Enhancing MATSim with capabilities of within-day re-planning [original]

This version is available at https://doi.org/10.14279/depositonce-7741

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

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

servers or lists, or reuse of any copyrighted component of this work in other works.

Illenberger, J.; Flötteröd, G.; Nagel, K. (2007). Enhancing MATSim with capabilities of within-day re-

planning. 2007 IEEE Intelligent Transportation Systems Conference.

https://doi.org/10.1109/ITSC.2007.4357810

Johannes Illenberger, Gunnar Flötteröd, Kai Nagel

Enhancing MATSim with capabilities of

within-day re-plannin

Accepted manuscript (Postprint)Conference paper |

Abstract— This paper presents a framework for simulation

of within-day re-planning for the MATSim project. Three

major building blocks are presented, each of which represents

specific aspects of driver behavior. These components comprise

(i) the provision of descriptive information in the form of link

travel costs, (ii) prescriptive information in the form of routes,

and (iii) a model of driver satisfaction. An exemplary model is

presented, which focuses on en-route re-planning under

different types of information provision. In this model driver

perception is constrained to link traversal costs and decisions

are made by application of a standard shortest path algorithm.

The satisfaction of a traveler is modeled with a scoring (utility)

function that evaluates routes as well as activities travelers are

aiming at. The framework’s applicability is tested with a simple

fictive network and a real-world network of Greater Berlin.

I. INTRODUCTION

n the field of transport planning, engineers agree that the

problems of transportation are no more a matter of

extending the infrastructure with concrete and steel, but

rather a matter of the efficient use of existing transport

networks [1]. Advanced Traveler Information Systems

(ATIS) are intended to fill in here by providing accurate

information through a variety of devices.

An important aspect is the response of drivers to provided

information. Since deployment of ATIS technologies is still

in an early state, practical experiences are limited. To gain

more insights into travelers’ decision making in-laboratory

experiments such as FASTCARS ([2] and [3]) and IGOR [4]

have been proposed.

Behavioral models derived from the results of these

laboratory experiments can be used in large-scale

simulations to evaluate ATIS technologies. Travel time

savings have been observed in several studies ([5], [6], [7]

and [8]), varying from three to 30 percent depending on

market penetration and network topology. Beside the

lowering of travel time itself, the reduction of its uncertainty

deserves to receive just as much attention.

As a contribution to the research in this field, this paper

presents a basic framework that enhances the MATSim

toolkit (“Multi-Agent Transport Simulation Toolkit”,

www.matsim.org) by capabilities of within-day and en-route

re-planning. The framework is integrated in the so-called

„mental layer” of agent behavior and defines a set of

1 Johannes Illenberger is with the group for Transport Systems Planning

and Transport Telematics, Technische Universität Berlin, D-10587 Berlin,

Germany, mail: illenberger@vsp.tu-berlin.de.

2 Gunnar Flötteröd is with the group for Transport Systems Planning and

Transport Telematics, mail: floetteroed@vsp.tu-berlin.de.

3 Kai Nagel is head of the group for Transport Systems Planning and

Transport Telematics, mail: nagel@vsp.tu-berlin.de.

interfaces in which specific implementations can model

specific behavioral patterns. In detail, the behavior of an

agent is represented by modules, each of which reflects a

certain behavioral aspect. The modularity allows to easily

exchange certain building blocks or even complete

behavioral implementations and thus to compare and

evaluate different models.

The framework allows to model ATIS strategies,

corresponding driver reactions, and finally to analyze the

interaction between ATIS, drivers and traffic conditions.

The remainder of this article is organized as follows: In

section II, a short introduction to MATSim is given and in

section III, the abstract agent model and its conceptual

background are presented. Section IV describes an

exemplary model and its implementation and section V

verifies the frameworks applicability by means of two

scenarios. The paper closes with a discussion and an outlook

in section VI.

II. MATSIM OVERVIEW

MATSim is a multi-agent based transport simulation

which originally envolved from TRANSIMS [9] and pursues

an activity-based approach to demand generation. Unlike

other transportation simulation packages MATSim is

throughout agent-based and generates individual activity

plans as input to the network loading rather than (time-

dependent) origin-destination matrices as typically used in

dynamic traffic assignment. More details about the demand

generation in MATSim can be found in [10] and [11].

