A Guidance System for Wide-area Complex Disaster Evacuation

based on Ant Colony Optimization

Hirotaka Goto

, Asuka Ohta

, Tomofumi Matsuzawa

, Munehiro Takimoto

Yasushi Kambayashi

and Masayuki Takeda

Department of Information Sciences, Tokyo University of Science, Chiba, Japan

Department of Computer & Information Engineering, Nippon Institute of Technology, Saitama, Japan

Keywords:

Ant Colony Optimization, Route Guidance System, Swarm Intelligence, Disaster Simulation, Seismic

Disaster.

Abstract:

This paper reports the results of applying our approach discovering safe evacuation routes to practical situa-

tions. Our approach is based on the ant colony optimization (ACO) and it is practical in the light of a real case

with a tsunami. ACO have been often employed for ﬁnding evacuation routes in traditional approaches, which

only take advantage of ants’ behavior more frequently following traces of other ants through pheromone com-

munications. We assume that there are a lot of danger zones in the damaged area. For example Rikuzentakata

is a city that extensively damaged in the 2011 Great East Japan Earthquake. In such a case, the traditional

approaches may present some unsafe routes through the danger zones. We have proposed an ACO based

approach that calculates evacuation routes avoiding danger zones. In our approach, evacuees can deposit de-

odorant pheromone around danger zones, which makes normal pheromone ineffective, so that our approach

gives routes not passing through the danger zones. We have implemented our approach as a simulator, con-

ducting experiments in the same situation as the Rikuzentakata case. Through the results of the experiments,

we show that our approach decreases the number of people suffering from collapsed and burning buildings.

1 INTRODUCTION

The Great East Japan Earthquake in 2011 was

recorded as an earthquake with magnitude 9.0. The

earthquake destroyed all the anti-tsunami structures

such as breakwaters and tide embankments, so that

the tsunami following the earthquake caused enor-

mous damage to the coastal areas. The damage that

people suffer from such a wide-area disaster can be

classiﬁed into three patterns: the direct disaster casu-

alties, the damage caused by collapsing of structures

and ﬁre during evacuation, and the damage caused

by the disastrous tsunami. In general, these dam-

ages do not occur simultaneously. In most cases, peo-

ple saved from the ﬁrst damage have to escape from

the third damage while avoiding the second damage.

Therefore, ideally, evacuation to safe areas should be

completed before the third attack where the tsunami

strikes. However, several zones damaged by an earth-

quake restrict the number of evacuation routes, so that

it becomes difﬁcult for people to evacuate. Thus, it is

extremely important to dynamically ﬁnd safe evacua-

tion routes depending on the situation where a wide-

area disaster occurs.

In this paper, we propose an algorithm of discov-

ering safe evacuations based on the ant colony opti-

mization (ACO) algorithm (Dorigo et al., 1996) ex-

tended for avoiding danger zones, e.g., areas with ﬁre

and other damages. Traditional approaches based on

ACO presents the shortest routes to safe areas depend-

ing on strength of pheromone, which may include the

danger zones. In our approach, we introduce a new

repulsive pheromone called deodorant pheromone to

ACO (Ohta et al., 2016).

In order to quantitatively evaluate the algorithm

in realistic situations, we present complex disaster

scenarios such as tsunamis. In the numerical exper-

iments, we show that our evacuation system can ﬁnd

effective routes that decrease the number of people

suffering from collapsed or burning structures. Also,

we discuss the limitations of our algorithm.

The structure of the balance of this paper is as fol-

lows. In the second section, we describe the related

work. In the third section, we provide an overview

262

Goto, H., Ohta, A., Matsuzawa, T., Takimoto, M., Kambayashi, Y. and Takeda, M.

A Guidance System for Wide-area Complex Disaster Evacuation based on Ant Colony Optimization.

DOI: 10.5220/0005819502620268

In Proceedings of the 8th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2016) - Volume 1, pages 262-268

ISBN: 978-989-758-172-4

of the ACO and discuss issues that we need to con-

sider during an evacuation. In the fourth section, we

describe the simulation model. In the ﬁfth section,

we discuss the results of the experiments. In the sixth

section, we discuss the other scenario. In the seventh

section, we describe a future work and we conclude

our discussion.

2 RELATED WORKS

Disaster is known to make several infrastructures un-

available, in which communication infrastructures are

also included. Once access points for Wi-Fi become

unavailable, we cannot take advantage of mobile de-

vices to collect useful information for our evacuation.

