Você está na página 1de 10

Crowd Behavior Cognitive Model Architecture Design

Quynh-Anh (Mimi) H. Nguyen


Frederic(Rick) D. McKenzie, PhD
Mikel D. Petty, PhD
Virginia Modeling, Analysis and Simulation Center
Old Dominion University
Norfolk, VA. 23529
757.686.6200, 757.683.6367
qhnguyen@odu.edu, fmckenzi@ece.odu.edu, mpetty@odu.edu

Keywords:
Crowd behavior, cognitive model architecture, game technology

ABSTRACT: Our United States military is increasingly engaged in urban combat or peace-keeping
missions. As a result, soldiers are also increasingly engaged with the civilian non-combatant inhabitants of
various nations. Yet, current military simulation models have little or no representation of these effects
which can lead to suboptimal training or experimentation results. More realistic and sophisticated crowd
models are desired to address this growing need.

It is unlikely that one crowd model will meet all our militarys crowd requirements since models are needed
with a variety of behaviors depending upon the type of mission, the size of encounters, and the user
application. A federate that may be used to provide such variety of civilian behaviors in a crowd context
would need to be flexible, configurable, and extensible. In this paper, we report about one such
instantiation and the framework and cognitive model that support it. The framework is a layered
architecture that is composed of a physical layer in which movements and other actions of the crowd are
manifested; and also a cognitive layer in which the motivations of these activities are generated and
propagated. Connecting these two layers is an API layer that provides mapping and communication
services for the stimuli, activities, and accompanying parameters that are crowd behavior centric. We focus
especially on the details about the cognitive model architecture (CMA) and the design process used to
achieve it.

1 INTRODUCTION potential for shooting innocent civilians, or being


shot at by hostiles in the crowd; potential for an
Crowds of non-combatants play a large and increasingly incident at the tactical level to influence
recognized role in modern military operations, and often operations and policy at the strategic level. [2]
create substantial difficulties for the combatant forces In spite of the military challenges and risks imposed by
involved. crowds, models of crowds are essentially absent from
In Somalia, U. S. Marines often faced hostile current production military simulations. This omission
crowds of rock-throwing women and children. has been understandable in the context of legacy
In Bosnia, U. S. Army soldiers had to disperse simulations that were historically focused on large-scale
angry mobs of Serb hard-liners near the town of engagements between heavy mechanized forces in
Banja Luka. More recently, Danish, French, and primarily non-urban settings. However, in the last decade
Italian forces attempted to control riots between the threat has changed and future engagements are
ethnic Albanians and Serbs in Mitrovice, expected to often involve lighter forces in urban settings.
Albania. [1] In simulations of such scenarios the absence of crowds
All military operations, large or small, have a and of non-combatants in general is a more serious
crowd control/crowd confusion factor. departure from realism. The absence of models of crowds
[C]rowds are one of the worst situations you can in military simulation, and the need to include them, has
encounter. There is mass confusion; loss of been widely recognized.
control and communication with subordinates;
Military forces are increasingly called upon to behavior fidelity needed within the context of military
support operations other than war in which they simulations and to allow iterative refinement of the
come into contact with civilian populations. In requirements for the crowd behavior CMA. These initial
some cases, the interaction takes place with CMA prototypes were developed and incorporated into
crowds of civilians. Unfortunately, the computer the crowd federate and may be extended as additional
generated forces that support virtual training capabilities are needed.
systems do not yet support the simulation of
crowds of civilians. [3]
3 THE CROWD MODEL
Representations are needed for (neutrals or
civilians) to represent operations other than war As documented in [9], our crowd model has two main
and the interactions among these forces. [4]. layers, the cognitive model (CM) layer and the physical
VMASC is engaged in a project aimed at developing a model (PM) layer, which communicates with each other
crowd modeling capability for military simulation. The through a crowd behavior API. Each layer is responsible
project consisted of a number of phases: a requirements for modeling different aspects of the crowd model. The
analysis to identify military simulation crowd modeling CM layer is responsible for modeling the mind for the
requirements; a literature survey to examine crowd. In the simplest sense, the cognitive model will
psychological research relevant to crowd modeling; a receive stimuli from the physical model, process the
design study to explore design issues in the stimuli, and select a behavior to send back to the physical
implementation of a crowd simulation; and the model to carry out. Therefore, the PM acts as the body
development of a crowd simulation, implemented as a of the crowd and is responsible for accepting behaviors as
distributed simulation federate, that will be interoperable input, carrying out those behaviors, and then outputting
with existing military simulations and will have a credible stimuli to the cognitive model. If we view the CM, PM,
psychological basis for the crowd behavior it generates. and crowd behavior API as black boxes, we have a very
These phases have been described in past publications in simple interface between the components as depicted by
varying levels of detail [5] [6]. Figure 1.
Although the goals of the current VMASC project is not
to create a predictive model of crowd behavior, we are
focused on developing a suitable cognitive model
architecture that is capable of incorporating the latest
psychologically research to produce credible crowd
responses. This paper focuses on this task of developing
the psychologically-based cognitive model architecture.
The psychological research used as the basis for our
computational model of crowd behavior may be found in
previous papers [7][8].