Specifically, a plan contains the agent’s intended schedule

of activities for the day, and the travel legs connecting the

activities. A leg holds several attributes describing the travel

from one activity to another such as departure time, expected

arrival time, route and transportation mode. Activities

contain type attributes such as home, work, education,

leisure as well as further information regarding activity

timing.

The initial plans are generated by disaggregating census

data. Next, there is a mechanism that allows the agents to

learn and optimize their plans. The system iterates between

plan generation (the mental layer, also referred as strategic

layer) and traffic flow simulation (the physical layer). The

system remembers several plans per agent and scores the

performance of each plan with a fitness function. Between

two iterations agents are able to modify plans with the use

of genetic algorithms. Those plans are modified by mutation

and recombination, e.g. recalculating new routes or varying

departure times, while “bad” plan instances are eventually

discarded.

Enhancing MATSim with capabilities of within-day re-planning

Johannes Illenberger1, Gunnar Flötteröd2, Kai Nagel3

This day-to-day re-planning mechanism, or more generally

period-to-period re-planning since a plan must not

necessarily be constrained to one day, is continued until the

plans are “relaxed”, i.e. in an approximate user equilibrium.

Fig. 1. The system iterates between plan generation/modification and traffic

flow simulation until an approximate user equilibrium is reached.

The framework proposed in this paper enhances MATSim

with capability of within-day re-planning. More specifically,

agents are now not only able to adapt plans between one

execution in the traffic flow simulation, but also to modify

them during the traffic flow simulation. This enables the

agents to spontaneously react to unforeseeable incidents

during a day. Plans, which are modified during the traffic

flow simulation can now be regarded as a new mutated

instance for the plans generation and day-to-day learning

process, although such a post-processing has not yet been

investigated.

Fig. 2. The traffic flow simulation is enhanced with within-day re-planning

capabilities. Agents are now able to mutate their plans during the run of the

traffic flow simulation.

Technically, day-to-day and within-day re-planning are

quite similar. The agent’s capabilities about how it can

modify its plan are the same and will be discussed later.

However, there are differences considering the knowledge the

agent possesses. With day-to-day re-planning the agent acts

based on the knowledge it accumulated over previous runs of

the traffic flow simulation. Whereas with within-day re-

planning the agent accesses rather locally available

information, such as current link travel times, and thus can

react to unforeseeable fluctuations in the traffic system.

To clearly distinguish between the packages of MATSim,

the package for plans generation and day-to-day re-planning

(represented by the box “plans generation” and “day-to-day re-

planning” in fig. 2) will be referred as the demand-modeling

package and the part that enhances the within-day re-

planning capabilities (represented by the box “within-day re-

planning” in fig. 2) will be referred as the telematics

package. For the purpose of this paper the telematics

package will be used as a standalone system without the

feedback of mutated plans to the demand-modeling package

depicted in fig. 2.

The following section describes the abstract agent model

that enables the basic within-day re-planning capabilities.

III. ABSTRACT MODEL DESCRIPTION

A. Basic agent model

The behavior of an agent is fully determined through its

plan. The linkage between the mobility simulation in the

physical layer and the mental agent representation is done by

the so called basic agent class. The basic agent holds an

activity plan and extracts the information required by the

mobility simulation out this plan. In particular these are: (i)

departure time, (ii) departure link, (iii) destination link and

(iv) the agent’s desired next link (if the agent is en-route).

Altering an agent’s behavior is possible only by

modification of its plan.

B. Agent brain

Presented so far, agent behavior is constrained to

execution of a predefined activity plan. In order to alter its

behavior it needs to become intelligent. Intelligence is

provided to the agent by equipping it with a brain object.

The essential task of a brain is to do on the fly modification

to the plan. Additionally, the brain is able to determine the

agent’s desire to re-plan at all. Possible capabilities of a

brain are departure time choice, activity location choice,

variations in activity sequencing and route choice.

An agent brain comprises three further components, each

of which represents a certain aspect of the re-planning

process. Descriptive information in the form of link travel

costs are provided by a so called link cost provider,

prescriptive information in the form of predefined routes are

obtained from a route provider and the agent’s satisfaction is

modeled by the contentment module. However, the ways in

which these three components are combined are up to a

concrete implementation of an agent brain. Figure 3 provides

an overview.

