Integrating Agent Based Modeling Into a Discrete Event Simulation'

Abstract

Movement of entities in discrete event simulation typically requires predefined paths with decision points that dictate entity movement. Human-like travel is difficult to model correctly with these constraints because that is not how people move and large individual differences exist in capabilities and strategies. Agent based modeling is considered a better way to simulate the real-time interaction of people with their environment. In this paper we propose to integrate agent based modeling with discrete event simulation to simulate the movement of people in a discrete event system. An agent based module was constructed within the AutoMod simulation package, and a test case was modeled in which people (agents) at a theme-park interact with objects and people in their environment to get directions and then walk or take a tram to their final destination. We explain the details of model implementation and describe the verification and initial validation of the model.

Full text

Proceedings of the 2005 Winter Simulation Conference
M. E. Kuhl, N. M. Steiger, F. B. Armstrong, and J. A. Joines, eds.
ABSTRACT
Movement of entities in discrete event simulation typically
requires predefined paths with decision points that dictate
entity movement. Human-like travel is difficult to model
correctly with these constraints because that is not how
people move and large individual differences exist in capa-
bilities and strategies. Agent based modeling is considered
a better way to simulate the real-time interaction of people
with their environment. In this paper we propose to inte-
grate agent based modeling with discrete event simulation
to simulate the movement of people in a discrete event sys-
tem. An agent based module was constructed within the
AutoMod simulation package, and a test case was modeled
in which people (agents) at a theme-park interact with ob-
jects and people in their environment to get directions and
then walk or take a tram to their final destination. We ex-
plain the details of model implementation and describe the
verification and initial validation of the model.
1 INTRODUCTION
The benefits of Discrete Event Simulation (DES) are well
documented throughout industry, military and academia
(E.g., Banks et al. 2003; Banks, Carson, Nelson, and Nicol,
2001; Law and Kelton 2000). There are certain situations,
however, that are very difficult to simulate using DES.
Any model that requires free movement of entities or a
very detailed movement pattern is not easily simulated
with DES. As an example, service systems through which
humans flow in unpredictable paths, while quite common,
are difficult to model in DES. Very strict assumptions
about the human’s choices have to be made by the modeler
in order to fit their behavior into the format of DES model-
ing. Furthermore, the real-time decisions of individual en-
tities would be very difficult to model using DES.
The primary limitation of DES involves the generality
of movement and function of entities in the simulation
world. In all commercial DES packages, a path must be
drawn from point A to point B in order for an entity to
move between those two points. This requirement makes
the modeling of human movement with DES unrealistic in
situations where paths are not predetermined (e.g., model-
ing congestion in a gift shop or the exit of patrons leaving a
sporting event). It is difficult to map every path the entity
could take between two points, and even more difficult to
assign probabilities to each of those paths for the navigat-
ing between the two points. A second limitation of DES
involves the ability of an entity to make decisions at very
small time increments. To do this in a DES, an analyst
would need to place a large amount of decision points ex-
tremely close to one another along the path of travel, while
also programming the required decision logic to be proc-
essed at each point. This process is not only laborious, but
the results would be difficult to verify and validate. A
third limitation of DES that affects the modeling of human-
like travel is that processing is done at user-defined deci-
sion points and is not autonomous to the flowing entities.
Because any processing or routing decisions must be made
by the DES servers, intelligent entities cannot be modeled
directly. This, in turn, limits the entities’ human-like func-
tion and could have an effect on the interaction between
the entity and all other objects in the simulation world. It
should be noted that all of these limitations can be over-
come using DES functions and clever programming, but
that does not negate the fact that DES is not well-suited to
the modeling of human movement and decision patterns.
Agent Based Simulation (ABS) offers a more straight
forward and accurate solution to these issues. ABS at-
tempts to simulate intelligent, autonomous entities (agents)
as they interact to attain some goal in their environment.
This simulation approach has been used to model many
different situations such as social evolution, segregation,
disease propagation, and advertisement effectiveness.
ABS has become a highly researched topic in the simula-
tion industry over the past decade. Since 2000, several
Conferences have formed (MABS, ABS, and AAMAS)
and many well established conferences (WSC, SimSol, and
INTEGRATING AGENT BASED MODELING INTO A DISCRETE EVENT SIMULATION
Benjamin Dubiel
Omer Tsimhoni
Department of Industrial and Operations Engineering
1205 Beal Avenue
University of Michigan
Ann Arbor, MI 48109, U.S.A.