Figure 1: The Simplified Crowd Model


2 DESIGN PROCESS AND GOALS
As simplistic as this information interchange may seem at
The design process started with an investigative software first glance, it soon became evident that it was not so easy
development process to initially create an exploratory to create. We had to carefully consider and define what is
prototype of the Crowd Behavior cognitive model meant by a stimulus and what is meant by a behavior. As
architecture (CMA). The goals of developing this one starts to delve into defining these terms, it becomes
prototype were twofold. Firstly, the process of designing clear that trying to define these terms in an unambiguous
and implementing the exploratory prototype would allow way is difficult. This becomes even more difficult as we
us to elaborate the requirements of a crowd behavior try to abstract away from the individual level and define
cognitive model. Secondly, we would gain an these in terms of groups and crowds. In this section, we
understanding of how to bridge the gap between the attempt to define these terms in a way that fits in well
psychological research and a computational model of with our crowd federate architecture, utilizing the crowd
crowd behavior. behavior API to handle any disconnects.
After developing the exploratory prototype, we developed
an experimentation prototype that could be utilized for
running experiments to examine the level of crowd
3.1 Stimuli Some examples of sensory, state change and event stimuli
that might be of interest in a crowd federate have been
We define stimuli as anything that may elicit a response identified in Table 1. It is important to note that an
from the crowd behavior cognitive model. Using this assumption is made that sensors will filter and provide
definition, we have identified three types of stimuli that a particular stimulus. For example, real people may have
could be exchanged between the CM and the PM. These a vision sensor, such as their eyes, which can see
are sensory, state change, and event stimuli. A sensory everything around them and provide preprocessed data
stimulus could be thought of as a product of the to their brain for further processing. Likewise, in the
continuous excitation of a characters sensors, such as physical model, the assumption is that there will be a
vision, audio, etc., that produces a definitive type of particular sensor to look for particular objects, such as
information being obtained about the world, such as vision sensors to look for the nearest person who is a
distance to nearest character of interest, nearest character member of the soldier group and other such items of
of interest, etc. A state change stimulus could be interest.
considered a world knowledge type of stimuli and could
Stimuli known to one particular CM is associated with
be any modifications in the crowd situation, such as a
stimuli generated from within a given PM by the mapping
change in the crowd, group, and/or individual parameters,
data accessed by the Crowd Behavior API. An important
which in turn could elicit a response from the crowd. For
artifact to note is that although we may want to assume
example, if the leader is shot or runs away, this could
that there is a straight one-to-one mapping between the set
cause the leadership structure crowd/group parameter to
of stimuli that the CM can process and the set of stimuli
change from highly organized to disorganized. Event
that the PM can produce, this may not necessarily be the
stimuli are used to capture non-reoccurring, infrequent
case. There could also be a disconnection between the set
discrete events that might transpire, such as hearing an
of behaviors selectable and realizable between the CM
explosion or gunfire, or seeing someone near them shot,
and the PM. Therefore, this disconnection may cause
or feeling themselves be hit by others in the crowd.
some stimuli not to be processed and/or some behavior
STIMULI not to be realized. Although the mapping data can be
Sensory Event State change configured to account for appropriate substitutes, this
See nearest See burning tire Leader is shot
should always be taken into account when plugging-n-
member of Hear gunfire Leader runs away
the control Hear explosion Lose ability to playing different cognitive models with different physical
force See/smell gaseous communicate models.
See nearest substance amongst crowd
friend See crashing helicopter members
See nearest
3.2 Behaviors
See/hear firefight Appearance/actions
member of a See someone near you of organized agitator
group self is get shot group We define a behavior as the desired response to
sympathetic See someone near you Crowd members stimulation from the environment of the individual,
to hit fleeing
Sense Members joining the
group, or crowd. So, what constitutes a behavior as it
Get hit by others in
distance to crowd crowd exists between the physical model and the cognitive
above three Get hit by a member of model? At what level should a behavior be defined?
Hear the control force
chanting See downdraft of
Hear militia helicopter
broadcasting See militia
on demonstrating weapons
megaphones See militia loading
weapons

