An Artiﬁcial Immune Approach for Optimizing Crowd Emergency

Evacuation Route Planning Problem

Mohd Nor Akmal Khalid and Umi Kalsom Yusof

School of Computer Sciences, University Sains Malaysia, 11800 Georgetown, Pulau Pinang, Malaysia

Keywords:

Emergency Evacuation, Emergency Route Planning, Immune Algorithm.

Abstract:

Disastrous situations, either naturally (such as ﬁres, earthquake, rising tides, hurricane) or man-made (such

as terrorist bombings, chemical spills, and so on), have claimed the lives of thousands. As such, optimizing

the evacuation operations during an emergency situation would require an effective crowd evacuation plan,

which is acknowledged to be one of the vital studies of the societal research as well as emergency route

planning (ERP) community. Several descriptions of prior developed approaches for emergency evacuation that

encompassed the needs of a variety of public community as well as fulﬁlling the complexity of the situation, are

summed up and discussed. This paper introduces an immune algorithm (IA) to optimize the evacuation plan for

solving the ERP problems. The approach is ﬁrst validated against previous work while further experimentation

reveals the effectiveness of the proposed IA, with regard to certain parameter calibrations, in the context of

ERP problems. The ﬁndings have been summarized and presented, whereas the potential for future work is

identiﬁed.

1 INTRODUCTION

Extreme events or disasters, be it natural or man-

made, often lead to emergency situations that requires

immediate action. Examples of natural disasters in-

clude hurricanes, ﬂoods, landslides, and tsunamis.

Examples of man-made disasters include terrorist at-

tacks and hazard material releases. These critical

events affect populated areas, inducing an immedi-

ate or life-threatening situation that triggers an emer-

gency response. In many cases, evacuation is the

common response to risk mitigation, requiring imme-

diate mobilization and time-critical actions, primar-

ily efﬁcient coordination, space capacity utilization,

and ensuring availability of emergency response re-

sources (Alsnih and Stopher, 2004). Thus, an emer-

gency evacuation can be deduced as the practical op-

tion for human survivability which is paramount in

risk mitigation.

Emergency evacuation involves collective re-

moval of residents/populations as quickly as possible

and with utmost reliability from areas considered as

dangerous zones to safe locations (Alsnih and Sto-

pher, 2004). The most disastrous form of collective

human behaviour are stampedes, which induced by

panic that often leads to serious fatalities (Hajibabai

et al., 2007). The ability to contribute towards an ef-

ﬁcient movement of people in heavily populated en-

closures or structures is vital to the daily operation of

large and complex structures (Hajibabai et al., 2007).

More importantly, it is an essential design feature in

the event of emergency situations.

To support emergency evacuation operation, the

crowd model is an essential tool in providing ef-

fective decision-making, enhancing the capability of

response to disaster, and reducing any adverse im-

pacts on both human beings and surroundings (Lv

et al., 2012). However, comprehending such crowd

activity and characteristics requires an appropriate

crowd model to improve evacuation plan efﬁciency

and crowd survivability (Wang et al., 2008).

There are four important factors that have pio-

neered the main foundation of the emergency route

planning (ERP) problem (Alsnih and Stopher, 2004):

(1) deciding where to evacuate people (goal); (2) de-

ciding the best routes to take (routing); (3) deter-

mining the rate of evacuees egress (ﬂow rate); and

(4) determining how to regulate ﬂow rates on these

routes (schedule). These decisions are methodologi-

cally and computationally challenging due to the fol-

lowing reasons; decision interdependence, simultane-

ous decision making, and concurrency (Alsnih and

Stopher, 2004). Therefore, effective ERP approaches

are needed in order to address these challenging is-

503

Nor Akmal Khalid M. and Kalsom Yusof U..

An Artiﬁcial Immune Approach for Optimizing Crowd Emergency Evacuation Route Planning Problem.

DOI: 10.5220/0005275305030508

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence (ICAART-2015), pages 503-508

ISBN: 978-989-758-074-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

sues faced during emergency evacuation.

The paper organization is presented as follows:

Section 2 will highlight the existing crowd evacua-

tion approaches in the context of ERP problems. Dis-

cussion on the proposed IA approach and its features

is given in Section 3. Section 4 validates and eval-

uates the proposed IA approach as well as analyzes

the computational result. Finally, Section 5 summa-

rizes the paper contribution and suggests the potential

future works.