1029

Dubiel and Tsimhoni
the SCS conference series) have opened tracks focusing
strictly on this powerful simulation approach.
In “Growing Artificial Societies”, Epstein and Axtell
(1996) state that an ABS consists of three primary ele-
ments: single or multiple autonomous agents, an environ-
ment or space, and rules that govern the movement and in-
teractions of the agents. With this definition, developing
an agent based module to simulate human movement is
possible as long as the agents are kept autonomous and the
correct rules are implemented to control their actions.
There are several software packages that can be used
to build agent based models. Among these are: AnyLogic,
(X J Technologies 2005), the EXODUS packages, (FSEG
2003), RePast (RAST 2005), and Swarm (Introduction to
Swarm 2005). Each of these packages is very specialized
and is designed to produce a specific model type such as
models with a large number of individuals (AnyLogic or
Swarm) or models that observe the behavior of individuals
in emergency situations (Exodus packages). Some of these
agent based packages can model simple human-like travel,
but none of them can integrate easily with common DES
packages.
In this paper, we argue that ABS can be integrated
with DES to model humans traveling freely through a DES
system. Adding the functionality of human-like movement
to a commercial DES package is extremely relevant when
considering simulation in the service industry and particu-
larly the theme park industry. The ability to model scenar-
ios including free moving, pseudo-intelligent individuals is
a very desirable tool, as this would allow for the modeling
of flow and design scenarios while utilizing existing DES
models of service and transportation systems around theme
parks.
We explore the integration of ABS and DES through a
test case which represents a real-world problem from the
theme park industry. A customer with limited knowledge
of the surrounding environment searches and walks to an
information source to obtain the location of a goal object
and then either travels on a discrete movement system
(tram) or walks to the goal object. The model can be used
to answer questions about location and quantity of maps,
signs, and informed employees in order to minimize travel
time of customers and maximize flow around parks. This
model could also predict arrival patterns to the discrete
components in the system (tram or queues) for multiple
different configurations of information sources. Overall
this integration would benefit many different scenarios
where free movement of entities was required.
2 SOFTWARE IMPLEMENTATION
The student version of the AutoMod simulation package
was chosen for the building and experimentation portions
of this project. This decision was made due to the first au-
thor’s detailed knowledge of the package, along with the
fully defined 3-D coordinate modeling space available in
AutoMod. With this feature, it is possible to build the
physical portion of an agent based model that will allow
agents to move freely throughout the environment. As a
proof of concept, the cognitive and interaction modeling
was based on a very simplified model of human percep-
tion, decision making, and movement. More complex
computational algorithms representing human behavior can
be added at a later point.
For simplicity purposes, the Agent Based Module
(ABM) was designed to be implemented from within the
DES environment to allow for the well defined features of
AutoMod to interact seamlessly with the agents. It is cer-
tainly possible, however, to implement ABM as a separate
module that interacts with DES externally. This option
was not explored because the external ABS packages
available did not fit the needs of the authors for this pro-
ject.
A schematic diagram of the overall system design can
be found in Figure 1. The ABM module contains four
types of objects:
• Dynamic Seeking Agents - Models an individual
attempting to locate an object in the environment
• Dynamic Informative Agents - Models individu-
als who are familiar with the simulated world
• Stationary Informative Objects - Models maps
or signs in the environment
• Stationary Non-Informative Objects - Models
obstructions to agents path (e.g. Trees)
These objects can locate and identify each other by
storing their visible characteristics in a dynamic model of
the physical environment (physical information arrays) that
is accessible by all objects in the ABM. Any characteris-
tics that are not visible, such as knowledge of the surround-
ing area, are stored in the object’s internal information
space. The internal information space of the Dynamic In-
formative Agents and Stationary Informative Objects can
be accessed by the Dynamic Seeking Agent through an in-
teraction (as described in detail in section 5). Seeking
Agents can move in and out of the ABM to utilize DES
functions that are integrated into the model. It is important
to note that both types of agents in the module have the
same overall design, except that the Informative Agent
never interacts with DES functions.
The decision to include two forms of information
sources (Informative Agent and Informative Object) was
made in order to test two major functions of ABS.
1030

Dubiel and Tsimhoni
Figure 1: Integration of an Agent Based Module with the Discrete Event Environment
As Huberman and Glance (1993) describe in their paper on
evolutionary systems, if the initial action of an agent is the
same throughout all experiments, it becomes likely one
agent will have an advantage over another based on the
starting condition.