Table 1: Supported Stimuli


The cognitive model would be responsible for accepting
stimuli as input and selecting and outputting an
appropriate behavior response to the physical model. In
turn, the physical model would be responsible for
modeling the necessary sensors to detect the stimuli and
sending them to the cognitive model. It is also
responsible for providing a repertoire of behaviors that
can be realized by that physical model. Figure 2: Behavior Hierarchy
As can be seen in Figure 2, we find that behaviors could This section describes the initial design and
be considered hierarchical in that it can be realized in the implementation of our crowd behavior cognitive model
physical model by combining some set of simple actions architecture as it was applied for the first set of fidelity
and simple animations as well as lower-level behaviors, level requirement experiments [11].
which in themselves can also made up of simple actions
and animations. They can be composed of any 4.1 Bridging the Gap
combination of actions and animations to achieve a
desired effect. For example, if we wanted to perform the There are many current research efforts being conducted
simple behavior of standing still, this would entail both these days that requires the use of a sound,
an action and an animation to achieve. In particular, it psychologically based crowd model. The focus of our
would require initiating a maintain speed = 0 action and studies has been non-combatant crowds that interact with
activating a stand still animation. US military operations. The intended uses of these
models vary from needing to understand how crowds
3.2.1 Repertoire of Behaviors respond to non-lethal weapons, how to influence and
direct the behavior of potentially hostile crowds, how to
The repertoire of behaviors that we initially identified for
bring a mob situation under control, etc. The problem
the crowd model was based on examining the necessary
that we have encountered has been that psychologists and
behaviors to support the reference scenarios [10] selected
sociologists are still trying to refine their understanding
to test our crowd federate. In particular, our prototype
about individual human behavior, making it that much
was developed to support the Mogadishu scenario. The
more difficult in understanding crowds, crowd behavior,
table below shows a mapping of an aspect of the crowds
and how individuals behave in crowd situations [12].
mental state, captured by an aggression level variable, to
possible crowd behaviors. The bolded behaviors are Much research has been conducted on human behavior
those that have been identified as high-priority, or models that focus on individual autonomous human
necessary, for the Mogadishu scenarios. The behaviors agents. Fewer models have addressed group or crowd
marked w/ the * are those currently supported in the behavior. Crowd behavior models in the literature range
prototype. from purely physical representations to purely cognitive
representations and combinations of both have been
AGGRESSION
LEVEL
BEHAVIOR successful to varying degrees in the domain to which they
Neutral Wandering* were intended. Motivations range from improving
Avoidance Fleeing* graphical rendering performance to providing realistic
Hiding behavior representations with emotions and other drivers.
Curious non- Hanging out/watching*
aggressive posture Standing/climbing on elevated structures Crowd models include particle systems [13], flocking
Seeking to burning tire/smoke* systems [14], or behavioral systems [15] -- the difference
Chanting/flag waving being an increasing level of interaction among
Aggressive posture Swarming toward helicopter*
Burning tires
participants in the crowd and with the environment in
Building barricade general. Along with increased interaction comes
Taunting/yelling increased attention to modeling the social and emotional
Carrying sticks/raising firearms
interactions among crowd members.
Aggressive non-lethal Throwing rocks/projectiles*
actions Pushing/shoving A recurring theme throughout the literature is the
Hand-to-hand fighting
Dragging bodies through streets necessity to model behavior either at the individual level
Violent lethal actions Shooting* and allow crowd behavior to emerge as a result of many
Firing RPGs goal-oriented individuals within the same location or at a
Throwing molotov cocktails
Driving technicos (civilian vehicles w/
crowd level where the behavior of the crowd may be
mounted machine-guns) completely controlled by a crowd entity. Again, there is a
Table 2: Aggression Level to Behavior Mapping range of efforts between these two extremes. The
psychology of crowds is not fully understood and is
4 THE COGNITIVE MODEL DESIGN constantly evolving. There is no psychological or
sociological model of crowds in existence today that has
As mentioned earlier, the cognitive model has two main been universally accepted and validated.
responsibilitiesto receive stimuli from the physical As the development of the cognitive model progressed, it
model and to select and send behavior commands back to became clear that there was still much to be learned when
the physical model as a response to the received stimuli. it came to understanding crowds and crowd behavior. We
realized that there would be significant effort required in 4.3 The Cognitive Model Architecture
bridging the gap between the psychology of crowd
behavior and the simulation of crowd behavior.

