This version is available at https://doi.org/10.14279/depositonce-7731
Copyright applies. A non-exclusive, non-transferable and limited
right to use is granted. This document is intended solely for
personal, non-commercial use.
Terms of Use
This is a draft chapter / article. The final version is available in "Smart transport networks : market
structure, sustainability and decision making" edited by Vanoutrive, T., Verhetsel, A., published in 2013,
Edward Elgar Publishing Ltd http://dx.doi.org/10.4337/9781782548331.00014
The material cannot be used for any other purpose without further permission of the publisher, and is for
private use only.
Kickhöfer, B.; Hülsmann, F.; Gerike, R.; Nagel, K. (2013). Rising car user costs: comparing aggregated
and geo-spatial impacts on travel demand and air pollutant emissions. Smart Transport Networks, 180–
207. https://doi.org/10.4337/9781782548331.00014
Benjamin Kickhöfer, Friederike Hülsmann, Regine Gerike, Kai Nagel
Rising car user costs: comparing
aggregated and geo-spatial impacts on
travel demand and air pollutant emissions
Submitted manuscript (Preprint)Chapter in book |
Rising car user costs: comparing aggregated and geo-spatial
impacts on travel demand and air pollutant emissions
Benjamin Kickh¨ofera, Friederike H¨ulsmannb, Regine Gerikeb, Kai Nagela
aTransport Systems Planning and Transport Telematics
Berlin Institute of Technology (TU Berlin)
bResearch Centre Mobility and Transport
Technische Universit¨at M¨unchen
29.02.2012
Abstract
When estimating future transportation costs for car users, there is some agreement that
these costs are likely to increase over the upcoming decades. The reasons are multifaceted:
rising prices for crude oil, large investments in alternative energy supply and in corresponding
demand-side infrastructure as well as road pricing schemes or taxes in order to internalize
negative external effects of road traffic. Rising costs are likely to change aggregated air
pollutant emissions as well as their spatial distribution. Especially in urban areas where
demand for road traffic is high, a reduction of air pollutant emissions could mitigate the
negative impacts on human health and the environment. It is likely that rising car user costs
lead to such a reduction in car travel demand. However, it is still an open question whether
congestion relief, in addition to the impact of a decrease in demand, yields potentials for
lowering local air pollutant emissions.
For this purpose, a real-world scenario of the Munich metropolitan area in Germany is set
up and simulated with the large-scale multi-agent microsimulation MATSim. The software is
capable of simulating complete daily plans of several million individuals and allows emission
calculations on a detailed level, e.g. for a single street or a single vehicle over the time of day.
Varying emission levels resulting from different vehicle characteristics, road categories, speed
levels, and traffic situations are considered. By mapping emissions back to the emission
source, i.e. the road segment, a spatial analysis of air pollutant emissions is performed,
identifying areas with high car travel demand and the resulting changes in emissions due to
a decrease in demand.
We find that car user price elasticities of demand for different subpopulations, including
inner-urban travelers and commuters, are in a reasonable range. Price elasticities of emissions
turn out to be higher than those of demand. This implies that when car user costs rise, the
reduction of air pollutant emissions is higher than the decrease in car travel demand. In a
more disaggregated analysis, we obtain that congestion relief is likely to lower emissions per
vehicle kilometer on urban roads. However, we also find that congestion relief can lead to
higher emissions per vehicle kilometer for high-speed arterials or tangential motorways.
Keywords: emission modeling, price elasticities, spatial analysis, urban transport, user
costs, agent-based modeling
1
Preferred citation style: Kickhöfer, B., F. Hülsmann, R. Gerike, and K. Nagel (2013). “Rising car user costs:
comparing aggregated and geo-spatial impacts on travel demand and air pollutant emissions”. In: Smart
Transport Networks: Decision Making, Sustainability and Market Structure. Ed. by T. Vanoutrive and A.
Verhetsel. NECTAR Series on Transportation and Communications Networks Research. Edward Elgar Publishing
Ltd, pp. 180–207. ISBN: 978-1-78254-832-4.
1 Introduction
Our paper starts from the assumption that car user costs are about to increase in the upcoming
decades. This is likely to have impacts on aggregated air pollutant emissions and on the spatial
distribution of emissions. The concentration of some air pollutants still exceeds the limiting
values prescribed by the European Union, especially in urban areas. Thus, the main focus of
this paper is the question whether a decrease in car travel demand due to higher user costs
would result in a over-proportional reduction of air pollutant emissions. When it comes to the
discussion of cost-related transport policies, large-scale transport models are needed. However,
for the analysis of air pollutant emissions, a detailed investigation of the micro-level is also
necessary. In order to combine both objectives, we use a multi-agent transport model for our
simulations. The multi-agent transport simulation MATSim1is able to simulate large-scale
scenarios. It is also particularly suitable for calculating air pollutant emissions on a detailed
level as complete daily plans are modeled and the traveler’s identity is kept throughout the
simulation process. For illustration purposes of the impacts on air pollutant emissions, nitrogen
dioxide (NO2) is chosen. Furthermore, the transport sector is the main source of NO2emissions
and NO2concentration limits are still often exceeded.
The paper starts with a presentation of the transport model in Sec. 2.1, followed by a
description of the emission modeling tool in Sec. 2.2. Sec. 3consists of three parts: first, a
presentation of the Munich base scenario; second, a description of the simulation approach
and a definition of four policy scenarios; and third, the validation of the base scenario with
respect to modal split and traffic volumes. In Sec. 4, aggregated car user price elasticities of
different subpopulations (inner-urban traffic, commuter, and inverse commuter) are calculated
and discussed. Furthermore, car travel demand as well as NO2emissions are analyzed on a
spatially disaggregated level for all scenarios. The paper ends with a conclusion in Sec 5.
2 Methodology
This section (i) gives a brief overview of the general simulation approach of MATSim and (ii)
shortly describes the emission modeling tool that is developed by the authors. Within the present
paper, only general ideas will be presented. For further information please refer to Raney and
Nagel [2006] and the Appendix or to H¨ulsmann et al. [2011], respectively.
2.1 Transport Simulation with MATSim
In MATSim, each traveler of the real system is modeled as an individual agent. The approach
consists of an iterative loop that has the following steps:
1. Plans generation: All agents independently generate daily plans that encode among
other things their desired activities during a typical day as well as the transport mode for
every intervening trip.
2. Traffic flow simulation: All selected plans are simultaneously executed in the simulation
of the physical system.
3. Evaluating plans: All executed plans are evaluated by a utility function which encodes
in this paper the perception of travel time and monetary costs for the available transport
modes.
4. Learning: Some agents obtain new plans for the next iteration by modifying copies
of existing plans. This modification is done by several modules that correspond to the
available choice dimensions. In the present paper, agents adapt their routes only for car
trips. Furthermore, they can switch between the modes car and public transport (pt).
