You are on page 1of 20

12 Fire Technology, 45, 431–449, 2009

Ó 2008 Springer Science+Business Media, LLC. Manufactured in The United States


DOI: 10.1007/s10694-008-0064-6

An Artificial Neural-network Based


Predictive Model for Pre-evacuation
Human Response in Domestic Building Fire

S. M. Lo*, M. Liu and Richard K. K. Yuen, Department of Building &


Construction, City University of Hong Kong, Tat Chee Avenue, Kowloon
Tong, Hong Kong
P. H. Zhang, Faculty of City and Environmental Engineering, Shenyang
Jianzhu University, Shenyang, PeopleÕs Republic of China

Received: 21 February 2006/Accepted: 16 June 2008

Abstract. The post-1993 WTC attack study (Proulx and Fahy, In: Proceedings of
ASIAFLAMÕ95—An International Conference on Fire Science and Engineering,
Hong Kong, 1995, pp 199–210) revealed that occupants took 1–3 h to leave the
110-storey buildings, and the pre-movement reactions could account for over two-
thirds of the overall evacuation time. This indicates that a thorough understanding of
the pre-evacuation behavioral response of people under fire situations is of prime
importance to fire safety design in buildings, especially for complex and ultra high-
rise buildings. In view of the stochastic (the positions of the occupants) and fuzzy
(uncertainty) nature of human behavior (Fraser-Mitchell, Fire Mater 23:349–355,
1999), conventional linear and polynomial predictive methods may not satisfactorily
predict the peopleÕs response. An alternative approach, Adaptive Network based
Fuzzy Inference System (ANFIS), is proposed to predict the pre-evacuation behavior
of peoples, which is an artificial neural network (ANN) based predictive model and
integrates fuzzy logic (if-then rules) and neural network (based on back propagation
learning procedures The ANFIS learning architecture can be trained by structured
human behavioral data, and different fuzzy human decision rules. The applicability in
simulating human behavior in fire is worth exploring.

Keywords: pre-evacuation human behavior, predictive model, fuzzy-neural network

1. Introduction
In Hong Kong and many large cities in China, numerous high-rise buildings have
been recently been constructed. In view of the rapid growth of population, the
construction of high-rise buildings will continue. The USÕs 9/11 incident and many
other large fire tragedies in the past few years in Hong Kong and China, such as
the fires in Garley Building (40 death), Top-One Karaoke (15 death) and Luoyang
Dongdu Commercial Building (309 death), have attracted the attention of the
government on building fire safety issues.

* Correspondence should be addressed to: S. M. Lo, E-mail: bcsmli@cityu.edu.hk


432 Fire Technology 2009

Traditionally, the construction and spatial design of a building is governed by


the prescriptive building and fire codes. However, the prescriptive rules are usu-
ally enforced for individual components. They may not guarantee efficient fire
protection in modern high-rise buildings with complex setting. The use of per-
formance-based approach is thus advocated in recent years in many countries to
encourage a holistic approach for fire safety design. Fire safety engineering
approach, which involves the use of fire and evacuation models to predict the
fire and smoke effect on the occupantsÕ movement as well as the structure of the
buildings, has been adopted to determine the fire safety performance of a build-
ing. In many fire safety engineering studies, prediction of occupantsÕ behavioral
reactions and their egress process (by using egress models, such as the Spatial-
Grid Evacuation Model [1, 2]) are important tasks for assessing the safety of
people in buildings.
Previous studies indicated that behavioral reaction of people may have signifi-
cant effect on the fire evacuation process [3–15]. The pre-evacuation reactions of
people may in some incidents account for a large portion of evacuation time [9].
A comprehensive study on the pre-evacuation behavioral reaction of people
under fire is necessary, and the dense population and tall buildings will urge
such study to be performed in China and Hong Kong [14]. Although the studies
on pre-evacuation human behavior have been performed for many years, data
that can be adopted to establish predictive models are still insufficient. Many
studies in pre-evacuation human behavior focused on what certain people did
and in what sequence (i.e. who and what) [13]. However, in many of such stud-
ies, the data and information collected are mainly dependent on special fire
cases and concurrent building environment. Establishment of a quantitative
pre-evacuation behavioral predictive model on the basis of such data may be
difficult.