By including two different types of information
sources and randomizing the agent’s initial angle (field of
vision), the validation and verification of the decision logic
becomes easier and a fair starting condition is obtained.
Using these two information sources also gave the
ability to experiment with the different forms of interaction
between two agents that are possible in an ABS: those that
require effort by both entities (strong) and those that re-
quire only one object to complete the interaction (weak)
(Michel, Gouaïch and Ferber, 2003).
3 TEST CASE
The test case for this project is based on a single Seeking
Agent interacting with information sources in order to find
a goal object located in the simulation world. This case is
meant to model an individual attempting to obtain direc-
tions and then using those directions to navigate to a de-
sired location (goal object).
Figure 2 demonstrates the major features of the test
case. The Seeking Agent enters the model at a random
point with all the basic decision logic required to find an
information source and interact with it to obtain the loca-
tion of the goal object. Once the Seeking Agent has ob-
tained the location of the object from an information
source, it navigates through the simulation world to obtain
the object. There are two transportation methods shown
that the Seeking Agent can use to get to the goal object. If
the Seeking Agent uses a Stationary Informative Object
(map), they will walk to the location of the goal object,
which will take approximately 10 minutes to perform (de-
pending on the agent’s walking speed).
If the Seeking Agent obtains the directions from an In-
formative Agent (person), they will utilize a tram (a dis-
crete movement system) that is available. The tram has an
inter arrival time of 5 minutes and takes 1.5 minutes to
transport the agent. There is also the possibility that the
Seeking Agent will misinterpret the information given and
get lost. In this case, the Seeking Agent will travel to the
perceived location of the object, and if nothing is found,
will return to the original location of interaction. This test
case was developed in order to build and test the module
during the initial stages of development.
1031

Dubiel and Tsimhoni
Figure 2: Physical Layout of the Environment and Some Possible Test Case Scenarios
4 ENVIRONMENT MODELING
As seen in Figure 1, the environment is the DES model
space in which all actions and interactions take place. The
information available to each agents from the environment
can be broken down into two main segments: the visible
characteristics available to all agents in the model (physi-
cal), and the characteristics that are known only by the in-
dividual agent (internal) that can be obtained through inter-
action with other agents or objects. The environment
modeling was implemented using the grid method (Epstein
and Axtell 1996). This method uses (X,Y) coordinates of
each object as the basis for the physical interaction in the
model. Multiple global array were used to hold all physi-
cal and internal attributes of the agents and objects in the
simulation world. These arrays included (X,Y) coordi-
nates, current angle of rotation (head and body), object
type, object shape, current state, and knowledge of sur-
roundings. Although these arrays are global, great care
was taken to limit the information available to all agents in
the model in order to maintain the integrity of the physical
versus internal information.
Object type, shape, size and state are used to deter-
mine the type of action and interaction that will occur as
the simulation progresses. As the work for this module
continues, a larger variety of objects will be integrated in
order to further the functionality of the module.
The interaction of the agent based entities with the
DES features was implemented using a trigger location and
variable combination. If the Seeking Agent desired to ride
the tram presented in the test case, they were required to
travel to the trigger location and a Boolean value was set to
indicate to the ABM that the Seeking Agent should be
transferred to the discrete movement system. The Seeking
Agent was then moved out of the ABM and placed into the
discrete queue to wait for the arrival of the tram.
5 MODELING HUMAN TRAVEL
The degree of detail with which intelligent agents in the
model represent real humans is likely to vary with the spe-
cific requirements of each model. For the purposes of this
test case, it is desired that information seeking agents (visi-
tors at the park) will have several basic abilities. They
should be able to perceive their environment (mainly visu-
ally), have navigation goals (e.g., I want to get to Space
Mountain), know when they need to ask for directions
(meta knowledge) or use maps for directions, avoid obsta-
cles (guidance) as they move around, interact with other
people, and utilize the knowledge obtained from the inter-
actions to navigate to a goal object.
To achieve these goals, all agents were programmed to
have visual and auditory perception, cognitive abilities
such as searching and decision making, and the ability to
control their movement at a gross level.