The choice between plans is performed with respect to a Random Utility Model.
1“Multi-Agent Transport Simulation”, see www.matsim.org
2
The repetition of the iteration cycle coupled with the agent database enables the agents to
improve their plans over many iterations. This is why it is also called learning mechanism (see
Appendix). The iteration cycle continues until the system has reached a relaxed state. At this
point, there is no quantitative measure of when the system is “relaxed”; we just allow the cycle
to continue until the outcome is stable.
2.2 Emission Modeling Tool
There are several sources of air pollution that can be assigned to road traffic: Warm emissions
are emitted when the vehicle’s engine is already warmed-up, whereas cold-start emissions occur
during the warm-up phase. Warm emissions differ with respect to driving speed, acceleration and
stop duration as well as vehicle characteristics including vehicle type, fuel type, cubic capacity
and Euro class [Andr´e and Rapone,2009]. Cold emissions differ with respect to distance traveled,
parking time, average speed, ambient temperature and vehicle characteristics [Weilenmann et al.,
2009]. Furthermore, emissions also result from evaporation and air conditioning. Due to their
small contribution to the overall emission level, this last source is not considered in the present
paper.
The calculation of warm emissions is composed of two steps: first, kinematic characteristics
and vehicle attributes are deduced from the MATSim simulation output. Then, this information
is used in order to extract emission factors from a database. MATSim exhibits activity chains
for every agent over the entire day. Whenever an agent enters or leaves a road segment a time
stamp is created. Thereby, it is possible to calculate the free flow travel time and the travel
time in a loaded network for every agent and road segment. As MATSim keeps demographic
information until the system is relaxed, information about each agent’s vehicle is available at any
time. Vehicle attributes are derived from survey data (see Sec. 3.1) and comprise vehicle type,
age, cubic capacity and fuel type. They can, therefore, be used for very differentiated emission
calculations. Where no detailed information about vehicle type is available, fleet averages for
Germany are used.
Having identified the above kinematic characteristics for a road segment, specific travel be-
havior resulting from such data is assigned by using the detailed handbook of emission factors
called HBEFA2. For some European countries including Germany, the handbook contains coun-
try specific emission factors that can vary by vehicle characteristics, road category, gradient and
speed limit. The handbook provides further disaggregated emission factors depending on four
traffic situations: free flow, heavy, saturated and stop&go. Such traffic situations are described
by kinematic characteristics, which are deduced from driving cycles, i.e. time-velocity profiles.
Typical driving cycles form the basis for calculating traffic situations and, thus, typical emission
factors in HBEFA.
In order to assign emission factors to the traffic flows generated by MATSim, the driving
behavior of an agent on a certain road segment in the MATSim simulation is linked to the respec-
tive HBEFA driving cycle. Therefore, each road segment is divided into two parts representing
stop&go and free flow traffic situations. A similar methodology was developed by Hatzopoulou
and Miller [2010] who, in a more simple approach, assume fixed exhaust emissions per time
unit. The present paper uses a more detailed calculation based on different traffic situations: it
is based on the assumption that cars role in free flow until they have to wait in the queue where
a stop&go traffic situation applies. The length of the queue depends on the traffic demand on
the road. If demand is higher than the capacity of a road segment, a queue emerges where
stop&go is assumed. Another reason for the segmentation of a road segment into free flow and
stop&go parts is due to the marginal difference between the emission factors of free flow, heavy
and saturated. In contrast to these three traffic situations, the emission factors of stop&go are
around twice as high. The difference between actual travel time and free flow travel time per
road segment corresponds to travel time spent in stop&go. The average speed of stop&go that
represents a kinematic characteristic of the typical stop&go driving behavior can be obtained
from the HBEFA database. The stop&go average speed and travel time is used to calculate
2“Handbook Emission Factors for Road Transport”, Version 3.1, see www.hbefa.net
3
the queue length. The respective emission factors can be assigned to the resulting stop&go and
free flow fractions. The implementation of the approach has been evaluated in a test scenario,
which compared real traffic data with MATSim simulations for a single road segment. For a
more detailed description of the emission modeling tool, please refer to H¨ulsmann et al. [2011].
Regarding cold-start emissions, HBEFA provides the relevant factors for passenger cars only.
The application of the relevant cold-start emission factor depends on two attributes: distance
traveled and parking time. The latter is calculated by subtracting the time stamp when the
activity starts from the time stamp when the activity ends. The subsequent distance traveled
is determined by aggregating the lengths of all road segments the agent drives along until the
next activity is reached. The longer the parking time and the accumulated distance, the higher
the cold-start emission factor.
In order to further process the warm and cold-start emissions, so-called emission events are
generated and further segmented into a warm pollution and cold-start pollution emission event.
The former describes the warm emissions for each person and road segment and adds a time
stamp. Cold-start pollution is given for each person and the road segment on which the trip
starts. The definition of emission events follows the MATSim framework that uses events for
storing disaggregated information in XML-format (see Appendix).
3 Scenario: Munich, Germany
The methodology described in Sec. 2is now applied to a large-scale scenario of the Munich
metropolitan area with about two million individuals. For this purpose, a scenario needs to
be set up based on network and survey data. The process is described in Sec. 3.1, followed by
a specification of the simulation procedure in Sec. 3.2 and a validation in Sec. 3.3 where it is
discussed to what extent the simulation reproduces reality.
3.1 Setting up the Scenario
Network (supply side) Network data was provided by the municipality of Munich [RSB,2005].
The data matches the format of the aggregated static transport planning tool VISUM3. It
represents the road network of the federal state Bavaria, being more detailed in and around the
city of Munich and less detailed further away. It consists of 92’259 nodes and 222’502 connecting
edges (= links). Most road attributes, such as free speed, capacity, number of lanes, etc. are
defined by the road type. Only geographical position and length are attributes of each single
link. This data is converted to MATSim format by taking length, free speed, capacity, number
of lanes, and road type from VISUM data. VISUM road capacities are meant for 24-hour
origin-destination matrices. Since the network is almost empty during night hours, peak hour
capacity is set to VISUM capacity divided by 16 (not 24). This results in an hourly capacity of
about 2000 vehicles per lane on an urban motorway. In order to speed up computation, some
road categories corresponding to small local roads are removed from the network. Furthermore,
nodes with only one ingoing and one outgoing link are removed. The two resulting links are
then merged, bringing the size of the network down to 17’888 nodes and 41’942 links. When
merging, the two link lengths are summed up; free speed is calculated based on the minimal
time needed for passing the original links; capacity is set to the minimum of the two links; the
number of lanes is calculated based on the number of vehicles that fit on the two original links;
and finally the road type – important input for emission calculations – is set to the one of the
outgoing link.
