Agent-Based Simulation of Human Movements

During Emergency Evacuations of Facilities

Joseph L. Smith, PSP
Senior Vice President
Applied Research Associates, Inc., USA

Abstract
During and after any extreme event of natural or man-made origin, getting
people out of harms way and to an area of safety are primary objectives. Orderly
and efficient evacuations are the key to saving lives. This was clearly
demonstrated during the attacks on the Pentagon and World Trade Center towers
on September 11, 2001. One of the major recommendations from the US
Governments Draft Final Report on the World Trade Center Towers collapse
was that building evacuation modeling and procedures should be improved to
facilitate safe and rapid egress and support better occupant preparedness for
evacuation during emergencies. Performing actual physical evacuations of large
occupied facilities is expensive and disruptive to occupants and visitors.
Accurately evaluating the affects of various events such as loss of power, smoke,
fire, explosions, chemical contamination and structural collapse on human
movements is not feasible in real world evacuation exercises. This paper
examines a new technology, computer-aided, intelligent agent-based evacuation
modeling, that can model and help plan emergency evacuations; run numerous,
accurate event-driven evacuation scenarios; support the design of egress for new
facilities or facility modernizations; support research in the areas of human
behavior in disaster sociology; and model the movement of first responders and
security personnel.
Keywords: emergency evacuation, scenario-based simulation, agent-based
evacuation modelling, disaster management.

Introduction

Evacuating large crowds of people under any circumstance is a challenge.

Evacuation of large facilities during an emergency or disaster is a much more
complex task because of the added elements of chaos, panic, and the high density
of the population. As terrorists move more towards preying on soft targets such
as hotels, hospitals, airports, train stations and stadiums, the potential for large
government, commercial, residential and sports facilities to become targets is a
reality for which we must prepare. Non-terror related events and accidents
involving inadequate planning and training for mass evacuations are also all too

common. Examples include the February 2003 Rhode Island nightclub fire that
killed at least 96 people and the August 2, 2004 supermarket fire in Asuncion,
Paraguay, that killed over 300 shoppers. Determining the most effective
evacuation plan for a large public facility requires in-depth analysis of multiple
factors. Determining the best routes, foreseeing potential problems, addressing
the chaos/panic factor, and orchestrating the evacuation are all critical aspects
that should be evaluated in a well developed disaster management plan.

Figure 1. Recent evacuations have demonstrated the need to better understand

and plan for large-scale movements of people under emergency conditions
(World Trade Center evacuation, September 11, 2001 [left] and Evacuation of
the US Capitol May 11, 2005 [right]).
To date, most egress planning has been accomplished with so-called hydraulic
models where people are assumed to flow out of a facility much like water
flowing out of holes in a water filled tube. Such models include typical rulebased speed limits for the occupants and calculate the flow of people as they pass
through openings. These simplistic models are not generally capable of
simulating human behaviour or the interaction of people with their changing
environment, and are not capable of assessing the effects of various, perhaps
even multiple, events that may occur during an evacuation.
While many facility owners do conduct evacuation drills, such drills are most
often not accomplished during times of peak facility occupancy and do not
include event based scenarios such as a bombing followed by fire, smoke, partial

collapse and falling debris that would impair or preclude egress through various
areas of the facility. Until recently, many evacuation plans called for phased
evacuations of only affected floors or areas of large facilities (e.g., if a fire was
detected in a high rise building, that floor plus the adjacent floors would be
evacuated first followed by the adjacent areas). Hence, it has to date been
difficult to accurately understand and predict full building evacuations under
realistic emergency conditions. Mass or total building evacuations pose
additional problems of stairwell congestion, chaos and potential stampeding of
panicked occupants. This is even more important in the post-September 11th era
where most people, especially in high rise facilities, now seem to demonstrate a
higher propensity to evacuate quickly.

