Multiagent Approach for Effective Disaster Evacuation

Yasuki Iizuka

, Katsuya Kinoshita

and Kayo Iizuka

Dept. of Mathematical Sciences, Tokai University, Hiratsuka, Japan

School of Network and Information, Senshu University, Kawasaki, Japan

Keywords:

Disaster Evacuation, Distributed Constraint Optimization Problem.

Abstract:

At times of disaster, or immediately prior to such periods, smooth evacuation is a key issues. However, it is

difﬁcult to achieve, because people tend to panic when faced with disaster. This paper proposes a system that

supports effective evacuation from danger using the framework of the Distributed Constraint Optimization

Problem (DCOP). The use of the DCOP facilitates the assisted optimization of people’s evacuation timing

without a center server. This system enables assistance in terms of evacuation guidance to be given to relieve

congestion, by calculating evacuation timing via an ad-hoc network of evacuees’ mobile devices (phones, PCs,

etc.). In this paper, we focus on the formalization of the disaster evacuation problem and how to solve it using

the framework of the Distributed Constraint Optimization Problem.

1 INTRODUCTION

Much effort has been expended in improving dis-

aster prevention countermeasures. Although most

of the countermeasures that have been implemented

are classiﬁed as Public-help, which are implemented

by the public sector, many people survive in times

of disaster based on Self-help (countermeasures im-

plemented by individuals) and Mutual-help (counter-

measures implemented based on mutual help). There-

fore, Mutual-help and Self-help are attracting much

attention (CabinetOfﬁce, 2011).

At times of disaster, or immediately prior to such

periods, smooth evacuation is a key issues. However,

it is difﬁcult to achieve, because people tend to panic

when faced with disaster, crowding evacuation pas-

sageways of buildings in the event of ﬁre and con-

gesting roads with cars containingpeople ﬂeeing from

predicted hurricanes.

Therefore, although evacuation guidance is very

important, the disaster countermeasure ofﬁce would

be unable to guide all evacuees intensively at the time

of a disaster. Evacuees need to take refuge based on

Mutual-help.

We aim to develop a the system that provides op-

timal evacuation guidance autonomously at the time

of a disaster. The system uses the mobile devices of

evacuees, performs distributed calculation using the

frameworkof the Distributed Constraint Optimization

Problem (DCOP), and does not need a center server.

In this paper, we focus on the formalization of the dis-

aster evacuation problem and how to solve it using the

frameworkof the Distributed Constraint Optimization

Problem.

2 DISASTER EVACUATION

The authors designed and developed a real-time dis-

aster information mapping system aimed at conjugat-

ing facilities such as university campuses for the pur-

pose of assisting Mutual-help (Iizuka et al., 2011).

This system is a web based system that can handle

disaster situation information in places such as uni-

versity campuses by level of detail, such as class-

rooms or laboratories in the school buildings by us-

ing Wi-Fi devices. It can aggregate the information

that is sent by users (informers), store it in a situation

database on the server and display a disaster situation

map to users on request. However, issues caused by

heavily-congested and crowded conditions were not

given sufﬁcient consideration in the experiments de-

scribed above. In the case where an enormous num-

ber of evacuees have to move, or they have to pass

throughnarrowaisles, they may become confused and

evacuation may take longer. In such cases, a navigat-

ing system for evacuees would be effective, prevent-

ing them from rushing into a certain passageway, or

instructing them to wait just a few moments, in order

to meet the ﬂow rate allowance.

223

Iizuka Y., Kinoshita K. and Iizuka K..

Multiagent Approach for Effective Disaster Evacuation.

DOI: 10.5220/0004905002230228

In Proceedings of the 6th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2014), pages 223-228

ISBN: 978-989-758-016-1

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: Usage image of the disaster evacuation assist sys-

tem.

By using a system like that of (Iizuka et al., 2011),

people could know “which evacuation route is safe”

, the next issue is “how to evacuate safely and ef-

fectively.” However, smooth and effective evacua-

tion is not always easy. People tend to rush forward

to passageways that are perceived to be safe, which

results in congestion. The more people rush, the

greater the congestion. The provision of additional

appropriate information concerning evacuation guid-

ance may make it feasible to avoid congestion, and

shorten evacuation times.

Disaster countermeasures ofﬁces are set up in or-

ganizations in order to determine and provide ap-

propriate evacuation routes. However, planning and

providing appropriate information rapidly is not easy

with limited resources. Therefore, an effective disas-

ter evacuation assistance system is required to address

these issues.