Population (demand side) In order to obtain a realistic time-dependent travel demand, sev-
eral data sources are converted into the MATSim population format. The level of detail of
the resulting individual daily plans naturally depends on the information available from either
disaggregated stated preference data or aggregated population statistics. Therefore, three sub-
populations are created, each corresponding to one of the three different data sources:
3“Verkehr In St¨adten UMlegung” developed by PTV AG (see www.ptv.de)
4
•Inner-urban traffic (based on Follmer et al. [2004]):
The synthetic population of Munich is created on the base of very detailed survey data
provided by the municipality of Munich RSB [2005], named “Mobility in Germany” (MiD
2002). In the area of the Munich municipality, 3612 households (with 7206 individuals)
were interviewed. The data consists of different data sets such as household data, person
specific data and trip data. A detailed description of survey methods and data structure
can be found in Follmer et al. [2004]. Individuals were asked to report their activities
during a complete day including activity locations, activity start and end times as well
as the transport mode for the intervening trips. Due to privacy protection, not the exact
coordinates of activity locations are available, but only the corresponding traffic analy-
sis zones (1066 zones in total). For the generation of the synthetic MATSim population,
individual activity locations are distributed randomly within these zones. Furthermore,
all incomplete data sets are removed, e.g. when the location or the starting times of one
activity is missing in the survey. The transport modes train and bus are treated as public
transport trips, motorbikes and mopeds are treated as car trips. The transport modes ride
(= in car as passenger), bike and other (= unknown) are kept for the initial MATSim pop-
ulation. Overall, the data cleaning results in 3957 individuals, the representative sample
for demand generation. Finally, these agents are “cloned” while holding activity transport
analysis zones constant but finding new random locations within these zones for every
clone. This process is performed until the population reaches the real-world size of 1.4
million inhabitants. Thus, the synthetic population living inside the Munich municipality
boundaries consists for this study of 1’424’520 individuals.
MiD 2002 also provides detailed vehicle information for every household. Linking this
data with individuals makes it possible to assign a vehicle to a person’s car trip and thus,
calculating emissions based on this detailed information. As of now there is, however, no
vehicle assignment module which models intra-household decision making. It is, therefore,
possible that a vehicle is assigned to more than one person at the same time.
•Commuter Traffic (based on B¨ohme and Eigenm¨uller [2006]):
Unfortunately, the detailed data for the municipality of Munich does neither contain in-
formation about commuters living outside of Munich and working in Munich nor about
people living in Munich and working outside of Munich. The data analyzed by B¨ohme
and Eigenm¨uller [2006] provides information about workers that are subject to the social
insurance contribution with the base year 2004. Origin and destination zones are classified
corresponding to the European “Nomenclature of Statistical Territorial Units” (NUTS)4,
level 3. Thus, the origin-destination flows between Munich and all other municipalities in
Germany are available. Neither departure times nor transportation mode are, however,
provided. The total number of commuters tends to be underestimated since public servants
and education trips are not included in this statistic. Therefore, every origin-destination
relation is increased by the factor 1.29 [Guth et al.,2010]. Initially, car trips are assumed
to 67% of the total commuter trips, public transport to 33% [MVV,2007]. Departure times
are set so that people arrive at their working place, according to a normal distribution
with N(8 a.m., 2hours) when routed on an empty network. Work end times are set to
nine hours after the arrival at the working place. This results overall in 510’150 commuters
from which 306’160 people have their working place in Munich. All these MATSim agents
perform a daily plan that encodes two trips: from their home location to work and back.
Due to this simplification, they are the first contribution to “background traffic”, as it will
be addressed from here on.
•Commercial Traffic (based on ITP/BVU [2005]):
The second contribution to “background traffic” is given by commercial traffic with the
base year 2004. On behalf of the German Ministry of Transport, ITP/BVU [2005] pub-
lished the origin-destination commodity flows throughout Germany differentiated by mode
4See http://epp.eurostat.ec.europa.eu/portal/page/portal/nuts_nomenclature/introduction,
last access 18.02.2011
5
and ten groups of commodities. Origin and destination zones inside Germany are classi-
fied corresponding to NUTS 2, and outside Germany to NUTS 3 level, respectively. The
number of trucks (>3.5 tons) between two zones or within a zone is calculated based on
the commodity flow in tons and the average loading of trucks.5The starting and ending
points of the trips are — due to the lack of more detailed data — randomly distributed
inside the origin and destination zone, respectively. The resulting MATSim agents obtain
a plan that only consists of two activities with one intervening trip. Departure times are
set so that the number of “en-route vehicles” in the simulation matches a standard daily
trend for freight vehicles.5For this scenario, trips are only considered if they are carried
out at least once in Bavaria during the day. This results in 158’860 agents with one single
commercial traffic trip.
Overall, the synthetic population now consists of 2’093’530 agents. To speed up computa-
tions, a 10% sample is used in the subsequent simulations; other studies indicate that this seems
to be an appropriate percentage in order to achieve results close enough to reality (see e.g. Chen
et al. [2008]). For background traffic, no detailed vehicle information is available. Emissions
are, therefore, calculated with the help of fleet averages for cars and trucks from HBEFA.
3.2 Simulation Approach
Choice Dimensions For the mental layer within MATSim which describes the behavioral learning
of agents, a simple utility based approach is used in this paper. When choosing between different
options with respect to a Random Utility Model, agents are allowed to adjust their behavior
among two choice dimensions: route choice and mode choice. The former allows individuals
to adapt their routes on the road network when going by car. The latter makes it possible to
change the transport mode for a sub-tour (see Appendix) within the agent’s daily plan. Only a
switch from car to public transport or the other way around is possible. Trips that are initially
done by any other mode remain fixed within the learning cycle. From a research point of view,
this approach can be seen as defining a system where public transport is a placeholder for all
substitutes of the car mode.
Utility Functions In the calculations for the travel related part of utility (see Eq. 6in the
Appendix), travel time and monetary distance costs are considered as attributes of every car
and public transport trip. Due to the lack of data of the municipality of Munich, the utility
parameters are taken from Kickh¨ofer [2009] who based the estimations on data from Switzer-
land provided by Vrtic et al. [2008]. The initial formulation of the utility functions for these
estimations is as follows:
Vcar,i,j =β0+βtr,car ·ti,car +βcost,car ·ci,car
Vpt,i,j =βtr,pt ·ti,pt +βcost,pt ·ci,pt ,(1)
where tiis the travel time of the trip to activity iand ciis the corresponding monetary cost.