2 LITERATURE REVIEWS

The traditional crowd evacuation solution simply con-

veys warning and threat descriptions where the need

for evacuation is issued via mass media communica-

tions to the affected population (Lu et al., 2003). The

traditional crowd evacuation solution which conveys

warning during an event of ﬁre within a structure and

allow evacuees to continue their occupancy and assist

in ﬁnding a safe exit. However, this solution does not

provide any information on how to escape and cannot

give clear insights into the situation after the evacu-

ation notice is elicited. Thus, lack of proper plan-

ning and management cause unanticipated effects on

crowds such as massive congestion, confusion, and

chaos. These include lacks of ﬂexibility, insufﬁcient

information, less intelligence, dynamic and/or current

information, and lack of means of providing interac-

tivity (Lu et al., 2003). Since then, various ERP ap-

proaches had been introduced for an efﬁcient evacua-

tion plan during extreme events.

Some studies have proposed prominent

mathematical-based approaches for solving the

ERP problems. Wang et al. (Wang et al., 2008)

had proposed a stochastic programming model for

evacuating crowd in a building network. Hui et al.

(Hui et al., 2010) had used a stochastic programming

model to allocate rescue route, which is generated

from particle swarm optimization algorithm. Lv et al.

(Lv et al., 2012) had designed an integer program-

ming model for supporting emergency management

under uncertainties. Sayyady et el. (Sayyady and

Eksioglu, 2010) had designed a model for transit-

dependent residents during a no-notice disaster using

a mixed-integer linear model incorporating Tabu

search to simultaneously optimize the emergency

response.

Heuristics are among the more popular and emerg-

ing ERP approaches, where most of them are well-

known in the ﬁeld of computational optimization. An

examples of the ERP approach based on heuristics

is the Capacity Constrained Route Planner (CCRP)

(Lu et al., 2003; Kim et al., 2007). Lu et al. (Lu

et al., 2003) had proposed CCRP approach to min-

imize evacuation time of evacuees while reducing

computational cost. Kim et al. (Kim et al., 2007) had

proposed CCRP approach which is focused on load

reduction and scalability of the route plan.

Studies employing meta-heuristic algorithms had

been proposed by several researchers. Cepolina (Ce-

polina, 2005) had proposed a simulated annealing ap-

proach for an evacuation plan. Li et al. (Li et al.,

2010) had proposed genetic algorithm incorporating

congestion as an important aspect for an effective

evacuation plan. Xie et al. (Xie et al., 2010) had

proposed a combination of Lagrangian relaxation and

Tabu search algorithms for a bi-level network op-

timization model, where lane reversal (contraﬂow)

and crossing elimination strategies are incorporated

to optimize the network at the upper level and a cell

transmission-based dynamic trafﬁc assignment on the

lower level.

Although several approaches have been intro-

duced, there is no speciﬁc approach that encapsulates

both crowd ﬂow regulation and optimizes their desig-

nated route simultaneously, which is vital for an ef-

fective and time-critical evacuation plan. An artiﬁ-

cial immune system (AIS) had offered a number of

profound features, capable of encapsulating the com-

plexity of the ERP problem. Among them are; the

ability to detect changes, learning and memory, adap-

tation, self-organized, scalability, robust, and decen-

tralization (Dasgupta et al., 2011). This algorithm

was introduced to model and apply immunological

principles of vertebrate into solving problems for a

wide range of areas such as optimization, data min-

ing, computer security, and robotics (Dasgupta et al.,

2011). Therefore, these attractive and salient features

of the AIS act as the primary motivation for adopting

it as the proposed ERP approach for optimizing the

evacuation plan.

3 THE PROPOSED IA FOR ERP

OPTIMIZATION

This section will provide detailed description of the

proposed immune algorithm approach for optimizing

the ERP problem. The organization of this section are

as follows: Section 3.1 highlights the proposed crowd

and its correlation with the IA optimization approach.

Section 3.2 describes the inspiration of the IA opti-

mization approach from the natural immune system.

Section 3.2.1 highlights the initialization scheme and

population instance of IA. Lastly, Section 3.2.2 de-

scribes the IA’s improvisation (mutation) mechanism,

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

504

speciﬁc for the ERP problem.