Our proposed system facilitates the assisted opti-

mization of people’s evacuation timing, by estimat-

ing the location of evacuees. This system enables as-

sistance to be given in the form of evacuation guid-

ance to relieve congestion, by calculating evacuation

routes and timing via an ad-hoc network of evacuees’

mobile devices (phones, PCs, etc.), intercommunica-

tion function and location information. The mobile

devices must be equipped with wireless LAN (Wi-

Fi). Locations of evacuees are estimated by mobile

devices using the positional relationship between the

device and the wireless base station. The evacua-

tion route and timing of each evacuee are calculated

by distributed processing using the evacuees’ devices

connected by an ad-hoc network. The framework

of the Distributed Constraint Optimization Problem

(DCOP) is used in order to solve the problem. Figure

1 shows the concept of the system.

We assume the use of this system on a university

campus. University campuses have various unique

features and issues when considering disaster preven-

tion. For example, it is difﬁcult to determine how

many people there are on campuses. There will be

many students studying in libraries, spare rooms, or

Agent

Destination

Evacuation zone

Figure 2: Model of evacuation zone.

cafeterias. Visitors can also use the open spaces. In

addition, administrative structural issues exist, which

are different from those in companies. University

campuses are a prime example of places that require

Mutual-help.

3 FORMALIZATION OF

EVACUATION

In this section we formalize the problem. The evacu-

ation covered in this paper is local adaptation evacu-

ation; we do not consider the planned evacuation of a

wide area

3.1 Formalization of Evacuation

In order to use the framework of the DCOP, it is

necessary formalize the disaster evacuation problem.

Thus, we considered the formalization as follows.

Evacuees with a mobile device are considered to

be agent A = {a

, ..., a

P = ∪{p

, ..., p

} are the places from which peo-

ple must evacuate.

When agent a

is located in time t at place p

, it is

written as place(a

,t) = p

. L(t) is the set of places

of an agent who has not evacuated.

L(t) := ∪{place(a

,t) | place(a

,t) ∈ P } (1)

Agent a

is assumed to be capable of detecting the

number of nearby agents nr(a

At the time of a disaster, the situation is ﬂuid. Peo-

ple may be unable to pass along a passage.

All places P

have a score of val(P

) ∈ N. This

score is assumed to decrease toward the refuge direc-

tion. It is assumed that the agent can search the evac-

uation routes, and calculate the scores for the areas

on the routes. In this paper, it is assumed that all the

agents’ score allocation is the same for simpliﬁcation.

The above modeling can express P as a graph as

shown in Figure 2. At this time, utility function f of

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

224

Estimate of situation

planning

move

Time t

Using DCOP

Estimate of situation

planning

move

Time t+1

Estimate of situation

planning

move

Time t-1

Using DCOP

Figure 3: Real-time planning problem.

the entire agent is as follows:

f(A,L ,t) :=

∑

|place(a

,t)∈P }

val(place(a

,t))

nr(a

)

(2)

The numerator expresses the desire to evacuate, and

the denominator is a constraint to avoid congestion.

In this study, evacuation is considered to be a real-

time planning problem (shown in Figure 3.) The po-

sition after moving of the agent in time t + 1 may not

be the position that the system computed in the time t.

An agent may be unable to move according to the sit-

uation which a system cannot know, or an agent may

not follow the guidance of a system.

That is, the evacuation problem at time t is to solve

the following expression.

arg min

( f(A, L,t + 1) − f(A, L,t)) (3)

The above f(A,L ,t + 1) − f(A, L,t) is negative. By

solving this expression, the agent knows the appro-

priate evacuation timing. To solve this, we adopt the

DCOP framework.

It is necessary to consider the following conditions

further:

1. Short time to complete evacuation

2. Short continuous wait time of each agent

3. Fairness of the agent

3.2 Evacuation Problem as Distributed

Constraint Optimization Problem

The Distributed Constraint Optimization Problems

(DCOP) are the fundamental framework in dis-

tributed artiﬁcial intelligence and have recently at-

tracted considerable attention (Yokoo and Hirayama,

2000)D Algorithms used to solve DCOP include

ADOPT (Modi et al., 2005)C OptAPO (Mailler and

Lesser, 2004), DPOP (Petcu and Faltings, 2005)C

NCBB (Chechetka and Sycara, 2006), and distributed

stochastic search algorithms (DSA) (Zhang et al.,

2005). As for a complete algorithm, an optimum

solution is guaranteed, despite the extended comput-

ing time. When using DCOP for real-world prob-

lems, particularly when solving problems involv-

ing robotics and sensor networks, problems must

be solved in distributed environments with mini-

mal computationresources (Fitzpatrick and Meertens,

2001)(Zhang et al., 2005). Under such circumstances,

seeking an optimum solution with a complete algo-

rithm is not always the best method, and there is a

need for a fast and efﬁcient approximation algorithm.