C. The re-planning process

On can interpret the activity plan as the agent’s intention

and the link cost provider as the agent’s believes. However,

the abstract model does not specify any commitment rules

which defines when to re-plan. Also the agent’s desires

(maximizing utility, maintaining the plan, etc.) are not

defined and are to be specified in a particular model and its

implementation.

The re-planning mechanism as it is processed by the agent

brain can be separated in three quite typical steps [12]:

1. Perception. The agent observes its current

environment which basically is given by the

current traffic state or retains information out if

its memory. The sensor system through which

the agent perceives or accesses its memory are

represented by the link cost providers.

2. Deliberation. The agent follows a certain strategy

to fulfill its desires. Both, the strategy and the

desires are to be specified in a particular model

implementation. Note that the deliberation about

to re-plan or not has already been done prior to

these steps by a certain commitment rule. Once

the re-planning process has been triggered the

agent definitely wishes to do so.

3. Execution. The agent modifies its plan according

to its deliberation in the previous step so that the

mobility simulation will move it through the

network following its new plan.

Fig. 3. Abstract agent model architecture. An agent is fed with an activity

plan generated by the demand-modeling package. The plan can be modified

by the brain with the help of a link cost provider, route provider and

contentment module. The mobility simulation moves the agent through the

network according to its plan. Information about instantaneous link travel

times are provided by the traffic flow simulation. Additionally the brain can

access the knowledge the agent accumulated in previous iterations.

D. Types of information provision

As mentioned above, the link cost provider and the route

provider represent two types of information sources. From

the conceptual side one can distinguish between intrinsic and

extrinsic provision. However, technically these sources of

information are dealt with in a unified way.

An intrinsic link cost provider may represent the agent’s

observation (what it can see by looking out of the window)

or the historical knowledge an agent accumulated in

previous trips. An extrinsic link cost provider may represent

an in-vehicle device, which supplies the driver with travel

times or messages broadcasted via radio.

For route providers this differentiation is not as distinct as

above. One may imagine an intrinsic route provider as a

representation of the process of acquiring a route when the

agent thinks by itself. The extrinsic counterpart may be an

in-vehicle navigation device, internet-based services,

variable destination signs (VDS) or even static guide posts.

The simulation framework currently implemented

provides three different types of basic link travel time

information.

Historical travel times represent the „typical“ state of the

traffic network4 as expected by the traveler, reactive travel

times (in literature also referred as naïve or instantaneous

travel times) represent a current snapshot of the traffic

network, and predictive travel times represent a forecast of

the traffic state within a given time window.5

Depending on the type of link travel costs that are

supposed to be modeled, an implementation of a link cost

provider may compose the three basic types to a new

representation of link costs. Of course link cost providers are

not restricted to the information generated by the mobility

simulation. Static information may be read out of files or

generated by other modules a priori. It would be even

conceivable that a link cost provider operates its own sub-

simulation to generate specific information.

IV. EXEMPLARY MODEL AND IMPLEMENTATION

This section presents an exemplary model

implementation. Its purpose is to validate the frameworks

applicability rather than to model realistic travelers’

behaviour.

A. Model assumptions

The model focuses on route switching, i.e. re-planning is

only done en-route with modifications to the agent’s current

route.

The agents’ desire is to always maintain the timing of its

plan, only being early is not considered as undesirable. The

contentment module is now used as a part of a commitment

rule, which decides when to re-plan. If the agent notices that

it will be late (e.g. due to congestion) it becomes displeased

and whishes to re-plan. To maintain its plan, the agent tries

4 Historical travel times are the accumulated knowledge of previous

iterations. As the relaxation process advances travel times are getting close

to a user equilibrium.

5 Such a prediction is generated once for all agents in a rolling-horizon

manner by running the mobility simulation forward without re-planning and

then switching back to the previously marked state.

to find a faster route based on its believes of link costs.

However, its available information is based on individual

observations and estimation of its surroundings and thus is

limited by the extent of its perceptivity.

The re-planning process is triggered as follows: Agents

are asked at each intersection by the simulation controller

about their desire to re-plan. If an agent is displeased the re-

planning process is triggered. However, re-planning is

computationally rather expensive (due to route searching)

and thus the simulation controller does only allow a certain

fraction of agents to re-plan. As the agent’s desire to re-plan

increases the probability that the controller selects it

increases, too. I.e., the simulation controller always tries to

determine the fraction of agents in such a way that it will

select the most displeased agents.