However, even in the cases, we may be able to con-

struct a mobile ad-hoc network through connecting

mobile devices one another, which may be effective

for sharing the information. In previous works that

deal with evacuation assistance, there are a lot of ap-

proaches assuming the ad-hoc network. Especially,

ACO based approaches are effective to detect evacu-

ation routes on the ad-hoc network. Avil´es et al. pro-

posed an approach for sharing information of evacu-

ation gates on the ad hoc network. In their approach,

ants and pheromone of ACO are implemented as soft-

ware mobile agents, where the agents corresponding

to ants dynamically guide evacuees to the evacuation

gates in a ﬂoor (Avil´es et al., 2014).

Asakura et al. proposed an approach that calcu-

lates evacuation routes based on ACO on a simula-

tor, and showed the effectiveness of applying their

approach to wide-scale disaster areas (Asakura et al.,

2013a)(Asakura et al., 2013b).

On the other hand, Mas et al. applied their

simulation-based approach, which did not use ACO,

to more practical case of the Great East Japan Earth-

quake, and showed that the shortest evacuation routes

were detected (Mas et al., 2012).

These previous works highlight just one of two is-

sues important for evacuation support systems, which

are consideration of the secondary disaster and reality

of assumed scenario. Our reports in this paper provide

not only remediation for the evacuation in the disaster

but also experimental results based on realistic sce-

narios to show the effectiveness of our approach.

3 EXTENDED ANT COLONY

OPTIMIZATION

This section explains the details of deodorant

pheromone mentioned in Section 1 by extending the

basic algorithm corresponding to traditional ACO,

where the deodorant pheromone suppresses the effect

of traditional pheromone in our extended ACO. Also,

we describe how the deodorant pheromone is used to

avoid danger zones.

3.1 Basic Algorithm

During foraging, real ants secrete a volatile chemi-

cal substance called pheromone that encourages other

ants to behave cooperatively. Once an ant discov-

ers food, it brings the food to the nest. In this pro-

cess, it put pheromone on the ground to provide sign-

posts for following ants. As well, the ants carrying

the food along the pheromone also put pheromone

on the same ground, strengthening the effectiveness

of attraction. The strengthening of pheromone re-

sults in some routes between the nest and the food.

Conversely, pheromone on the ground unused by the

ants gradually decreases the density of pheromone by

evaporation. Thus, the accumulation of pheromone

along restricted routs causes positive feedback, so

that ants can ﬁnd the shortest route among multiple

routes with different lengths. If pheromone informa-

tion were not available, ants would forage by moving

at random.

Figure 1: Route generation by the ACO. S and G denote the

starting point and destination point, respectively.

Figure 1 illustrates a route established by foraging

ants. Consider that three ants explore on three routes

with different distances between a starting point S

and a destination point G. First, the ants simultane-

ously explore from S to G. Once the ants reach a

branch, they select their own directions randomly. If

all the ants select the middle route, they will arrive

at the destination at the same time. After that, in the

process where the ants return to S, they deposit their

pheromone. If they return to G again, they can use the

previously deposited pheromone as guidance. Notice

here that the ants preferentially select the middle

route, because the density of the pheromone on the

other routes is decreased by evaporation. In this way,

the ACO ﬁnds the shortest route. Summarizing the

A Guidance System for Wide-area Complex Disaster Evacuation based on Ant Colony Optimization

263

steps of ACO, they are as follows:

1. Ants seek a route between the starting point

(Start) and destination point (Goal).

2. Ants locally explore a branch along the route.

3. Ants secrete pheromone on the ground while they

travel their routes.

4. Pheromone evaporates at a constant rate.

5. The above steps are repeated through the speciﬁed

number of iterations.

3.2 Extended Pheromone Behaviors

Assume that we simply search evacuation routes

based on the algorithm of Section 3.1 in the event of a

disaster. In this case, even if the suggested evacuation

route is the approximately shortest route, it may in-

clude or touch danger zones. In order to avoid danger

zones, we have introduced deodorant pheromone that

decreases the pheromone laid around danger zones.

Deodorant pheromone erases the traces of previously

deposited pheromone and attenuates the pheromone

in surrounding regions, and imparts new information

about danger zones.

In our ACO, normal pheromone contributes to

constructing shorter evacuation routes, and deodor-

ant pheromone contributes to adjusting the evacua-

tion routes to avoid danger zones. Thus, the two kinds

of pheromone cooperativelyconstruct safe evacuation

routes for evacuees.