4.2 Design Decisions

In the pursuit of developing both the psychological and


computational models for crowd behavior and in the
analysis and considerations of established research, we
have made several design decisions of import. These
decisions are detailed below:
We are incorporating Musse & Thalmanns [16] [17]
concept of a crowd hierarchy. There will be crowds,
groups, and individuals. Crowds are made up of
groups and groups are made up of individuals. A
group can be made up of only one individual, but a
crowd has to be made up of two or more groups.
Our architecture is focused at the group and crowd
levels. We are using the crowd parameters identified
during the Crowd Modeling Workshops being
conducted by Brooks Air Force Base in support of
the Joint Non-Lethal Weapons Directorates
(JNLWD) goals of understanding crowds and crowd
control. Since the crowd model is focused at the
group and crowd levels, we do not need to apply any
of the individual cognitive models and approaches Figure 3: Crowd Cognitive Model Architecture Prototype
available such as SOAR, ACT-R, BDI, etc.
After multiple meetings between the cognitive model
Our premise is that you can get realistic crowd
implementation and psychological teams, an initial crowd
behavior without needing to model the cognitive
behavior model was produced. Figure 3 depicts the
processes of all of the individuals. In fact, our
prototype crowd cognitive model architecture that was a
cognitive model architecture will be able to simulate
more crowd members because it does not have to result of these design meetings. In the most basic sense,
calculate so many individuals cognitive processes. the CM will accept stimuli as input and produce a
However, the individual parameters will still be behavior response as output.
important, especially as it deals with how they may As can be seen, the crowd CM will track multiple
impact group decisions and how group decisions may parameters and mental states. The parameters are used to
impact an individual. describe the state of the crowd situation, such as its size,
The mental states of the crowd & of the group have leadership structure, etc. The mental states are variables
been simplified to a single state, called the crowd that are used to describe internal states of the crowds
aggression level (CAL) & group aggression level mind. Please note that we are not suggesting that there
(GAL), respectively. The individual mental state will
exists a crowd mind that controls how the individuals
be called the PAL, or personal aggression level. This
behave; these mental states are more to store internal
is an assumption taken to simplify the model and
crowd information that might somehow influence the
because at the basic level, we are interested in what
makes crowds turn violent. However, we should behavior of the individuals and groups within that crowd.
realize that there should possibly be more mental
states modeled and that the mental state could be
capturing different emotions, such as jealousy level,
fear level, etc.
Cognitive Model (Initialization) Cognitive Model (Execution)

Calculate new CAL


Set initial CP

For n=1 to N-1,


Calculate initial Crowd where N=# of groups, Increment n
Aggression Level (CAL) select GroupnGroupx

Is Groupn
Distribute GAL part of Groupxs
to Groups crowd?

Yes

Weight CAL to Group No


impact w/ Groupn S

Federate Execution

Figure 4 Crowd/Group Parameters Physical model simulation Update Groupns GAL

As mentioned in section 4.2, the crowd/group parameters


were adopted from those identified by the Crowd Did Groupns GAL
No
Stimuli
Modeling Workshop attendees during the first and second change?

workshops as being important to consider when modeling Cognitive Model (Execution)


Yes

crowds. These parameters can be separated into three Update GAL of Select new behavior

categories: crowd parameters, control force parameters, Groupx receiving stimuli for Groupn

and situational parameters, and are shown in Figure 4.


Weight Group to CAL
The individual parameters have yet to be identified, but impact w/ Groupx LC
Send behavior to PM Process next Groupn