In this case the simulation logic diverts from a truly agent-

orientated approach, since a traveller that does feel the need

to re-plan would definitely do so in the real world. Truly

agent-oriented implementations would use independent

computing threads for each agent, and these threads might

decide by themselves when they become active. Current

computers are, however, not able to process as many threads

as agents (approx. 200,000 agents for large scenarios).

Systems with much more lightweight micro-threads might

be a way out and are a topic of ongoing research [13].

B. Link cost perception

An agent has two ways to acquire information: It can

observe its direct surroundings and it can estimate

information that cannot directly be perceived. Estimation is

based on current observation in combination with historical

and common knowledge. Observations are based on the

current state of the traffic network but are distorted by

individual errors of perception. Consequently, to model the

driver’s perception and estimation of link travel times the

historical and reactive travel times are required.

Observations are spatially limited to the agent’s current

link and on its immediately succeeding links (i.e. all

outgoing links of its downstream node). For these links the

agent is aware of the reactive (instantaneous) travel times.

However, to create behavioral diversity in a simple way the

agent’s perception is distorted by two individual perception

errors. The first is a white noise added to the reactive travel

times while the second represents the uncertainty of

appraising the correct travel time when the current traffic

state differs from the habitual known. If, e.g., link travel

times are twice as high as historically observed, an agent is

more unsure about the real travel time than if observations

match its historical knowledge.

Travel times of unobservable links are scaled according to

the agent’s current observation of link travel times compared

to the historically learned. Additionally, scaled and

perceived link travel times are corrected if they fall below

the free flow travel time.

C. Routing

The routing is done with a time variant Dijkstra best path

algorithm. The routing algorithm is supplied with link travel

costs by a link cost provider as described in the previous

section B.

In the literature, route choice models are often realized as

random utility models that account for the non-deterministic

behavior of humans [14]. However, MATSim focuses on

large-scale scenarios, and a discrete choice model for route

choice appears quite expensive in terms of computational

performance. Thus we choose a purely simulation-based

approach, apply strict shortest path algorithms, and realize

the non-deterministic behavior by randomization of link

costs as described in the previous section.

D. Contentment

Our exemplary implementation of the agent brain uses a

contentment module to determine the agent’s need to re-plan.

Contentment is represented as a scalar value out of the

interval [-1,1], where 1 means the agent is pleased, 0 the

agent is indifferent, and -1 the agent is displeased. As an

agent becomes displeased (i.e. values less than 0), its need to

re-plan increases.

The implementation for the contentment module used here

is based on a scalar scoring (utility) function for plans

introduced by Charypar and Nagel in [11]. This scoring

function evaluates the quality of a plan by summing the

utilities of all activities that are performed and all travel

(dis)utilities.

The utility for performing an activity is a logarithmic

function of activity duration whereas the penalty, or more

precise the negative utility for travel is modeled as a linear

function of trip duration.

The contentment of an agent is defined as the quotient of

the expected plan score (based on the agent’s current beliefs

of travel times) and the initial plan score (calculated from

the demand-modeling package).

V. SIMULATION

A. Introduction

To validated the applicability of the proposed framework,

two scenarios with the above introduced model were run

which will be described in the following: first a synthetic

corridor example (called the “simple” scenario), and then a

scenario based on real-world data from Berlin (called the

“Berlin” scenario). For both scenarios, the set-up is as

follows:

• As base case, a set of initial plans is given. Plans

contain departure times and routes for every agent.

• These plans are then run through the simulation

described in this paper. A number of experiments have

been conducted where the fraction of agents that are

allowed to re-plan has been varied. This set of agents is

chosen according to both, a global re-planning

probability and the agents’ need to re-plan. This is done

in such a way, that agents with high re-planning needs

are preferred for re-planning.

For the “simple” scenario, the initial plans are manually

constructed. For the “Berlin” scenario, they are taken from

the MATSIM demand-modeling package. The plans for the

“Berlin” scenario can be considered relaxed, i.e.

approximately in an user equilibrium. However, since the

equilibrium is only approximate agents are still able to

somewhat improve their performance.

B. “Simple” test scenario

Fig. 4. Simple test network. The gray arrows denote the routes as defined in

the initial plans. According to their initial plans agents depart at the six

leftmost horizontal links and travel via the middle route to the three

rightmost links.

