Simulation Modelling Practice and Theory 36 (2013) 44–59

Contents lists available at SciVerse ScienceDirect

Simulation Modelling Practice and Theory

journal homepage: www.elsevier .com/locate /s impat

Integration of Smoke Effect and Blind Evacuation Strategy(SEBES) within fire evacuation simulation

1569-190X/$ - see front matter � 2013 Elsevier B.V. All rights reserved.http://dx.doi.org/10.1016/j.simpat.2013.04.001

⇑ Corresponding author at: Posts and Telecommunications Institute of Technology (PTIT), Hanoi, Viet Nam.E-mail addresses: nmhufng@yahoo.com (M.H. Nguyen), ho.tuong.vinh@auf.org (T.V. Ho), jean-daniel.zucker@ird.fr (J.-D. Zucker).

Manh Hung Nguyen a,b,⇑, Tuong Vinh Ho a, Jean-Daniel Zucker a,c

a IRD, UMI 209, UMMISCO, IFI/MSI, Vietnam National University of Hanoi, Viet Namb Posts and Telecommunications Institute of Technology (PTIT), Hanoi, Viet Namc UPMC Univ Paris 06, UMI 209, UMMISCO, F-75005 Paris, France

a r t i c l e i n f o

Article history:Received 6 August 2012Received in revised form 16 February 2013Accepted 2 April 2013Available online 5 June 2013

Keywords:Multi-agents systemFire evacuation simulationEvacuation strategyBlind evacuationSimulation modelling

a b s t r a c t

Many fire evacuation models have been proposed in recent years to better simulate such asan emergency situation. However most of them do not respect a recommendation of fireevacuation experts regarding the fact that evacuees should follow the boundaries of obsta-cles or wall to find the exits when their visibility is limited by smoke. This paper presentsan agent-based evacuation model with Smoke Effect and Blind Evacuation Strategy (SEBES)which respects that recommendation by integrating a model of smoke diffusion and itseffect on the evacuee’s visibility, speed, and evacuation strategy. The implementation ofthis model enables us to optimise the evacuation strategies taking into account the levelof visibility. The obtained simulation results on a realistic model of the Metro supermarketof Hanoi confirm the important impact of smoke effect and blind evacuation strategy onthe number of casualties.

� 2013 Elsevier B.V. All rights reserved.

1. Introduction

Fire is increasingly a cause of casualties in modern life. For instance, the Myojo 56 building fire in Tokyo (Japan) on Sep-tember 1st 2001 has killed 44 people and 291 people killed in Mesa Redonda shopping center fire in Lima (Peru) on Decem-ber 29th 2001. There were also 11 people who died in a fire at the detention center of Amsterdam Schiphol Airport(Netherlands) on October 27th 2005. The Moscow (Russian) hospital fire killed 46 people on December 9th 2006. The SantikaClub fire in Bangkok (Thailand) killed 66 people on January 1st 2009. The ABC daycare center fire killed 47 people in Her-mosillo (Mexico) on June 5th 2009. The 2010 Dhaka fire was a fire in the city of Dhaka (Bangladesh) on 3rd June 2010 thatkilled at least 117 people. And this list could infinitely grow up.

The huge loss in these fires leads to at least two important questions: (1) Were people trained to practice the best strategyto fire evacuate? and (2) Were the building designed with the best inside configuration regarding to fire evacuate? These twoquestions show common issues: how can we assess which strategy is best among the fire evacuation strategies? More spe-cifically, given a particular building which strategy is the best one? The real answer does not exist unless we could exper-iment in the real environment! One truthful approach is to rely on simulation environment modelling as close as the realworld fire evacuation conditions.

Once a fire evacuation simulation model is proposed, it has to comply with at least two modelling points of view. First,from the point of view of fire evacuation experts, the model should take into account the smoke diffusion and its effect on theevacuation, the observable range, the evacuation speed, and the toxic poisoning level of evacuees. In particular, it should

M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59 45

respect the recommendation of these experts by modelling the movement of evacuees in a limited visibility condition withthe strategy of tracking the wall or obstacles. Moreover, the proposed model should accommodate to the model differentevacuation strategies in order to see which is the best suitable strategy for a given building. Second, from the point of viewof building architecture designers, it should enable to apply with several realistic building evacuation plans (with GIS data)to see which is the best evacuation plan for a given building.

Our objective in this paper is to propose a model for fire evacuation simulation based on agents. This model does not onlysimulate the effects of fire/smoke on the abilities to move, to observe of evacuees, but also takes into account the given ad-vise of fire evacuation experts, called Smoke Effect and Blind Evacuation Strategy (SEBES). We thus developed this model as atool which could help fire evacuation training experts to visually demonstrate what evacuation strategy is better in a givenenvironment. Our contribution is thus three-folds:

� First, a proposal of a new agent-based model for fire evacuation simulation is given. This model respects the recommen-dation of experts in fire evacuation by taking into account their recommendation that evacuees should follow the bound-aries of obstacles or wall to find the exits when their visibility is limited due to the smoke.� Second, an implementation of the proposed model based on an agent based integrated GIS support platform (GAMA [1])

supporting the development of an useful tool for two groups of users:– The first group is experts in fire evacuation. They could use this tool as a visual demonstration to illustrate what strat-

egy is the best for evacuees to evacuate by applying all considered strategies into the model and run it, then comparethe output parameters to see which is the best among them. This could lead their evacuation training courses to bemore intuitive and convince. For instances, in the case study of the Metro supermarket of Hanoi, we compare threestrategies of fire evacuation: following the evacuation signs, following the crowd, and following the own’s path whenevacuees could observe still, and following the boundaries of obstacles and/or wall when their visibility is limited. Thesimulation results show that following the evacuation signs is the best strategy in that situation.

– The second group is building architects, constructors, interior designers, etc. They could use this tool to choose the bestinternal configuration of a given building regarding the effect in fire evacuation by applying their different designs intothis model and run it, then compare the output parameters to see which is the best configuration.

This paper is organised as follows: Section 2 presents some related works in the field of crowd evacuation modelling andsimulation. Section 3 presents our agent-based model including a Smoke Effect and Blind Evacuation Strategy (SEBES) mod-ule for fire evacuation simulation. Section 4 presents the application of our model to a real case study, including two types ofscenario: scenarios comparing three bind evacuation strategies, and scenarios comparing three other evacuation strategies innormal condition. Finally Section 5 presents a discussion of the simulation results and some conclusions as well as a discus-sion about future research.