the initial belief is that they will correspond closely to


those parameters identified by the Human Effects Figure 5: Computational Model Algorithm
Advisory Panel in conjunction with the JNLWDs study What follows is a more detailed description of the
on crowd behavior [18]. These individual parameters are computational model algorithm. The portions of the
emotion, focus, stress, confidence, and motivation. algorithm not implemented yet are italicized.
4.4 The Computational Model Algorithm Initialize the Crowd Parameters
Calculate an initial crowd aggression level (CAL)
Figure 5 depicts the computational model algorithm for From this CAL, use an appropriate algorithm to
the prototype. For this iteration of the prototype, we have distribute and initialize the groups and their GALs
implemented those portions of the algorithm that are in (see Section 4.4.1 for an example of such a
outlined in solid lines. Those steps outlined with dashed distribution).
lines have not yet been implemented in the prototype, but Once the user specifies (or the federate automatically
are planned improvements for the next iteration of the generates) the group types & their distributions, the
cognitive model prototype. We can see that the federate will begin execution.
computational model is relatively straightforward. There Stimuli will be sent to the individual members. Only
is an initialization phase, where the crowd parameters the stimuli for the leader of the group will be
(CP) are used to calculate an initial crowd aggression analyzed and the stimuli x crowd parameter matrix
level (CAL), which in turn is distributed down to the will be used to calculate a new GAL for the group
groups aggression levels (GALs). How this CAL is that leader is a part of. By only using the leaders
distributed to the groups is a design decision that can be stimuli, we achieve two objectives. We avoid having
the same stimuli possibly having multiple effects on
changed based on the requirements of the scenario. For
the groups GAL since they will all probably receive
the initial prototype, we decided to model three types of
the same stimuli and we reduce the amount of CM
groups, one that had a GAL equal to the CAL, one that
processing involved since stimuli of non-leaders are
was less aggressive and the last more aggressive. not processed.
After the cognitive model is initialized, the next phase is Calculate a new CAL. For the prototype, the CAL is
the federation execution phase. During this phase, the directly impacted by the incoming stimuli. Ideally,
crowd federate will be executing, with the physical model the impact on the CAL will be weighted by some
continuously feeding stimuli into the cognitive model. combination of the group parametersLeadership,
The cognitive model will process the received stimuli to Cognitive Conflict, and Social Influence (LCS).
select a new group action to be sent to the physical model.
For all groups that are members of the same it is used to dynamically create enumerated types for each
crowd as the group who received the stimuli, of the parameters. The format of the labels is a string
modify their GALs based on some weighting followed by a period followed by another string. The first
determined by their proximity to the stimulated group string is the name of the enumerated type while the
and to some subset of the LCS weighting. second is a value for that type. So with the following
Update the Groups GAL. For the first iteration labels: Terrain.STEEP, Terrain.FLAT,
prototype, the CAL is distributed between the three CrowdSize.SMALL, CrowdSize.MED, and
groups, one with the same value, one less aggressive, CrowdSize.BIG, there would be two enumerated types:
and the last more aggressive. Terrain and CrowdSize. The Terrain type will have an
If the groups GAL changed aggression level, select a attribute list storing STEEP & FLAT while the
new behavior for the group and send it to the physical CrowdSize type will have an attribute list of SMALL,
model to be realized. MED, & BIG.
4.5 The ImplementationTables This table is used in the CMA in a couple of ways. In
particular, it is used to calculate the initial crowd
One of the goals of the crowd behavior CMA is to aggression level (CAL). The initial CAL is the average of
develop a flexible architecture that can be easily adapted the aggression level changes added to 0.5. For example,
to express the most current psychological understanding lets say that the current values for the crowd parameters
of crowds and crowd behavior. In order to achieve this are Terrain = Flat, CrowdSize = Med, and Mood =
flexibility, the current cognitive model implementation Curious. If we go down the CParm X CParm table, we
utilizes three tables to configure itself. This section see from the highlighted values that the sum of the
provides details of the tables that were used and how they changes would be 1 (or 1+1-1). The average changes
were implemented. would be (1+1-1)/3 = 1/3. So, the initial aggression level
would be 0.5+0.33, or 0.83.
4.5.1 Crowd Parameters x Crowd Parameters From this initial CAL, we assume that we have three
Table
representative groups around that CAL of which the first
This table is a pair-wise comparison of the crowd group will be less aggressive, the second would have the
parameters vs. crowd parameters. This is a head-to-head same aggression level, and the third group will be slightly
comparison of each crowd parameter and its current more aggressive. The current delta that the groups
condition with all of the other crowd parameters and their aggression levels differ from the mean CAL is one
current conditions. If any of the two parameters have the aggression level interval, but this value can be modified
values as indicated by the row and column headings, then as more data becomes available. Therefore, with an
a point is assigned for that slot with one of the three initial CAL = 0.83 and interval of 0.16, we can compute
possible measures of 0, -1, or 1. A 0 is assigned if the the following GALs:
combination of those two parameters would cause the
GAL1 = CAL 0.16 = 0.67
crowds aggression level to remain the same; a -1 if it
GAL2 = CAL = 0.83
would cause the crowd to become less aggressive; and a 1
GAL3 = CAL + 0.16 = 0.99
if it would cause the crowd to become more aggressive.
Table 3 below is a sample of such a comparison. Mapping the CMA to this CM proved to be interesting
considering there are only two groups of Somalis
Terrain.
Steep
Terrain. Flat
Crowd Size. Crowd Size. Crowd Size.
Small Med Big
Mood.
Neutral
Mood.
Avoidance
Mood.
Curious
Mood.
Aggressive
Mood.
Aggressive
Mood.
Aggressive
modeled, the Somali citizen and the Somali militiaman,
yet three groups. Therefore, it was assumed that all
Posture Non-lethal Lethal Action
Terrain.
Steep