1032

Dubiel and Tsimhoni
5.1 Perception
The vision of the agents is modeled using physical infor-
mation arrays along with trigonometric functions to deter-
mine the agent’s field of vision. Once visible objects are
determined, the peripheral and foveal vision memory of the
Seeking Agent is populated with physical information of
each visible object. The peripheral vision extends 180 de-
grees horizontally around the agent’s direct line of sight,
while the foveal vision spans two degrees. If an object is
in the agent’s peripheral vision, the Seeking Agent can per-
ceive the general characteristics of the object (detect its ex-
istence, size, and movement). In order to determine if the
object is a target or distracter, the Seeking Agent must look
at the target (using foveal vision).
The vision memory processing is performed at regular
intervals (once every 50 milliseconds) as long as the Seek-
ing Agent is not moving their eyes. The visual processing
continues while the Seeking Agent is in transit. This paral-
lel processing allows agents to make decisions such as
avoiding obstacles or changing route while moving .
As a further addition to the vision function of agents,
all objects in the model have a defined salience (promi-
nence) level which represents the features of an object that
make it stand out against its background. Some of these
features include contrast, luminance, color, size and shape
(Wickens, Gordon and Liu, 1998). In the current imple-
mentation, all of these factors are simplified and summa-
rized as a single salience level defined from 1 to 10. This
attribute is a main factor in the searching method of the
Seeking Agent as described later in this section. Once an
agent’s vision memory has been updated with the charac-
teristics of the visible objects, it runs the decision logic.
A separate factor to further model the perception ca-
pabilities is the agent’s range of vision. This attribute de-
scribes the maximum distance where the Seeking Agent
can differentiate between objects. In all experiments, this
factor is set to be 400 feet, but can be changed to simulate
visibility conditions and differences in individual charac-
teristics.
5.2 Decision
The decision logic is implemented based on the current de-
cision state of the agent. A simplified list of the decision
states and the decision progression is shown in Table 1.
A Seeking Agent proceeds through these states, in the
described order, until it has reached the goal object. The
starting state of the Seeking Agent is determined by the
confidence the agent has in their personal knowledge of the
location of the goal object. Each Seeking Agent enters the
model with this confidence factor assigned to them ran-
domly. This factor along with the perceived location of the
goal object determine the agent’s flow through the decision
logic. If the agent’s confidence is greater than a high
threshold (set at 85%), they do not ask for directions and
proceed directly to step 11 using their perceived location of
the goal object. A Seeking Agent has a 50% probability of
knowing the correct location of the goal object upon enter-
ing the model. The agent’s knowledge of the location was
not made to correlate with their confidence in order to de-
termine the modules function when an Seeking Agent
becomes lost.
Table 1: Description of the Agent Decision States
Step Description Next Step
1 Searching for IS 2
2 Decide on Action 1, 5 or 6
3 If nothing found => rotate body 1 or 4
4 If nothing found => travel 1
5 Found Informative Object => travel 6
6 Found Informative Agent => travel 9
7 At Informative Object => interact 11
8 At Informative Agent => interact 11
9 Emit verbal signal 8
10 Respond to verbal signal 8
11 Go to goal object End/Next
Object
If the Seeking Agent does not have the required
amount of confidence to proceed to the goal object, it
flows through the decision logic as shown above. Once the
Seeking Agent has determined the location of the goal ob-
ject with an adequate confidence by interacting with in-
formation sources, it walks or travels to the location.
A sociability parameter accounts for the agent’s ten-
dency to approach a person rather than look for a map.
This parameter is given a value of zero if that Seeking
Agent will not approach an Informative Agent until it is
their last option and a value of one if the Seeking Agent
prefers human interaction.
5.2.1 Search
The search functions are performed to locate information
sources or targets that are close to the object. The Seeking
Agent must do this by distinguishing between targets (an
information source or the goal object) and distracters. A
distracter is defined as any object in the environment that is
of no interest to the agent.
During the search, the Seeking Agent performs two
visual processes in parallel, using the foveal and peripheral
fields of vision. If a Seeking Agent has a target object in
its foveal vision, any information that is received by the
peripheral scan is ignored and the Seeking Agent will
travel to the target object in sight. If no object is located in
the foveal field of vision, the Seeking Agent continues a
visual search of its surroundings.
