Multi-agent Transport Simulation for Regional
Mohamed Bakillah1, Hubert Klüpfel2, Gregor Lämmel3, Georg Walenciak1
1Rupprecht-Karls-Universität Institute for GI-Science, Heidelberg, Germany
2TraffGo HT GmbH, Duisburg, Germany
3Technical University of Berlin, Institute for Traffic Systems Planning, Berlin, Germany
Abstract. As periodic wildfires in the southern California, USA, or the Fuku-
shima Daiici nuclear disaster have shown, man-made and natural disasters
might make human areas or regions (temporarily) inhabitable or the safety of
inhabitants might be threatened severely, and the evacuation of this area re-
quired. One major decision in facing disasters is whether to evacuate or not.
Criteria for making that decision on a civil defence level are: the time available
for evacuation, the socio-economic situation, the warning systems available, the
severity and time evolution of the thread, and so on, and so forth. One major
question is whether the available safe evacuation time (ASET) is larger than the
required safe evacuation time (RSET). Another one is the feasibility traffic pat-
terns that will result from an evacuation warning. The method presented in this
paper is able to determine the RSET.
1 The GRIPS Research Project
The GRIPS-Project (GIS-Based-Risk-analysis-Information and Planning System)
(http://www.geog.uni-heidelberg.de/forschung/gis_grips.html) aims at providing use-
ful and objective information for the decision makers facing questions like whether to
evacuate or not. Within GRIPS, the multi-agent transport simulation toolkit MATSim
(http://www.matsim.org) is used to simulate the evacuation regarding two use-cases.
The first use case is simulation of the evacuation of the district Wilhelmsburg in
Hamburg, Germany, in consideration of an upcoming flooding (details are described
by Durst. et. al. in these proceedings). The second use-case examines the simulation
of the evacuation of an area affected by an industrial accident in the city of Essen,
Germany. In neither case, the disaster is modelled directly, but handled as an external
parameter “threat” that affects a certain region. This is the region potentially to be
One of the main goals of the project is the development of a computer tool that can
be used by disaster managers as a decision support system. For that matter, both uses-
cases are specified in cooperation with local authorities. In this context, two central
aspects are of interest: on the one hand the tool might be used shortly after a disaster
occurred. Then, civil defence authorities have to decide whether a district should be
evacuated or not. On the other hand, the tool might be used in pre-disaster planning,
i.e. adapting the traffic and civil defence infrastructure (like shelters).
From the scientific perspective, there are two central issues: extending the existing
agent based queue model for transport simulation by a 2D simulation for simulating
pedestrians on the small spatial scale (see Laemmel et. al. in these proceedings) and
the development of a generic data model which describes a wide variety of evacuation
2 Model Components and Behavioural Categories
We will only briefly summarize the state of the art here. Pel et.al. provide a compre-
hensive and current review of the state of the art in dynamic evacuation traffic simula-
tion . They describe three model components:
1. dynamic travel demand model,
2. dynamic trip distribution model, and
3. dynamic traffic assignment model
and three behavioural categories described by these three model components:
4. evacuation participation and departure time choice,
5. destination choice, and
6. route choice.
The parts 2 and 3 are part of the MATSim traffic simulation. The travel demand is
currently assumed to be a certain percentage of the overall population. Demand can
be assigned to buildings or blocks. The challenge is the fine-grained estimation The
participation and departure time choice can be modelled by statistical distributions.
Further details can be found in .
3 Multi-Agent Transport Simulation for Regional Evacuation
Existing models are either focusing on the simulation of large areas but simplifying
geometric details or they are focusing on small-scale scenarios (like single buildings)
with high geometric resolutions. The latter are usually not suited for large areas be-
cause of their computational complexity. In the GRIPS project, multimodal traffic is
simulated. Currently, the modes are pedestrians, busses, and cars. For vehicular traffic
it seems to be reasonable to use a simulation model with a coarse resolution like the
queue model in MATSim. However, for more complex situations the queue model is
too coarse. This is usually the case when it comes to pedestrians who are navigating
through complex environments like train stations. This is a particular problem in
Hamburg, since the city railway is part of the authorities’ evacuation plan.
When doing multimodal simulations there is also a need to have mode change op-
tions. For example a pedestrian might start at home and walk to the next bus stop.