Event Simulation

It is now possible through the use of computer simulations to model the

behaviour of human movements during evacuations that include events that
impact the simulated humans ability to successfully egress. Applied Research
Associates developed an advanced computer program called E-SIM (Event
Simulator) that employs an agent-based approach to modelling the behaviour of
people attempting to evacuate a facility. Each agent (or simulated human) has
unique properties such as size and speed that allow the realistic simulation of
various physical and behavioural characteristics. The simulation updates agent
positions and paths each time step during the model run. The following
discussion presents some of the major features that are considered in such an
advanced simulation tool. Since the E-SIM tool is one of the most advanced
evacuation simulation tools currently available, it is used as an example to
demonstrate the major considerations required for such analysis.
2.1 Model Basis
Calculating the movement of people in a virtual environment poses considerable
problems. Where do they go? How do they know where to go? What if they
encounter an unforeseen obstacle on the way? What if something impedes their
movement?
While humans may instinctively react and adapt to their
environments, getting a simulated human to do so is a difficult task.
Agent routing in E-SIM is computed using a two level hierarchical path finding
approach based on the A* algorithm. The A* search algorithm (pronounced "Aystar") is a tree search algorithm that finds a path from a given initial node to a
given goal node. It employs a heuristic estimate which ranks each node by an
estimate of the best route that goes through that node. Using this approach, a
high level pathfinder determines the route from the agents current position to
the destination using a graph of rooms connected by exits. Once the high level
path is defined, a low level pathfinder is executed each time an agent transitions
from one room to another. The low level pathfinder is a basic A* approach
which treats each room as a triangular mesh with holes where obstacles prevent

agent movement (see Figure 2). This allows the agent to intelligently navigate
around corners or obstacles such as furniture that may be in a room.

Figure 2. Basic path finding algorithm uses a triangular mesh.

Each agent also has the ability to remember problems or obstructions that they
encountered while attempting to exit the facility. This memory is used by the
pathfinder to avoid routing an agent back to a blocked exit. Agents that pass
within a threshold distance can also communicate their knowledge of blocked
exits and other observations/problems to other simulated humans. If the agent
gains information from another agent that they did not previously know, and the
information identifies a problem along their current path, the high level
pathfinder is invoked to determine an alternate route to their destination.
Most people in the real world often do not choose the optimal path out of a
facility. Studies clearly indicate that most people normally will first select the
path they used to come into the facility even when other paths are shorter. Hence
it is important to consider alternates that are not necessarily optimal. An E-SIM
user can specify which exit(s) are primary exits and determine what percentage
of the population in the building will use these exits. During an emergency,
people will likely use the exit they are more familiar with and/or the exit they
used to enter the building that day. Since people are continually exposed to exit
signs but are rarely, if ever, required to use non-familiar exits, they may have
learned to filter out this information. During an emergency, it is unlikely that
occupants will be prepared to try a route they have never used before to leave the
building. (Reference 1).

Agent Information

In a model such as E-SIM, agents may be defined by their position, condition,

size, speed, and response time. The condition of the agent is whether or not they
are dead, waiting because someone is in their way, waiting because of their
response time to external stimuli, etc. When modelling an individual or a few
simulated humans it is readily possible to manually describe individual unique
characteristics for specific agents. However in large simulations involving
hundreds or thousands of agents, it is important to be able to statistically describe