Travel times and monetary costs are mode dependent, indicated by the indices. The utilities
Vcar,i,j and Vpt,i,j for person jare computed in “utils”. Estimating the parameters6
ˆ
βtr,car =−2.26/h , ˆ
βtr,pt =−2.36/h , ˆ
βcost,car =−0.2/mU , ˆ
βcost,pt =−0.0535/mU
and splitting the time related parameters into opportunity costs of time and additional disutility
caused by traveling (see e.g. Kickh¨ofer et al. [2011]), leads to the functional form7for the overall
5Estimations are based on personal correspondence with Dr. Gernot Liedke from Karlsruhe Institute of
Technology (October, 2010).
6Estimated parameters are in this paper flagged by a hat. his one hour and mU is a unit of money.
7The alternative specific constant β0(see e.g. Train [2003]), is estimated not significantly different from zero
and is, therefore, not considered in the functional form of the utility functions. This essentially means that no
general a-priori preference for one of the transport modes can be found in the data.
6
utility of an activity:
Vcar,i,j = +2.26
ht∗,i ·ln tperf ,i
t0,i −0.2
mU ·ci,car
Vpt,i,j = +2.26
ht∗,i ·ln tperf ,i
t0,i −0.0535
mU ·ci,pt −0.1
h·ti,pt ,
(2)
In this paper, ci,car and ci,pt are calculated for every trip by multiplying the distance between
activity locations i−1 and iby a specific out-of-pocket distance cost rate for car and public transit
(see below). For the functional form of the positive utility earned by performing an activity,
please refer to Eq. 7in the Appendix. Because of the argument regarding the opportunity cost
of foregone activity time when arriving early (see Appendix), the effective marginal disutility of
early arrival is ˆ
βearlyeff =−ˆ
βperf ·t∗,i/tperf ,i≈ − ˆ
βperf =−2.26/h which is equal to the effective
marginal disutility of traveling with car ˆ
βtr,careff . The effective marginal disutility of traveling by
pt is, by the same argument, ˆ
βtr,pteff =−ˆ
βperf ·t∗,i/tperf ,i−| ˆ
βtr,pt | ≈ −ˆ
βperf −|ˆ
βtr,pt | ≈ −2.36/h.
As a result of this simulation approach, it is possible to observe mode reactions to price
increases and to derive price elasticities of demand.
Simulation Procedure For 800 iterations, 15% of the agents perform route adaption (discovering
new routes), 15% change the transport mode for a car or pt sub-tour in their daily plan and 70%
switch between their existing plans. Between iteration 801 and 1000 route and mode adaption
is switched off; in consequence, agents only switch between existing options. The output of
iteration 1000 is then used as input for the continuation of the base case and the four different
policy cases:
•Base case: car user costs remain constant at 10 ct/km
•Policy case 1: increasing car user costs by 25% to 12.5ct/km
•Policy case 2: increasing car user costs by 50% to 15 ct/km
•Policy case 3: increasing car user costs by 75% to 17.5ct/km
•Policy case 4: increasing car user costs by 100% to 20 ct/km
User costs for public transport are assumed to be constant at 17 ct/km for all policy cases.
Please note, that the term “user costs” is referred to as out-of-pocket costs for the users. All
simulation runs are continued for another 500 iterations. Again, during the first 400 iterations
15% of the agents perform route adaption while another 15% of agents choose between car and
public transport for one of their sub-tours. The remaining agents switch between existing plans.
For the final 100 iterations only a fixed choice set is available for all agents. When evaluating
the impact of the car user cost increases, the final iteration 1500 of every policy case is compared
to iteration 1500 of the base case.
3.3 Verification of the Base Case
Modal Split While converting the input data described by Follmer et al. (2004) into the MAT-
Sim synthetic population (see Sec. 3.1), some individuals were omitted due to a lack of coor-
dinates or activity times. Therefore, Table 1 shows differences in the modal split over all trips
comparing the input data with the synthetic subpopulation at iteration 0 and 1500. Note that
only the mode share of the subpopulation traveling within Munich is shown. As one can see,
the initial synthetic population overestimates the percentage of walk trips by 2.55% and of bike
trips by 2.05%, while underestimating the percentage of car trips by 3.52% and of ride trips by
1.61%. Public transport trips remain almost unchanged and the unknown mode is not discussed
further due to the small number of trips. The error seems to be acceptable since no major
differences occur.
7
Table 1: Trips per transport mode as percentage of total trips; Comparison between input data
(Follmer et al., 2004) and the MATSim synthetic subpopulation.
mode Follmer et
al. (2004)
Synthetic
population
it.0
Synthetic
population
it.1500
difference
it.0
difference
it.1500
bike 10 12.05 12.05 +2.05 +2.05
car 26 22.48 20.88 −3.52 −5.12
pt 22 21.98 23.59 −0.02 +1.59
ride 13 11.39 11.39 −1.61 −1.61
undefined 0 0.55 0.55 +0.55 +0.55
walk 29 31.55 31.55 +2.55 +2.55
When the system is in a relaxed state, car trips are even more underestimated, whereas
public transport trips are overestimated compared to iteration 0. Reasons might be the missing
location choice module and the assumptions regarding the specification of the utility function.
Overall, the additional increase in public transport and decrease in car trips amounts only to
1.6%. Thus, the synthetic MATSim population seems to be a good starting point for analyzing
the change in travel demand and air pollutant emissions due to rising car user costs.
Comparison to Counting Stations Before analyzing demand and emission reductions, the realism
of the executed plans in the simulation is verified. The interaction of individuals on the physical
representation of the road network is simulated over 1500 iterations as described in Sec. 2.1.
After reaching a stable outcome, some kind of measurement must exist to determine the quality
of the simulation output. For the Munich region, data from 166 traffic counting stations is
available and aggregated for every hour over time of day.
(a) Comparison for one hour (2 p.m to 3 p.m.) (b) Hourly analysis over time of day
Figure 1: Realism of the simulation results. 166 traffic counting stations provide real-world
traffic counts for the Munich municipality area.
The best quality of this data is available for Thursday, January 10th 2008. It is now used
to compare simulated traffic volumes to real-world values. Different statistical values can be
calculated, such as mean relative error or mean absolute bias. Fig. 1shows two examples of
standard reports that MATSim automatically generates: Fig. 1a depicts the comparison for one
hour and all counting stations. If all data points were on a 45 degree line, the simulation would
nicely reproduce reality. However, as one can see, there are errors between simulated and real
values. The mean relative error for every sensor is a good indicator for the overall fit of the
simulation. It is calculated as:
MRE =
Qsim −Qreal
Qreal
,(3)
8
where Qsim indicates the simulated and Qreal the real-world vehicle flow over the corresponding
counting station in the corresponding hour. Averages for a given hour are obtained by averaging
over all sensors. In the example shown in Fig. 1b, the simulation deviates strongly from reality
during night hours, i.e. from midnight until 7 a.m. During daytime, i.e. from 7 a.m. until
the evening, the hourly mean relative error is between 30% and 50% with better values in the
afternoon.