3.1 The Proposed Crowd Model

The crowd model adopted is the mesoscopic model,

where the relationship between local inter-individual

interactions of the evacuees (micro) and collective

patterns of the crowd (macro) are considered. One of

the efforts conducted was the work in (Wang et al.,

2008). Although the local inter-individual interac-

tions of the evacuees are captured, the collective pat-

terns of the crowd are vaguely deﬁned and a clear

boundary of the group is not raised. Therefore, the

ﬁrst aspect of this study will seek to get the apparent

group characteristic, namely, group cohesion. Group

cohesion can be deﬁned as the tendency for a group

to be in unity while working towards a goal or to ful-

ﬁll the demands of its members (Carron and Brawley,

2000). The size of the group may affect and prolong

the evacuation procedure, which potentially increases

the evacuee’s exposure to danger.

In this particular setting, the group size, with the

assumption of a full compliance level, is initialized.

The group is differentiated with an identiﬁcation g

which consists of evacuees k of different sizes d

where each of the groups are located at the start-

ing or source location η

similar or dissimilar from

the other groups. From this formulation, the evacuee

k is assumed to belong in a particular group g that

strongly cooperate (100% compliance of individuals

within the group) towards achieving their goal (i.e.,

escaping from danger). The evacuation network’s ca-

pacity (C

i j

) will be checked against the group sizes

), where the allowance of passing through and pos-

sible delays are determined by two rules: (1) the travel

time for the group remains the same, if group size ≤

capacity, or (2) the travel time for the group is equal

to the (group size / capacity) * travel time, if group

size > capacity.

3.2 The Natural Immune System

Vertebrates (organisms having internal bones), have

developed a highly complex and effective immune

system composed of a vast array of cells, molecules,

and organs that work together to maintain life. To-

gether with other bodily systems, the immune system

maintains a stable state of the organism’s vital func-

tion, named homeostasis. The most exceptional roles

of the system are the protection against the attack of

disease-causing agents (pathogen), and the elimina-

tion of malfunctioning cells. Other roles of the sys-

tem also include its capability of recognizing the cells

(or molecules) within the organism as either harmful

(non-self ) or harmless (self ) (Bagheri et al., 2010).

The clonal selection theory is used to elaborate

the adaptive immune system which involves respond-

ing to recognized pathogens and enhance its capabil-

ity of recognizing and eliminating future encounters

(Bagheri et al., 2010). When a non-self antigen in-

vades an organism, the immune system starts by re-

acting to pathogens that invade the organism in such a

way that the immune cells recognize this antigen with

different degrees of afﬁnity. These immune cells then

undergo afﬁnity maturation, proliferation, and clonal

selection.

3.2.1 Population Representation and

Initialization

Information encoded in each population instance con-

sists of arrays of antibody-antigen bindings, which

represents the emergency route (antibody receptor)

and the crowd model with their considered features

(antigen). The antibody can vary in length and cap-

ture certain parts of the nodes and edges of the log-

ical graph structure considered; a source-destination

pair (a single evacuation path P). On the other hand,

the antigen may be ﬁxed to the crowd features which

considers the size of the group d

and the source loca-

tion/node η

. A single population consists of ﬁxed

size antibody-antigen bindings which make up the

overall evacuation plan (the scheduling of the crowd

egress). Thus, if the available group count is g and the

available paths of the network are p, then the length

of a single population would be g × p.

The initialization scheme involves generating the

antibody cells by the aggregation of all available

paths or routes in the evacuation network considered.

This collection is generated by performing a recur-

sive depth ﬁrst search algorithm where all the possible

source-destination pairs are recorded regardless of the

capacity and travel time involved. From this collec-

tion, the available antibody receptors are made. With

the information of the crowd starting location, crowd

size, and size of the crowd in a group, the antigen

cells are generated and tagged with speciﬁc identiﬁca-

tion information. From these collections of antibod-

ies binding to antigens, the entire available crowd’s

groups (antigen) and their randomly assigned source-

destination path or route (antibody), form a popula-

tion instance that represents the complete evacuation

plan (schedule). When a population instance is gen-

erated, the evaluation process is conducted by simu-

lating the evacuation plan. The NCT of the popula-

tion instance, will form the afﬁnity of the antibody-

antigen bindings; higher afﬁnity means lower NCT,

and vice-versa. Thus, this population initialization

process will be repeated to generate the collection

AnArtificialImmuneApproachforOptimizingCrowdEmergencyEvacuationRoutePlanningProblem