Terrain. Flat

Crowd Size.
Small
-1 1
Somali militiamen would be assigned to the most
Crowd Size.
Med
Crowd Size.
-1 1 aggressive group, GAL3, while the Somali citizens would
be equally split between the average and less aggressive
-1 1
Big
Mood.
0 0 0 0 0
Neutral
Mood.
Avoidance
Mood.
-1

-1
-1

1
-1

-1
-1

-1
-1

-1
groups, GAL1 & GAL2.
Curious
Mood.
Aggressive
Posture
1 1 1 1 1
Once the aggression levels have been initialized and the
crowd members separated into groups, the crowd
Mood.
Aggressive 1 1 1 1 1
Non-lethal

Mood.
Aggressive
Lethal Action
1 1 1 1 1 cognitive model can now begin its steady-state operation.
Table 3: Crowd Parameter x Crowd Parameter In this state, the CM will accept stimuli and produce a
behavior. The following two input tables, the Stimuli X
The top row specifies the parameters being used in our CAL table and the Mood X Behavior table support this
crowd model. This row of labels is very important in that operation.
4.5.2 Stimuli X Crowd Aggression Level Table since those are the only behaviors implemented in the
physical model in the prototype. However, the intention
This table is used to determine how incoming stimuli
is to choose certain behaviors based on a probability or on
might affect the crowds aggression level. In the initial
some more complex algorithm using the current crowd
CM prototype, we are only concerned with the various
parameters and/or stimuli received.
stimuli outlined in the table. However, it would be just as
simple to include the effects of other types of stimuli such Mood: Aggression Behavior Repertoire
as the state change and event stimuli. Table 4 is a Level
Avoidance Flee From
sampling from the table developed for the prototype. Neutral Wander Hang Out
From the figure, we can see that there are two types of Curious Stare At Seek To
crowd members modeled for the Mogadishu scenario Aggressive Posture Burn Tire Swarm To
[19], the Militia and the Citizen. For each stimuli and Aggressive Non lethal Throw Rock at
Aggressive Lethal Action Shoot At
current aggression level, we see that there are again three
Table 5: Mood X Behavior Table
possible measures, 0, -1, and 1. Similar to the Crowd
Parameter X Crowd Parameter table, these represent 4.6 Whats Next?
changes to the crowds aggression level. If the Militia
receives a stimuli called See Control Force and is We were successful in implementing a computational
currently in a Curious state, then we see that the 1 model for the cognitive model and incorporating it into
means that his aggression level will increase. As we can the crowd federate. This prototype is currently being used
see, the response to the same stimuli is dependent on both to run fidelity experiments [11] to determine at what level
the crowd member type and the current aggression level. the crowds should be modeled in order to have a relevant
impact on the training objectives of the expected users.
CROWD AGGRESSION LEVEL We hope to be able to take lessons learned from these
MEMBER STIMULUS Aggressive Aggressive Aggressive
TYPE Neutral Avoidance Curious Posture Non-lethal Lethal Action experiments to incorporate changes into the next iteration
Militia See control force -1 -1 1 1 1 1
See friend 1 -1 -1 -1 -1 0 of the prototype. Along with those changes, we have
identified other portions of the cognitive model
See neutral 0 0 0 0 0 0
See burning tire 1 -1 1 1 1 1
Citizen See control force
See friend
0
0
1
0
1
1
1
-1
1
-1
0
0 architecture that could be improved or expanded. These
See neutral
See burning tire
0
-1
0
1
0
1
0
1
0
1
0
0 enhancements are outlined below:
Table 4: Stimuli x Aggression Level Not only are there the parameters identified by the
JNLWDs Crowd Modeling Workshop being used as
So, when the CM receives stimuli, it will process all
the group parameters for computing the groups
stimuli received by a particular character and
mental state, there will also be other parameters that
decrease/increase the CAL by the appropriate amount
will be used to capture the within-group and
based on which stimuli was processed. From this new between-group interactions. The initial set of
CAL, new GALs are redistributed to each of the defined interaction group parameters that have been
groups. If the GAL "level" of the group that this identified is the Leadership, Cognitive Conflict, and
individual belongs to changes, generate a new action for Social Influence (LCS) parameters. The LCS group-
this character. Otherwise, return a null action. The action level parameters will be used to weight how the
that is selected is taken from the following table, the groups changing aggression levels may also affect
mood X behavior table. the aggression levels of other groups within the
crowd.
4.5.3 Mood X Behavior Table The crowd parameters from the Crowd Modeling
The final table used in the implemented cognitive model Workshops are applicable at the group level, too.