In order to reach this accuracy, some adjustments were done, e.g. varying the parameters of
the normal distribution that describe work arrival time peak and variance for commuters (see
Sec. 3.1). For now, since this is meant to be a research scenario, the quality of the simulations
seems to be adequate. However, by further optimizing travel demand and network information,
better values for the mean relative error can be obtained as Chen et al. [2008]orFl¨otter¨od et al.
[2011] showed for a scenario of Zurich, Switzerland.
4 The Relationship between Car Travel Demand and Air Pollutant Emissions
This section aims at investigating two research questions: (i) “Are price elasticities of emissions
higher than those for car travel demand?”, and if yes, (ii) “Can a spatial effect be observed?”.
In Sec. 4.1, we derive overall price elasticities of car travel demand from the simulation and then
compare these to price elasticities of NO2emissions. In Sec. 4.2, we first identify areas with high
travel demand in the city of Munich using a more disaggregated approach. In a second step, we
present a spatial analysis of absolute changes in demand and NO2emissions due to policy case
4. Then, we investigate the role of absolute changes in emissions per vehicle kilometer following
the same spatial analysis.
4.1 Aggregated Price Elasticities
Possible reactions of car users to increasing distance costs comprise, in the present paper, either
choosing shorter but eventually more time consuming routes or changing the transport mode to
public transport, the placeholder for all substitutes to car.
Fig. 2shows the daily demand for vehicle kilometers traveled (vkm) over different distance
cost factors (from 10 ct/km for the base case up to 20 ct/km for the highest policy case). The
reduction in demand is presented for three different subpopulations (see Sec. 3.1): black circles
correspond to inner-urban traffic, red rectangles and green crosses to inverse commuter and
commuter, respectively. The inner-urban demand for vehicle kilometers traveled drops from
about 4000000 vkm in the base case by 18% to roughly 3330000 vkm in the highest policy case.
Much larger reductions in car travel demand are observed for the other subpopulations: car
travel demand of inverse commuters drops from 106500000 vkm by 54% to 7540000 vkm, and for
commuters from 406240000 vkm by 72% to 102900000 vkm. The big difference between inner-
urban demand and (inverse) commuters is due to the much longer distances traveled by the
last two groups where the car mode gets extremely unattractive. Travel demand reactions for
freight traffic is not shown since this subpopulation is not allowed to change from car (or truck)
to public transport. The figure also provides linear regression lines including their functional
forms for every subpopulation. Even though, especially for commuter traffic, a linear regression
obviously does not lead to the best fit (one can nicely see the “inverse-S-shape“ produced by
the logit model), it is still quite appropriate in order to derive constant price elasticities.
Choosing p0= 10 ct/km as operating point, price elasticities of demand can directly be
derived for every policy case i, using:
ηq,p =
qi−q0
q0
pi−p0
p0
,(4)
where qiis the travel demand at price level pi. In order to describe the overall relationship
between user costs and car travel demand, a constant price elasticity can be derived using the
regression functions:
9
●●●●●
10.0 12.5 15.0 17.5 20.0
10 20 30 40 50
car user costs [ct/km]
vehicle kilometers traveled [x100'000 vkm]
●
Subpopulation:
Urban
Inverse.Commuter
Commuter
Figure 2: Overall daily vehicle kilometers traveled for the base case and the four
policy cases by subpopulation: simulated values and estimations as linear regression
functions; values for a representative 10% sample.
●●●●●
10.0 12.5 15.0 17.5 20.0
0 200 400 600
car user costs [ct/km]
NO2 emissions [kg]
●
Subpopulation:
Urban
Inverse.Commuter
Commuter
Freight
Figure 3: Overall daily NO2emissions in kilograms for the base case and the four
policy cases by subpopulation: simulated values and estimations as linear regression
functions; values for a representative 10% sample.
10
ηq,p =dq
dp ·p0
ˆq0
,(5)
where dq
dp is the gradient of the corresponding regression function and ˆq0is the estimated initial
demand for car trips at p0= 10 ct/km. Applying Eq. 5to the three subpopulations leads to the
following estimated constant price elasticities of car travel demand:
ˆηUrban
q,p =−0.173 ,ˆηInverse.Commuter
q,p =−0.502 ,ˆηCommuter
q,p =−0.692 .
These estimations indicate that e.g. a car user cost increase of 10% (at the operating point
p0= 10 ct/km) leads to a reduction in car trips by 1.73% for inner-urban traffic, by 5.02%
for inverse commuter and by 6.92% for commuter. Graham and Glaister [2002] present a wide
range of fuel price elasticities collected from different studies. When summarizing the different
studies, the authors find short-term fuel price elasticities in the range from −0.2 to −0.5, for
Germany around −0.45. However, the range within Germany goes from −0.25 to −0.86. The
fuel price elasticities found in the present paper are somewhat smaller for inner-urban traffic
and within the range for inverse commuter and commuter. Obviously, introducing more choice
dimensions into the model, such as location choice or the possibility of dropping activities, is
likely to influence the results. At this point, it can be stated that, overall, the model produces
reasonable behavioral reactions to car user price increases.
Similarly to Fig. 2, overall NO2emissions are shown in Fig. 3, again for the base case and
the four policy cases. Linear regression lines and functional form are also provided. In this
figure, freight traffic emissions are indicated by blue crosses in order to show the big impact
of freight traffic emissions on overall emission levels. Since freight demand is not allowed to
change the mode to public transport, its emissions stay more or less stable for all policy cases.
Only a small reduction can be observed, probably resulting from shorter distances chosen by the
router module. Equally to the price elasticities of demand, price elasticities of NO2emissions
are calculated:
ˆUrban
q,p =−0.219 ,ˆInverse.Commuter
q,p =−0.608 ,ˆCommuter
q,p =−0.792 .
The price elasticities are found to be roughly the same for other exhaust emission types
under consideration (PM and SO2). When comparing them to the price elasticities of car
travel demand from above, one can notice a higher elasticity of emissions than of demand for
all subpopulations. Thus, an increase in car user costs leads to a higher reduction in emissions
than in demand. Two explanations come to mind:
1. An over-proportional fraction of travelers who performed long car trips with high speed
levels now change from car to public transport (“biased mode switch effect”)
2. Travelers are driving faster on formerly congested roads (“congestion relief effect”)
Both explanations are based on the fact that emission levels are usually the lowest for speed
levels around 60 km/h [see e.g. Maibach et al.,2008, p.58]. Emissions per vehicle kilometer
increase for lower but also for higher speed levels, forming an “U-shaped” function with its
minimum around 60 km/h. That is, when mainly trips with high speed levels are reduced (in
this case by changing to another mode), overall emissions drop more than demand. The same is
true when traffic flow becomes more fluid on formerly congested roads. It seems that the second
effect can be observed since our model includes spill-back effects, different traffic states, and
individual vehicle characteristics. The following section will address this hypothesis by looking
at spatial patterns of changes in travel demand and in air pollutant emissions.