3.2.1 Normal Pheromone

The normal pheromone is applied along the passable

evacuation route. Once an evacuee arrives at a safe

area, the normal pheromone is updated by Eq. (1),

where τ

(t) is the pheromone value at coordinates

(i, j) at time t, and G

denotes the evacuees who ar-

rived at the safe area in time t.

(t + 1) = (1− ρ)τ

(t) +

∑

k∈G

∆τ

(1)

(t) is decreased by an evaporation rate ρ, and is

increased by ∆τ

for evacuee k that has reached the

safe area, step by step.

∆τ

is determined by α and T

, where α is the

amount of pheromone applied at coordinates (i, j),

and T

denotes the trafﬁc when the evacuee k has

reached the safe area as follows:

∆τ

(t) =



α if (i, j) ∈ T

0 otherwise

(2)

(t) is limited by upper and lower bounds, τ

max

and τ

min

well as traditional ACOs based on MAX-

MIN ant system (Sttzle and Hoos, 2000). The up-

per bound prevents solutions from being trapped in

local minimums, which is one of characteristics of lo-

cal neighborhood searching. The lower bound per-

mits searching of any solutions.

0 < τ

min

< τ

(t) ≤ τ

max

(3)

In the case where the normal pheromone is only used,

our ACO works in the same way as traditional ACOs.

We call it normal ACO (nACO).

3.2.2 Deodorant Pheromone

The deodorant pheromone, which exerts a repul-

sive force on evacuees, causes normal pheromone to

be updated whenever an evacuee encounters a dan-

ger zone. At this time, the density of the normal

pheromone is decreased.

Also, the deodorant pheromone decreases the

density of the pheromone not only at the center of

the danger zone but also at the surrounding area.

The update of normal pheromone in a certain range

including a danger zone is performed based on the

following equation:

(t + 1) = (1− σ

(t)+1

)τ

(t) (4)

Where σ is the rate of the deodorant, and n

(t) is the

distance from the coordinate (i, j) marked as a dan-

ger zone by evacuee k. Under Eq. (4), the deodorant

pheromone spreads only within a certain range.

In the simulation, we limit the spread of the de-

odorant pheromone as n

(t) ≤ N.

We call our ACO that includes both normal and

deodorant pheromones the extended ACO (eACO).

4 SIMULATION MODEL

4.1 Disaster Scenario

This section shows the scenario that we assume in

our simulation. The scenario is based on the situ-

ation of the city of Rikuzentakata, which was ex-

tensively damaged in the 2011 Great East Japan

Earthquake. The peak inundation and evacuation

locations in Rikuzentakata have been published in

the Rikuzentakata Earthquake Veriﬁcation Report

(Rikuzentakata, 2014). In addition, the onset times of

the earthquake and tsunami as well as the peak time of

ICAART 2016 - 8th International Conference on Agents and Artiﬁcial Intelligence

264

the inundation have been summarized in chronolog-

ical order (Ushiyama and Yokomaku, 2012). From

these references, we have determined the time al-

lowed for evacuation is approximately 45 minutes af-

ter the earthquake occurs. Then the ﬂooding gets to

be maximized. Table 1 shows the time schedule of

each event.

Table 1: Timing of each event.

Time Event

14:46 Earthquake occurs

15:24 Inundation start

15:29 Inundation peak

Figure 2 chronologically shows the situations of

tsunami. By using the information and assumptions,

we have succeeded in reproducing the ﬂood situa-

tions of tsunami induced by the earthquake in the city

of Rikuzentakata. However, refugees must have also

encountered ﬁre and collapse of rubble as secondary

disasters at several locations during their evacuation,

though the details of them were not recorded. We sim-

ply assume those secondary disasters occur randomly.

Figure 2: Chronological order of the tsunami damage situ-

ation.

As mentioned in Section 4.1, we assume that the

evacuation time is restricted to the period from the

earthquake outbreak to arrival of the tsunami. The

start and end of the simulations are depicted in Fig-

ures 3 and 4, respectively. Evacuation activities are

simulated by a multi-agent simulator in which the

agents correspond to people. In the simulation, the

city of Rikuzentakata is mapped onto a 200 200 grid

(Figure 3). Each grid cell denotes a passable area,

an impassable area, a safe area or a danger zone.

Once the simulation starts and the earthquake breaks

out, the simulator generates evacuees randomly in the

passable areas on the map. After that, evacuees deter-

mine their evacuation routes based on local informa-

tion along ACO based guidance.