We first set up a simple test scenario with a grid shaped

network including 41 links. All links are equal in their

attributes (1000 m length, 1800 vehicles per hour max. flow

and 7.5 m/s free speed). The demand consists of 6000 agents

departing at 7:00 and traveling from the left to the right side

(see fig. 4).

According to their initial plans, all agents use the middle

route. Because of its limited capacity, spillback occurs

shortly behind the demand entry points. The resulting travel

times are used as the historical traffic pattern. It is

questionable if this is realistic since the historical travel

times are not in user equilibrium. However, these simulation

runs are to demonstrate the capabilities of the presented

framework and rather than to simulate realistic travel

behaviour. Furthermore this extreme case shows more

clearly the effects of within-day re-planning.

We run several simulations with different types of link

cost providers to investigate the impact of descriptive

information provision. Beside the model of agents’ link

travel time perception presented in IV.B, we additionally

provided the agents directly with historical, reactive and

predictive travel times. As a comparison criterion we use the

average deviation from user equilibrium, i.e. the difference

of the route’s duration the agents actually experienced and

the best route calculated a posteriori, and averaged this over

all agents. In the “simple” scenario the average deviation is

9:50 for a simulation run without within-day re-planning.

Figure 5 shows results that one may not expect at first

glance. Predictive information provision leads to less

deviation from a user optimum than reactive information

provision. But re-planning with historical travel times leads

to better results than with reactive (instantaneous) travel

times. The link cost provider which models the agent’s

perception produces results that are between the ones of the

historical and the reactive link cost provider. At this point

the cause of the peak in the graph for historical travel times

is unknown.

Fig. 5. Comparison of link cost provider in the simple test scenario. ‘hist’ =

historical, ‘react’ = reactive, ‘pred’ = predictive (with 30 mins prediction

window) and ‘perceived’ = perceived travel times.

The results with reactive information provision

demonstrate the problem of overreaction. Overreaction

describes the situation in which drivers overcompensate in

response to information, again causing sub-optimal traffic

conditions. This effect can be well observed with high re-

planning probabilities.

C. “Berlin” test scenario

To investigate the applicability of the framework in real

world applications, a large-scale scenario with a reduced

road network representing the metropolitan area of Berlin

(Germany) has been set up. The network includes

approximately 2400 links and is bounded by the Berlin

beltway (fig. 6).

Fig. 6. Reduced road network representing the metropolitan area of Berlin.

Activity plans are now taken form the MATSim demand-

modeling package and represent a 10 percent sample of

Berlin’s population (approx. 170,000 agents).

The same investigations as with the simple test scenario

have been conducted for the Berlin scenario. The average

user equilibrium deviation without re-planning is 5:06.

On qualitative inspection of fig. 7 we now observe the

expected results: Better information leads to better results

(“predictive” < “reactive” < “perceived” < “historical”). The

phenomenon of overreaction with reactive information

provision does not occur in this scenario.

But in contrast to the simple scenario, increasing re-

planning probability does not always decrease the

equilibrium deviation. With the use of historical or perceived

travel times it even dramatically impairs results. Recall that

the initial plans for the Berlin scenario have undergone

several iterations in the MATSim demand-modeling package

and are thus close to the user equilibrium. Accordingly, it is

not possible to significantly improve the traffic state by

providing additional information to drivers and as more

agents are allowed to divert from their original route the

traffic state moves further away from the equilibrium.

Fig. 7. Comparison of link cost provider in the Berlin test scenario. ‘hist’ =

historical, ‘react’ = reactive, ‘pred’ = predictive (with 30 mins prediction

window) and ‘perceived’ = perceived travel times.

Considering the results of the predictive information

provider, the increasing values of user equilibrium deviation

seem to be related to the accuracy of the prediction. For

remembrance, the prediction is done by running the

simulation forward without within-day re-planning. It is

obvious that if more agents are allowed to re-plan, the

experienced traffic state potentially differs more from the

prediction as if fewer agents are allowed to re-plan.

VI. CONCLUSION

This work presented an agent-based framework to

enhance MATSim with capability of within-day re-planning.