4.2 Spatial Analysis of Changes in Car Travel Demand and Air Pollutant Emissions
This section analyzes car travel demand and NO2emissions on a spatially disaggregated level.
Using the features of the emission modeling tool, demand and NO2emissions can be aggregated
per road segment and for any desired time interval. For visual presentation of the spatial effect
11
within the urban area of Munich, emissions are spatially smoothed using an Gaussian distance
weighting function with a radius of 500m. Starting with the base case shown in Fig. 4, one
notices a high level travel demand (in vehicle kilometers traveled) for the inner ring road, the
middle ring road, the main arterial motorways, and the tangential motorway in the north-west
of Munich (see Fig. 4a). Travel demand is highly correlated with the level of exhaust emissions
(see Fig. 4b).8The population exposure of NO2emissions near these road sections is critical
which is also found in the air pollutant concentration levels at monitoring stations, e.g. at
Landshuter Allee [LFU,2011]. Fig. 5shows the absolute change in NO2emissions between
the base case and the 100% price increase (policy case 4). As already presented in Sec. 4.1,
the increase in car user costs leads to an important reduction in emission levels. This finding
is now confirmed by Fig. 5a which decomposes the overall effect in a spatial distribution. The
lesson learned when comparing that picture to Fig. 4is that roads with the highest potential
for emission reductions are located along the corridors with the highest travel demand (and
therefore the highest emissions). Fig. 5a also shows that potential gains are considerably larger
at the medium and high speed roads than, for example, in the inner urban area. Our approach
allows to show such effects on a detailed single-street level while still being applicable to large-
scale scenarios. This allows both the identification of relevant corridors (“hot spots”) and the
spatially disaggregated analysis of the consequences of policy measures.
In order to answer the question whether spatial patterns of higher emission elasticities com-
pared to demand elasticities can be observed, Fig. 5b is analyzed. Similar to Fig. 5a, it depicts
the absolute difference in emission levels between the base case and the 100% price increase
(policy case 4), but now the absolute change in emissions per vehicle kilometer traveled. Values
above zero imply that vehicles produce more emissions per km traveled, whereas values below
zero indicate that vehicles emit less emissions for the same distance. Again, two effects can be
observed that correspond with those presented in Sec. 4.1:
1. The “biased mode switch effect” is most important for the main arterial motorways and the
tangential motorway in the north-west of Munich: Fig. 5a indicates that overall emissions
(and demand) go down on these road segments. But following Fig. 5b, average emissions
per vehicle kilometer go up (light gray areas).
2. The “congestion relief effect” seems to be less coherent in Fig. 5b than the first effect. How-
ever, dark gray areas indicate that average emissions per vehicle kilometer go down. This
means that a reduction in travel demand leads to lower emissions per vehicle kilometer.
The first effect can be interpreted as follows: in the base case, average speeds on motorways
were closer to the (emission) optimal speed of 60 km/h. Fewer vehicles on these roads lead
to higher emissions per vehicle kilometer, since travelers drive faster. That is, congestion relief
leads to higher emissions per vehicle kilometer. A similar finding was obtained by Newman
and Kenworthy [1989], who state that the average traffic speed is correlated positively, and not
negatively, with gasoline consumption per capita. The second effect might be interpreted as
follows when combining the aggregated and the disaggregated observations: it is likely that due
to the reduction in demand, average travel speeds in the corresponding areas get closer to the
(emission) optimal speed of 60 km/h. Emissions along a congested urban road are about twice
as high as when traffic is flowing. When car travel demand is reduced and, thereby, the traffic
situation on the road segment changes from stop&go to saturated or even heavy, emissions are
more reduced than the flow on that road segment. That is, congestion relief leads to lower
emissions per vehicle kilometer especially in urban contexts.
8Our method currently localizes all emissions on a road segment at the center coordinate. This explains why
the tangential motorway in the north-west of Munich is shown as a sequence of filled circles rather than an
uninterrupted line.
12
(a) Vehicle kilometers traveled in vkm/km2
(b) NO2emissions in g/km2
Figure 4: Base case: areas with high car travel demand and areas with high NO2emissions.
Plots based on spatial averaging for all road segments. Values for a representative 10% sample.
13
(a) Absolute change in NO2emissions in g/km2
(b) Absolute change in NO2emissions per vehicle kilometer in (g/vkm)/km2
Figure 5: Absolute changes in NO2emissions between the base case and the 100% price increase
(policy case 4). Plots based on spatial averaging for all road segments. Values for a representative
10% sample.
14
5 Conclusion
In this paper, we set up a real-world scenario of the Munich metropolitan area and simulated
travel demand of a 10% sample (around 200’000 individuals) with a large-scale multi-agent
simulation. We coupled the simulation with detailed emission factors from the “Handbook
Emission Factors for Road Transport”, considering the kinematic characteristics derived from
the simulation and vehicle attributes obtained from survey data. Since the simulation keeps
track of the approximate position and attributes of every traveler’s vehicle during every time
step, it was possible to map the kinematic characteristics to different traffic situations such as
free flow or stop&go. Thereby, emissions were calculated every time a traveler leaves a road
segment, or starts her engine. The mapping of demand (in vehicle kilometers) or emissions back
to the road segments was therefore quite straightforward.
We then introduced four policy cases, where user costs for car are rising from 10 ct/km
in four steps up to 20 ct/km. Aggregated price elasticities of demand were found to be in a
reasonable range for all subpopulations. Commuters reacted more sensitive to the price increase
than inner-urban travelers, e.g. by changing from car to public transport. Price elasticities of
NO2emissions turned out to be higher than those of demand. Two possible explanations were
given: first, it might happen that an over-proportional fraction of travelers who performed long
car trips with high speed levels changed from car to public transport. We called this the“biased
mode switch effect”. Second, it seems that travelers are driving faster on formerly congested
roads, referred to as the “congestion relief effect”.
A spatially more disaggregated analysis allowed to identify so called “hot spots” that bear
high potentials for emission reduction: absolute emissions dropped most in many, but not all,
areas where travel demand was high. Furthermore, the spatial analysis showed that the “biased
mode switch effect” was most important for high-speed arterials and tangential motorways since
absolute emissions (and demand) goes down on these road segments, but average emissions per
vehicle kilometer go up. Due to higher speeds, fewer vehicles — in this case — lead to higher
emissions per vehicle kilometer. The “congestion relief effect” was found to be less coherent in
terms of the type of road segment. Nonetheless, some areas showed a reduction in emissions per
vehicle kilometer caused by a reduction in demand. Due to higher speeds, fewer vehicles — in
this case — lead to lower emissions per vehicle kilometer. Possibly, both effects stem from the
fact that the emission optimal speed is usually around 60 km/h. Measures that allow travelers
to drive faster than that will result in higher emissions per vehicle kilometer. However, the relief
of congestion seems to bear some potential to reduce emissions on urban roads.