There are two main forms of human visual search;
parallel (or feature) and serial (Wickens, Gordon and Liu,
1033

Dubiel and Tsimhoni
1998; Wolfe 1998). Parallel searches occur when the tar-
get object is distinguishable from distracter objects by a
single characteristic such as color, size or motion. A serial
search is characterized by the systematic search of each
item in the field of vision of the individual. The parallel
search is much more effective than a serial search. To
simulate the parallel search, two factors were taken into
account; motion of the object, and salience level. Motion
was assumed to be the most prominent factor in the search
algorithm, and therefore any moving object in the agents
peripheral vision would immediately draw the attention of
the agent. When there was no moving object in sight, ob-
jects with a salience level that stands out among the other
object in the Seeking Agent field of vision (salience differ-
ence greater than the threshold) will draw the attention of
the agent. The salience threshold for each Seeking Agent
in all experiments was set to four (out of 10). If there are
no objects that stood out in the agents peripheral vision, a
serial search would begin.
When performing a serial search, the Seeking Agent
scans the environment by systematically foveating each
object in its field of view. Once an object is in the foveal
field of vision, its characteristics are processed. If it is not
the target, the eyes are moved to a different object. If the
target is not found with eye movements, the Seeking Agent
turns its head to a new direction and continues a systematic
search until a target is found. If a target is still not found,
the Seeking Agent starts walking in a randomly selected
direction until it encounters an information source or the
goal object. This scenario, while not explored during ex-
perimentation, could offer some valuable insights as to
where to place the information sources to achieve the
greatest coverage with the fewest amount of resources.
5.3 Movement
The movement section of the module was separated into
two phases: obtaining and chasing. The obtaining phase
was implemented when the destination was stationary and
the chasing phase was implemented when the destination
of the Seeking Agent was moving (a moving Informative
Agent).
When obtaining a target that is in view, the Seeking
Agent follows the shortest path available to the object.
This navigation method is used when traveling to all sta-
tionary objects including stationary Informative Agent,
Stationary Informative Object and the goal object. The
visual perception while moving allows the Seeking Agent
to avoid any obstruction in its path. Navigating around ob-
structions was implemented so as to minimize the agent’s
distance traveled, as a simulation of real human behavior.
To do this, the Seeking Agent begins to alter its path as
soon as it recognizes the obstruction. This implementation
method ensures the Seeking Agent will travel the shortest
possible distance without causing a collision.
When chasing an object, the Seeking Agent perceives
the direction of travel of the object (by accessing the
physical characteristic array of the object) with a given
level of error. This information allows the Seeking Agent
to travel on a path of intersection intelligently rather than
chasing the objects past position.
The chasing process is terminated when the Agents are
in comfortable hearing distance. The Seeking Agent emits
a verbal signal (“excuse me”) that is perceived by the in-
formative agent when the distance between them is within
a given radius (set at 10 feet). The receiving agent then re-
sponds by focusing on the emitting agent and engaging in a
verbal interaction.
6 INTERACTIONS
The interaction portion of the module was implemented to
give the agents a method to obtain detailed and non-visual
information from their environment. A discussion of weak
vs. strong interactions (Michel, Gouaïch and Ferber, 2003)
follows.
6.1 Weak Interactions
A weak interaction typically occurs between an agent and
some other (inanimate) object in the environment. This
type of interaction requires the attention and consent of
only one of the objects involved (the agent) and therefore
the inanimate object can interact with multiple agents at
one time. Some examples of this type of interaction are a
fruit tree as a food source, a map as an information source
or a garbage can as a disposal unit.
This interaction type was simply implemented by re-
quiring that the Seeking Agent be at a minimal distance
(two feet) from the DIO (map) to obtain the needed infor-
mation. Once this requirement has been met, the Seeking
Agent interprets the needed information (currently imple-
mented as a time delay and a download of information) and
proceeds on to the location. In the test case, the agent had
a 70% chance of interpreting the information correctly.
6.2 Strong Interactions
A strong interaction is defined to occur between two
agents. This type of interaction requires the attention and
consent of both objects involved and therefore an agent
will first need to gain the attention of the other party before
beginning the interaction. Some examples of this type of
interaction are any type of society growth or reproduction
model, an agent asking for directions, or a call center agent
answering a phone.
This interaction type is implemented in the module by
utilizing the perception attributes to trigger a change in the
decision state of each agent. Once both agents recognize
the need for an interaction (verbal signal), their states
1034

Dubiel and Tsimhoni
change as an implementation of the strong interaction logic
included in the module. After the interaction is initiated,
the Informative Agent tells the Seeking Agent the location
of the goal object with a 90% accuracy level. At this time
the Seeking Agent terminates the interaction and proceeds
towards the Goal object.