2. ANN Predictive Technique


Predictive modeling can be drawn from statistics, database techniques, optimiza-
tion techniques, machine learning, etc. In the past two decades, the use of neu-
ral network approaches for data mining has been well established [16–18]. An
artificial neural network (ANN) consists of an interconnected group of artificial
neurons and processes information using a connectionist approach to computa-
tion. In most cases, an ANN is an adaptive system that changes its structure on
the basis of external or internal information that flows through the network dur-
ing the learning phrase. Training a neural network model basically performs by
selecting one model with minimum cost from the set of allowed models. The
input/output training data are fundamental in neural network technology
because they convey the necessary information to ‘‘discover’’ the optimal operat-
ing point. An input is presented to the neural network and a corresponding
desired or target response is set at the output. An error exists owing to the dif-
ference between the desired response and the system output. The error informa-
tion is then fed back to the system and the system parameters are adjusted in a
Pre-evacuation Human Response in Domestic Building Fire 433

system fashion. The process is repeated until the performance is acceptable.


When training is completed, the ANN parameters are fixed. However, a critical
issue concerning neural networks is the over-fitting problem [18]. This phenome-
non may be due to the fact that a neural network training phrase captures not
only useful information contained in the given data set, but also unwanted
noise. Thus, the validation of the output of the trained ANN becomes neces-
sary. It could be done by comparing the output with a set of new data which
have not been not employed for the training.

3. Adaptive Network Fuzzy Inference System


Most of the human behavioral reaction data in fire scenarios are information
dependent and described in terms of actions depending on the state of the envi-
ronment [19]. The decision rules may be represented by the probabilistic values
assigned. As the data acquired are situational dependent and the sample size may
not be extensive, the probabilistic values derived from the limited samples may be
imprecise (with low confidence level). Training for an ANN model becomes diffi-
cult. The use of fuzzy inference system, which adopts the fuzzy if-then rule, is an
alternative approach to predict the special human behavioral data [20–23]. When
employing the fuzzy inference system, the major key tasks are to formulate an
appropriate approach for:

(a) transforming the human behavioral responses into the rule-base data; and
(b) calibrating the membership functions (MF) so as to minimize the output error.

In this perspective, we adopt the architecture—Adaptive Network Based Fuzzy


Inference System (ANFIS), advocated by Jang [24], to serve as a basis for con-
structing a set of fuzzy if-then rules with appropriate MF to establish the predic-
tive model for human pre-evacuation behavioral response in fire.
Adaptive networks have been widely applied in the field of modeling, decision-
making, signal processing and control [24]. It is established with initial pre-estab-
lished MFs of input–output parameter pairs of the network and pre-defined fuzzy
if-then rules. On the basis of back propagation learning procedures, ANFIS can
complete the training and learning process on the adaptive network. During the
training process, it can improve the pre-established MFs and refine the fuzzy
inference if-then rules. The trained algorithm can be adopted to deduce the
human pre-evacuation behavioral actions (output parameter) based on different
human behavioral characteristics and different environmental situations (input
parameter).
For example, we may assume that the input parameters of an ANFIS is x, y,
and the output parameter is z. The pre-defined fuzzy inference if-then rules on the
basis of Takagi and Sugeno fuzzy rule set [24] can be illustrated in Figure 1
below:
434 Fire Technology 2009

A1 B1
W1

f1 p1 x q1 y r1 w1 f 1 w2 f 2
X Y f
w1 w2
_ _
A2 B2 w1 f 1 w2 f 2
W2

f2 p 2 x q 2 y r2
X Y
x y

Figure 1. Fuzzy inference system [23].

Rule 1: if x is A1 and y is B1, then f1 = p1x + q1y + r1


Rule 2: if x is A2 and y is B2, then f2 = p2x + q2y + r2

which, A1, A2: fuzzy set corresponding to x; p1, p2: membership degree for input
signal x belongs to fuzzy set A1, A2; B1, B2: fuzzy set corresponding to y; q1, q2:
membership degree for output signal y belongs to fuzzy set B1, B2; r1, r2: con-
stants; f1, f2: output signal rule 1 and rule 2.
The architecture and learning procedures, on the basis of JangÕs works [24], are
shown in Figure 2 and Table 1.

4. Data Collection
Previous studies [3–5, 15] have found that evacuation is not necessarily the first
action of people under fire. They tend to affirm the threat before deciding on their
action. Bryan [11] has identified in his study that the behavioral reaction of people

Layer 1 Layer 3 Layer 5


Layer 2 Layer 4

A1
W1
1 ϕ1
1
x Π Ν

A2
ϕ
Σ
B1
y
Π Ν 2 ϕ2

B2 W2 2

Figure 2. ANFIS architecture [23].