The abstract model distinguishes between a module for link

cost perception, route searching and contentment. Together,

these modules represent an agent’s behavior. The framework

provides flexible options for adjusting the behavior by

choosing different implementations of the modules. For the

particular model presented here, Dijkstra’s best path-

algorithm has been used for the route searching model, a

scoring function to model the contentment, and historical,

reactive, predictive travel times as well as combinations of

them to model the perception of link costs. Beside the

exemplary implementation, additional implementations

modeling simple within-day destination choice [15] and

guidance by means of variable message signs [16] exist in

this within-day re-planning framework.

Two test scenarios demonstrated the frameworks

applicability. Although the exemplary model does not claim

to mimic realistic behaviour the simulation results appear to

be reasonable. Of course, further applications will require a

more careful calibration and validation of the model.

Altogether, it can be expected that the presented

framework provides valuable insights into the effects of ITS

measures not only in current and future traffic conditions,

but also on driver contentment itself.

Our future research will concentrate the decision making

process considering travel time uncertainty and risk

aversion. In this context it will be practicable to also deal

with departure time choice which has been neglected in our

early studies.

REFERENCES

[1] G. Wolfgang Heinze. Kurskorrektur - Eine Ortsbestimmunt der

Raumordnung aus Verkehrssicht. Working Paper 06-6, Inst. for Land

and Sea Transport Systems, TU Berlin, Germany, 2006.

[2] Adler, J.L., Recker, W.W., and McNally, M.G. "A Conflict Model

And Interactive Simulator (FASTCARS) for Predicting En-route

Driver Behavior in Response to Real-Time Traffic Condition

Information". Transportation 20 (2), 83-106, 1993.

[3] Adler, J.L, Recker, W.W., and McNally, M.G. "Using Interactive

Simulation to Model Driver Behavior under ATIS". Proceedings of

the ASCE 4th International Conference on Microcomupters in

Transportation, Baltimore, MD, 1992.

[4] Bonsall, P.W. and Parry, T. "Using an Interactive Route-Choice

Simulator to Investigate Drivers’ Compliance with Route Guidance

Advice". Transportation Research Record 1306, pp 59-68, 1991.

[5] Adler, J. L., V. J. Blue & T. L. Wu. “Assessing Network and Driver

Benefits From Bi-Objective In-vehicle Route Guidance”, presented in

the 78th TRB Annual Meeting, TRB, 1999.

[6] Emmerink, R. H. M., K. W. Axhausen, & P. Rietveld. “Effects of

Information in road transport networks with recurrent congestion”,

Transportation 22, pp.21-53, 1995.

[7] Wunderlich, K. E. “An Assessment of Pre-Trip and en route ATIS

Benefits in a Simulated Regional Urban Network”, in the 3rd world

Congress on Intelligent Transport Systems, Intelligent Transportation:

Realizing the Future, Orlando, Florida, 1996.

[8] Levinson, D., Gillen, D., Chang, E. “Assessing the Benefits and Costs

of Intelligent Tranportation Systems: The Value of Advanced Traveler

Information Systems”, California PATH Research Report, UCB-ITS-

PRR-99-20, 1999.

[9] TRANSIMS http://transims.tsasa.lanl.gov/. TRansportation ANalysis

and SIMulation System, accessed 2007. Los Alamos National

Laboratory, Los Alamos, NM.

[10] B. Raney, K. Nagel; An improved framework for large-scale multi-

agent simulations of travel behavior; in: P. Rietveld, B. Jourquin and

K. Westin (eds.), Towards better performing European

Transportation Systems.

[11] D. Charypar and K. Nagel. Generating complete all-day activity plans

with genetic algorithms. Transportation, 32(4):369–397, 2005.

[12] J. Ferber. Multi-agent systems. An Introduction to distributed artificial

intelligence. Addison-Wesley, 1999.

[13] L. Bläser. A Component Language for Structured Parallel

Programming, JMLC 2006, LNCS 4228 Springer 2006.

[14] M. Ben-Akiva and M. Bierlaire. Discrete choice methods and their

applications to short-term travel decisions. Kluwer, 1999.

[15] Gunnar Flötteröd and Kai Nagel. Bayesian modeling and estimation of

combined route and activity location choice. Working Paper 06-3,

Inst. for Land and Sea Transport Systems, TU Berlin, Germany, 2006.

[16] Carl Rommel. Automatic Feedback Control Applied to

microscopically Simulated Traffic - The potential of route guidance in

the Berlin traffic network. Master’s thesis, Inst. for Land and Sea

Transport Systems, TU Berlin, Germany, 2007.