505

of population instances based on user-deﬁned popu-

lation size (pop

size

3.2.2 Proliferation, Clonal Selection, and

Afﬁnity Maturation

During this stage, every population instance is sub-

jected to proliferation, but special condition is speci-

ﬁed for populating instances to undergo afﬁnity mat-

uration. The proliferation involves cloning the origi-

nal parent cell, where it will undergo further process

(afﬁnity maturation). Afﬁnity maturation is where

the cloned cell will undergo somatic hyper-mutation

process based on the user-deﬁned mutation steps

(mutate

steps

) where it will determine how many repe-

titions are needed based on the population instance’s

afﬁnity value (pop

a f f inity

). Two types of afﬁnity mat-

uration are adopted: Type-1 Ab-Mutate and Type-2

Ag-Mutate.

For Type-1 Ab-Mutate, all the population in-

stances that have undergone previous proliferation,

will be mutated by exchanging the allowable range

of available antibody receptors with certain anti-

gens. This means the source-destination pair (anti-

body receptor) for a speciﬁc crowd group (antigen) is

changed into another similar source-destination pair

(the same source but different destination, or the

same source and destination). This mutation method

is adopted to produce a variety of population in-

stance’s antibody-antigen binding. This process can

be broadly related to the differentiation of an immune

cell into an effector cell.

In contrast, for Type-2 Ag-Mutate, a clonal selec-

tion mechanism is ﬁrst conducted based on the user-

deﬁned clonal selection rate (select

rate

) (within the

range of [0,1] and the value is very small). When

the clonal selection condition is met, a mutation rate

(mutate

rate

) which is also a user-deﬁned parameter,

will determine how many population instances un-

dergo Type-2 Ag-Mutate that are selected randomly.

Next, the randomly selected population instances will

proliferate (cloned) and undergo Type-2 Ag-Mutate

where the antibody receptor will be tested with a

different “closely-shaped” or similarly-signature anti-

gen. This means a source-destination pair (antibody

receptor) for a speciﬁc crowd group (antigen) will

be exchanged into another crowd group of similar

starting location or source. This mutation method is

adopted to simulate a small tweak of the population

instance’s antibody-antigen binding. Since this muta-

tion type is rarely happens, this process can be con-

sidered the event of immune cells differentiating into

memory cells.

For the case of better afﬁnity, the “mutated”

antibody-antigen binding will replace the parent cell.

These two types of somatic hyper-mutations will

basically automate the ordering of the evacuation

schedule (antibody-antigen binding) and ensure the

diversity of the population instances. In addition,

the somatic hyper-mutation performed based on the

user-deﬁned mutation steps (mutate

steps

) will deter-

mine how many times the mutation process is re-

peated based on the population instance’s afﬁnity

value (pop

a f f inity

). Thus, the mutation process is re-

peated pop

a f f inity

× mutation

steps

times, which is in

accordance with the term hyper-mutation (high rate

mutation).

4 COMPUTATIONAL RESULT

This section involves evaluating the approach perfor-

mance study and highlights the ﬁndings of the study.

This section is divided into three sections: Section 4.1

describes the evaluation of the approach; Lastly, Sec-

tion 4.2 describes the ﬁnding of group cohesion.

4.1 Evaluating the Proposed Approach

The main purpose for this experiment is to determine

the optimal parameter calibrations and evaluate the

performance of the proposed IA. The performance

measure considered is the Network Clearance Time

(NCT) which, in this study’s context, refers to the last

exit time of the evacuees. 9 options of the IA parame-

ter settings are considered where 50 samples are col-

lected from each options. The parameter settings of

each options is given in Table 1. Exhaustive computa-

tion is conducted where the best NCT value out of the

50 samples is taken, as well as their average and stan-

dard deviation before and after optimization. From

the options, the optimal parameter of the proposed IA

is determined from the obtained results.