The Distributed Constraint Optimization Problem

(DCOP) is deﬁned as a tuple < A, X, D, F > (Modi

et al., 2005)(Petcu and Faltings, 2005)D A set of vari-

ables X = {x

, x

, ..., x

} exists, each of which is as-

signed a value taken from a ﬁnite and discrete domain

D = {D

, D

, ..., D

}, and each of which is also as-

signed to multiple agents A = {a

, a

, ..., a

}. (n is

handled as n = m here for simpliﬁcation purposes.)

Constraint function F = { f

×D

→ R} is deﬁned

between x

and x

. The agent a

only has the follow-

ing information: information about x

, which is as-

signed to a

, and the cost function f

k∗

. In this case,

the purpose of DCOP is to obtain an assignment for

variable A that minimizes the summation of the cost

function F (A) =

∑

(A) In DCOP, an assignment

that offers the minimum F (A

) amongst all pos-

sible assignments A is deﬁned as the optimum solu-

tion. In DCOP, agents solve problems by exchanging

values of the variable through message transmission

with other agents whose variable are associated by

constraints. The framework of DCOP does not need a

center server in order to solve problems.

When treated as a distributed constraint optimiza-

tion problem, the evacuation problem can be consid-

ered as follows.

An agent has a variable to store the place to which

the agent should move. And the agent decides the po-

sition of the time t + 1 using DCOP in time t. This is

continuously performed as a real-time planning prob-

lem (Figure 3). The situation conﬁrmation for every

step is indispensable. At this time, minimization of

Formula(2), i.e., minimization of Formula(3), is the

objective function. The denominator of Formula(2) is

the constraint of an agent’s move and can be described

by binomial constraint. The molecule of Formula(2)

becomes a unary constraint.

The above illustrate a simple application of

DCOP.

MultiagentApproachforEffectiveDisasterEvacuation

225

Figure 4: Multi-agent simulation of a disaster evacuation.

4 EXPERIMENT USING

MULTI-AGENT SIMULATION

A multi-agent simulation is often used for disaster

evacuation experiments, not only for disaster preven-

tion planning, but also building or city planning. We

conducted an experiment using multi-agent simula-

tion in order to investigate the validity of the proposed

system. In the experiment, we also investigated the

inﬂuence of an approximation algorithm to solve the

problem.

4.1 Experimental Conditions

We set up a 2-story school building with six rooms

for the experimental conditions. It was assumed that

the evacuation route was limited to only one accord-

ing to the disaster. The agents shall move rationally to

an evacuation place and agents move according to the

model of the crowd walking (Kaneda and Okayama,

2007). The evacuee’s psychological model is not used

in this experiment; the agent shall follow guidance if

guidance is available. The number of agents was set

to 400. The location of the classrooms and the simu-

lation image are shown in Figure 4.

Evacuation is started from the situation where

people are randomly distributed throughout the class-

rooms and passages. The evacuation guidance in this

experiment negotiates only the evacuation start timing

from each classroom by DCOP. If the population den-

sity of a passage becomes below the threshold value,

evacuation will be started from any one classroom on

each ﬂoor. DCOP shall be used for the selection of

the classroom that may start evacuation. In this ex-

periment, in order to solve DCOP, the approximation

algorithm DSA (Zhang et al., 2005) was used. It is

important to solve the problem in a short time in an

urgent situation such as disaster evacuation. In such

a case, an approximation algorithm is more suitable

250 254 258 262 266 270 274 278 282 286 290 294 298 302 306 310 314 318 322

Frequenc

Time(seconds)

Without guidance

With guidance

Figure 5: The frequency distribution of the evacuation com-

pletion time (with/without guidance).

250 254 258 262 266 270 274 278 282 286 290 294 298 302

Frequency

Time(seconds)

approximate solution

strict solution

Figure 6: The frequency distribution of the evacuation com-

pletion time (approximate solution/strict solution).

than complete algorithm. DSA is a randomized algo-

rithm.

