Kỷ yếu Hội nghị KHCN Quốc gia lần thứ XIV về Nghiên cứu cơ bản và ứng dụng Công nghệ thông tin (FAIR), TP. HCM, ngày 23-24/12/2021
DOI: 10.15625/vap.2021.0043
NAVIGATING EMERGENCY CROWD EVACUATION
USING MAS-GIG MODEL
Dinh Thi Hong Huyen1, Hoang Thi Thanh Ha2, Michel Occello3
1
Quy Nhon University
2
Da Nang University of Economics, The University of Da Nang
3
Grenoble Alpes University
[email protected],
[email protected],
[email protected]
ABSTRACT: In this paper, we consider the problem of modeling complex systems at several levels of abstraction. We
proposed MAS-GiG, a multi-level multi-agent model for modeling crowd navigation systems. Our approach is based on the
formation of different levels to enhance the effectiveness of managing distributed multi-level systems. The mechanism of formation of
levels is carried out from the bottom up. Thanks to the multi level observation, the complexity of the system is reduced on each level,
and the system can also provide instructions that guarantee an effective monitoring. For experimentation, the MAS-GiG model was
applied in the case of navigation the crowd following the safe path for minimizing the number of casualties.
Keywords: Complex system, multi-agent system, crowd navigation, emergency evacuation.
I. INTRODUCTION
Over the past several decades, complex systems have been the subject of extensive research in various
approaches. The concept of “complexity” is defined depending on the field of study. Some examples such as: human
brain, world economy, ecology system, internet, human society, global climate, power grid, transportation system,
crowd move. In these systems, there are always some characteristics that make the study of it more and more
challenging, such as: many participants, heterogeneous components, the system is always changing, diverse and special
is its multi-level, multi-proportionality. It is because of such properties that in complex systems there are always
unpredictable phenomena arising from local interactions that can affect the whole system [24]. Local interactions
between parts give rise to spontaneous activity. This can be a process that is not controlled by any external agent. This
process is completely decentralized, distributed in the system.
Modeling complex systems allows us to understand the system and predict its changes. According to Michael
Batty et al [1], the author argues that it is not possible to take a holistic view of a complex system to reproduce its
diverse behaviors. Therefore, many studies are performed and simulation based on partial representations. Multi-agent
systems (MAS) provide a suitable model for modeling such complex phenomena.
The passenger system at locations such as train stations, airports, and shopping malls is also considered a
complex system. This system has a large number of passengers whose behavior is constantly changing, a very large
number of interactions including passengers interacting with each other and they interacting with the environment. If
there is an emergency situation such as a fire, each passenger does not know the environment well, so he panics, he
tries to get out at all costs, he may jostle and trample with others. That's why the death rate is high. As for the
emergency management agency, they could not guide individual evacuation, because of the large number of evacuees,
plus the panic. Automated guidance systems are not effective. Therefore, the problem of managing and navigating the
crowd according to each level in an emergency situation is really necessary.
From the characteristics of complex systems, specifically the crowd system, we choose the MAS approach to
model the system for a number of reasons such as. First, MAS uses separate entities to represent components such as
autonomous, intelligent, flexible, interactive agents. Specifically, agents interact with their neighbors or with the
environment to gain knowledge. Then the agents use their knowledge to decide and perform an action on the
environment to solve the assigned task. Second, MAS is a kind of intelligent distributed system in which autonomous
agents exist in a world without global control [2]. Third, MAS describes the system in a natural way, specifically
describing pedestrian behavior. At any given time, each agent in the MAS is situated at any location in the environment
and is capable of acting autonomously in that environment. Example, an agent can move to another location by
changing its current location or it can change its state such as move to another place, return to the previous position etc.
Another advantage of MAS is that it can use multi-level organization. Conte et al [3] argue that for a given
complex system, expert knowledge is available at different levels of detail depending on the domain. The author also
argues that expert knowledge is largely unusable if we consider it only at the micro level. In MAS models, entities
corresponding to different levels of abstraction will be represented by agents at different levels in the MAS
organization [4].
In this paper, we propose the MAS-GiG multi-level multi-agent model to model the crowd. The main problem
of the research is to build different levels in the model to control and manage the system at each level. The
60
NAVIGATING EMERGENCY CROWD EVACUATION USING MAS-GIS MODEL
organizational structure in the model is determined by the group structure and the interaction relationships. The model
is applied in crowd navigation during a fire in a flight lounge. The experimental results include the total evacuation
time and the total number of people unable to get out of the dangerous place to the safe place. If the smaller the total
evacuation time and the remaining number of evacuees then the model is more optimal.