After she has reached the bus stop, she has to wait for the bus. When the bus arrives
she has to get on the bus. Now let’s assume the bus takes her to the next train station,
where she leaves the bus and walks to the platform to get on a city railway. The city
railway then finally takes her out of the evacuation area.
4 Focus of Application: Flooding and Industrial Accident
The development of the tool includes two basic use-cases, the flooding scenario for
Hamburg-Wilhelmsburg and the industrial accident scenario for Essen. Nevertheless,
GRIPS is intended to be applicable to a variety of possible cases. Therefore, an exten-
sible data model is developed containing the types of scenarios and their scope. One
important aspect regarding the scope may be the scale. For that, the evacuation of a
plane or a single building is considered to be out of scope for GRIPS because the tool
addresses regional evacuation. For other tools the scope might be building evacuation.
Within the data model, the input required to specify a scenario is analyzed. For
now, the minimum information for performing an evacuation simulation in GRIPS is
the infrastructure (e.g., road network), the population characteristics and the evacua-
Fig. 1. Screenshot of the GRIPS User Interface.
In MATSim, the road network is represented as directed graph. In any case, the
network must be suitable for routing. However, different information is required for
pedestrian routing, vehicular routing or public transport routing. Different scenarios
require different input. The data will therefore distinguish between mandatory input
and optional input data. Further the data model will have to define important scenario
types like pedestrian evacuation.
5 The GIS and ABM based Simulation Tool
The simulation module is based on MATSim, which represents the street and road
network as a queuing network. The agents are located initially on a link close to their
initial position, i.e. “on the street”. The time it takes for the agents to get onto the
street could be calculated by heuristics based on the respective values from the litera-
ture [10-21]. This time will then be part of the reaction time. The reaction time distri-
bution has a strong influence on the formation of congestion. If all persons react im-
mediately, i.e. start evacuating just after the alarm is triggered, this will result in more
congestion. In summary, the pre-movement time can be divided into
t pre-movement = t detection + t alarm + t decision + t preparation (1)
The times for detection and for alarm are global, the time for decision and prepara-
tion are individual parameters. It is sufficient to impose a statistical distribution for
the pre-movement time. The results presented below have been obtained using the
prototype version of GRIPS. The workflow comprises three steps
1. Creating the scenario
(a) Road network from OSM + network change events (XML)
(b) Population from GIS layers (e.g. land use, buildings, currently shape-file)
(c) Evacuation area as a polygon (e.g. from hazard calculations, currently shape-
file, i.e. static)
2. Running the simulation
3. Analysis of the results
Fig. 2. Screenshot of the Network (visualized with SenozonVia, www.senozon.com)
An example for the network is shown in Fig. 2. This network can be dynamically
changed when the simulation is running. There is currently no on-trip re-routing of
the agents, i.e. the agents do not change their direction when the network changes.
This is only taken into account when the agents re-plan their route in the next iteration
(for details on network change events for the case of inundation after a Tsunami,
please refer to ).
6 Simulation Results
In this section we present simple simulation results as a proof of concept. The focus is
on the workflow suitable for the application at the civil defense center of command
and control and for planning purposes (cf. Fig. 1). The road network shown in Fig. 2
was created from open street map (www.openstreetmap.org). First, the map.osm file
is imported into QuantumGIS (an open source program). The evacuation area and the
population polygons are drawn on separate GIS layers and saved as shape-files. All
three files, map.osm, population.shp, and the evacuationarea.shp are processed by the
MATSim scenario generator (code and jar-file available at www.matsim.org). The
scenario generator creates network.xml and population.xml, i.e. standard MATSim
input files. These can be processed by the MATSim evacuation simulation module.
The evacuees are located on links close to their initial position. In the first iteration,
they are routed to the evacuation node (currently visible in the center of the network,
cf. Figs. 2 and 3). Figure 3 shows a snapshot of the simulation, i.e. the situation ten
minutes after all the agents have started to evacuate.
Fig. 3. Screenshot after 10 minutes (visualized with SenozonVia, www.senozon.com)
The simulation is performed iteratively . The agents are searching for a route
which leads to shorter travel times (based on the cost function, e.g. the travel time).
This process is illustrated in Fig. 4: the length of the paths increases with each itera-
tion. Commuters analyze their decisions and try alternative routes to save time. These
alternative routes might be longer (in distance) but faster (in time).
Fig. 4. Travel distances increase with the iteration.
