-wvf7
[21] Centraal Bureau voor de Statistiek (CBS); Rijkswaterstaat (RWS). 2017. Onder-zoek Verplaatsingen in Nederland 2017 - OViN 2017. h�ps://doi.org/10.17026/ dans-xxt-9d28. (27 July 2017). DOI:h�p://dx.doi.org/h�ps://doi.org/10.17026/ dans-xxt-9d28
[22] Jan Vuurstaek, Luk Knapen, Bruno Kochan, Tom Bellemans, and Davy Janssens. 2018. First steps towards a state-of-the-art parking simulator. Elsevier. [23] Rashid Waraich and Kay Axhausen. 2012. Agent-based parking choice model.
Transportation Research Record: Journal of the Transportation Research Board 2319 (2012), 39–46.
A PARAMETERS PARKING CHOICE MODEL
In this table the parameters estimated by the behavioral parking choice from Khaliq [6] estimated for SimPark with data from Bel-gian residents, is shown.
A�ributes Levels On-street O�-street
Constant 1,43498*** 0,74734**
On-street parking costs Free 0,375** -0,45503**
1.00 euro/hr 0 0
2.00 euro/hr -0,66473*** -0,04891 Dist parking to dest 100 meter 0,26537 0,06244
200 meter 0,13903 -0,1062
300 meter 0 0
Expected parking Less 60 mins -0,22427 -0,31095
duration 60-120 mins 0 0
More 120 mins -0,50326*** -0,57673*** Number of street visited No. streets 0,11948 0,06118
1 street -0,0759 -0,11476
2 street 0 0
Distance to O�-street 100 meter 0,14493 0,19799 200 meter 0,05733 0,24008
300 meter 0 0
Cost O�-street 0.50 euro/hr 0,13813 1,46904***
1.50 euro/hr -0,04468 0,80949***
2.50 euro/hr 0 0
Parking spaces/100 meter 10 spaces 0 0
15 spaces 0,00115 0,05839 20 spaces -0,12578 0,08517
Speed limit 20 km/hr 0,35714 0,04826
40 km/hr 0,4822** 0,2446
60 km/hr 0 0
Occupancy rate 50 percent% 0,43625* 0
75 percent 0 0,1506
100 percent -0,13304 -0,02007 Maximum parking duration 2 hours 0,32897 0,36201
4 hours -0,10779 0,0024
Unlimited 0 0
Goodness-of-�t Log-likelihood of the null model, LL(0): -2098 Log-likelihood of the optimal model, LL(B): -1934 LRS=-2[LL(0)-LL(B)]: 328 (df = 42) McFadden�s Rho-Square : 0.275 McFadden�s adjusted Rho-Square MNL: 0.264