The paper is organized as follows: Section II, related works to multi-level multi-agent models. Section III, our
approach to building the MAS-GiG model. Section IV, experiment the proposed model for crowd navigation when the
fire occurred in a lounge with a structure similar to lounge 5, floor 1st, Danang International Airport. Finally, Section
V, focuses on our conclusions and discussion of the results obtained and future research.
II. RELATED WORKS
According to Wood et al [5], the authors proposed a methodology for the design of multi-agent systems, with a
simple approach of building a MAS through an entire software development life cycle. From problem description to
implementation including 6 steps: Capturing Goals, Apply Use Cases, Refining Roles, Create Agent Classes, Build
Conversations ), Assembling Agent Classes. The authors do not mention groups and multi-level structures. Nahid
Salehi et al [6], proposed a velocity-based method to simulate the actual behavior of the moving group. From the initial
position to move to the common destination, in the process of moving, agents in a group must try to maintain the
formation and cohesion of the group. Groups are created by combining multiple agents in a specific shape based on
adjusting their orientation and position. Each group has a leader and followers. The authors did not mention multi-level
structure. Noureddine et al. proposed a multi-agent model with AGR (Agent/Group/Role) architecture [7,8] with three
concepts: agent, group and role. An agent is an entity that acts, communicates, has a role, and is a member of a group.
A group is a set of agents, the group structure is an abstract representation of a group described by a set of roles. A role
is an abstract representation of an agent's position and function in a group. Navarro et al. [10] studied the multi-level on
MAS to increase simulation speed. In this model, only one level is active at a time while other levels in the same model
do not co-exist (virtual). In the model, the author does not mention the group. Nguyen Thi Ngoc Anh and et al.
proposed a synthetic model that combines a mathematical model and a multi-agent model [13]. In this model, the
authors combine many types of heterogeneous models to represent several levels. Agents are synthesized and routed
using equations of fluid flow. Interaction between levels allows the combination of models with information sharing
functions. This model did not use the concept of group. The experiment of the model was applied for tsunami warning.
Thomas Huraux et al. proposed a method to construct a holonic multilevel MAS [4]. In the model the agents are
organized as nested groups, but the author does not mention the structure and formation of groups. The levels in the
model all exist during the simulation. Stefania Bandini et al. proposed a multi-level model that combines three different
MAS, including a cognitive MAS (MASC) is at the tallest level, a reflex MAS (SMAR) is at the lowest level, and a
recursive MAS located in the middle level [14]. Recursive MAS communicates between SMACs and SMARs. Hoang
Thi Thanh Ha and et al. proposed a multi-level multi-agent recursive (MAS-R). The purpose is to increase the
effectiveness of monitoring distributed multi-level systems [15]. The levels in the model are virtual abstraction, only
the base level is real in the model. Afra Khenifar et al. presented the background of a multi-agent multi-level model
[16] to solve the problem of cooperation between groups of multi-agent to observe and control collective products. In
this study, the collective product is the result of the interaction of different agents.
We appreciate the collective products in complex systems that are merged from interaction from multi elements.
We are interested in the group’s behavior because group members usually have maken collective decisions and actions.
In this paper we propose a multi-level multi-agent model (MAS-GiG) with the purpose to enhance the effectiveness of
managing distributed multi-level systems that are suitable for multi-scale complex systems. Our model emphasizes the
group structure; group formation; multi-level formation and interaction. The group structure at each level consists of a
group leader and members, the members of the group at the upper adjacent level are representatives of the groups at the
lower adjacent level. The different levels in the model are closely related. We firstly apply this model for modeling the
human crowd. Therefore, the MAS-GiG model has some more features than existing models: group formation based on
social relationships. This is a main feature of the public crowds. We created the group based on this criteria because it
is highly practical. Moreover, members of a family, friends, and colleagues often go together to the same goal. They
rarely separate when moving. Because of this feature, the stability of the group is higher. Therefore, the crowd is less
affected, the evacuation is more convenient.
We choose the study in the document [39] to compare with the proposed model, the MAS-GiG model and the
navigation algorithm [39] which have the same things as planning the agent navigation in crowded environments, avoid
obstacles, use a pre-calculated route, and provide universal connectivity. However, the MAS-GiG model has some
differences compared to the algorithm in [39] and these are also its advantages such as: first, MAS-GiG is a multi-level
model. It manages and navigates the crowd by each level, each of which manages and navigates the crowd within its
scope and performs the navigation under the direction of the superiors; second, the navigation plan based on the
available physical environment. The physical environment is represented by a path map 2D. The path map is created
from intersection points and paths connecting two intersecting points. The intersection point is called a node and the
path connecting two nodes is called an edge. The navigation plan is managed and coordinated by the highest level in
the model; thirdly, combine shortest path finding Dijkstra algorithm and safest route in choosing the route to evacuate.