groups of agents. E-SIM was develop to handle this by allowing a group of

agents to be added to a room or to building floor level in either a random or a
specific pattern. When adding a group, the size and speed of the group can have
either a uniform distribution or a normal distribution of characteristics. A
normal distribution assigns agent attributes along a bell shaped curve. The
uniform distribution assigns values within the interval of possible values, with
any value as likely to occur as any other value. An agent type can also be given
a unique color in order to visually distinguish one agent type from another (i.e.,
male, female, child, handicapped, and visitor).
3.1 Size
In an agent-based simulator, the physical size of the simulated humans is
important especially when egress is considered in close quarters such as
elevators, stairwells, and corridors. A good source of information on standard
sizes is the 2000 Architectural Standards (Reference 2). Typical agents are
assumed to have sizes defined by minimum and maximum diameters (see Table
1). E-SIM users are not limited to using standard dimensions. First responders
and emergency personnel should be assumed to be equipped with material that
will effectively make them larger in confined spaces (i.e., they may be carrying
axes, oxygen equipment, fire fighting equipment, etc.). Hence, provisions must
me made in agent-based simulations to account for the additional space required
by such agents. Fully equipped first responders may be assumed to be 50 to
100% larger other than normally dressed adults.
3.2 Speed
E-SIM has built in speed ranges for predefined types of people. A maximum
and minimum speed may be specifically assigned to individual agents or
randomly distributed to each simulated person in that particular category. The
user may change the maximum and minimum or define their own class of
people. The maximum speed a user can enter for any one person is normally
limited to 1.8 m/s (Reference 3). It is important that agent-based simulations
also take into account that the speed of an individual will decrease when
ascending or descending stairs or if the individual is disabled. Further, when a
person approaches another person in their path, the person behind will try to go
around the person in the way. That persons speed will decrease as he/she
approaches the other person. Typical human speeds are given in Table 2.

Table 1. Typical simulated human dimensions (inches).

Men
Women
Children
Age

Minimum
(2.5 percentile)
16
14.4

Maximum
(97.5 percentile)
19.4
17.7

9.5
9.1
9.9
10.5
11.9
14.8

9.5
11.5
12.6
13.8
15.6
18.2

1
3
5
7
9
12
15

Table 2. Typical movement speeds (meters/second) for simulated agents.

Men
Women
Disabled
Locomotion Disability
Disability No Aid
Disability Crutches
Disability Walking Stick
Disability Walking Frame
Disability Electric
Wheelchair
Disability Wheelchair
Disability Assisted
Wheelchair
Disability Assisted Walk

Min.

Max.

1.10
1.05
0.71
0.57
0.70
0.67
0.49
0.34

1.60
1.45
1.25
1.02
1.02
1.24
1.08
0.83

0.89

0.89

0.38

0.94

1.02

1.59

0.58

0.92

Min
Down
Stairs
0.85
0.85

Max
Down
Stairs
1.05
1.05

Min
Up
Stairs
0.85
0.85

Max
Up
Stairs
1.05
1.05

0.22
0.20
0.22
0.24

0.45
0.47
0.22
0.46

0.26
0.35
0.26
0.35

0.52
0.55
0.45
0.35

0.53

0.90

0.55

0.78

3.3 Response Time

The response time of an agent varies depending on their particular situation.
Response time is defined as the time span between when a person first hears an
alarm or signal until he/she starts their exit out of the building (Table 3). For

instance, a person asleep in a hotel will have a longer response time than if
he/she were awake because he/she will have to first wake up, get oriented, get
dressed, and then leave. Another factor that effects response time of a person is
the type of alarm. An alarm with a live voice giving instructions is more
effective with a shorter response time than just an alarm sounding. E-SIM
allows the user to define any value for the response time of any individual.
Some agents may take tens of minutes to respond (e.g., the guy who just has to
make that last phone call or who just doesnt believe this is real alarm).
Table 3. Typical ranges of response time (seconds).

Awake Familiar
Awake Unfamiliar
Asleep Familiar
Asleep Unfamiliar
Required Assistance

Min.
0
0
15
30
60

Max.
60
120
120
180
600

For example, a typical male office worker could have the following
characteristics. His size would be that of a typical male; therefore his diameter
would be between 16 and 19.4 inches. His speed would be between 1.10 to 1.60
m/s on a flat surface and between 0.85 to 1.05 m/s going up and down stairs. His
response time would be that of an awake and familiar person, between zero and
60 seconds. If this typical male office worker has broken his leg and is now
using crutches, his speed would reduce to that of a person who uses crutches. So
his speed on a flat surface would be between 0.67 to 1.24 m/s, 0.22 m/s going
down stairs, and between 0.26 to 0.45 m/s going upstairs.
Special cases must also be considered in advanced evacuation simulations. For
example, the speed and behaviour of a mother with small children will be
significantly different than if the same woman were alone. Another example
would include caregivers and patients. Any advanced event/evacuation
simulation tool must be able to account for such behaviours and speed
restrictions. Finally, in order to populate large facilities, an event simulator
should provide a method for groups of agents to be added to each layer or floor
of the facility using either a uniform or normal distribution incorporating the
minimum and maximum speed and response time.