4.2 Experimental Results

We compared evacuation completion time in cases

with and without evacuation guidance. If there is no

evacuation guidance, evacuation will be simultane-

ously started from all the classrooms, but if there is

evacuation guidance, evacuation will be started from

one classroom.

The simulation was performed 300 times under

this condition, and the frequency distribution of the

evacuation completion time of the result is shown in

Figure 5. The evacuation completion time was plot-

ted on the X-axis, and frequency was plotted on the

Y-axis. In this experimental result, when there was

evacuation guidance by DCOP, evacuation comple-

tion time decreased by about 10%. This effect should

change according to conditions such as the location of

the classrooms, width of the passage, etc.

The approximation algorithm DSA was used for

this experiment. For this reason, although evacua-

tion should have been ideally performed from any one

classroom of each ﬂoor, evacuation might have been

simultaneously performed from two classrooms. This

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

226

is a case where a strict solution cannot be found by

DSA.

The case where evacuationis performed only from

one classroom is called a strict solution. The fre-

quency distribution of only a strict solution is shown

in Figure 6. In the ﬁgure, the case of the approximate

solution was also plotted for comparison. Although a

signiﬁcant difference was observed in these two as a

result of the t-test, the difference was very slight. In

this experiment, there were only three classrooms on

one ﬂoor, so the difference between the strict solution

and the approximate solution might be small.

5 RELATED WORK

(Lass et al., 2008) discussed the application of DCOP

to coordination in a disaster management situation.

Authorities must assign tasks and resources in dis-

aster scenarios; unfortunately accomplishing this in

real time is currently difﬁcult. They argue the frame-

work of DCOP is uniquely suited to meet the require-

ments imposed of coordination mechanisms in these

settings. (Nguyen et al., 2012) extended DCOP to

Stochastic DCOP(SDCOP) in order apply it to dis-

aster management proposed by (Lass et al., 2008)D

In SDCOP, the constraint rewards are deterministic

values but are sampled from known probability dis-

tribution function called reward functions. And they

proposed an algorithm that solves SDCOP. These re-

searches are specialized in the resource (or shelter) as-

signment problem at the time of a disaster, and have

not made reference to evacuation guidance.

Evacuation problems can be modeled in dynamic

network ﬂows (Hamacher and Tjandra, 2002)D The

standard approach to solving dynamic ﬂow problems

is to transform the graph into a time-expanded net-

work. However, the expanded graph is larger.The

major computational bottlenecks are the time and

memory required to construct the expanded network.

Some heuristic algorithms have been proposed for

this problem (Hamacher and Tjandra, 2002)D (Lu

et al., 2005) considered capacity constrained routing

heuristics. (Hadzic et al., 2011) considered the prob-

lem of planning evacuation routes in deteriorating net-

works, where nodes become unavailable over time. In

these researchs of heuristic algorithms, it is assumed

that problem solving is performed in a non-distributed

environment.

There are also some related works concerning dis-

aster evacuation simulation systems using multi-agent

simulation (Burstedde et al., 2001) (Helbing et al.,

2000)(Shi et al., 2009).

6 CONCLUSIONS

We aim to develop a the system that provides optimal

evacuation guidance autonomously at the time of a

disaster. This system enables assistance to be given in

the form of evacuation guidance to relievecongestion,

by calculating evacuation routes and timing via an ad-

hoc network of evacuees’ mobile devices, without a

center server.

In this paper, the problem of disaster evacua-

tion was formalized and how to solve it using the

framework of the Distributed Constraint Optimiza-

tion Problem was examined. In the experiment us-

ing a multi-agent simulation, when evacuation guid-

ance using DCOP existed, the evacuation completion

time decreased by about 10%. And even when it was

solved with an approximation algorithm, the effect on

the evacuation completion time was small.

As the formalization in this paper is very simple, it

is necessary to introduce a more realistic model. The

fairness of an agent’s latency time, and minimizing

the total evacuation completion time should also be

written as a constraint. The fairness of latency time

will be realizable if an agent’s continuation latency

time is minimized. Therefore, it is necessary to write

the continuation latency time as a unary constraint.

Moreover, it is necessary to examine the guidance

method itself to ensure that people are able to follow

the guidance.

ACKNOWLEDGEMENTS

This work was supported in part by a JSPS Grant-in-

Aid for Scientiﬁc Research (25350481).