Dinh Thi Hong Huyen, Hoang Thi Thanh Ha, Michel Occello
61
Fourth, MAS-GiG does not calculate collisions during evacuation because it has a pre-planned plan to guide crowd
evacuation.
III. MULTI-LEVEL MULTI-AGENT MODEL MAS-GIG
In this section, we introduce agent model, multi-agent system, architecture of MAS-GiG model and its
application to navigate emergency crowds at a flight lounge.
A. Agent model
Each agent in MAS-GiG has autonomous, responsive, and interactive behavior. It is characterized by a set of
attributes P, knowledge K, a set of roles R, and a set of actions A.
1. Agent properties
Properties agents are the variables that can be
manipulated by the agent. An agent's properties describe its
characteristics, including his identifier, shape, and size. An
attribute can be a variable to perform calculations such as:
perception score, decision time. Or more complex like a list, a
collection like: a list of agents in the same group. Every agent
controls autonomously and locally a set of scopes for
perception, communication and action. In this study, an agent is
represented by an individual and a group is represented by a
group of individuals that have a social relationship with each
other. A group is a collective of individuals.
2. Agent knowledge
Figure 1. Agent architecture
Agent’s knowledge, also known as cognitive score
(cognitiveGrade), consists of the understanding of the environment (knownExitNo), personal experience (experience),
decision-making time (decisionTime). In each level, based on the agent's knowledge to define the role. The agent with
the highest knowledge in the group is chosen as the group leader. The role of group leader is managing the group.
3. Agent role
Each agent has a role to perform its respective actions. There are types of roles such as: LeaderRole,
MemberRole, IndieRole. LeaderRole is the role of a group leader. MemberRole is the role of a group member.
IndieRole is the role of an individual that is not in any group. A LeaderRole performs a number of corresponding
actions such as: interacting with group representatives in the upper level, interacting with group representatives in the
same level, interacting with group members in the same group, and interacting with individuals to which they do not
belong to any group. A MemberRole performs a number of corresponding actions such as: interacting with the group
leader in the same group, interacting with group members in the same group, and interacting with individuals to which
they do not belong to any group. An IndieRole performs a number of corresponding actions such as: interacting with a
group leader in the same level, interacting with a group member in the same level, and interacting with other
individuals in the same level.
4. Agent action
The agent's action depends on its decision. For each specific situation, an agent makes a decision for himself or
he interacts with others and then makes a decision. An agent acts according to his decision. In the model, actions are all
performed by sending/receiving messages.
B. The multi-agent system
The multi-agent system is organized according to the AEIO approach [34]. It includes agents (A), environment
(E), interactions (I), and organizations (O).
1. Environment (E)
A simulated environment is a two-dimensional space. Environment objects are categorized and located in the
environment for simulating. The objects are categorized based on their functions, such as obstacles, paths, exits, etc.
2. Interactions (I)
The interaction is represented by sending and receiving messages. It includes sending/receiving messages
between agents of the same level and between agents of different levels. The structure of a message is based on the
interaction protocol in MASH [27]. The mechanism to actually interact is based on two classes of Messages and
Frames. They represent the Network and Datalink layers in the OSI seven-layer model of computer networks,
respectively [31,32]. The structure of the message: Message (senderID,receiverID, content). Protocol for
sending/receiving messages in MASH (see figure 2).
62
NAVIGATING EMERGENCY CROWD EVACUATION USING MAS-GIS MODEL
There are two types of interactions that are horizontal interaction and vertical interaction. Horizontal
interactions are interactions performed by sending/receiving messages between agents at the same level. Vertical
interactions are interactions performed by sending/receiving messages between agents at different levels. We specify
five interactions to be used in the model:
(1) Indie ↔ Indie: the individuals communicate with each other.
(2) Members ↔ Members: the group members in the same group communicate with each other.
(3) GroupLeader ↔ Member: a group leader communicates with a member in the same a group.
(4) GroupLeader ↔ Indie: a group leader communicates with an individual.
(5) GroupLeader ↔ GroupLeader: a group leader communicates with a group leader.
int
Message
int
senderID
int
byte()
Frame
header
receiverID
senderID
content
int
receiverID
byte()
data-msg
byte()
trailer
Figure 2. The interaction protocol in MASH
3. Organizations (O)
Organization (O) is represented by the organizational structure and the relations. The organizational structure
mentioned here is the group organizational structure [17].
C. Multi-level architecture of the MAS-GiG model
In the MAS-GiG model, each agent represents an
intelligent and autonomous agent. It has knowledge that can
decide for itself to take personal action to achieve personal
goals. It can interact with other agents to perform actions and
achieve common goals.
- Level 0: this is the real level, the agents at level 0 are
base agents, it represents an individual in the application. At
this level, the agents have the same role. The interaction is
carried out between individuals. There is no organization.