Pre-evacuation Human Response in Domestic Building Fire 435

Table 1
Description of the ‘‘Layers’’ in ANFIS Architecture with Respect to
Figure 1

Layer Description

1 Every node i in this layer is a node with a node function representing the membership function
lAi(x) of Ai and it specifies the degree to x which the given satisfies the quantifier
Ai (i = 1, 2 here).
2 Every node in this layer is a node labeled P which multiplies the incoming signals and sends
the products out. Each node output represents the firing strength of a rule.
3 Every node in this layer is a node labeled N. The i-th node calculates the ratio of the i-th ruleÕs
firing strength to the sum of all rulesÕ firing strengths (i = 1, 2 here).
4 Every node i in this layer is a node with a node function: Ŵiui = Ŵi (pi x + qi y + ri )
where Ŵi is the output of layer 3, and (pi + qi + ri) is the parameter set. Parameters in
this layer will be referred to as consequent parameters (i = 1, 2 here).
5 The single node in this layer is a node labeled R that computes the overall output as the
summation of all incoming signals.

under a fire situation is complicated and a series of decision-making processes


involving recognition, validation, definition and evaluation may be undertaken.
Figure 3 shows the simplified process.
Post-fire investigation is a common approach to collect human reaction data. It
may be argued that this technique relies on what people have said rather than
what they have done (i.e. less reliable for results based on autobiographical mem-
ory). However, the reasons for using this method to obtain the behavioral infor-
mation of people under real fire situations are as follows:

(i) This is the only possible method to obtain the behavioral information about
people in real fire situations. It is difficult to observe what people have done
during a real fire because it is unlikely that any building has a full CCTV sys-
tem to ‘‘record’’ the reactions of people in every part of the building and that
it is not destroyed in the fire! And the view remains unobstructed throughout
the fire.
(ii) If information is based on the observations of firemen, the information con-
cerning the situation in the period between ignition and the arrival of the fire-
men may not be collected. Furthermore, the observation of firemen cannot
completely reflect the real status of the evacuees as they have been engaged in
firefighting.
(iii) A fire drill exercise does not reflect the real world situation.
(iv) It is impossible to set fire in a building to ‘‘test’’ the behavioral reaction of
people as this may endanger the people concerned.

Both interview and questionnaire techniques are reasonable approaches in post-


fire investigation, while the reasonability and effectiveness of these approaches
become the key factors. In order to collect accurate information and to assist the
436 Fire Technology 2009

Previous
Role experience in Education level
fire

Awake Asleep

Commitment behaviour
Pre-fire activity

YES YES

Initial stage Recognition Validation and definition Evaluation

NO NO

Type of cue Ignore the cue Seek for more information Decision
Making

Evacuate

Fight the fire

Alert others

Other actions

Figure 3. simplified human reaction process.

interviewees in their recall of the fire situation, interviews conducted in person1


shortly after the fires are preferred. Interviewees are required to answer structured
questions which are established in accordance with the decision-making process as
shown in Figure 3. Questions are shown in Table 2.
The scope of this study is limited to investigate the human reaction in fires of
high-rise domestic buildings. In Hong Kong, over 90% of the people live in the
high-rise buildings. Three cases were summarized in Table 3 and all the buildings
were developed by the Hong Kong Housing Authority. The building age, building
form and layout are very similar. At each floor of the buildings, the flats are
accessed through an internal corridor and the occupantsÕ living patterns are typi-
cal for low-income families in Hong Kong. In view of the similarities, the fires
happened in these buildings were selected for conducting the survey. A total of
150 occupants in three fire cases were interviewed, and 65% of the interviewees
were female and 35% were male.

1
The advantages of using face-to-face interviews can be summarized as: (i) it provides an opportunity to
establish rapport with the subject and to stimulate the trust and cooperation; (ii) it provides an oppor-
tunity to help the subject in their interpretation of the questions; and (iii) it allows flexibility in determining
the wording and subsequence of the questions by giving greater control over the situation [25]
Pre-evacuation Human Response in Domestic Building Fire 437