7 VERIFICATION AND VALIDATION
As a first step in understanding the benefit of integrating
ABS into a DES model, verification and validation was
performed. The focus of this step was on the following
questions: “Is the ABM (Computerized Model) accurate to
the Conceptual Model presented in the test case?” and
“Does the function of the Computerized Model match the
expected result of the actual system?”.
7.1 Computerized Model Verification
In approaching the first question, both static and dynamic
verification techniques (Sargent 2003) were used. Struc-
tured walk-throughs were used to ensure that the module
was operating correctly in the initial stages of develop-
ment. This was done by including only a single option that
the Seeking Agent was required to take and following the
Seeking Agent through its decision pattern. This allowed
the authors to verify the Seeking Agent was implementing
the decision logic correctly during experimentation, be-
cause the correct results were known previous to the ex-
periment being conducted. A repeatability test was also
conducted during this verification to ensure there were no
unexpected factors affecting the model. Once the logic
was operating to the satisfaction of the authors, dynamic
verification was conducted.
The dynamic verification was completed by including
the random decisions in the model and tracing the Seeking
Agent through its decision logic. The operational graphic
technique (Sargent 2003) was used to visually illustrate the
decision logic used by the Seeking Agent to obtain the goal
object. To implement this technique, the decision states of
the Seeking Agent (1-11) were recorded during experimen-
tation and were then plotted by time to demonstrate the
flow of the agent’s decision logic (Figure 3).
In this example, the Seeking Agent first switched be-
tween State 1 (searching for an information source) and
State 2 (deciding on action) until it found an object it rec-
ognized as an information source. The small amounts of
time the Seeking Agent was in State 2 represent the deci-
sion time the agent required to recognize and record the
objects in memory. Once the Seeking Agent recognized
that the second object examined was an Informative Agent,
the decision state was changed to State 6 (Walk to Moving
Object) and the Seeking Agent proceeded to the Informa-
tive Agent. When the Seeking Agent arrived at the infor-
mation source, it took 26 seconds (set as a parameter)
Figure 3: Decision State Logic Flow
1035

Dubiel and Tsimhoni
to interact with the Informative Agent. The Seeking Agent
then proceeded to the goal object (State 11). This flow of
this decision logic was deemed to be satisfactory and many
more analyses like this were performed to validate the de-
cision logic flow. This process was done several times
with the initial position of the Seeking Agent and place-
ment of the other objects in the simulation world begin
changed with each run of the simulation. The results of
these experiments were as expected, and therefore the
model was verified to operate according to the specifica-
tions of the conceptual model.
7.2 Subjective System Validation
A subjective validation process was conducted to compare
the operation of the module with the expected results from
an actual system. The Animation Observation validation
technique was used to complete this task. The Animation
Observation technique was used to visually confirm that all
of the agents’ movements and decisions appeared to be re-
alistic. This process was performed with 50 different start-
ing positions and the Seeking Agent performed as expected
in all trials. The scenarios observed include:
• Interaction with Informative Object
• Interaction with Informative Agent
• Obtaining Goal Object via tram
• Obtaining Goal Object via walking
• Misinterpretation of information received
• Locate Goal Object without information source
• Emission of verbal signal
• Reaction to verbal signal
Based on this technique the high level agent decision and
movement patterns are accurate to the point of visually
mimicking the real life system.
8 CONCLUSIONS
The overall purpose of this project was to show that ABS
can be integrated with a common DES package to model
humans traveling freely through a DES environment. The
module was successful in mimicking the decision pattern
and movement of individuals navigating the simulation
world, without a defined path or decision points. With the
defined concept and implementation of this module, the
integration can now be used to run experiments that can
demonstrate its full benefit.
An example test case demonstrated a potential useful
application for this integration. By combining DES and
ABS, we were able to simulate aspects of the system that
cannot be simulated by either of the simulation methods
alone. As work on this project proceeds, real world scenar-
ios will be used so that quantitative validation can be per-
formed to show the actual benefit of this approach.
9 FUTURE WORK
Future work will consist of (a) further validation of model
elements, (b) addition of features to the ABM to add accu-
racy and scope, and (c) applying this model to a real world
problem with empirical data.
In building the ABM module, it was found that em-
pirical data on human behavior and decision making are
not readily available in a form that is easy to simulate. For
example, walking speeds, times for making decisions, and
the probabilities of strategies are rarely reported in the lit-
erature and need to be further explored.