The results obtained is analyzed and depicted as in

Figure 1. The percentage of the relative standard de-

viation (RSD) is computed for each option, where it

is compared with the best and the average of NCT. As

depicted in Figure 1, option 2 and option 4 have the

lowest RSD for NCT value which means the solution

had obtained the best value all the time (no variabil-

ity). Option 3, option 5, and option 7 also obtained

low RSD for NCT, which means the solution obtained

contain a small amount of variation where small in-

consistencies are present. This is caused by the dif-

ferent pop

size

, gen

size

, and mutate

steps

values, since all

of them shared similar values of mutate

rate

, select

rate

and d

. The remaining options (option 1, option 6,

option 8, and option 9) do not share any similarity in

their parameter settings, all of them used the d

value

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

506

Table 1: The IA parameter options.

Parameters

Options

1 2 3 4 5

pop

size

50 70 100 50 70

gen

size

300 300 300 500 500

mutation

steps

100 75 50 75 50

mutation

rate

0.05 0.01 0.03 0.01 0.03

selection

rate

0.3 0.5 0.7 0.5 0.7

1 5 3 5 3

Parameters

Options

6 7 8 9

pop

size

100 50 70 100

gen

size

500 700 700 700

mutation

steps

100 50 100 75

mutation

rate

0.05 0.03 0.05 0.01

selection

rate

0.3 0.7 0.3 0.5

1 3 1 1

of 1. This causes the RSD of NCT value to be high

and variation of the solution becomes very inconsis-

tent. Thus, option 2 and option 4 parameter settings

are believed to be the best setting for the proposed IA

algorithm.

1 2 3 4

5 6

7 8 9

Option No.

Value

Average and relative standard deviation of NCT

Before Average

After Average

Before RSD

After RSD

Figure 1: NCT analysis before and after optimization for

every options.

The options also reveal that the smaller size of the

pop

size

(e.g. 50 or 70) and the gen

size

(e.g. 300 or

500) are able to obtain optimum results for this public

data set. The possible routes of the evacuation net-

work itself are 15 in total, make the search space lim-

ited. Thus, using larger pop

size

would not be feasi-

ble and will burden the search process with redun-

dant routes. In addition, the crowd size of 30 in total

(medium scale) which is further divided into smaller

groups, had lesser need of greater gen

size

due to a

small amount of variation and limited search space

(only 15 route options). From the two best options

(option 2 and option 4), their pop

size

is balanced with

the gen

size

. For example, in option 2, lower pop

size

(e.g. 50) should have higher gen

size

(e.g. 500) and

vice-versa. This would cause optimum result to be

achieved with acceptable convergence rate. The evac-

uation solution improves when gen

size

≤ 500 and is

believed to be affected by the group size (d

) as well.

4.2 The Effect of Group Cohesion

An additional options of 2a and 2b, relative to op-

tion 2 with changes in the parameter setting of d

and 3, respectively) are conducted. The value of of

best and average values of the solution is computed

as well as the RSD for each additional options. Fig-

ure 2 depicts the best, average, and RSD of NCT for

the additional options. Although there are not many

indications present in the best values of NCT, the av-

erage and RSD had showed to inversely proportional

relationship with d

values when the d

increases (see

Figure 2). The steadily declining value of RSD had

proven that increases in the value of d

(within a cer-

tain size) results in lower variations of the solution,

where better NCT value is obtained with low incon-

sistency.

2a 2b 2

Option No.

Value

Average and relative standard deviation of NCT

Average

RSD

Figure 2: Analysis of NCT for option 2a, 2b, and 2.

The group size (d

) affects the solution quality,

where an increase in the ﬂow rate happens when the

is increased. When evaluating the ﬂow rate, d

≥ 3

indirectly decrease the NCT value of a particular so-

lution. This is possible due to the group cohesion

assumption in the crowd model, where evacuees that

move in larger groups tend to move better (better ﬂow

rate) while conforming to the network capacity. This

AnArtificialImmuneApproachforOptimizingCrowdEmergencyEvacuationRoutePlanningProblem

507

particular situation simulates a strong bond between

the evacuees within a group and cooperative behavior

is elicited. Although the proposed crowd evacuation

model does not considers the dynamic behavior(s) of

groups, the general group behavior with certain de-

grees of assumptions (assuming 100% compliance of

individual within group and 0% cooperation of inter-

groups relation) during evacuation has been success-

fully demonstrated.