Table 2
Outline of the Structured Questions

Questions Structured answer

Initial conditions 1. Gender 0. Female, 1. Male


2. Age 1. <16, 2. 16–30, 3. 31–45, 4. 46–55, 5 56–65, 6.
>65
3. Education level 1. Primary or below
2. Secondary
3. Post-secondary/post-graduate
0. Others (please describe)
4. Accompany situation 1. Alone
when fire occurred? 2. With family or friend, small group size of
members less than three at similar age level
3. With family, large group size of members more
than three, with children or elderly persons
0. Others (please describe)
5. Initial location in 1. Far below the fire floor
building when fire 2. Immediate below the fire floor
occurred? 3. At the fire floor
4. Immediate above the fire floor
5. High above the fire floor
0. Others (please describe)
6. Are you familiar with 1. Know nothing
the building evacuation 2. Know a little
facility and emergency 3. Familiar
plan? 0. Other (please describe)
Pre-fire activities 7. What was your initial 1. Drunk
status when fire hap- 2. Asleep
pened? 3. Entertainment action
4. Stayed leisurely
5. On duty for security or etc.
0. Others (please describe)
Role 8. What is your ‘‘position’’ 1. Father
in your family? 2. Mother
3. Grandfather
4. Grandmother
5. Children
0. Others (please describe)
9. Are you a management 1. Yes
staff of the building? 2. No
Past experience 10. Did you ever experi- 1. Always experienced false alarm and no fire/
ence fire event or par- training experience
ticipate in fire safety 2. Never experience fire/training experience/
training such as fire knowledge
drill exercise? 3. Acquire a little fire knowledge from TV, etc. but
no fire experience
4. Have formal fire training (including fire drill
exercise) but no fire experience
5. Have formal fire training and have real fire
experience
0. Others (please describe)
438 Fire Technology 2009

Table 2
continued

Questions Structured answer

Recognition 11. What was the initial 1. Alerted by fire fighters


cue? 2. Feel extremely hot/heat
3. By noises/abnormal sound (except fire bell)
4. Burnt smell
5. Alerted by adjacent occupiers
6. Alerted by family members at the same flat
7. By smoke
8. By flame
9. By fire alarm/warning system
0. Others (please describe)
Validation 12. Did you attempt to 1. Yes
and definition seek for more informa- 2. No
tion? 0. Others (please describe)
Evaluation/ 13. What was your initial 1. Attempt to fight against the fire
decision reaction after confirm- 2. Inform Others/call for help
ing the fire cue? 3. Collect personal belongings/get dress, etc.
4. Escape immediately
5. Stay in the flat and wait for help
0. Others (please describe)
14. What did you do after 1. Try to fight against the fire
the initial action? 2. Inform Others/call for help
3. Collect personal belongings/get dress, etc.
4. Escape immediately
5. Stay in the flat and wait for help
0. Others (please describe)
Aftermath 15. Did you leave the 1. Evacuated to safety region by myself
building and how? 2. Rescued by fireman or others
3. Stay in the flat
0. Others (please describe)

Remarks: occupants in the room of fire origin have not been included in the surveys; the interviews were conducted
in the occupantsÕ native language and verbal explanation about the meanings of the questions has been given (i.e. the
structured questions were presented in Chinese).

5. Input and Output Membership Functions for the ANFIS


Data collected from the post-fire surveys were used for the training of the ANN.
The data set was divided into two batches. One was used for training the net-
works and the other served as ‘‘checking data’’.
People will start to react to an incident when they have perceived a fire cue,
such as the fire bell, the smoke smell, noise from other people, etc. In modern
multi-ownership residential buildings in Hong Kong, people commonly live in a
fire-fight enclosed units—the flats. It is not easy for them to know the activities in
other units or to have any response to an incident in other units unless they have
been ‘‘informed’’ by the cue or they preceive the dangerous scenario (e.g. flame or
smoke) themselves. Accordingly, in Hong KongÕs multi-ownership residential
Pre-evacuation Human Response in Domestic Building Fire 439

Table 3
Summary of the Fire Cases

Location of Building Fire broke


fire case Building age Height form Date out time Detailed scenes

Lei Muk Constructed About 20 Rectangular 16 Oct, Around Fire broke out in a flat
Shue in 1970s storey block with 1996 9:15 p.m. at 4th floor and 80
Estate internal people evacuated.
corridor
Wah Fu Constructed About 20 Rectangular 27 Oct, Around Fire broke out in a flat
Estate in 1970s storey block with 1997 2:00 p.m. at 16th floor. Adja-
internal cent occupiers noted
corridor the unusual smell,
sought further infor-
mation and called the
emergency centre.
Sau Mau Constructed About 20 Rectangular 31 Dec, 8:30 a.m. Fire broke out in a flat
Ping in 1970s storey block with 1996 at 3rd floor. Adjacent
Estate internal occupiers noted the
corridor unusual smell, the
smoke and called the
emergency centre.