We think that this paper can add valuable information to the transport planning and policy
decision making process by providing insights into a new emission calculation model for large-
scale scenarios. In future studies, we plan to account for more choice dimensions than just route
and mode choice. This is likely to influence the results. Also, the robustness of the results needs
to be tested by performing sensitivity analysis. A possible extension would be the modeling of
air pollutant concentration which could be used to validate simulation results with measured
concentration values.
Acknowledgements
This work was funded in part by the German Research Foundation (DFG) within the research
project “Detailed evaluation of transport policies using microsimulation”. Important data was
provided by the Municipality of Munich, more precisely by ‘Kreisverwaltungsreferat M¨unchen’
and ‘Referat f¨ur Stadtplanung und Bauordnung M¨unchen’. Our computer cluster is maintained
by the Department of Mathematics at Berlin Institute of Technology (TU Berlin). The authors
would like to thank two anonymous reviewers for their valuable comments.
15
References
M. Andr´e and M. Rapone. Analysis and modelling of the pollutant emissions from european
cars regarding the driving characteristics and test cycles. Atmospheric Environment, 43(5):
986–995, 2009. ISSN 1352-2310. doi: 10.1016/j.atmosenv.2008.03.013.
S. B¨ohme and L. Eigenm¨uller. Pendlerbericht Bayern. Technical report, IAB, 2006.
N. Cetin, A. Burri, and K. Nagel. A large-scale agent-based traffic microsimulation based on
queue model. In Proceedings of the Swiss Transport Research Conference (STRC), Monte
Verita, CH, 2003. See www.strc.ch. Earlier version, with inferior performance values: Trans-
portation Research Board Annual Meeting 2003 paper number 03-4272.
Y. Chen, M. Rieser, D. Grether, and K. Nagel. Improving a large-scale agent-based simulation
scenario. Technical Report 08-15, 2008.
G. Fl¨otter¨od, Y. Chen, and K. Nagel. Behavioral calibration and analysis of a large-scale travel
microsimulation. Networks and Spatial Economics, 2011. ISSN 1566-113X. doi: 10.1007/
s11067-011-9164-9.
R. Follmer, U. Kunert, J. Kloas, and H. Kuhfeld. Mobilit¨at in Deutschland – Ergebnisbericht.
Technical report, infas/DIW, 2004. See www.kontiv2002.de.
C. Gawron. An iterative algorithm to determine the dynamic user equilibrium in a traffic
simulation model. International Journal of Modern Physics C, 9(3):393–407, 1998.
D. J. Graham and S. Glaister. The Demand for Automobile Fuel: A Survey of Elasticities.
Journal of Transport Economics and Policy, 36(1):1–25, 2002.
D. Grether, Y. Chen, M. Rieser, and K. Nagel. Effects of a simple mode choice model in a large-
scale agent-based transport simulation. In A. Reggiani and P. Nijkamp, editors, Complexity
and Spatial Networks. In Search of Simplicity, Advances in Spatial Science, chapter 13, pages
167–186. Springer, 2009. doi: 10.1007/978-3-642-01554-0.
D. Guth, C. Holz-Rau, V. Killer, and K. Axhausen. R¨aumliche Dynamik des Pendelverkehrs in
Deutschland und der Schweiz: Das Beispiel M¨unchen und Z¨urich. 2010.
M. Hatzopoulou and E. Miller. Linking an activity-based travel demand model with traffic
emission and dispersion models: Transport’s contribution to air pollution in Toronto. Trans-
portation Research Part D: Transport and Environment, 15(6):315–325, 2010. ISSN 1361-9209.
doi: 10.1016/j.trd.2010.03.007.
F. H¨ulsmann, R. Gerike, B. Kickh¨ofer, K. Nagel, and R. Luz. Towards a multi-agent based
modeling approach for air pollutants in urban regions. In Proceedings of the Conference
on “Luftqualit¨at an Straßen”, pages 144–166. Bundesanstalt f¨ur Straßenwesen, FGSV Ver-
lag GmbH, May 2011. ISBN 978-3-941790-77-3. Also VSP WP 10-15, see www.vsp.tu-
berlin.de/publications.
ITP/BVU. Prognose der deutschlandweiten Verkehrsverflechtungen 2025. Technical report,
2005. URL http://daten.clearingstelle-verkehr.de/220/.
B. Kickh¨ofer. Die Methodik der ¨okonomischen Bewertung von Verkehrsmaßnahmen in Mul-
tiagentensimulationen. Diplomarbeit (Diploma Thesis), TU Berlin, Institute for Land and
Sea Transport Systems, Berlin, Germany, June 2009. Also VSP WP 09-10, see www.vsp.tu-
berlin.de/publications.
B. Kickh¨ofer, D. Grether, and K. Nagel. Income-contingent user preferences in policy evaluation:
application and discussion based on multi-agent transport simulations. Transportation, 38:
849–870, 2011. ISSN 0049-4488. doi: 10.1007/s11116-011-9357-6.
16
N. Lefebvre and M. Balmer. Fast shortest path computation in time-dependent traffic net-
works. In Proceedings of the Swiss Transport Research Conference (STRC), Monte Verita,
CH, September 2007. See www.strc.ch.
LFU. ¨
Uberschreitungsh¨aufigkeit f¨ur Feinstaub – PM10, Stickstoffdioxid und Schwefeldioxid.
Technical report, Bayerisches Landesamt f¨ur Umwelt, 2011. URL http://www.lfu.bayern.
de/luft/lueb/ueberschreitung_pm10_so2_no2/index.htm. Last access 30.08.2011.
M. Maibach, D. Schreyer, D. Sutter, H. van Essen, B. Boon, R. Smokers, A. Schroten, C. Doll,
B. Pawlowska, and M. Bak. Handbook on estimation of external costs in the transport sector.
Technical report, 2008.
MVV. Regionaler Nahverkehrsplan f¨ur das Gebiet des M¨unchner Verkehrs- und Tarifverbundes.
Technical report, Munich Local Transport Provider, 2007.
P. W. G. Newman and J. R. Kenworthy. Gasoline Consumption and Cities. Journal of the
American Planning Association, 55(1):24–37, 1989. doi: 10.1080/01944368908975398.
B. Raney and K. Nagel. An improved framework for large-scale multi-agent simulations of travel
behaviour. 2006.