This of course poses the question, whether this iteration process produces realistic
results. There are basically two answers to this question:
1. Iteration is interpreted as anticipation
2. No iteration, i.e. all agents use the geometrically shortest path.
Additionally, an iteration in between can be used. The first iteration leads to the
longest overall evacuation time. The agents are just taking the shortest path. There
might be considerable congestion and no anticipation or reasoning is used concerning
the influence of traffic on the agents’ route choice decision. If the agents don’t have
any knowledge or experience concerning the evacuation. On the other hand, the final
iteration – after relaxation into the Nash equilibrium or user optimum – shows a situa-
tion where the agents take into account the traffic situation and search iteratively (or
by clever anticipation) for an optimal route. Optimal here means that unilateral
changes do not increase the individual performance, e.g. do not decrease the travel
time (if the cost function just takes into account the travel time).
Other strategies might also be suitable. For example, the risk of being hit by the in-
coming wave after a Tsunami might be part of the cost function. In this case, no part
of the route should go downhill, if risk aversion is very high. A longer route is then
accepted because it does decrease the risk in a situation where the warning time is not
know (see  for details.
Fig. 5. The scores increase with the iteration.
Fig. 6. : The vehicles “en route” decrease with time.
The parameters for the road network in our simulation are the standard ones. The
capacities depend on the road category . The results shown in Fig. 6 are for itera-
tion 0, i.e. for the shortest path. The curves depend on each other, via
N en route + N arrivals = const. = N total (2)
In this example, the total number (cf. eq. 1) is 1000 agents. They all depart immedi-
ately, which is of course an artificial assumption and used for the purpose of illustra-
tion, i.e. to produce a high demand. For iteration 0 (Fig. 6, the overall time is four
hours) which is due to the sever congestion (Fig. 2). Basically all agents use the same
7 Summary, Conclusion, and Outlook
We have presented an approach to the traffic simulation for regional evacuation. The
scope and aim of the GRIPS research project are described. The paper focused on the
work flow and presented some simple results. The reader is invited to reproduce the
results. All components used in the workflow are freely available, i.e. open source
software (MATSim and QGis) and open access data (www.openstreetmap.org).
Acknowledgement. This project was funded in part by the German Ministry for Educa-
tion and Research (BMBF) under grants 13N1138[1-3] (“GRIPS”).
1. Freie und Hansestadt Hamburg, Behörde für Inneres und Sport (2008): Sturmflutbroschü-
re, http://www.hamburg.de/contentblob/569738/data/sturmflutbroschuere.pdf. (05.032010)
2. G. Lämmel, H. Klüpfel, and K. Nagel: Risk minimizing evacuation strategies under
uncertainty. 5th International Conference on Pedestrian and Evacuation Dynamics,
Washington, DC, Mar 8-10, 2010.
3. Nagel, K. et. al. (2010): MATSim - Multi-Agent Transport Simulation Toolkit, TU Berlin
und ETH Zurich, URL: http://www.MATSim.org/ [Stand: 03.03.2010]
4. H. Klüpfel: Models for Crowd Movement and Egress Simulation. In: S. Hoogendoorn et.
al.: Traffic and Granular Flow 2003. Springer, Berlin, 2004.
5. W.M. Predtetschenski and A.I. Milinski (1971): Pedestrian Flow in Buildings. Müller,
6. S. Buchmüller and U. Weidmann: Parameters of pedestrians, pedestrian facilities, and
walking facilities. IVT, Technical report 132, ETH Zurich, 2007.
7. RiMEA-Projekt: Richtlinie für Mikroskopische Entfluchtungsanalysen, www.rimea.de
8. A. Schadschneider, W. Klingsch, H. Klüpfel, T. Kretz, C. Rogsch, and A. Seyfried.
Evacuation dynamics: Empirical results, modeling and applications. In: Encyclopedia of
Complexity and System Science. Springer, Berlin and Heidelberg, 2009.
9. Pel, Adam J.; Bliemer, Michiel C. J.; Hoogendoorn, Serge P. (2012): A review on travel
behaviour modelling in dynamic traffic simulation models for evacuations. In:
Transportation 39 (1), S. 97–123.
10. Gregor Lämmel: Escaping the Tsunami: Evacuation Strategies for Large Urban Areas
Concepts and Implementation of a Multi-Agent Based Approach. Dissertation, Technical
University of Berlin, Fakultät V, Verkehrs- und Maschinensysteme, Berlin, 2011.