Remark: the flats are occupied by families of similar income group as they are government subsidiary housing.

building fires, the awareness of the occupants would be a major factor affecting
their subsequent responses.
Whether a person will response to a fire cue will depend on his or her level of
recognition to the cue. The recognition level will depend on the intensity of the
cue, and is inversely related to the level of attention on the activities that he or
she is performing at that time. In other words, the degree of recognition is a
function of ‘‘personal characteristics’’, ‘‘motivational factor’’ and ‘‘signal factor’’
(Figure 4 refers). Personal factor refers to factors such as age, physiological char-
acteristics, and etc. Motivational factor is considered as the peopleÕs pre-fire activ-
ity and their previous fire training and/or fire experience. Signal factor is the
means by which the occupants receive the cue.
In order to facilitate the adaptive network training and checking, the above-
mentioned influencing parameters are grouped into four fuzzy input parameter
categories, namely age/physiological characteristics, previous fire experience/train-
ing, pre-fire activity and cue acquirement. The parameters of input–output pairs
are discussed as follows.

5.1. Age, etc.


Reaction of people is a cognitive process, which involves perception, memory and
thinking. People use cognition to understand what is going on [26]. The difference
in age level will represent differential living experience and responsibility/role in a
family. This implies that the age may have some influence on the reaction pattern
440 Fire Technology 2009

Responses Recognition/ attention

Depends upon

Categories of
parameters Signal factors
Personal factors Motivational factors

Basic Age, Previous


parameters physiological Cue acquirement means
experience, Pre-fire activity
affecting the characteristics, etc training, etc
reaction

Negative effect Positive effect


(concentrating in (duty of the Fire bell/ sign Smoke, flame, Alerted by
completing a person, e.g. as from warning heat, etc others
task) safety officer) system, etc

Figure 4. Influence parameters of recognition/attention process.

of people. It was proved by the information [2, 19] that large proportion of people
committing fire fighting would increase with age excepting children and old peo-
ple. It was also believed that children and older people will likely follow others
action [8]. Accordingly, x1 is introduced to represent the different age level of
occupants, of which three groups are specified—young, adult and old [(1) refers],
and their corresponding MF are shown in Figure 5.
8
< 1 - - - - young
x1 ¼ 2 - - - - adult ð1Þ
:
3 - - - - old

Figure 5. Membership function of age.


Pre-evacuation Human Response in Domestic Building Fire 441

5.2. Previous Fire Experience/Training, etc.


x2 is the parameter used to represent the occupantsÕ fire experience/training level
[shown in (2) below]. It is considered [11] that people will attempt to validate the
cue and ‘‘match’’ the information that they have received with their memory.
They tend to react to what they believe is happening in accordance with their past
experience and an optimistic expectation. People with past experience in fire inci-
dents may have longer first reaction times [27] and they may believe that they can
cope with the threat and so be less likely to leave immediately. They may be no
more likely to contact the fire department than those who do not have experience
before. They may be, however, more likely to fight the fire or direct families or
friends to evacuate rationally. On the other hand, if one who has frequently expe-
rienced false alarms may tend to ignore the cue (Figure 6).
8
< 1 - - - - poor
x2 ¼ 2 - - - - general ð2Þ
:
3 - - - - well

5.3. Pre-fire Activity, etc.


Canter [8] has mentioned that cue reception can be a function of pre-fire activity.
When a person is engaged in a specific activity (e.g. playing computer games,
watching TV, etc.), he or she may tend to complete the activity before reacting to
something else unless the intensity of the cue is so huge that can arouse his or
her attention. This commitment behavior is an important factor that affects the

Figure 6. Membership function of fire experience/training.


442 Fire Technology 2009

Figure 7. Membership function of pre-fire status.

reaction of people [10]. On the other hand, when people are awake and not
‘‘immerged in special activities’’, they seem to have more rational reactions and be
more efficient in making first decision as compared with those who are asleep or
unawake. For people who are asleep or unawake, they may be alerted to ‘‘semi-
awake’’ status, and their first reactions would mostly be to seek more information
or simply follow others action. Accordingly, pre-fire status is represented by x3
and is shown in (3) (Figure 7).
8
< 1 - - - - unawake
x3 ¼ 2 - - - - involved ð3Þ
:
3 - - - - awake

5.4. Cue Acquirement/Signal Level


The means of cue acquirement is represented by x4 (shown in (4)). It is the recog-
nition process and can be defined into two categories: direct and indirect. The rec-
ognition process involves ‘‘attention’’ in the cognitive process. It is obvious that
reaction may be delayed if people do not perceive the cue as indicative of a dan-
gerous situation. If someone acquires the cue directly (by fire bell, burning smell
or flame/smoke), he or she will make the decision on the basis of his or her per-
ception and cognition or the role in the group. The ambiguous nature of threat
cues indicates that individuals who do not have specific fire prevention or fire pro-
tection education and experience recognise only large amount of smoke or threat-
ening flame as indicative of a dangerous situation [9]. Accordingly, if an
ambiguous cue (e.g. fire bell, noise, burnt smell, etc.) is received, searching for fur-
ther information may be the peopleÕs first reaction. Moreover, the ‘‘avoidance
behaviour’’ of the people may cause them to assume that the threat is of mild nat-
ure [8] and they may continue their work or simply ignore the cue unless they
have past fire safety training. Accordingly, it is considered whether the cue can
Pre-evacuation Human Response in Domestic Building Fire 443

Figure 8. Membership function of direct/indirect cue acquirement.

arouse the attention of the people and whether it can have sufficient ‘‘quality’’ to
assist the people to ‘‘define’’ the level of danger (Figure 8) .