- Level 1: is the first group level in the model. A
group at level 1 is formed from base agents at level 0. The
formation of the group based on the social relationships such
as: family, couple, friends, colleagues [17]. Each group has a
structure. The group structure consists of a group
representative and group members. The role of a group
representative is LeaderRole, the role of a group member is
MemberRole, and the role for individuals who are not part of
the group is IndieRole.
Figure 3. The general architecture of MAS-GiG
The interactions include: the group members interact
with each other in the same group, the group leader interacts with group members in the same group, the group leader
interacts with each other at the same level, the group leader interacts with the group leader at different levels. The
group leader interacts with the individuals who are not part of the group.
- Level 2: this is the second group level. At this level, the agents are mostly representatives of groups at level 1.
The agents belong to the same group when the distance between them is less than or equal to a constant r. The group
structure at level 2 also includes a group representative and group members. The group representative is chosen based
on the knowledge score. At this level, the role of agents, the interactions between the agents are similar to level 1. The
number of agents at level 2 is much reduced compared to level 1.
Dinh Thi Hong Huyen, Hoang Thi Thanh Ha, Michel Occello
63
- Level n-1: this is the (n-1)th group level. The
members at level (n-1) th are mostly representatives of groups
at level n-2. At this level, the formation of the group, the
group structure, group's representative selection, the
interaction, and kinds of roles are similar to the level 2.
- Level n: level n is the highest level and is also the
general operating level for the entire system. At this level
there is only a representative. The interaction is performed
with the representatives of the lower levels.
Figure 3 illustrates the general architecture of the
MAS-GiS model. Figure 4 illustrates the formation, structure,
and interaction of levels in the MAS-GiG model.
D. The MAS-GiG model in crowd navigation
1. The steps are performed for simulating the application.
The steps are performed for simulating the application
(figure 5). Initially, requirements are analyzed about physical
structure and estimated the number of levels of model for
application. The physical structure includes the distribution of
doors, paths, walls, obstacles, exits, etc. Based on this
distribution, we define the levels of the model for the
application and make the scenarios.
Figure 4. The formation of levels in MAS-GiG
Next, we develop the evacuation simulation model.
There are two agent-related factors such as: fire perception and
physical disability. There are some building-related factors
such as: exit width, exit location, fire location, path width. The
occupancy density is representing a relation between the agent
and the building. Finally, the output datas include the number
of total evacuation times and remaining occupants are
statistically analyzed.
- Fire perception: we assume that 100% of the
occupants perceive a fire emergency and they respond after 10
seconds [45].
- Physical disability: we assume that 100% of the
occupants are not physically disabled.
- Exit width: the exit door pertains to the building's
characteristic. In this application, we assume that the width of
the exit door is 80 cm.
Figure 5. The steps are performed for simulating
li i
- Exit locations, fire locations: these factors depend on
the physical structure of the application and they are described
in the scenarios.
- Occupant density: this factor depends on the number
of passengers in each scenario.
The last, the third is statistical analysis about total
evacuation time, the number of remaining evacuees. These data
can be compared with the results of other methods to evaluate
the proposed model.
2. Some levels of MAS-GiG model are described for crowd
navigation at a flight lounge
According to [33] guidelines for emergency evacuation
Figure 6. The levels for the application
at airports in case of fire, we propose the levels of the model
for the application. There are five levels: indie level (Indie) is the base level, each member at this level represents a
passenger in the lounge. The group level (Group) is the group of passengers. The area level (Area) is also a group level
that a group at this level includes the representatives of the groups at group level. The lounge level (Lounge) is also a
group level that a group at this level includes the representatives of the groups at area level. The AirportOperator level
64
NAVIGATING EMERGENCY CROWD EVACUATION USING MAS-GIS MODEL
(AirportOperator) is an emergency management
department. In this application, there is only one
lounge so the AirportOperator manages the
lounge directly (figure 6).
The mechanism of formation of levels is
carried out from the bottom up: a group consists
of many Indies who have a social relationship
with each other [17]. At the group level, there are
many groups, each group has a group
representative that is called a group leader. An
Area consists of many Groups, these Groups are
in a range of the Area. At the Area level, there are
many Areas, each Area has a representative that
is called a Guide. A Lounge consists of many
Areas, these Areas are in the same lounge. At the
Lounge level in the application, there is only a
Lounge and the representative is called a
GuideLeader.
Figure 7. Class diagram agent levels
The relationships between agent classes in the application are presented in figure 7.
IV. EXPERIMENTING MAS-GIG MODEL IN CROWD NAVIGATION AT A FLIGHT LOUNGE
Apply the proposed model in the application of crowd navigation when the fire occurred in a lounge with a
structure similar to the lounge 5, floor 1, Danang International Airport.