2. Related works

Recently, there has been an increasing number of models proposed for fire evacuation modelling in buildings. Table 1summaries a partial collection of recent proposed agent models for fire evacuation. We consider models at two levels:

� At the level of modelling, we consider the modelling of agent types involving a fire evacuation: the evacuees (eV. – col-umn), the group or crowd of evacuees (g/c – column), the fire (fi. – column), the alarm or voice system (a/v – column), andthe smoke (sm. – column).� At the level of optimisation, we consider the optimisation on the building design and the evacuation plan design (de. –

column), the optimisation on evacuation strategies in normal condition (visible evacuation strategy – v.e column), andthat in limited visibility condition (blind evacuation strategy – b.e column).

More detail, let us analyse at the level of modelling. Evacuee and fire are two objects modelled in most of the listed mod-els. There is only a small number of models modelling the smoke [6,7,12,39]. In these smoke models, the authors took intoaccount the fact that smoke affects the visibility and speed of evacuee. Furthermore, no model does respect a recommenda-tion of fire evacuation experts on the fact that evacuees should follow the boundaries of obstacles or wall to find the exitswhen their visibility is limited by smoke.

At the level of optimisation, there are many models built to choose the best floor designs or evacuation plan for a givenbuilding [3,5,6,10,29,30,35,40]. There are also some models optimised evacuation strategies in normal (visible) condition[25,31,34]. But there is no model to optimise evacuation strategies in limited visibility condition. We do not aim to builda better model of existing over of all aspects but to focus on smoke modelling and taking into account expertrecommendations;

Our model will model many kinds of agent: evacuee, fire, alarm, smoke, etc. in which the behaviour of evacuee is mod-elled based on a recommendation of fire evacuation experts on the fact that evacuees should follow the boundaries of obsta-cles or wall to find the exits when their visibility is limited by smoke. This enables us to optimise on many aspects: optimisethe evacuation plans, optimise the evacuation strategies in both conditions: visible and invisible.

Table 1Summary of recent proposed models (eV = evacuee, g/c = group or crowd, fi = fire, a/v = alarm or voice, sm = smoke, de = design, v.e = visible evacuation,b.e = blind evacuation).

Models Modelling Optimisation

eV. g/c fi. a/v sm. de. v.e b.e

Alavizadeh et al. [2] U U U

Averill and Song [3] U U U

Garca-Cabrera et al. [4] U U U

Chaturvedi et al. [5] U U U

Daito and Tanida [6] U U U U U

Filippoupolitis et al. [7,8] U U U

Gianni et al. [9] U U

Hanea et al. [10] U U U U

Helbing et al. [11] U U U

Hu et al. [12] U U U

Huang et al. [13] U U

Lin et al. [14] U U

Luo et al. [15] U U U

Korhonen and Hostikka [16] U U U

Kuligowski et al. [17,18] U U

Okaya and Takahashi [19,20] U U U

Pan et al. [21] U U U

Patvichaichod et al. [22,23] U U U

Qiu and Hu [24] U U

Rahman et al. [25] U U U

Ren et al. [26] U U

Ruppel et al. [27,28] U U

Sagun et al. [29] U U U U

Said et al. [30] U U U

Schneider and Konnecke [31] U U U

Shen and Chien [32] U U

Shendarkar et al. [33] U U U

Suryotrisongko and Ishida [34] U U U

Tang and Ren [35] U U U

Tingyong et al. [36] U U

Tsai et al. [37] U U U

Wang et al. [38] U U

Weifeng and Hai [39] U U U

Yi and Shi [40] U U U

Our model U U U U U U

46 M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59

3. SEBES: an agent-based simulation model

This section presents our agent-based model including a Smoke Effect and Blind Evacuation Strategy (SEBES) module forfire evacuation simulation: Section 3.1 presents the general architecture of the model; Section 3.2 presents the modelling ofevacuee agents; Section 3.3 presents the modelling of fire agents; Section 3.4 presents the modelling of smoke agents; Sec-tion 3.5 presents the modelling of fire alarm agents; and Section 3.6 presents the modelling of sign and plan agents.

3.1. Architecture of the model

The architecture of our simulator could be seen at three levels as depicted in Fig. 1. At the platform level, the model isdeveloped on the simulation platform GAMA [1]. GAMA provides a simulation development environment for building spa-tially explicit agent-based simulations. It enables: (i) to use arbitrarily complex GIS data as environments for the agents; (ii)to run simulations composed of vast numbers of agents; (iii) to conduct automated controlled experiments on various sce-narios, with a systematic, guided or ‘‘intelligent’’ exploration of the space of parameters of models; and (iv) to let users inter-act with the agents in the course of the simulations.

The second level is the simulator which relies on a multiagent system. This is the core of our approach which includes thefollowing types of agent:

� Evacuee agent: representing an evacuee. This agent could see the fire/smoke, hear the alarm, and evacuate to one of theemergency exits by avoiding the obstacles and other evacuees.� Alarm agent: representing a fire alarm. This agent could detect fire/smoke in its detection range and ring in a ringing dura-

tion of time.� Fire agent: representing fire. The fire agent could propagate within the building space.

Fig. 1. The three levels architecture of the model SEBES.

M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59 47

� Smoke agent: representing smoke. The smoke agent is created from fire agents. It could propagate inside the buildingspace and therefore increase the smoke intensity at a give position by time.� Sign and plan agent: representing of evacuation signs and plan. This is a non-movable agent. This provides the information

about the direction to emergency exits.

The modelling of these agents will be presented in the next sections.The third level is the visualisation level. This level supports displaying the realistic status of the simulation as well as the

values of the output parameters.The details of the classes of the model are depicted in Fig. 2: all agent classes inherit from the species agent which is the

highest in the hierarchy of agent in the GAMA language. At the level of agent skills, the Evacuee agents are able to move, sothey have a Moving skill. Other agents have Situated skill.

Fig. 2. The different classes of the model SEBES: the diffusion of smoke is modelled by the method propagate().

48 M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59

3.2. Evacuee agents

This agent represents an evacuee, he has the following attributes:

� Observable range: the space around an evacuee that can be observed and perceived by the evacuee. An evacuee agentcould only observe evacuation signs and/or other evacuees within this range.� Toxic level: the level of toxicity poisoning an evacuee. This is initially as zero, and then increased due to the effect of

smoke. An agent is considered as to be died if his toxic level reaches 100%.� Fire exposure level: this represents the sensitive level in the fire of the evacuee. The higher this value is, the more the evac-

uee agent is affected by fire/smoke.� Speed: speed of an evacuee in evacuation. This speed is changed according to the effect of visibility, and the toxicity level.� Passed position list: The time stamped list of positions that an evacuee agent has had during evacuation.