Our ABM module can be enhanced with several fea-
tures based on existing literature and some experimenta-
tion. For example, individual differences in preference for
the display of information, additional object types in the
simulated environment and outside of it (as accessible by
mobile phones), additional modes of transportation, and
additional interactions such as transfer of money in stores
and in ATMs.
Once this module is further validated and somewhat
expanded, it will be used to model a real-world problem
and make decisions concerning factors that affect the
agent’s behavior
REFERENCES
Banks, J., Carson, J. S. II, Nelson, B. L., Nicol, D. M.
2001. Discrete Event Simulation, Third Edition.
Prentice Hall, Inc. Upper Saddle River, New Jersey.
Banks, J., Hugan, J., Lendermann, P., McLean, C., Page,
E. H., Pegden, C. D., Ulgen, O., and Wilson, J. R.
2003. The Future of the Simulation Industry. In the
Proceedings of the 2003 Winter Simulation Confer-
ence, eds. S. Chick, P. J. Sánchez, D. Ferrin, and D. J.
Morric. 2033-2043. IEEE, Piscataway, New Jersey.
Available via <
http://www.informs-sim.org
/wsc03papers/263.pdf> [Accessed March 1,
2005]
Epstien, J. M. and Axtell, R. L. 1996. Growing Artificial
Societies: Social Science from the Bottom Up. Cam-
bridge, MA.: MIT Press
FIRE SAFETY ENGINEERING GROUP website. 2003.
Software Information [online]. Available online via
<
http://fseg.gre.ac.uk/EXODUS/air.ht
ml>[accessed March 1, 2005].
Huberman, B. A. and Glance, N. S. 1993. Evolutionary
Games and Computer Simulations. In the Proceed-
ings of the National Academy of Science USA 90.
7716-18
Introduction to Swarm website. 2005. Main Page. Avail-
able online via
<http://www.swarm.org/wiki
/Introduction_to_Swarm>. [accessed March
1, 2005]
1036

Dubiel and Tsimhoni
Law, A. M., and W. D. Kelton. 2000. Simulation Modeling
and Analysis. 3rd ed. New York: McGraw-Hill.
Michel, F., Gouaïch, A. and Ferber J. 2003. Weak Interac-
tion and Strong Interaction in Agent Based Simula-
tion. In the Proceedings of the Multi-Agent Based
Simulation III, eds. D. Hales, B. Edmonds, E. Norling,
and J. Rouchier. 43-56. Volume 2927., Springer-
Verlag Berlin Heidelberg 2003
Repast Agent Simulation Toolkit website. 2005. Package
Overview. Available online via
<
http://repast.sourceforge.net>. [ac-
cessed March 1, 2005]
Sargent, R. G. 2003. Verification and validation of simula-
tion models. In Proceedings of the 2003 Winter Simu-
lation Conference, eds. S. Chick, P. J. Sánchez, D.
Ferrin, and D. J. Morrice, 37-48. Piscataway, New Jer-
sey: Institute of Electrical and Electronics Engineers.
Available via
<www.informs-sim.org/
wsc03papers/006.pdf> [accessed September 1,
2004].
Wickens, C. D., Lee, J., Liu, Y., and Gordon-Becker, S. E.
2004. Introduction to Human Factors Engineering.
Prentice Hall, Inc. Upper Saddle River, New Jersey.
X J Technologies. 2005. Company Information [online].
Available online via <
http://www.xjtek.com>
[accessed March 1, 2005].
AUTHOR BIOGRAPHIES
BENJAMIN DUBIEL received his bachelors of science
in Industrial Engineering from the University of Michigan
in 2005. The work presented here was performed as one of
many undergraduate projects. He has nine years of ex-
perience with multiple simulation packages including
Automod, Arena, and Promodel. His primary interest is
using simulation to solve operational problems in transpor-
tation, manufacturing, and theme park related fields. His
e-mail address is: <
ish6868@hotmail.com>.
OMER TSIMHONI is an Adjunct Assistant Professor at
the Department of Industrial Engineering of the University
of Michigan and an Assistant Research Scientist at the
University of Michigan Transportation Research Institute
(UMTRI). He received his Ph.D. from the University of
Michigan in 2004. His primary research interest is in the
application of simulation to cognitive modeling and human
factors. His e-mail address is
<omert@umich.edu>.
1037