4.2 Simulator Overview

Each evacuee obtains pheromone information from

the eight cells surrounding the cell he or she resides,

determining his or her next move along the guidance

of the pheromone information.

In each simulation step, the evacuees move from

their current cells to one of the eight surrounding the

Figure 3: Start of the simulation.

Figure 4: End of the simulation. Red squares mark danger

zones.

cells. At this time, the direction where to move is

stochastically determined by the pheromone informa-

tion described in Section 3.2. The probability p

(t)

of an evacuee’s moving to (x, y) at time t is given by

the following equation:

(t) =

(t)

∑

(i, j)∈X

(t)

(5)

Where X

(t) indicates the movable locations sur-

rounding the evacuee k. Equation (5) probabilistically

promotes the movement toward neighbor cells with

relatively high pheromone values.

Sometimes, the guidance may give undesirable

suggestion such as going towards a danger zone. In

such cases, the evacuee stop moving at that time, and

then, deposits the deodorant pheromone to mark it for

other evacuees. Once the evacuee reaches a safe area,

pheromone information put along his trace is reset.

A Guidance System for Wide-area Complex Disaster Evacuation based on Ant Colony Optimization

265

Table 2: Simulation parameters.

Parameter Value

Number of evacuees 1000

Evaporation rate ρ 0.0005

Amount of pheromone adding α 1.0

Upper bound of pheromone τ

max

30.0

Lower bound of pheromone τ

min

1.0

Constant of deodorant pheromone

′

-50.0

Inﬂuence range of deodorant

pheromone N

Deodorant rate σ 0.5

Figure 5: The number of evacuees caught in danger zones

versus the number of victims.

Figure 6: The number of evacuees involved in the tsunami

versus the number of victims.

5 EXPERIMENTAL RESULTS

This section compares the simulation results of the

cases using nACO and eACO.

The simulation used the settings listed in Table 2,

and randomly located 10 to 25 danger zones (incre-

mented by ﬁve). Each experiment was run 500 times,

and the mean values are plotted. In the following ﬁg-

ures, the solid and dashed lines indicate the results of

the cases using nACO and eACO, respectively.

Figure 7: The number of approaches of evacuees to danger

zones, as a function of simulation time.

5.1 Results

Figure 5 shows the relationship between the number

of danger zones and the number of victims encounter-

ing danger zones. We have observed that more evac-

uees are entrapped in the case for nACO than eACO.

Figure 6 shows the relationship between the num-

ber of danger zones and the number of victims en-

countering tsunami. In this case, the number of en-

dangered evacuees is higher in the case for eACO than

nACO.

Figure 7 plots the cumulativenumber of times that

evacuees approach the danger zones. Overall, us-

ing the eACO decreases the number of evacuees ap-

proaching the danger zones.

5.2 Discussion

Although the number of at-risk evacuees increases as

the number of danger zones increases in both algo-

rithms, we can observe that the number of those in the

case for eACO smaller than nACO. This reﬂects the

effect of the deodorant pheromone that repels agents

from danger zones and their vicinities. However, this

beneﬁt was partially offset by the increased number

of evacuees affected by the tsunami (Figure 6). The

deodorant pheromone in eACO enforces a detour-like

evacuation activity on evacuees and have them expo-

sure to the tsunami risk.

We therefore investigate how much duration time

the evacuees need to escape the tsunami through safe

routes while avoiding dangerous area by using eACO

in the next experiment.

6 ADDITIONAL SCENARIO

We have already experienced another devastating

tsunami in 2004 that followed an earthquake off the

ICAART 2016 - 8th International Conference on Agents and Artiﬁcial Intelligence

266

coast of Sumatra, which was recorded as magnitude

9.1. In this disaster, there was a relatively long lapse

of time between the outbreak of the earthquake and

the arrival of the tsunami. Given the pattern of the

earthquake, we can vary the timing of the tsunami at-

tack. In this experiment, we have changed the lapses

of time between the outbreak of the earthquake and

arrivals of the tsunami whereas the other parameters

is left unchanged.

Figure 8: The number of evacuees involved in the tsunami,

versus delay time of the tsunami.

6.1 Results

Figure 8 plots the number of evacuees involved in the

tsunami, versus the delay time of the tsunami. If there

is a time lapse of approximately 20 minutes, the num-

ber of tsunami victims became smaller in the case for

eACO than nACO.