The evacuee agent’s behaviours are presented in Fig. 3a: in a normal condition, the evacuee agent normally moves insidethe building. He starts to evacuate if and only if either he sees a fire/smoke or he hears the fire alarm’s ringing. His evacuationmovement is finished when he gets out of the building. During his evacuation, he moves following the evacuation movementprinciple. His observable range is reduced and the toxic level is increased by time due to the intensity of fire/smoke. Theseprinciples are presented in Sections 3.2.1 and 3.2.3.

3.2.1. Evacuee movement principleThe fire evacuation experts of Hanoi Fire Evacuation Association have suggested us to respect the fire evacuation guide-

lines when the evacuee meets obstacles: the evacuee should move along the border of the obstacle until the door (target) orthere no more obstacle in front of the evacuee. We take into account this principle of evacuee movement.

We use priority direction approach for modelling of agent’s movement. In this approach, agent chooses the direction hav-ing the highest priority to move. Other directions will then be prioritised relatively to the one having the highest priority.There are two movement strategies: 4 directions or 8 directions (as depicted in Fig. 4). We use the 8 directions strategy inall simulations. The more the direction is near the highest priority direction, the more the direction has high priority. At eachstep, the agent considers the highest priority direction to move. If it is not possible, the agent will consider the next lowerpriority direction, and so on. A candidate direction is not considered if only if: either it is on an obstacle, or it leads to a posi-tion which is in the recent passed positions list of the agent.

In order to avoid the infinite loop of agent movement in the case having obstacles on the agent direction, we use a recentpassed positions list which contains the n last positions of agent. Agent thus considers the next position to move which is notin its recent passed positions list. We do not save all the passed positions of agent because the dynamic environments: somekind of obstacles, such as fire, can dynamically change. There may be a fire at the position x at the moment t1, but may be no

Fig. 3. Behaviour modelling for agents.

Fig. 4. The two movement strategies based on 4 or 8 directions.

M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59 49

more fire at x at the moment t2 > t1. So we limit the size of the list to give agent a possibility to return to the positions whichit passed in long time.

An agent determines its own recent movement tendency by considering m last positions (m < n, n is the size of recent passedpositions list). Therefore, the priority direction is the arc from the m-latest position to the current position of agent. Fig. 5illustrates the movement principle of an agent when there is an obstacle on its evacuation way. At the time t = t0, the agenthas not meet the obstacle yet, so it continues to move to its target. Next step, t = t0 + 1, the agent meets the obstacle, it findsits recent movement tendency which is still direct to target because the two latest positions are on the same line (m = 2). Butthe 1st, 2nd and 3rd direction are impossible (in the obstacle, case of 8 directions), so the 4th and 5th direction are possible.Assume that the agent turns right. At the time t = t0 + 2, the priority direction is the arc from the position at t = t0 to the one att = t0 + 2. In this direction, the first four priorities are not possible, the 5th is possible (the blue arc), and so on. At the timet = t0 + 4, there is no more obstacles in front of the agent, so its priority direction is the direct line to its target.

3.2.2. Blind movement principleIn a blind situation, an evacuee uses the same recent movement tendency principle to move, except that he does not know

exactly where is the target. Therefore, his movement is based on following rules (Fig. 6):

� Blind movement rule 1: if there is not any obstacles or walls near him, the evacuee moves ahead (straight, right-straight, orleft-straight).� Blind movement rule 2: if there is some obstacles or walls in front of him, the evacuee changes its movement direction as

follows:– if the current direction is perpendiculars to the surface of an obstacle/wall, the new direction could be either right or

left (Fig. 6, case of ‘‘90 touch’’),– if the current direction is not perpendicular to the surface of an obstacle/wall, the new direction will be the nearest

direction to the current one which enables the evacuee to follow the obstacle/wall (Fig. 6, case of ‘‘normal touch’’).� Blind movement rule 3: if an evacuee is tracking an obstacle/wall, he continues to track until the end of the obstacle/wall

(Fig. 6, case of ‘‘during tracking’’).� Blind movement rule 4: at the end of an obstacle, the evacuee continues to move along the current movement direction

(Fig. 6, case of ‘‘getting out of obstacle’’).� Blind movement rule 5: at the ‘‘end’’ of a wall, the evacuee continues to follow the next face of the wall until he reaches an

exit (Fig. 6, case of ‘‘getting out of wall’’).

Note that in the blind movement rule 5, there could be an exception when the wall is a closed block in the building, theevacuee thus could repeat his movement around the wall forever. Therefore, he could not get out of the building. This

Fig. 5. The movement principle to avoid obstacles.

Table 2Simulation parameters (based on [41,43,44]).

Parameter Values

Number of simulations for each scenario 100Number of evacuee agents 1000Length of recent passed positions list 20Influence factor of smoke (b) 0.01Safety smoke intensity threshold (h) 12.5%

Fig. 6. The blind movement principle.

50 M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59

problem could be solved by using the proposed recent passed positions list technique: if the evacuee is aware that he is repeat-ing the evacuation path around a wall, he will change to the blind movement rule 4: at the end of the closed block wall, hecontinues to move as the current direction by considering the closed block wall as an obstacle.

3.2.3. Toxic level evolutionIf an evacuee agent is in the smoke, he is poisoned by toxic fumes. The higher the smoke’s intensity at the agent’s position

is, the more it is poisoned. Assume that pti and ptþ1

i are the toxicity poisoned of evacuee i at the time t and t + 1, we have:

1 In aare (10100 < =

ptþ1i ¼

0% if pti þ b � ðItðpÞ � hÞ < 0%

pti þ b � ðItðpÞ � hÞ if 0% 6 pt

i þ b � ðItðpÞ � hÞ 6 100%

100% if 100% < pti þ b � ðItðpÞ � hÞ

8><>:

ð1Þ

where b is the influence factor of smoke; It(p) is the intensity of smoke at the evacuee’s position p (of the agent i) at the timet; h is a safety smoke intensity threshold.