T. Reinhold. Konzept zur integrierten Optimierung des Berliner Nahverkehrs. In ¨
Offentlicher
Personennahverkehr. Springer Berlin Heidelberg, 2006. doi: 10.1007/3-540-34209-5\8.
M. Rieser, D. Grether, and K. Nagel. Adding mode choice to a multi-agent transport simulation.
Transportation Research Record: Travel Demand Forecasting 2009, 2132:50–58, 2009. doi:
10.3141/2132-06.
RSB. Municipality of Munich: Referat f¨ur Stadtplanung und Bauordnung, 2005.
K. Train. Discrete choice methods with simulation. Cambridge University Press, 2003.
M. Vrtic, N. Sch¨ussler, A. Erath, M. B¨urgle, K. Axhausen, E. Frejinger, M. Bierlaire, R. Rudel,
S. Scagnolari, and R. Maggi. Einbezug der Reisekosten bei der Modellierung des Mo-
bilit¨atsverhaltens. Schriftenreihe 1191, Bundesamt f¨ur Strassen, UVEK, Bern, CH, 2008.
Final Report for Project SVI 2005/004.
M. Weilenmann, J.-Y. Favez, and R. Alvarez. Cold-start emissions of modern passenger cars
at different low ambient temperatures and their evolution over vehicle legislation categories.
Atmospheric Environment, 43(15):2419–2429, 2009. ISSN 1352-2310. doi: 10.1016/j.atmosenv.
2009.02.005.
Appendix: Simulation Details
The following paragraphs are meant to present more information about the MATSim simulation
approach that is used in this paper. Every step of the iterative loop in Sec. 2.1 is in the following
illustrated in more detail.
Plans Generation An agents daily plan contains information about his planned activity types
and locations, about duration and other time constraints of every activity, as well as the mode,
route, the desired departure time and the expected travel time of every intervening trip (= leg).
Initial plans are usually generated based on microcensus information and/or other surveys. The
plan that was reported by an individual is in the first step marked as “selected”.
17
Traffic Flow Simulation The traffic flow simulation executes all selected plans simultaneously
in the physical environment and provides output describing what happened to each individual
agent during the execution of its plan. The car traffic flow simulation is implemented as a
queue simulation, where each road (= link) is represented as a first-in first-out queue with two
restrictions [Gawron,1998,Cetin et al.,2003]: First, each agent has to remain for a certain
time on the link, corresponding to the free speed travel time. Second, a link storage capacity is
defined which limits the number of agents on the link; if it is filled up, no more agents can enter
this link. The public transport simulation simply assumes that traveling takes twice as long as
traveling by car on the fastest route in an empty network9and that the travel distance is 1.5
times the beeline distance between the activity locations. public transport is assumed to run
continuously and without capacity restrictions [Grether et al.,2009,Rieser et al.,2009].
All other modes are modeled similar to public transport: travel times are calculated based
on mode specific travel speed and the distance estimated for public transport. However, the
attributes of these modes are not relevant for the present paper since agents are only allowed
to switch from car to public transport and the other way around. Trips from the survey that
are no car or public transport trips, are held fixed during the learning cycle, thus not changing
mode share in any direction.
Output of the traffic flow simulation is a list that describes for every agent different events,
e.g. entering or leaving a link, arriving or leaving an activity. These events are written in
XML-format and include agent ID, time and location (link or node ID). It is, therefore, quite
straightforward to use this disaggregated information for the calculation of link travel times or
costs (which is used by the router module), trip travel times, trip lengths, and many more.
Evaluating Plans In order to compare plans, it is necessary to assign a quantitative measure
to the performance of each plan. In this work, a simple utility-based approach is used. The
elements of our approach are as follows:
•The total utility of a plan is computed as the sum of individual contributions:
Vtotal =
n
X
i=1 Vperf ,i+Vtr,i,(6)
where Vtotal is the total utility for a given plan; nis the number of activities; Vperf ,iis the
(positive) utility earned for performing activity i; and Vtr,iis the (usually negative) utility
earned for traveling during trip i. Activities are assumed to wrap around the 24-hours-
period, that is, the first and the last activity are stitched together. In consequence, there
are as many trips between activities as there are activities.
•A logarithmic form is used for the positive utility earned by performing an activity:
Vperf ,i(tperf ,i) = βperf ·t∗,i ·ln tperf ,i
t0,i (7)
where tperf is the actual performed duration of the activity, t∗is the “typical” duration of
an activity, and βperf is the marginal utility of an activity at its typical duration. βperf is
the same for all activities, since in equilibrium all activities at their typical duration need
to have the same marginal utility. t0,i is a scaling parameter that is related both to the
minimum duration and to the importance of an activity. As long as dropping activities
from the plan is not allowed, t0,i has essentially no effect.
•The disutility of traveling used for simulations is estimated from survey data which is
explained in Sec. 3.2.
9This is based on the (informally stated) goal of the Berlin public transport company to generally achieve
door-to-door travel times that are no longer than twice as long as car travel times. This, in turn, is based on the
observation that non-captive travelers can be recruited into public transport when it is faster than this benchmark
Reinhold [2006].
18
In principle, arriving early or late could also be punished. For the present paper, there
is, however, no need to do so, since agents are not allowed to reschedule their day by changing
departure times. Arriving early is already implicitly punished by foregoing the reward that could
be accumulated by doing an activity instead (opportunity cost). In consequence, the effective
(dis)utility of waiting is already −βperf ·t∗,i/tperf ,i≈ −βperf . Similarly, that opportunity cost
has to be added to the time spent traveling.
Learning After evaluating daily plans in every iteration, a certain number of randomly chosen
agents is forced to re-plan their day for the next iteration. This learning process is, in the
present paper, done by two modules corresponding to the two choice dimensions available: a
module called “router” for choosing new routes on the road network and a module called “sub-
tour mode choice” for choosing a new transport mode for a car or public transport trip. The
router module bases its decision for new routes on the output of the car traffic flow simulation
and the knowledge of congestion in the network. It is implemented as a time-dependent best
path algorithm [Lefebvre and Balmer,2007], using generalized costs (= disutility of traveling)
as input. The sub-tour mode choice module changes the transport mode of a car sub-tour to
public transport or from a public transport sub-tour to car. A sub-tour is basically a sequence
of trips between activity locations. However, the simulation needs to make sure that a car can
only be used if it is parked at the current activity location. Thus, a sub-tour is defined as a
sequence of trips where the transport mode can be changed while still being consistent with the
rest of the trips. It is e.g. assured that a car which is used to go from home to work in the
morning needs to be back at the home location in the evening. If the car remains e.g. at the
work location in order to use it to go for lunch, then the whole sub-tour of going to work and
back needs to be changed to public transport.
19