Scenarios and Results

The types of scenarios that may be

simulated
in
an
agent-based
event/evacuation
simulation
are
numerous. The significant advantage
of an advanced simulation tool such

Figure 3. Sample simulation

showing injured agents.

as E-SIM is the ability to introduce multiple events within a single evacuation

scenario that allows the user to evaluate potential real world situations that could
not be safely or economically reproduced in an actual evacuation drill. E-SIM is
configured to simulate any combination of two types of events. These are
classified as either environmental that will slow or impede an agents progress,
or standard that will kill all agents within the designated event area and deny
all access to others attempting to pass through an area. Examples include:
Environmental Event
Loss of power
Loss of light (darkness)
Water (flooding)
Smoke (light)
Biological agent exposure

Standard Event
Explosion
Toxic chemical exposure
Structural collapse
Smoke (heavy)
Fire blocking egress

4.1 Example
Figure 4 shows a sample output from an E-SIM analysis. Data can be presented
as static snapshots or as movie media. Such graphical display of the information
allows the users to readily identify areas of concern and take corrective actions
or make plans as necessary. In one case analyzed for a large facility, a client
recently discovered unforeseen problems. In that case, a shelter-in-place plan
was evaluated. The simulation clearly showed that the facility had inadequate
interior shelter areas and potentially produced a crowd control problem due to
the large number of visitors expected in the facility. It would not have been
feasible to quickly and clearly identify this as an issue without an advanced
simulation of the plan.
4.2 Validation
E-SIM has been validated at the component level, through planned comparison
to evacuations conducted on small facilities, and through comparison to
anecdotal information available from actual large scale evacuation events. Data
such as that shown in Figure 5 from a simulation of a small facility evacuation
have been shown to be useful in comparing to actual observed data. Figure 5
shows the number of people on a particular floor as well as the number of people
entering and leaving that floor level as a function of time during the evacuation.
There is no known substantial documentation of full building evacuations of
large and heavily populated facilities at this time.

Figure 4. Sample simulation. Top normal evacuation where most people

exit from main exit (oval). Bottom agents exit from alternate exit
(rectangle) due bombing in lobby that killed simulated agents.

Figure 5. Data from a sample simulation (agent count on a specific floor).

Conclusion

Modelling emergency evacuations through the use of computer aided, agentbased simulation offers significant advantages over more simplistic modelling
approaches. Such tools may be used to evaluate egress and events in existing

facilities as well as in the design and layout of new facilities. Future versions of
the E-SIM tool will include enhancements to address evolving egress problems.
One such new issue is the use of hardened window systems in hurricane and
terror resistant construction. Such windows (Figure 6) are, by design, difficult to
break and will impede emergency egress by occupants or ingress by first
responders into a facility. Without planning and training for such contingencies,
an unsuspecting person may panic and be trapped in a dangerous
situation.

Figure 6. Woman unsuccessfully attempting to pass through a hardened

window system by using office furniture (computer) as egress tools.

References
[1] Engineering Guide to Human Behavior in Fire, SFPE, Review Draft, August
2002.
[2] Young, Architectural GRAPHIC standards CD-ROM, John Wiley & Sons,
Inc. New York, NY 2000.
[3] Yang, Fang, Huang, and Deng, Occupant Evacuation Model Based on
Cellular Automata in Fire, Chinese Science Bulletin, Vol. 47 No. 17
September 2002.