1 - - - - indirect
x4 ¼ ð4Þ
2 - - - - direct

6. ANFIS Predictive Model


An output vector y in six constant MF is assigned to the ANFIS based pre-move-
ment predictive model, which represents the peopleÕs initial reaction with respect
to the input parameter, set I, and their corresponding membership function, set S.
As the gender is not an imprecise input, the training and checking data are clus-
tered by the gender at the starting point of training, where female is assigned 0
and male is assigned 1. Thus two ANFIS models are established and the training
results are subsequently integrated to perform further analysis.
8
>
> 1 - - - - Attempt to fight the fire
>
>
>
> 2 - - - - Inform others/ call for help (e.g. dial emergency call)
<
! 3 - - - - Collect personal belongings
y = F ( I ; S) =
>
> 4 - - - - Escape immediately
>
>
>
> 5 - - - - Seek for additional information
:
6 - - - - Other actions
ð5Þ
444 Fire Technology 2009

7. Fuzzy if-then Rules


Rules are selected by judgment at the beginning of the training. Fuzzy if-then
rules (fuzzy conditional statements) are adopted and they are expressions of the
form IF A THEN B, where A and B are labels of fuzzy sets characterised by
appropriate MF. Fuzzy if-then rules, in view of their concise form, can be
employed to capture the imprecise modes of reasoning that play an essential role
in the human ability to make decision in an environment of uncertainty and
imprecision [24].
Typically, there are two types of fuzzy if-then rules. One type has both its input
and output variable as fuzzy set characterized by MF. Another type of fuzzy if-
then rules, proposed by Takagi and Sugeno [23], has fuzzy sets involved only in
the premise part. By using the Takagi and SugenoÕs fuzzy if-then rules, we
describe the peopleÕs pre-movement reaction, the initial rules, as follows:

Rule 1: If Age is young, fire experience is limited and fire cue is intensive flame
or smoke and acquired directly, then initial reaction is to inform others.
Rule 2: If fire experience is good, pre-fire activity is awake and not concentrate
at a task, and fire cue is intensive flame or smoke and acquired directly then
initial reaction is positive, fire fighting or alert others, and not searching for
more information.
Rule 3: If fire experience is good, pre-fire activity is awake and not concentrate
at a task and fire cue is ambiguous (e.g. noise, burnt smell, etc.—acquire indi-
rectly) then initial reaction is to search for more information.

Each rule is assigned initial weighting of 1 at the beginning of training. When


training is completed, the fuzzy inference system will learn and generate more
rules of different weightings. It is not necessary to introduce too many rules ini-
tially as the training process will generate a number of additional rules. Too many
initial rules may cause the network to be hung up. The obtained fuzzy rule sets
can be used to predict peopleÕs pre-movement reaction.

8. ANFIS Architecture for Learning and Learning Result


The back propagation algorithm [18, 24] is selected for adjusting the weights of
the feed-forward network such that the network learns the functional mapping of
a set of inputs to the desired outputs. The back propagation approach is an opti-
mization procedure based on the gradient descent method. There are two phases
to the back propagation training algorithm: a forward pass followed by a reverse
pass. In the first phase, the forward pass, the input is presented at the input layer
and is propagated forward through the network. The current output of the net-
work is compared with the target or desired output and the ‘‘error’’ is determined.
In the second phase, the backward pass, the ‘‘error’’ is propagated back through
the network so that the weights may be adjusted to minimize the error. Besides
that, logical operation ‘‘and’’ adopts the ‘‘product’’ method, while operation ‘‘or’’
adopts the ‘‘probabilistic’’ method. The whole ANFIS structure after training is
Pre-evacuation Human Response in Domestic Building Fire 445

Input Rules
Weight sum
parameter
Input MF Output MF
Output

x4: Cue acquirement

x3: Pre-fire Activities

x2:Fire experience

x1: Age

Normalization factor

Figure 9. ANFIS architecture for the predictive algorithm on human


response in fire.