Following Europe guideline (CFPA-E No. 19:2009 [41]), every inspiration has 16% of oxygen concentration. An evacueeshows serious symptoms if the oxygen concentration is lower than 14%. When the oxygen concentration decreases to14%, the smoke in the air must be 100 � 14 � 100/16 = 12.5%.1 Therefore the safety smoke intensity threshold is chosen in thismodel is h = 12.5% (Table 2). It means that if the smoke intensity is over 12.5%, the evacuee starts to be poisoned.

Note that, following formula (1), the toxic level of an evacuee will be decreased if the evacuee enters in a zone having thesmoke intensity lower than h. And inversely, the toxic level will be increased if the evacuee enters in a zone having the smokeintensity higher than h. The more the smoke intensity is high, the faster the toxic level is increased. The evacuee will be sup-ported to be died if his or her toxic level is equal to 100%.

normal condition: there are 16 l of oxygen in 100 l of normal air. Assume that now there are x litres of smoke in 100 l or polluted air, it means that there0 � x) litre(s) of normal air. So the oxygen in these (100 � x) litre(s) of normal air are 16⁄(100 � x)/100. It takes less than 14% means that 16⁄(100 � x)/14. It equals to x > = 100⁄(1 � 14/16) = 12.5.

Fig. 7. Relation between smoke intensity and evacuee’s observable range and speed.

M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59 51

3.2.4. Evacuee observable range reduce principleLike the evacuee’s power, the evacuee’s observable range is decreased if the evacuee is in the smoke. The higher the inten-

sity of the smoke at the position of the evacuee is, the less the evacuee is able to observe around it (this is well modelled inthe model of Kang [42]). The relationship between the smoke’s intensity and the observable range of evacuee is depictedin Fig. 7. Assume that r0

i and rti are the observable range of evacuee i at the beginning (without any smoke) and at the time

t. In order to keep the model as simple as possible, we use a function which states that the observable range is inverselyproportional to the smoke intensity:

rti ¼ ð1� ItðpÞÞ � r0

i ð2Þ

where It(p) is the intensity of smoke at the evacuee’s position p (of the agent i) at the time t (normalised in the interval of[0,1]).

3.3. Fire agents

This agent represents fires. The smoke agent is created from fire agents. Fire agent could propagate.

� Duration: its time to live. Normally, the duration of a smoke is longer than that of a fire.� Propagation speed: the speed of propagation of the fire or the smoke. This speed changes by time and by the quantity of

fire agents in the building.� Affected zone: the space around a fire which can affect evacuees inside it.� Smoke creation speed: The smoke could not create other smoke, but fire could do it. This attribute of fire defines the speed

to create smoke of a fire.

The fire agent’s behaviours are presented in Fig. 3b: From its start, a fire burns until its ‘‘time to live’’ is equal to zero.During its burning, a fire continues to create smoke with its smoke creation speed. And the fire could also propagate by cre-ating other fire near by its position with its propagation speed.

3.4. Smoke agents

This agent represents smoke. The smoke agent is created from fire agents: (i) smoke, once being born from fire, is rela-tively independent from fire and (ii) smoke could move in some unpredictable directions: During moving around inside thebuilding, smoke changes the smoke intensity at a given position by time. We therefore model smoke as an agent.

� Direction: The direction to propagate. The direction is determined based on the following principle: the smoke movesfrom the position with higher smoke intensity to the position with lower smoke intensity.� Propagation speed: the speed of propagation of the smoke. The speed of smoke is determined based on the following prin-

ciple: the more the difference of smoke intensity at the two positions, the higher the speed of smoke

The smoke agent’s behaviours are presented in Fig. 3c: From its creation, a smoke updates its speed and direction at everysimulation step. After updating these two attributes, the smoke moves to the next position. If the position is already outsideof the building, it dies. Otherwise, it continues to update its attributes.

3.5. Alarm agents

This agent represents a fire alarm with following attributes: Main properties:

� Ringing duration: the duration of ringing when fire/smoke was detected.� Detection range: this agent rings if there is fire or smoke appearing in this zone.

Fig. 8. The evacuation plan of the Metro supermarket of Hanoi.

52 M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59

The alarm agent’s behaviours are presented in Fig. 3d: In a normal condition, the alarm agent does not ring. It starts toring if and only if either it detects the fire/smoke inside its detection range. This ringing leads evacuee agents to evacuate. Itstops ringing when the ringing duration is over passed.

3.6. Sign and plan agents

They represent the evacuation signs and plan. They are non-movable agents. They provide the information about thedirection to emergency exits.

4. Case study: simulation with the Metro supermarket of Hanoi

In this section, we apply the proposed model to simulate the fire evacuation in the supermarket Metro of Hanoi. Ourobjective is to use simulation experiments to study what is the best evacuation strategy in the supermarket environment.This section is organised as following: Section 4.1 presents the setting up of environment for simulations; Section 4.2 vali-dates the model of smoke and blind evacuation; Section 4.3 validates the fire expert’s recommendation; Section 4.4 opti-mises the evacuation strategies in normal condition.

4.1. Simulation setup

This section presents the setting up of environment for simulations: Section 4.1.1 presents the setting up of the evacu-ation plans; Section 4.1.2 sets up simulation parameters; Section 4.1.3 presents analysis and evaluation criteria.

4.1.1. The evacuation plan of Metro supermarket of HanoiThe environment of simulation is a representation of GIS data as shown in Fig. 8. The Metro supermarket of Hanoi is sit-

uated on one floor, with eight emergency exits: three on the front, two on the left, and three on the right. In the simulation,the emergency exits are represented by red rectangles. People can directly get to the left and right emergency exits frominside. While in order to go to the front emergency exits, people have to pass two more gate layers: first, the cashiers layerwith 12 main outputs, each is divided by two to have 24 cashiers in total; second, the security layer with four doors, and thenthe two main exits to the parking. Another exit is the entrance which can use as an emergency exit.

The inside configuration of Metro2 supermarket can be decomposed into three main zones. First, the left zone is the one forclothes and electronic materials. There are two rows of shelves. One is tall so people cannot look over to see the signs, thereforethe evacuation signs are putted on the shelves. One other is short so people can look over shelves to see the evacuation signs:there is only one common evacuation sign above them. The evacuation signs in this zone indicate the direction to the two leftemergency exits.

2 This plan is taken based on the inside configuration of the Metro supermarket of Hanoi on the September 12th 2011

M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59 53

Second, the central zone presents in majority dry and cosmetic goods. There are three identical rows of shelves. The nineshelves on the left are tall so the evacuation signs are also pasted on them. The three remain on the right are short for frozengoods and therefore there is only one evacuation sign above them. The evacuation signs in this zone represent the directionto three front emergency exits.