5 CONCLUDING REMARKS

This paper offers an immune algorithm (IA) ap-

proach, that incorporates new ideas in designing the

solution representations and their respective opera-

tors in solving the ERP problems. A crowd model

that considers group cohesion with a certain degree

of assumptions is presented and evaluated, while IA

approach is also evaluated by performing various ex-

periments which constitute the parameter calibration

in order to attain an optimal result. The insights and

ﬁndings of the observed results from the experiments

have been discussed and presented with respect to the

main interest of the study.

The group cohesion is assumed on the basis that

everyone within a group completely cooperates with

each other without capturing individual compliance

rate. In addition, the inter-group relation is also ne-

glected (no social interaction between group) which

is shown to exhibit discrete delays in the ﬂow rate

of the evacuation. In a real evacuation, group be-

havior tends to have varying compliances due to dif-

ferent needs and goals, as well as a certain amount

of interaction between groups (causing greater delays

and even blocking). Therefore, further enhancement

of the proposed IA approach considering dynamics

of group is expected while considering the effects of

larger crowd sizes is recommended to further support

the ﬁndings presented in this paper.

REFERENCES

Alsnih, R. and Stopher, P. R. (2004). Review of proce-

dures associated with devising emergency evacuation

plans. Transportation Research Record: Journal of

the Transportation Research Board, 1865(1):89–97.

Bagheri, A., Zandieh, M., Mahdavi, I., and Yazdani, M.

(2010). An artiﬁcial immune algorithm for the ﬂex-

ible job-shop scheduling problem. Future Generation

Computer Systems, 26(4):533–541.

Carron, A. V. and Brawley, L. R. (2000). Cohesion concep-

tual and measurement issues. Small Group Research,

31(1):89–106.

Cepolina, E. M. (2005). A methodology for deﬁning build-

ing evacuation routes. Civil engineering and environ-

mental systems, 22(1):29–47.

Dasgupta, D., Yu, S., and Nino, F. (2011). Recent advances

in artiﬁcial immune systems: models and applications.

Applied Soft Computing, 11(2):1574–1587.

Hajibabai, L., Delavar, M., Malek, M., and Frank, A.

(2007). Agent-based simulation of spatial cognition

and wayﬁnding in building ﬁre emergency evacua-

tion. In Geomatics Solutions for Disaster Manage-

ment, pages 255–270. Springer.

Hui, F., Zi, Z., and Gang, H. (2010). Developing a model for

the emergency rescue routing problem using stochas-

tic programming theory. In Logistics engineering and

intelligent transportation systems (LEITS), 2010 In-

ternational Conference on, pages 1–4. IEEE.

Kim, S., George, B., and Shekhar, S. (2007). Evacuation

route planning: scalable heuristics. In Proceedings

of the 15th annual ACM international symposium on

Advances in geographic information systems, page 20.

ACM.

Li, Q., Fang, Z., Li, Q., and Zong, X. (2010). Multiob-

jective evacuation route assignment model based on

genetic algorithm. In Geoinformatics, 2010 18th In-

ternational Conference on, pages 1–5. IEEE.

Lu, Q., Huang, Y., and Shekhar, S. (2003). Evacuation plan-

ning: a capacity constrained routing approach. In In-

telligence and Security Informatics, pages 111–125.

Springer.

Lv, Y., Huang, G., Sun, W., and Chao, D. (2012).

A scenario-based modeling approach for emergency

evacuation management and risk analysis under mul-

tiple uncertainties. Journal of hazardous materials.

Sayyady, F. and Eksioglu, S. D. (2010). Optimizing the use

of public transit system during no-notice evacuation

of urban areas. Computers & Industrial Engineering,

59(4):488–495.

Wang, P., Luh, P. B., Chang, S.-C., and Sun, J. (2008). Mod-

eling and optimization of crowd guidance for building

emergency evacuation. In Automation Science and

Engineering, 2008. CASE 2008. IEEE International

Conference on, pages 328–334. IEEE.

Xie, C., Lin, D.-Y., and Travis Waller, S. (2010). A dynamic

evacuation network optimization problem with lane

reversal and crossing elimination strategies. Trans-

portation research part E: logistics and transportation

review, 46(3):295–316.

Zeng, M. and Wang, C. (2009). Evacuation route planning

algorithm: Longer route preferential. In Advances

in Neural Networks–ISNN 2009, pages 1062–1071.

Springer.

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

508