shown in Figure 9 and some fuzzy inference if-then rules are obtained through
network learning and training. The comparison between checking data and AN-
FIS output after 450 epochs of training is shown in Table 4.
As the data for training the ANFIS networks are limited, the input and output
responses are grouped respectively into four and six categories. The result pre-
sented in this article serves as an illustration in that ANFIS algorithm can be a
reasonable approach to predict human responses. With continuous collection of
data in Hong Kong and Mainland China, we can refine the MF and the learning
rules. Moreover, other parameters, such as the peopleÕs initial position before the
outbreak of fire, the social relationship between occupants and the role in the
family, etc., can also be included in the ANFIS model.
Table 5 summarizes the percentage of output responses obtained by the post-
fire surveys and ANFIS-based predictive model output. On the basis of the post-
fire surveys of the three cases, the percentage (42.8% from the predictive results)
of people seeking for further information once fire cue had been received was
high. The amount of people attempting to fight against the fire was very little. The
results are different from the surveys carried out by Wood in UK [4] (about 15%
of people may attempt to fight against the fire). This may be due to the fact that
446 Fire Technology 2009

Table 4
Comparison of ANFIS Output and Checking Data

Fire Pre-fire Cue Pre-movement reaction ANFIS


P-No. Sex Age experience activity acquirement obtained by survey output

1. 0 4 2 2 5 4 4
2. 0 3 4 2 9 6 5
3. 1 1 3 2 6 4 3
4. 0 1 3 2 6 4 3
5. 0 3 2 3 9 5 5
6. 1 3 4 2 6 4 4
7. 0 3 2 2 3 5 4
8. 0 3 4 2 6 4 5
9. 1 3 4 3 3 5 5
10. 0 3 2 2 6 5 5
11. 1 3 3 3 7 3 5
12. 0 2 3 5 7 2 2
13. 0 2 3 4 7 3 3
14. 1 3 3 4 7 5 4
15. 1 3 4 5 6 1 2
16. 0 4 4 2 3 5 5
17. 1 2 4 4 3 5 6
18. 1 3 3 2 5 5 5
19. 0 3 2 2 5 5 4
20. 0 3 2 2 5 6 6

Remarks: P-No.: Numbering of persons; Pre-movement reaction obtained by survey: checking data not used for
training of the network; For illustration purpose, only 20 sets of data are shown in the table.

Table 5
Summary of Response

Average percentage of people undertaking this as their first


action
Pre-movement initial
behaviour category Survey resulta ANFIS output ±%b

1. Fight against the fire 4.8 0 4.8


2. Inform others/call for help 9.6 14.3 4.7
3. Collect personal belongings 4.8 9.5 4.7
4. Escape immediately 28.5 23.9 4.6
5. Seeking further information 42.8 42.8 0
6. Others actions 9.5 9.5 0

a
The survey results are from a set of checking data (randomly selected from the whole sample) that have not been
used for the training.
b
With reference to the whole people that are affected by the fire.

the people in Hong Kong live in multi-ownership apartments in which residential


units are accessible through a common fire protected corridor (i.e. the corridor is
enclosed by fire rated walls, floors, doors, etc.). The first cue that the people will
Pre-evacuation Human Response in Domestic Building Fire 447

receive may be ambiguous and their reaction to seek further information is not
surprising. Moreover, in multi-storey apartments, most of the people may be diffi-
cult to locate the fire and thus only a few of them, who may be the management
staff or the residents at the vicinity, may attempt to fight against the fire.

9. Concluding Remarks
In this study, an Adaptive Network Fuzzy Inference System based algorithm has
been proposed to predict the pre-movement reaction of people in fire. Human
behavioral data are imprecise in nature [20–22]. The uncertainty of human behav-
ior is due to vagueness or fuzziness rather than due to randomness alone. It is
understood that the decision of a person for an action can be regarded as a ran-
dom event. However, the information or the situation affecting his or her decision
is in most cases imprecise. Therefore, the description to the data cannot be precise
and is fuzzy in nature. Thus, the use of fuzzy sets, which expressed by gradual
transition membership, to represent the data value appears possible. The ANFIS
based predictive model, which is based on fuzzy inference algorithm, is therefore
adopted.
The training results show that such approach can be applied in manipulating
the limited data of human reaction in fire and can serve as part of an evacuation
model to predict the initial response of the evacuees.
It is well-known that noisy sparse data sets with highly complex nonlinear rela-
tionships, such as the human reaction data, are better candidates for ANN than
traditional statistical models [24]. The ANFIS algorithm is a common architecture
of ANN for manipulating vague and imprecise data. However, there is still no
established procedure for determining if the ‘‘best’’ ANN models have been
achieved. Finding the best combination of a network topology and parameters is
still an art. It is taken that better performing models may be developed by more
successful selection of parameters and architecture. This will require further inves-
tigation on the use of ANN predictive models in fire safety field. Establishing a
reliable predictive model for human reaction in fires requires extensive data collec-
tion. Insufficient data will render the most promising predictive model inefficient.
Apart from developing effective predictive algorithm, organizing extensive data
collection surveys should also be performed. In order to refine the ANFIS based
predictive model, more human reaction data will be collected. Establishing reason-
able MF, which represents the possibility level of an input/output, will be another
task to improve the effectiveness of the predictive model.