Therefore, these parameters will be used to determine
prototype is the Mood X Behavior table (Table 5). As the
the general aggression level impact on the groups.
name implies, this table maps a particular mood level to a
The LCS group-level parameters will be used to
list of possible behaviors to select based on that mood.
weight how the groups aggression levels will affect
This table was produced from the violent behavior the crowd.
continuum described in Table 2 in the Behaviors section
The algorithm for distributing the initialized CAL to
above. the different groups should be modified. Instead of
From the table we see that if the current character is at a having only three groups, we must identify more
curious aggression level, then it could choose to either group types, such as familial groups, friends,
participate in a stare at or seek to behavior. acquaintances, agitators, peaceful protesters, etc. The
Currently, only the first column behaviors are selected federate should provide a list of possible crowd
distributions of groups and group types that the user Representation, Orlando FL, May 7-9 2002, pp. 87-
can select from. This will somehow be based off of 95.
some of the crowd parameter values, such as crowd [4] Pew, R. W., et al, Modeling Human and
size, demographics, initial mood, propensity for Organizational Behavior: Application to Military
violence, etc. Each group type will have a particular Simulation, National Research Council, National
initial GAL associated with it. If you want to achieve Academy Press, Washington D.C., 1998.
a particular CAL, you can select a predefined set of [5] Petty, M. D., McKenzie, F. D. and Gaskins, R. C.,
group makeup with different GALs, distributed in a Requirements, Psychological Models, and Design
particular way that combines together to produce the Issues in Crowd Modeling for Military Simulation,
desired CAL. Proceedings of the Huntsville Simulation Conference
In [18], the experts in crowds, crowd control, and 2003, Huntsville AL, October 29-31 2003.
crowd behavior specify that the crowd should not be [6] Petty, M. D., McKenzie, F. D., Gaskins, R. C., and
treated as an entity, but rather as a 3-phase process Weisel, E. W., Developing a Crowd Federate for
namely the assembling phase, the rallying phase (we Military Simulation, Proceedings of the Spring
prefer the term rallying rather than gathering, as 2004 Simulation Interoperability Workshop,
specified), and the dispersal. We plan to incorporate Arlington VA, April 18-23 2004, pp. 483-493.
this concept of the crowd as a process by identifying [7] Boone, C. M., Gaskins, R. C. and Petty, M. D.,
and weighting the affects of the various parameters Observations of Crowd Behavior, Proceedings of
on the overall aggression levels based on which the 2004 Conference on Behavior Representation in
phase the crowd is currently in. Modeling and Simulation, Arlington VA, May 17-20
2004, pp. 397-398.
5 CONCLUSION [8] Gaskins, R. C., Boone, C. M., Verna, T. M., Bliss, J.
P. and Petty, M. D., Psychological Research for
As can be gleaned from this paper, much work still needs Crowd Modeling, Proceedings of the 2004
to be done in the development of a cognitive model for Conference on Behavior Representation in Modeling
and Simulation, Arlington VA, May 17-20 2004, pp.
crowd behavior. However, we were able to create a CM
401-402.
architecture that processes incoming stimuli to produce
[9] McKenzie, F. D., Xu, Q., Nguyen, Q. H. and Petty,
credible crowd behavior. Due to time constraints, we
M. D., Crowd Federate Architecture and API
were only able to implement a simplified cognitive model Design, Proceedings of the Fall 2004 Simulation
that directly responds to stimuli. However, the CM Interoperability Workshop, Orlando FL, September
framework can easily support a more complex algorithm 19-24, 2004.
for processing the stimuli and their effect on the crowd [10] Weisel, E. W. and Petty, M. D., Reference
parameters & crowd aggression level to produce a Scenarios for Crowd Federate Validation,
credible crowd-level behavior. Proceedings of the Fall 2004 Simulation
Most of the data that we have collected during our Interoperability Workshop, Orlando FL, September
research is at the crowd and group level. We realize that 19-24 2004.
[11] Hunter, K., Petty, M. D. and McKenzie, F. D.,
a lot more research still needs to be done at the individual
Experimental Evaluation of the Effect of Varying
levels. However, once the research has been completed
Levels of Crowd Behavior Fidelity on the Outcome
and the psychological model more refined, we could
of Certain Military Scenarios, accepted for
easily incorporate these into the CM to produce more publication at the Spring 2005 Simulation
complex behavior selection. Interoperability Workshop, San Diego, CA, April 3-8
2005.
6 REFERENCES [12] McPhail, Clark. The Myth of the Madding Crowd.