6.2 Discussion

Our eACO algorithm is beneﬁcial when there is suf-

ﬁcient duration time for escape between the outbreak

of an earthquake and the following tsunami. How-

ever, neither ACO is effective in a case that immedi-

ate evacuation is needed, because considerable time

is required to construct safe evacuation routes. On

the other hand, the eACO is applicable to many real

scenarios in which evacuees can ﬂee dangerous situ-

ations, such as building collapses and ﬁre, before a

tsunami strikes.

7 CONCLUSIONS

This paper proposes an evacuation support system and

analyzes its effectiveness by simulation experiments

of a case of the wide-scale disaster area in the city of

Rikuzentakata. The city was extensively damaged by

the 2011 tsunami.

The proposed evacuation system that implements

our approach extends the ant colony optimization

method (ACO), which allows agents to share their

routing information after reaching safe areas. In the

model, evacuees take the routes traveled in previous

evacuation activities while they avoid dangerous area

by using the information given by preceding evac-

uees.

Comparing with the normal ACO, our ACO model

decreases the numbers of evacuees caught in danger

zones. The result indicates that our approach assists

to construct safe evacuation routes that bypass dan-

gerous zones. However, it increases the number of

evacuees who were caught in the tsunami. We can ob-

serve that our method is beneﬁcial when there is suf-

ﬁcient time lapse between the outbreak of the earth-

quake and the arrival of the tsunami. In that case, the

additional deodorantpheromonerepels evacueesfrom

danger zones, while guiding them away from the in-

undation of the tsunami.

As a future work, we plan to evaluate and gener-

alize the various experimental parameters. We also

need to investigate how to disperse the escape routes

for mitigating the trafﬁc congestion of the evacuees. It

is necessary because in actual evacuation scenes peo-

ple may concentrate on a particular path toward safe

areas and hinder the evacuationactivities. The present

evacuation guidance system focuses on the direction

of movement. To simulate precise evacuation behav-

ior, the agents must undertake more human-like intel-

ligent actions.

ACKNOWLEDGEMENTS

This work is supported in part by Japan Society for

Promotion of Science (JSPS), with the basic research

program (C) (No. 25330089 and 26350456), Grant-

in-Aid for Scientiﬁc Research.

REFERENCES

Asakura, K., Fukaya, K., and Watanabe, T. (2013a). Con-

struction of navigational maps for evacuees in disaster

areas based on ant colony systems. International Jour-

nal of Knowledge and Web Intelligence, 4:300–313.

Asakura, K., Fukaya, K., and Watanabe, T. (2013b). A map

construction system for disaster areas based on ant

colony systems. Procedia Computer Science, 22:494–

501. 17th International Conference in Knowledge

Based and Intelligent Information and Engineering

Systems.

Avil´es, A., Takimoto, M., and Kambayashi, Y. (2014).

Distributed evacuation route planning using mobile

A Guidance System for Wide-area Complex Disaster Evacuation based on Ant Colony Optimization

267

agents. In Transactions on Computational Collective

Intelligence XVII, volume 8790, pages 128–144.

Dorigo, M., Maniezzo, V., and Colorni, A. (1996). Ant sys-

tem: optimization by a colony of cooperating agents.

IEEE Transactions on Systems, 26(1):29–41.

Mas, E., Suppasri, A., Imamura, F., and Koshimura, S.

(2012). Agent-based simulation of the 2011 great east

japan earthquake/tsunami evacuation: An integrated

model of tsunami inundation and evacuation. Journal

of Natural Disaster Science, 34(1):41–57.

Ohta, A., Goto, H., Matsuzawa, T., Takimoto, M., Kam-

bayashi, Y., and Takeda, M. (2016). An improved

evacuation guidance system based on ant colony op-

timization. In Intelligent and Evolutionary Systems,

volume 5 of Proceedings in Adaptation, Learning and

Optimization, pages 15–27.

Rikuzentakata (2014). The city of rikuzentakata the great

east japan earthquake veriﬁcation report: Rikuzen-

takatashi higashi nihon daishinsai kenshou houkoku

sho (in japanese).

Sttzle, T. and Hoos, H. H. (2000). Maxmin ant system. Fu-

ture Generation Computer Systems, 16(8):889–914.

Ushiyama, M. and Yokomaku, S. (2012). Estimation of sit-

uation in rikuzentakata city just before tsunami atack

based on time stamp data. Japan Society for natural

disaster science, 31(1):47–58.

ICAART 2016 - 8th International Conference on Agents and Artiﬁcial Intelligence

268