Third, the right zone presents the vegetables and foods. This is divided into three small blocks, each has an right emer-gency exit. The evacuation signs in this zone represent the direction to the three right emergency exits.

4.1.2. Simulation parametersIn order to make the results comparable, we use the same values for input parameters of all simulations: steps of simu-

lations; number of people; initial distribution. These values are estimated based on the Europe guideline on Fire safety engi-neering concerning evacuation from buildings (CFPA-E No. 19:2009 [41]), the Human factors: Life safety strategies Occupantevacuation, behaviour and conditions (PD7974-6:2004 [43]), and the Fire and Smoke: Understanding the Hazards of The Com-mittee on Fire Toxicology, Board on Environmental Studies and Toxicology, National Research Council [44]. These parametersare shown in Table 2.

4.1.3. Analysis and evaluation criteriaFor each evacuation strategy, we run the simulations many times (100 times at least) with the same value of initial

parameters: the number of people (N = 1000), and the speed of people. At the output, we need to calculate the followingparameters:

� Percentage of survivals. The simulated environment is one floor building, so a person is considered as escaped if s/hepassed one of emergency exits.� Percentage of death. A person is considered as dead when his or her toxic level reaches 100%.� Average time for a person to be escaped. It is the average time duration from the moment when s/he starts to evacuate

until s/he escapes.� Average rate of toxic level of survivals.

In comparing these parameters among strategies for each experiment, we will see which strategy of occupants is better inthis realistic environment of the Metro supermarket of Hanoi. A strategy is considered as better if three following observa-tions are true: (1) the % of survivals is higher (% of death is lower); (2) the average time to escape is shorter; and (3) theaverage rate of toxic poisoned of survivals is lower.

4.2. Validation of smoke and blind evacuation model

The development of smoke during fire is presented in some snapshots of simulation interfaces (Fig. 9). Following the timeof fire, the smoke increases the propagation space and intensity. These are correspond to the results of the modelling ofsmoke in Section 3.4.

In order to indicate the effect of smoke on the movement of evacuees, at the visual level, we tracked the evacuation pathsof some evacuees in the condition of limited visibility due to smoke (Fig. 10a). These evacuees are modelled with the evac-uation strategy recommended by fire evacuation experts: tracking the walls or obstacles. The results in this case show that inmajority time of movement, evacuee really tracks the walls and/or obstacles.

Fig. 9. The visualisation of fire and smoke during simulation.

Fig. 10. Comparison of the movements of three evacuees between two cases: with and without smoke model.

54 M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59

Let consider further, we compare the evacuation paths of the same evacuee (at the same start position) in two conditions:with smoke (limited visibility – Fig. 10a) and without smoke (the evacuee could observe the evacuation signs to follow them– Fig. 10b). The results show that these two paths are different. That is reasonable because in the limited visibility, evacueedoes not see the evacuation signs to follow them, it has to follow the wall or obstacle, so, its path could be significantly longerthan in the case that evacuee could observe the signs to go to the approximate nearest exit.

4.3. Validation of fire expert’s recommendation

This section validates the fire expert’s recommendation: Section 4.3.1 presents the evacuation contexts and the consid-ered evacuation strategies; Section 4.3.2 presents and analyses the obtained results.

4.3.1. Evacuation context and strategiesIn order to verify if our model simulate correctly the advice of fire evacuation specialists and/or experts in a blind evac-

uation condition, we compare the given strategy of tracking the wall to the wander moving strategies, including: going aheadand random moving.

� Blind evacuation strategy 1: Tracking the wall. The evacuee goes straight until touching an obstacle or a wall. It then tracksfollowing boundaries of the obstacle. At the end of these boundaries, it continues to go straight. These actions arerepeated until the evacuee gets to an exit.� Blind evacuation strategy 2: Going ahead. The evacuee goes ahead until touching an obstacle or a wall. Once it touches an

obstacle, it changes its direction and continues to move. These actions are repeated until the evacuee gets to an exit.� Blind evacuation strategy 3: Random moving. The evacuee randomly moves until finding an exit.

4.3.2. ResultsWe analysis the results at two levels: vision and statistics. At the level of vision, our objective is to verify whether the

agent of each strategy behaves as expected (described). We did track the evacuation paths of three agents which are ran-domly chosen as representatives of agents corresponding to three blind evacuation strategies as depicted in Fig. 11: the cycleline corresponds to the path of tracking the wall, the square line corresponds to the path of going ahead, and the rectangleline corresponds to the path of random moving. Intuitively, each agent evacuates corresponding to its defined strategy: thepath given by the strategy of tracking the wall seem to be based on the walls; that of going ahead seem to be straight untilthe evacuate direction is changed when the agent touches an obstacle; and that of random moving seem to be unpredictablebecause its evacuate direction is randomly changed. We could consider that the simulated behaviours of three kinds of agentare real enough to be acceptable.

At the statistic level, our objective is to quantitatively compare the output parameters to see whether the recommendedstrategy of tracking the wall is better than the two others in case of blind evacuation. The results of the % of survivals by thesimulation steps are depicted in Fig. 12. For more detail, the % of survivals in the case of tracking the wall is higher than thosein the case of going ahead or random moving.

The % of survivals (or death) at the end of simulation is shown in Fig. 13a: the % of survivals in the case of tracking the wallis higher than that in the case of going ahead (M(tracking) = 49.12%, M(ahead) = 35.37%, significant difference with p-va-lue < 0.001) or random moving (M(tracking) = 49.12%, M(random) = 21.75%, significant difference with p-value < 0.001).

Without any contradiction, the average rate of toxicity level (in %) of survivals is also significant (as depicted in Fig. 13b):in the case of tracking the wall, people is poisoned lower toxic fumes than going ahead (M(tracking) = 69.85%,

Fig. 11. The different paths among three blind evacuation strategies.

Fig. 12. The % of survivals by simulation steps among three blind evacuation strategies.

Fig. 13. Comparison of output parameters among three blind evacuation strategies.

M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59 55

M(ahead) = 77.44%, significant difference with p-value < 0.001) or random moving (M(tracking) = 69.85%, M(random) =83.68%, significant difference with p-value < 0.001).