A. Introduction about MASH (Hardware/Software MAS Simulation)
MASH [46] is a simulator developed by
Professor Michel Occello et al. to experiment the
works related to multi-agent based simulation in
three ways: software simulation, hardware
simulation and hybrid simulation (combining
software and hardware simu-lation). The simple
architecture of MASH is shown in figure 8.
MASH provides a mechanism to emulate both
a virtual agent and an embedded agent. Each agent
has the corresponding properties and behaviors to
perform its own role. The virtual agent and
embedded agent can be integrated in the same
simulation, which is abstracted through the
Individual Agent Manager.
Figure 8. The simple architecture of MASH
In addition, to manage the communication
and interaction between agents through the Society Manager. Agents interacting with the environment are managed by
the Environment Manager. All information is communicated through the event mechanism, which is managed through
the Event Manager.
B. Experiment design
The Random (i.e. scenario 0) scenario is
designed to compare with some scenarios of the
proposed model. The random scenario is tested with
three fire locations corresponding to the three
proposed scenarios F1, F2, F3. The relevant
parameters are used in general for all scenarios. When
all occupants are assumed to have 100% fire
awareness and the occupants are not disabilities. The
width of all exit doors are 80 cm. The fire origin was
established at three locations as in the three MASGiG model test scenarios. There are about 200 to 300
evacuees in the lounge. Three scenarios are
independently compared with the Random scenario.
Figure 9. A design for a navigation scenario
For each scenario, 10 trials are run, the average total
evacuation time and average total remaining evacuees are calculated.
e
exit
Dinh Thi Hong Huyen, Hoang Thi Thanh Ha, Michel Occello
65
We assume that the fire location is at the entrance (Figure 9), the passengers are not allowed to choose the exit
as the entrance, they can only choose the remaining two exits to move out. For the random scenario, an agent chooses
randomly the exit, if he moves to the exit where it has a fire then he has to move back to the original place, then he
chooses randomly one of the two remaining exits to move. In figure 9, the blue arrows are the directions to be moved,
the red arrows are the directions that are not moved.
C. Some scenarios
We experiment on three scenarios, based on the
actual flight lounge structure [44]. The lounge 5 is one of
4 lounges on the 1st floor of Da Nang International
Airport. The following is a brief description of the lounge
structure. Passengers enter the lounge from the stairs on
the 2nd floor. They boarded the plane at the boarding
gate. In addition, there is an exit to move to the ground
floor, which is located in the hallway next to lounge 6. In
the lounge there is an area for passengers to sit called the
waiting area. An area for souvenirs is called Shop, an area
is called Restaurant and an area is for Toilet. Based on the
actual waiting room structure, we create a 2D map for
experimental simulation. In each area are numbered
correspondingly such as: (1): Safe area, (2): Boarding
gate, (3): Entrance to lounge from 2nd floor, (4): Exit to
Figure 10. The map for experimental environment
go down to the ground floor, (5): Waiting area, (6): Shop,
(7): Restaurant, (8): Toilet, between the rows of seats in the waiting area are paths. In the three experimental scenarios,
the entrance (3), boarding gate (2) and exit (4) are used as three exits E3, E2, E1. Three fire positions are set at three
positions respectively and they are noted F3, F2, F1 (figure 10).
Scenario 1: The fire occurs at the position F1 which does not coincide with one of the three exits. Each agent
can choose one in three exits E1, E2, E3 to move to a safe place.
Scenario 2: Fire occurs at the exit E2. Each agent can choose one of the two exits E1 or E3 for moving. In this
case, the agents are not allowed to move through the exit E2.
Scenario 3: Fire occurs at the exit E3. Each agent can choose one of the two exits E1 and E2 for moving. In this
case, the agents are not allowed to move through the exit E3.
- For position F1, fire occurs at a location not close to any emergency exits, so F1 represents the fire locations
where it is not located at the exits.
- For position F2, fire occurs next to the boarding gate. Only a small number of passengers know exit E2. So F2
represents fire locations where a small number of passengers know the exit at these locations.
- For position F3, fire occurs next to the entrance to the lounge. All passengers know exit E3. So F3 represents
fire locations where all passengers know the exit at these locations.
In the simulation, the individual passengers are represented by blue dots, the group passengers are represented
by the yellow dots, the group leaders are represented by the red dots, the area leaders are represented by the black dots
and the lounge leader is represented by a gray dot.
D. Experimental results
To evaluate the proposed model, we present two types of results. First, the results are visually represented by the
multi-level evacuation instructions. The groups in the areas of the lounge that are close to the fire are evacuated first.
Then the groups in the areas of the lounge that are close to the exit are evacuated next (see figures 11a, 11b, 11c).
Second, the average total evacuation time and average total remaining evacuees (table 1).