REFERENCES

Burstedde, C., Klauck, K., Schadschneider, A., and Zittartz,

J. (2001). Simulation of pedestrian dynamics using a

two-dimensional cellular automaton. Physica A: Sta-

tistical Mechanics and its Applications, 295(3):507–

525.

CabinetOfﬁce (2011). Disaster Management In Japan.

Government of Japan.

Chechetka, A. and Sycara, K. (2006). No-commitment

branch and bound search for distributed constraint op-

timization. In Proceedings of the ﬁfth international

joint conference on Autonomous agents and multia-

gent systems, pages 1427–1429. ACM.

Fitzpatrick, S. and Meertens, L. (2001). An experimen-

tal assessment of a stochastic, anytime, decentralized,

soft colourer for sparse graphs. In 1st Symposium on

Stochastic Algorithms: Foundations and Applications,

pages 49–64.

MultiagentApproachforEffectiveDisasterEvacuation

227

Hadzic, T., Brown, K. N., and Sreenan, C. J. (2011). Real-

time pedestrian evacuation planning during emer-

gency. In Tools with Artiﬁcial Intelligence (ICTAI),

2011 23rd IEEE International Conference on, pages

597–604. IEEE.

Hamacher, H. W. and Tjandra, S. A. (2002). Mathemat-

ical modelling of evacuation problems–a state of the

art. Pedestrian and evacuation dynamics, 2002(227-

266):1–2.

Helbing, D., Farkas, I., and Vicsek, T. (2000). Simu-

lating dynamical features of escape panic. Nature,

407(6803):487–490.

Iizuka, K., Iizuka, Y., and Yoshida, K. (2011). A real-time

disaster situation mapping system for university cam-

puses. In Ozok, A. and Zaphiris, P., editors, Online

Communities and Social Computing, volume 6778

of Lecture Notes in Computer Science, pages 40–49.

Springer Berlin / Heidelberg.

Kaneda, T. and Okayama, D. (2007). A pedestrian agent

model using relative coordinate systems. In Agent-

Based Approaches in Economic and Social Complex

Systems IV, pages 63–70. Springer.

Lass, R. N., Kopena, J. B., Sultanik, E. A., Nguyen, D. N.,

Dugan, C. P., Modi, P. J., and Regli, W. C. (2008).

Coordination of ﬁrst responders under communication

and resource constraints. In Proceedings of the 7th

international joint conference on Autonomous agents

and multiagent systems-Volume 3, pages 1409–1412.

International Foundation for Autonomous Agents and

Multiagent Systems.

Lu, Q., George, B., and Shekhar, S. (2005). Capacity con-

strained routing algorithms for evacuation planning:

A summary of results. In Advances in spatial and tem-

poral databases, pages 291–307. Springer.

Mailler, R. and Lesser, V. (2004). Solving distributed con-

straint optimization problems using cooperative me-

diation. In in Proceedings of the Third International

Joint Conference on Autonomous Agents and Multia-

gent Systems, pages 438–445. IEEE Computer Soci-

ety.

Modi, P. J., Shen, W.-M., Tambe, M., and Yokoo, M.

(2005). Adopt: asynchronous distributed constraint

optimization with quality guarantees. Artif. Intell.,

(161):149–180.

Nguyen, D. T., Yeoh, W., and Lau, H. C. (2012). Stochas-

tic dominance in stochastic dcops for risk-sensitive

applications. In Proceedings of the 11th Interna-

tional Conference on Autonomous Agents and Multia-

gent Systems-Volume 1, pages 257–264. International

Foundation for Autonomous Agents and Multiagent

Systems.

Petcu, A. and Faltings, B. (2005). A scalable method for

multiagent constraint optimization. In IJCAI, vol-

ume 5, pages 266–271.

Shi, J., Ren, A., and Chen, C. (2009). Agent-based evacua-

tion model of large public buildings under ﬁre condi-

tions. Automation in Construction, 18(3):338–347.

Yokoo, M. and Hirayama, K. (2000). Algorithms for dis-

tributed constraint satisfaction: A review. In Au-

tonomous Agents and Multi-Agent Systems, volume 3,

pages 198–212.

Zhang, W., Wang, G., Xing, Z., and Wittenburg, L. (2005).

Distributed stochastic search and distributed breakout:

properties, comparison and applications to constraint

optimization problems in sensor networks. Artif. In-

tell., (161):55–87.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

228