Walter de Gruyter, Inc., New York, 1991.
[1] Kenny, J. M. and Gilpin, W. L., Insertion of Crowd [13] Bouvier, E. and Guilloteau, P., Crowd simulation in
Behavior Models into the INIWIC Course, immerse space management, Proceedings of
Proceedings of the 2002 Interservice/Industry Eurographics Workshop on Virtual Environments
Training, Simulation, and Education Conference, and Scientific Visualization 96, 1996, pp. 104-110.
Orlando FL, December 2-5 2002, pp. 1064-1074. [14] Reynolds, C. W., Flocks, herds, and schools: A
[2] Ferguson, M., Questionnaire response, May 8 2003. distributed behavioral model, Proceedings
[3] Reece, D. A., Crowd Modeling in DISAF, SIGGRAPH 87, pp. 25-34.
Proceedings of the Eleventh Conference on [15] Brogan, D. and Hodgins, J., Group behaviors for
Computer-Generated Forces and Behavior systems with significant dynamics, Autonomous
Robots, Vol. 4, 1997, pp. 137-153.
[16] Musse, S. R., Babski, C., Capin, T. and Thalmann, Dominion University where he currently serves as
D., Crowd modelling in collaborative virtual Principal Investigator (PI) and Co-PI on projects
environments, Proceedings of the ACM Symposium involving medical modeling and simulation, behavior
on Virtual Reality Software and Technology, Taipei representation in simulations, and simulation
Taiwan, November 1998, pp. 115-123. architectures. Prior to joining ODU, he held a senior
[17] Musse, S. R. and Thalmann, D., Hierarchical model scientist position at Science Applications International
for real time simulation of virtual human crowds, Corporation (SAIC), serving as Principal Investigator for
IEEE Transactions on Visualization and Computer several distributed simulation projects. At SAIC he was a
Graphics, Vol. 7, No. 2, April-June 2001, pp. 152- Team Lead on a large distributed simulation system. He
164. has over 10 years of research and development experience
[18] Kenny, J. M., McPhail, C., Farrer, D. N, Odenthal,
in the software and artificial intelligence fields, including
D., Heal, S., Taylor, J., Ijames, S. and Waddington,
object-oriented design and knowledge-based systems.
P., Crowd Behavior, Crowd Control, and the Use of
Both his M.S. and Ph.D. work have been in artificial
Non-Lethal Weapons, Technical Report, Penn State
Applied Research Laboratory, January 1 2001. intelligence focusing on knowledge representation and
[19] McKenzie, F. D., Garcia, H., Nguyen, Q. H., model-based diagnostic reasoning.
Seevinck, J. and Petty, M. D., Mogadishu Terrain Mikel D. Petty is Chief Scientist of the Virginia
Generation and Correlation for Crowd Modeling, Modeling, Analysis and Simulation Center at Old
Proceedings of the Spring 2004 Simulation Dominion University. He received a Ph.D. in Computer
Interoperability Workshop, Arlington VA, April 18- Science from the University of Central Florida in 1997.
23 2004, pp. 944-950. Dr. Petty has worked in modeling and simulation research
and development since 1990 in areas that include
7 ACKNOWLEDGEMENT simulation interoperability, computer generated forces,
multi-resolution simulation, and applications of theory to
This research described in this paper is sponsored by the simulation. He has published over 110 research papers
Defense Modeling and Simulation Office and managed by and has been awarded over 50 research contracts. He has
the Air Force Research Laboratory. That support is served on a National Research Council committee on
gratefully acknowledged. We would also like to thank the modeling and simulation and is currently an editor of the
psychology team Ryland Gaskins and Carlotta Boone for
journals SIMULATION: Transactions of the Society for
tireless meetings and endless tables.
Modeling and Simulation International and Journal of
Defense Modeling and Simulation.
8 AUTHORS BIOGRAPHIES

Quynh-Anh (Mimi) H. Nguyen is a Graduate Research


Assistant and PhD student in the Engineering Modeling &
Simulation program at the Virginia Modeling, Analysis,
& Simulation Center at Old Dominion University. She
received her M.E. degree in Engineering Modeling &
Simulation from Old Dominion University in 2003, and a
B.S. in Electrical Engineering from George Mason
University in 1997. Prior to coming to ODU, Ms.
Nguyen held a staff engineer/scientist position at
Lockheed Martin Naval Electronics & Surveillance
SystemUndersea Systems, where she worked on
several modeling & simulation projects, including
designing & implementing a collaborative engineering
environment and developing complex system
performance models using discrete-event simulations.
While at VMASC, Ms. Nguyens research has focused on
using military simulation, distributed simulation and
gaming technology to develop a Crowd Federate and
crowd cognitive model.
Frederic (Rick) D. McKenzie, Ph.D. is an Assistant
Professor of Electrical and Computer Engineering at Old

Você também pode gostar