Table 1. The experimental results
66
NAVIGATING EMERGENCY CROWD EVACUATION USING MAS-GIS MODEL
Each scenario is run 10 times and averaged to compare with the Random scenario. The experiment results for
MAS-GiG model: the result of scenario 1 is presented in figure 11a, the result of scenario 2 is presented in figure 11b,
the result of scenario 3 is presented in figure 11c.
The experimental results show that the total evacuation time for the scenarios of the MAS-GiG model is always
less than the total evacuation time of the Random scenario. Specifically, the average evacuation total time of Random
scenario with fire location F1 is 127.4s while scenario 1 of MAS-GiG model is 113.8s. Similarly, the average
evacuation time of the Random scenario with fire location F2 is 139s while that of the MAS-GiG scenario 2 is 127s,
and the average evacuation time of the Random scenario with the fire location F3 is 130.4 s while scenario 3 of the
MAS-GiG model is 119 s. The average total remaining evacuees of Random scenario with fire location F1 is 16.2
people while scenario 1 of MAS-GiG model is 8.2 people. The average total remaining evacuees of Random scenario
with fire location F2 is 19.8 people while scenario 1 of MAS-GiG model is 10.28 people. The average total remaining
evacuees of Random scenario with fire location F3 is 20.6 people while scenario 1 of MAS-GiG model is 12.4 people.
Figure 11. The experimental results
V. DISCUSSION AND CONCLUSION
As we have presented in the previous sections, each group has at least 2 members, so after each new level
formation step, the number of agents is reduced by more than half compared to the number of agents at the previous
adjacent level and the number of interactions is also reduced by more than half. Therefore, it is easier to observe the
system, and the computational complexity is reduced. This is an advantageous feature of the MAS-GiG model
compared to other studies in the same field.
The proposed model has achieved a number of results: first, the navigation of passenger movement is level-bylevel and collision-free. The groups in the areas of the lounge that are close to the fire are evacuated first. Then the
groups in the areas of the lounge that are close to the exit are evacuated next. Second, the total evacuation time and the
total number of evacuees remaining are lower than the result of the Random method. These results answer the question
why we built a multilevel model? If there is only one operating level that coordinates the entire system then the
guidance information cannot be transmitted to all evacuees, each evacuee finds the exit himself. In this situation, the
evacuation is similar to the Random evacuation method. Therefore, the total evacuation time and total remaining
evacuees are large. This leads to a high death rate.
In the next studies, we will continue to study individual behavior, collective behavior. Apply proposed model in
navigating emergency crowd evacuation in many applications with different physical structures such as nightclubs,
shopping malls. For the airport application, we will extend the model to multiple levels and apply on multiple
lounges.There are more than one fire spot in an area that has different lounges.
REFERENCES
[1] Michael Batty & Paul M. Torrens. “Modeling Complexity: The Limits to Prediction”. ISSN: 1467-1298. 2001.
[2] Ahmad Esmaeili, Nasser Mozayani, Mohammad Reza Jahed Motlagh, “Multi-level Holonification of Multi-agent Networks”,
12th Iranian Conference on Intelligent Systems, 2014.
[3] Conte, R., Edmonds, B., Moss, S., and Sawyer, R. K. (2001), “Sociology and social theory in agent based social simulation: A
symposium”, Computational & Mathematical Organization Theory, 7(3):183-205.
Dinh Thi Hong Huyen, Hoang Thi Thanh Ha, Michel Occello
67
[4] Thomas Huraux, Nicolas Sabouret, and Yvon Haradji, “A Multi-Level Model for Multi-Agent Based Simulation”, 6th
International Conference on Agents and ARTificial intelligence (ICAART 2014), Mar 2014, Angers, France. hal-01835404.
[5] Wood, M. F. and DeLoach, S. A. (2001), “An overview of the multiagent systems engineering methodology. In Agent-Oriented
Software Engineering”, pages 207-221. Springer.
[6] Nahid Salehi, and Mankyu Sung, “Realistic Multi-Agent Formation Using Discretionary Group Behavior”, Connecting
Research and Researchers, DOI: 10.3390/app10103518, 2020.
[7] Noureddine Seddari, Mohamed Belaoued, Salah Bougueroua, “Agent/Group/Role Organizational Model to Simulate an
Industrial Control System”, World Academy of Science, Engineering and Technology International Journal of Industrial and
Manufacturing Engineering Vol:11, No:10, 2017.
[8] O. Kaz´ık, “Role-based Approaches to Development of Multi-Agent Systems: A Survey”, WDS'10 Proceedings of Contributed
Papers, Part I, 19-24, 2010. ISBN 978-80-7378-139-2 © MATFYZPRESS.
[9] Tranouez, P. (2005). Contribution `a la mod´elisation et `a la prise en compte informatique de niveaux de description multiples,
PhD thesis, Universit´e du Havre.