In the same line, the average time to escape (Fig. 13c) in the case of tracking the wall is significantly shorter than goingahead (M(tracking) = 138.91, M(ahead) = 166.45, significant difference with p-value < 0.001) or random moving (M(track-ing) = 138.91, M(random) = 192.55, significant difference with p-value < 0.001).

The first simulation’s results show that in the case of blind evacuation due to smoke, tracking the wall brings the higher %of survivals, the lower toxic level, and the shorter time to escape than going ahead or random moving (Table 3). These

Table 3Summary on 95% confidence interval of output parameter values among three blind evacuation strategies.

Parameter Tracking of wall Going ahead Random moving

% Of survivals 47.7–55.5 33.8–36.9 20.0–23.7Toxicity level (%) 68.9–70.7 76.2–78.8 82.4–85.0Time to escape 136.0–141.7 162.9–169.9 188.8–196.2

56 M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59

simulation results confirm the advice of fire evacuation experts that in the condition of limited visibility, evacuees shouldtrack the wall instead of wander moving.

4.4. Optimisation of evacuation strategies in normal condition

This section optimises the evacuation strategies in normal condition: Section 4.4.1 presents the evacuation contexts andconsidered evacuation strategies; Section 4.4.2 presents and analyses the obtained results; Section 4.4.3 discusses about theobtained results regarding our setting up of simulations.

4.4.1. Evacuation context and strategiesIn the first experiment, we considered in a tide situation with assumption that the smoke has completely filled inside of

the building. The evacuee thus is no more able to observe anything in around to choose the direction. However, in a realisticsituation, especially at the beginning of the fire, the evacuee is still able to find the direction to evacuate because there is nottoo much of smoke. So the evacuee could dynamically change its evacuate strategy depended on the real situation: obser-vable or not. In the blind condition, the experiment 1 pointed out that it is better if the evacuee uses the strategy of trackingthe wall. We aims to see what is the best strategy in visible condition considering three main strategies to evacuate in a fireevacuation: follow the self (the own’s path), or follow the crowd, or follow the evacuation signs.

� Normal evacuation strategy 1: Evacuees follow their own path (Fig. 14a). The own path is the direction to the emergencyexit which the agent has seen the last time. Therefore the goal exit of agent does not change during evacuation. If agentdoes not see anything because of smoke, it tracks following the wall to find an exit.� Normal evacuation strategy 2: Evacuees follow the crowd (Fig. 14b). If agent can still see around, it observes the crowd and

follows it. If agent does not see anything because of smoke, it tracks following the wall to find an exit.

Fig. 14. Three normal evacuation strategies.

M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59 57

� Normal evacuation strategy 3: Evacuees follow the evacuation signs (Fig. 14c). If agent can still see around, it observes thesigns and follows its indication direction. If agent does not see anything because of smoke, it tracks following the wall tofind an exit.

In order to compare the three strategies in a more general and realistic situation, we run the simulation with the randominitiation of these parameters: the initial position of fire, the people’s fire exposure level, speed, observable range.

4.4.2. ResultsFollowing the time of simulation, the % of survivals in the case of following the evacuation signs is higher than those in the

case of following the crowd or following the own’s path (Fig. 15).The % of survivals (or death) at the end of simulation is shown in Fig. 16a: the % of survivals in the case of following the

evacuation signs is higher than that in the case of following the crowd (M(signs) = 82.48%, M(crowd) = 70.97%, significant dif-ference with p-value < 0.005) or following the own’s path (M(signs) = 82.48%, M(own path) = 58.55%, significant differencewith p-value < 0.001).

Likely, the average of toxicity level (in %) of survivals is also significant (as depicted in Fig. 16b): in the case of followingthe evacuation signs, people is poisoned lower toxic fumes than following the crowd (M(signs) = 18.46%, M(crowd) = 29.93%,significant difference with p-value < 0.005) or following the one’s own path (M(signs) = 18.46%, M(own path) = 42.22%, signif-icant difference with p-value < 0.001).

In the same tendency, the average time to escape (Fig. 16c) in the case of following the evacuation signs is significantlyshorter than following the crowd (M(signs) = 104.61, M(crowd) = 129.72, significant difference with p-value < 0.001) or fol-lowing the one’s own path (M(signs) = 104.61, M(own path) = 149.20, significant difference with p-value < 0.001).

To summary, these results show that in the case of following the evacuation signs, the % of survivals is higher, the averageof toxic level is lower, and the time to escape is shorter than following the crowd or following the own’s path (Table 4).

4.4.3. DiscussionsThese results are just obtained in only one case of initial population. The population for these simulations is initiated as

following aspects.First, the proportion of people using one of three evacuation strategies is approximately equal. It means that among 1000

people for each simulation, there are approximately one third people following each strategy: following the signs, followingthe crowd, and following the own’s path. The evacuation strategy of each evacuee is randomised, in each simulation, as longas the rate of people using each strategy is about one third.

We aim to run simulations with different proportions of people following each strategy. This enables us to test the effectsof initial proportion of people on the goodness of these three evacuation strategies. In other word, this leads us to verify thequestion whether following the signs is always better to fire evacuate in this supermarket with any proportion of peopleusing these three strategies. This is one of our works in the next period.

Second, the initial positions of people are also randomised. This may make us, in some circumstances, out of the realisticspatial distribution of people inside the supermarket. We thus aim to initiate the people spatial distribution as realistic aspossible by apply some intuitive rules such as: there are more people at the shelves of promotion or entertainment zones,there are more children at the shelves of toys, and more women at the shelves of food and cooking products, etc. By applyingthese facts, we could evaluate how the spatial distribution of people effect on the goodness of these three evacuation strat-egies. This is also one of our work in the near future.

An other field that this paper does not yet address is the optimisation at the configuration level. We also aim to run sim-ulations with different configurations of signs distribution, configurations of shelves positioning, configurations of emer-gency exits, etc. These potential simulations could help us to find out which configuration (or inside design of building) isbetter than the current one. This is also one of our next works in the line.

Fig. 15. The % of survivals by simulation steps among three standard evacuation strategies.

Fig. 16. Comparison of output parameters among three standard evacuation strategies.

Table 4Summary on 95% confidence interval of output parameter values among three evacuationstrategies in standard condition.