[10] Navarro, L., Flacher, F., and Corruble, V. (2011), “Dynamic level of detail for large scale agent-based urban simulations”,
Proc. of 10th Int. Conf. on Autonomous Agents and Multiagent Systems (AAMAS 2011), pages 701-708.
[11] Minar, N., Burkhart, R., Langton, C., and Askenazi, M. (1996), “The swarm simulation system : a toolkit for building multiagent simulations”, GEMAS Studies in Social Analysis, Working Paper 96-06-042.
[12] Gil-Quijano, J., Piron, M., and Drogoul, A. (2007), “Mechanisms of automated formation and evolution of socialgroups : A
multi-agent system to model the intraurban mobilities of Bogota city”, Idea Group Inc.
[13] Nguyen, T. N. A., Zucker, J.-D., Du, N. H., Drogoul, A., and Vo, D.-A. (2011),”An hybrid equation-based and agent-based
modeling of crowd evacuation on road network”, International Conference on Complex Systems.
[14] Stefania Bandini, Sara Manzoni & Giuseppe Vizzari, “Toward a platform for multi-layered multi-agent situated system
(mmass)-based simulations: focusing on field diffusion”, Applied Artificial Intelligence 2006.
[15] HOANG Thi Thanh Ha, NGUYEN Thanh Binh, Michel OCCELLO, “Applying theory of types to formal specification of
recursive multiagent systems”, IEEE-RIVF International Conference on Computing and Telecommunication Technologies,
Danang : Viet Nam (2009). No: 3. Pages: 57-64. Year 2009.
[16] Afra Khenifar, Jean-Paul Jamont, M. Occello, Mouloud Koudil, “About cooperation of multiagent collective products: An
approach in the context of cyberphysical systems”, Conference: 2016 IEEE RIVF International Conference on Computing &
Communication Technologies, Research, Innovation, and Vision for the Future (RIVF), 2016.
[17] Dinh Thi Hong Huyen, Hoang Thi Thanh Ha, Michel Occello, “Detect the formation of groups in a crowd based on social
relations” The 10th conference on information technology and its application. CITA 2021.
[18] Y. Demazeau, “From Interactions to Collective Behavior in Agent-Based Systems”, European Conference on Cognitive
Science, Saint-Malo France, 1995.
[19] Moussaïd, M., et al., “The walking behaviour of pedestrian social groups and its impact on crowd dynamics”, PloS One 5(4),
e10047 (2010)
[20] J. Sochman and D. C. Hogg:, “Who knows who - inverting the social force model for find-ing groups”, In Proc. of IEEE
ICCVW, 2011.
[21] H. Singh, R. Arter, L. Dodd, P. Langston, E. Lester, andJ. Drury, “Modelling subgroup behaviour in crowd dy-namics dem
simulation. Applied Mathematical Modelling”, 33(12):4408-4423, 2009.
[22] Manon Prédhumeau, Julie Dugdale, Anne Spalanzani, “Modeling and Simulating Pedestri-an Social Group Behavior with
Heterogeneous Social Relationships”, SCS 2020 - Spring Simula-tion Conference, May 2020, Virtual event, United States.
ffhal-02514963f.
[23] Fasheng Qiu*, Xiaolin Hu, “Modeling group structures in pedestrian crowd simulation”, Simulation Modelling Practice and
Theory. 2010.
[24] Hoàng Thị Thanh Hà, “Un modèle multi-agent récursif générique pour simplifier la supervision de systèmes complexes
artificiels décentralisés”, These de doctorat informatique, Université de Grenoble, 2012.
[25] Hoàng Thị Thanh Hà, Hoàng Thị Bích Ngọc, “Mô hình đa tác tử hỗ trợ ra quyết định dựa trên sự hòa nhập dữ liệu rời rạc: ứng
dụng phát hiện cháy rừng”, Đề tài khoa học và công nghệ cấp cơ sở. Đại học Đà Nẵng, 2015.
[26] Huang C-Y, Tzou P-J, Sun C-T, “Collective opinion and attitude dynamics dependency on informational and normative social
influences”, Simulation. 2010:0037549710387940.
[27] Jean-Paul Jamont, Michel Occello, “A Multiagent Tool to Simulate Hybrid Real/Virtual Embedded Agent Societies”,
Proceedings of the 2009 IEEE/WIC/ACM International Conference on Intelligent Agent Technology, IAT 2009, Milan, Italy,
15-18 September 2009.
[28] Xiao-Feng Xie, Stephen F. Smith and Gregory J. Barlow, “Schedule-Driven Coordination for Real-Time Traffic Network
Control”, Vol. 22 (2012): Twenty-Second International Conference on Automated Planning and Scheduling.
[29] Raul Hector Cajias, Antonio Gonzalez-Pardo, David Camacho, A Multi-agent Traffic Simulation Framework for Evaluating
the Impact of Traffic Lights, International Conference on Agents and Artificial, 2011.
[30] Ranbo ZhuLiangjie XuFu Li, “Simulation Modeling of Car-Following in Mixed Traffic Flow Based on Multi-Agent System”,
20th COTA International Conference of Transportation Professionals, 2020.
[31] Depavath Harinath, OSI Reference Model - A Seven Layered Architecture of OSI Model, Published in Volume 3, Issue 8,
August 2013.
[32] Piyush Saxena, OSI Reference Model - A Seven Layered Architecture of OSI Model, International Journal of Research (IJR)
Vol-1, Issue-10 November 2014 ISSN 2348-6848.
[33] https://www.emergency-live.com/of-interest/emergency-in-airports-how-is-an-airport-building-evacuation-carried-on/
68
NAVIGATING EMERGENCY CROWD EVACUATION USING MAS-GIS MODEL
[34] Y. Demazeau: “From Interactions to Collective Behavior in Agent-Based Systems”, European Conference on Cognitive
Science, Saint-Malo France, 1995.
[35] Enrico Ronchi, “Developing and validating evacuation models for fire safety engineering”, Fire Safety Journal, journal
homepage: http://www.elsevier.com/locate/firesaf, 2020.
[36] Terrence Macy et al. , “Fire Safety and Emergency Guidelines”, 2009.
(https://portal.ct.gov/-/media/DDS/Safety/FireSafetyandEmergencyGuidelinesMain.pdf?la=en).
[37] James Lord, Brian Meacham, Amanda Moore, “Guide for evaluating the predictive capabilities of computer egress models”,
NIST GCR 06-886, 2005.
[38] Peter van de Leur, “Building evacuation, rules and reality”, HERON, Vol. 50, No 4 (2005).
[39] Jur van den Berg, Sachin Patil, Jason Sewall, Dinesh Manocha, Ming Lin, “Interactive navigation of multiple agents in
crowded environments”, Conference: Proceedings of the 2008 Symposium on Interactive 3D Graphics, SI3D 2008, February
15-17, 2008, Redwood City, CA, USA.
[40] N. T. N. Anh, Z. J. Daniel, N. H. Du, A. Drogoul, and V. D. An, “A hybrid macro-micro pedestrians evacuation model to speed
up simulation in road networks,” in Advanced Agent Technology, vol. 7068, pp. 371-383, Springer, 2012. View at: Publisher
Site | Google Scholar.
[41] E. Ronchi, E. D. Kuligowski, P. A. Reneke, R. D. Peacock, D. Nilsson, “The Process of Verification and Validation of
Building Fire Evacuation Models”, National Institute of Standards and Technology, 2013. Technical Note 1822.
[42] E. Ronchi, E. D. Kuligowski, D. Nilsson, R. D. Peacock, P. A. Reneke, “Assessing the verification and validation of building
fire evacuation models”, Fire Technol. 52 (Sep. 2014) 197-219.
[43] H. Frantzich, D. Nilsson, K. Rød, “Utrymning och tekniska installationer i vagtunnlar med dubbelriktad trafik€”, Lund Brand,
Lunds Tek. Hogsk., 201€6.
[44] https://danangairportterminal.vn/vi/maps/#
[45] Shi, L.; Xie, Q.; Cheng, X.; Chen, L.; Zhou, Y.; Zhang, R. “Developing a database for emergency evacuation model”. Build.
Environ. 2009, 44, 1724-1729. [CrossRef].
ĐIỀU HƯỚNG SƠ TÁN ĐÁM ĐÔNG KHẨN CẤP BẰNG MÔ HÌNH MAS-GIG
Dinh Thi Hong Huyen, Hoang Thi Thanh Ha, Michel Occello
TÓM TẮT: Trong bài báo này, chúng tôi xem xét vấn đề mô hình hóa các hệ thống phức tạp ở một số cấp độ trừu tượng.
Chúng tôi đã đề xuất MAS-GiG, một mô hình đa tác tử đa mức để mô hình hóa hệ thống điều hướng đám đông. Phương pháp tiếp
cận của chúng tôi dựa trên việc hình thành các mức khác nhau nhằm nâng cao hiệu quả của việc quản lý hệ thống đa mức phân tán.
Cơ chế hình thành các mức được thực hiện từ dưới lên. Nhờ quan sát đa mức, mức độ phức tạp của hệ thống được giảm bớt ở mỗi
mức, hệ thống cũng có thể cung cấp các hướng dẫn đảm bảo giám sát hiệu quả. Để thử nghiệm, mô hình MAS-GiG đã được áp dụng
trong trường hợp điều hướng đám đông đi theo lối đi an toàn để giảm thiểu số lượng thương vong.