Parameter Signs Crowd Own’s path

% Of survivals 78.8–86.2 64.2–77.7 53.3–63.8Toxicity level (%) 14.5–22.4 23.3–36.6 36.8–47.6Time to escape 99.7–109.5 121.2–138.3 145.8–152.5

58 M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59

5. Conclusion and future works

This paper presented an agent-based model with Smoke Effect and Blind Evacuation Strategy (SEBES) for simulation offire evacuation inside buildings. This model respects a recommendation of fire evacuation experts on the fact that evacueesshould follow the boundaries of obstacles or walls to find the exits once their visibility is limited by smoke, by integrating themodel of smoke and its effect on the evacuee’s visibility, speed, and evacuation strategy.

The proposed model is then applied to simulate of human behaviours in a fire evacuation in a realistic plan of the Metrosupermarket of Hanoi with GIS data. The simulations are carried out with the optimisation in two conditions. First, normalevacuation strategies include: following the crowd, following the one’s own path, or following the evacuation signs. Second,blind evacuation strategies include: tracking the wall, go ahead, and random moving.

The simulations results indicated that following the evacuation signs is the best visible evacuation strategy in three givenones, and tracking the wall is the best invisible evacuation strategy in three given ones. This results confirm one more timethe effectiveness of a recommendation given by fire evacuation experts.

The simulations also provided an animation tool for fire evacuation experts to train people about fire evacuation in morevisible ways.

Developing a framework for simulation of various building realistic plan in order to optimise the evacuation plan or theevacuation configuration are some of our future research directions.

References

[1] E. Amouroux, C. Quang, A. Boucher, A. Drogoul, GAMA: an environment for implementing and running spatially explicit multi-agent simulations, in:10th Pacific Rim International Workshop on Multi-Agents (PRIMA), Thailand, 2007.

[2] F. Alavizadeh, C. Lucas, B. Moshiri, Adaptation of personality-based decision making to crowd behavior simulations, in: Proceedings of the 13thInternational Conference on Artificial Intelligence: Methodology, Systems, and Applications, AIMSA ’08, Springer-Verlag, Berlin, Heidelberg, 2008, pp.385–389.

[3] J.D. Averill, W. Song, Accounting for Emergency Response in Building Evacuation: Modeling Differential Egress Capacity Solutions, 2007.[4] C. Garca-Cabrera, P. Morillo, J.M. Ordua, Analyzing large-scale crowd simulations for building evacuation, in: Y. Demazeau, F. Dignum, J.M. Corchado, J.

Bajo, R. Corchuelo, E. Corchado, F.F. Riverola, V. Juliaacuten, P. Pawlewski, A. Campbell (Eds.), Trends in Practical Applications of Agents andMultiagent Systems – 8th International Conference on Practical Applications of Agents and Multiagent Systems, PAAMS 2010 Special Sessions andWorkshops, Salamanca, Spain, 26–28 April 2010, Ad vances in Soft Computing, vol. 71, Springer, 2010, pp. 591–598.

M.H. Nguyen et al. / Simulation Modelling Practice and Theory 36 (2013) 44–59 59

[5] A. Chaturvedi, A. Mellema, S. Filatyev, J. Gore, DDDAS for fire and agent evacuation modeling of the rhode island nightclub fire, in: Proceedings of the6th International Conference on Computational Science – Volume Part III, ICCS’06, Springer Verlag, Berlin, Heidelberg, 2006, pp. 433–439.

[6] M. Daito, N. Tanida, The influence of mental factors on fire evacuation behavior and the effect of introducing rfid and cellular phone, RCSS DiscussionPaper Series 100 (2010) 1–14.

[7] A. Filippoupolitis, L. Hey, G. Loukas, E. Gelenbe, S. Timotheou, Emergency response simulation using wireless sensor networks, in: Proceedings of the1st International Conference on Ambient Media and Systems, Quebec City, Canada, 2008.

[8] A. Filippoupolitis, An adaptive system for movement decision support in building evacuation, in: Proceedings of the 25th Int. Symposium on Computerand Information Sciences, London, UK, 2010, pp. 389–392.

[9] D. Gianni, G. Loukas, E. Gelenbe, A simulation framework for the investigation of adaptive behaviours in largely populated building evacuationscenarios, in: Organised Adaptation in Multi-Agent Systems workshop at the 7th International Conference on Autonomous Agents and MultiagentSystems, 2008.

[10] D. Hanea, H. Jagtman, S. Sillem, J. Wiersma, Modelling self-reliance, evacuation and fire fighting activities during a large fire in a public building – adynamic approach, Safety Science Monitor 11 (2) (2007) 1–17.

[11] D. Helbing, I. Farkas, P. Molnàr, T. Vicsek, Simulation of pedestrian crowds in normal and evacuation situations, in: M. Schreckenberg, S.D. Sharma(Eds.), Pedestrian and Evacuation Dynamics, Springer, Berlin, 2002, pp. 21–58.

[12] X. Hu, Z. Wang, F. Jia, E. Galea, M. Patel, Simulating smoke transport in large scale enclosure fires using a multi-particle-size model, in: Fire SafetyScience – Proceedings of the Tenth International Symposium 10, paper to appear in Fire Safety Science – Proceedings of the Tenth InternationalSymposium, University of Maryland, USA, held 19–24 June, 2011.

[13] P. Huang, J. Kang, J.T. Kider, B. Sunshine-Hill, J.B. McCaffrey, D.V. Rios, N.I. Badler, Real-Time Evacuation Simulation in Mine Interior Model of Smokeand Action.

[14] Y. Lin, I. Fedchenia, B. LaBarre, R. Tomastik, Agent-based simulation of evacuation: an office building case study, in: W.W.F. Klingsch, C. Rogsch, A.Schadschneider, M. Schreckenberg (Eds.), Pedestrian and Evacuation Dynamics 2008, Springer, Berlin Heidelberg, 2010, pp. 347–357.

[15] L. Luo, S. Zhou, W. Cai, M.Y.H. Low, F. Tian, Y. Wang, X. Xiao, D. Chen, Agent-based human behavior modeling for crowd simulation, ComputerAnimation and Virtual Worlds 19 (3–4) (2008) 271–281.

[16] T. Korhonen, S. Hostikka, Fire dynamics simulator with evacuation: FDS + Evac technical reference and user’s guide, English (119) (2010) 30–57.[17] E.D. Kuligowski, The process of human behavior in fires, Technology 1632 (May) (2009). NIST Technical Note.[18] E.D. Kuligowski, S.M.V. Gwynne, The need for behavioral theory in evacuation modeling, in: W.W.F. Klingsch, C. Rogsch, A. Schadschneider, M.

Schreckenberg (Eds.), Pedestrian and Evacuation Dynamics 2008, Springer, Berlin Heidelberg, 2010, pp. 721–732.[19] M. Okaya, T. Takahashi, BDI agent model based evacuation simulation, in: L. Sonenberg, P. Stone, K. Tumer, P. Yolum (Eds.), Proceedings of the 10th

International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2011), Taipei, Taiwan, May 2–6, 2011, vol. 1–3, 2011, pp. 1297–1298.

[20] M. Okaya, T. Takahashi, Human relationship modeling in agent-based crowd evacuation simulation, in: D. Kinny, J.Y. jen Hsu, G. Governatori, A.K.Ghose (Eds.), Proceedings of Agents in Principle, Agents in Practice – 14th International Conference, PRIMA 2011, Wollongong, Australia, November16–18, 2011, Lecture Notes in Computer Science, vol. 7047, Springer, 2011, pp. 496–507.

[21] X. Pan, C.S. Han, K. Dauber, K.H. Law, A multi-agent based framework for the simulation of human and social behaviors during emergency evacuations,AI and Society 22 (2) (2007) 113–132.

[22] W. Jirasingha, S. Patvichaichod, Modeling fire evacuation of a library building based on the numerical simulation, American Journal of Applied Sciences8 (5) (2011) 452–458.

[23] T. Saelao, S. Patvichaichod, The computational fluid dynamic simulation of fire evacuation from the student dormitory, American Journal of AppliedSciences 9 (3) (2012) 429–435.

[24] F. Qiu, X. Hu, Modeling group structures in pedestrian crowd simulation, Simulation Modelling Practice and Theory 18 (2) (2010) 190–205.[25] A. Rahman, A.K. Mahmood, E. Schneider, Using agent-based simulation of human behavior to reduce evacuation time, in: Proceedings of the 11th

Pacific Rim International Conference on Multi-Agents: Intelligent Agents and Multi-Agent Systems, PRIMA ’08, Springer-Verlag, Berlin Heidelberg,2008, pp. 357–369.

[26] A. Ren, C. Chen, J. Shi, L. Zou, Application of virtual reality technology to evacuation simulation in fire disaster, in: H.R. Arabnia (Ed.), Proceedings of the2006 International Conference on Computer Graphics and Virtual Reality, CGVR 2006, Las Vegas, Nevada, USA, June 26–29, 2006, CSREA Press, 2006,pp. 15–21.

[27] U. Rppel, K. Schatz, Designing a BIM-based serious game for fire safety evacuation simulations, Advanced Engineering Informatics (2011).[28] U. Ruppel, P. Abolghasemzadeh, Bim-based immersive evacuation simulations, in: 18th International Conference on the Application of Computer

Science and Mathematics in Architecture and Civil Engineering Digital Proceedings IKM 2009 CDROM Weimar Germany 7 (09).[29] A. Sagun, D. Bouchlaghem, C.J. Anumba, Computer simulations vs. building guidance to enhance evacuation performance of buildings during

emergency events, Simulation Modelling Practice and Theory 19 (3) (2011) 1007–1019.[30] H. Said, A. Kandil, H. Cai, Agent-based simulation of labour emergency evacuation in high-rise building construction sites, in: Construction Research

Congress (ASCE 2012), 2012, pp. 1104–1113.[31] V. Schneider, R. Konnecke, Egress route choice modelling concepts and applications, Fire Safety Engineering (2010) 619–625.[32] T.-S. Shen, S.-W. Chien, A building evacuation simulation model (EMS) for building evaluation, International Journal on Architectural Science 6 (1)

(2005) 15–30.[33] A. Shendarkar, K. Vasudevan, S. Lee, Y.-J. Son, Crowd simulation for emergency response using bdi agents based on immersive virtual reality,

Simulation Modelling Practice and Theory 16 (9) (2008) 1415–1429.[34] H. Suryotrisongko, Y. Ishida, Emergence of cooperation as the impact of evacuee’s solidarity, in: Proceedings of the 2011 IEEE International Symposium

on Safety, Security and Rescue Robotics. Kyoto, Japan, November 1–5, 2011.[35] F. Tang, A. Ren, Agent-based evacuation model incorporating fire scene and building geometry, Tsinghua Science Technology 13 (5) (2008) 708–714.[36] F. Tingyong, X. Jun, Y. Jufen, W. Bangben, Study of building fire evacuation based on continuous model of FDS and EVAC, in: International Conference on

Computer Distributed Control and Intelligent Environmental Monitoring, 2011, pp. 1331–1334.[37] J. Tsai, N. Fridman, M. Brown, A. Ogden, I. Rika, X. Wang, S. Epstein, A. Zilka, M. Taylor, M. Tambe, E. Bowring, S. Marsella, G.A. Kaminka, A. Sheel,

ESCAPES – evacuation simulation with children, authorities, parents, emotions, and social comparison, in: Proceedings of the Tenth International JointConference on Autonomous Agents and Multi-Agent Systems (AAMAS-11), 2011.

[38] Y. Wang, L. Zhang, J. Ma, L. Liu, L. Zhang, D. You, Combining building and behavior models for evacuation planning, IEEE Computer Graphics andApplications 31 (2011) 42–55.

[39] Y. Weifeng, T.K. Hai, A model for simulation of crowd behaviour in the evacuation from a smoke-filled compartment, Physica A: Statistical Mechanicsand its Applications (2011).

[40] S. Yi, J. Shi, An agent-based simulation model for occupant evacuation under fire conditions, in: Proceedings of the 2009 WRI Global Congress onIntelligent Systems, GCIS ’09, vol. 01, IEEE Computer Society, Washington, DC, USA, 2009, pp. 27–31.

[41] E. Guideline, Fire Safety Engineering Concerning Evacuation from Buildings, Tech. Rep., CFPA-E No. 19:2009, 2009.[42] K. Kang, A smoke model and its application for smoke management in an underground mass transit station, Fire Safety Journal 42 (2007) 218–231.[43] B.S. Institute, The Application of Fire Safety Engineering Principles to Fire Safety Design of Buildings. Part 6: Human Factors: Life Safety Strategies

Occupant Evacuation, Behaviour and Conditions (Sub-system 6), Tech. Rep., PD7974-6:2004, 2004.[44] B.o.E.S. Committee on Fire Toxicology, N.R.C. Toxicology, Fire and Smoke: Understanding the Hazards, The National Academies Press, 1986.