Acknowledgements
The work described in this paper was fully supported by a grant from the
Research Grant Council of the Hong Kong Administrative Region, China [Project
No. CityU1207/02E]. The authors also acknowledge the assistance of Ms. S. Y.
Chan and Ms. V. Wong for the collection of data.
448 Fire Technology 2009

References
1. Proulx G, Fahy R (1995) A study of the New York Trade Centre evacuation. In: Pro-
ceedings of ASIAFLAMÕ95—an international conference on fire science and engineer-
ing, Hong Kong, pp 199–210
2. Lo SM, Fang ZA (2000) Spatial-grid evacuation model for buildings. J Fire Sci
18(5):376–394
3. Wood PG (1972) The behaviour of people in fires. Department of Environment and
Fire OfficersÕ Committee, Joint Fire Research Organization, UK
4. Wood PG (1980) A survey of behavior in fires. In: Canter D (ed) Fires and human
behaviour. John Wiley & Sons, Ltd, Chichester, pp 83–97
5. Bryan JL (1977) Smoke as a determinant human behaviour in fire situation, Depart-
ment of Fire Protection Eng., University of Maryland, College Park, MD
6. Canter D, Breaux J, Sime J (1980) Domestic, multiple occupancy and hospital fires.
Fires and human behaviour, John Wiley & Sons, Ltd, Chichester 117–137
7. Sime J (1980) The concept of panic. Fires and human behaviour, John Wiley & Sons,
Ltd, Chichester 63–81
8. Canter D (1985) Studies of human behaviour in fire: empirical results and their implica-
tions for education and design. Building Research Establishment Report, UK
9. Sime J (1992) Human behaviour in fires—summary report. Report No. 45, Building
Use & Safety Research Unit, School of Architecture, Portsmonth Polytechnic, UK
10. Proulx G (1994) Human response to fires. Fire research news. Natl Res Counc Can
71:1–3
11. Bryan JL (1995) Behavior response to fire and smoke. The SFPE handbook of fire pro-
tection engineering2NFPA, Quincy, Mass 3-241–3-262
12. Proulx G (1995) Evacuation time and movement in apartment building. Fire Saf J
24:229–246. doi:10.1016/0379-7112(95)00023-M
13. Pauls J (1999) A personal perspective on research, consulting and codes/standards devel-
opment in fire-related human behaviour, 1969+1999, with an emphasis on space and
time factors. Fire Mater 23:265–272. doi:10.1002/(SICI)1099-1018(199911/12)23:6<
265::AID-FAM698>3.0.CO;2-O
14. Lo SM, Lam KC, Yuen KK, Fang ZA (2001) Pre-evacuation behavioural study for the
people in a high-rise residential building under fire situations. Int J Eng Performance-
Based Fire Codes 2(4):143–152
15. Zhao CM, Lo SM, Liu M, Zhang SP (in press) A post-fire survey on the pre-evacua-
tion human behavior. Fire Technol (available online)
16. McCulloch WS, Pitts W (1943) A logical calculus of the ideas immanent in nervous
activity. Bull Math Biophys 5:115–133. doi:10.1007/BF02478259
17. Rosenblatt F (1958) The perception: a probabilistic model for information storage and
organization in the brain. Psychol Rev 65:386–408. doi:10.1037/h0042519
18. Schüürmann G, Müller E (1994) Back propagation neural networks—recognition vs.
prediction capability. Environ Toxicol Chem 13(5):743–747. doi:10.1897/1552-8618
(1994)13[743:BNNVPC]2.0.CO;2
19. Fraser-Mitchell JN (1999) Modelling human hehaviour within the fire risk assessment
tool CRISP. Fire Mater 23:349–355. doi:10.1002/(SICI)1099-1018(199911/12)23:6<349
::AID-FAM710>3.0.CO;2-3
20. Zadeh LA (1965) Fuzzy sets. Inf Contr 8:338–353. doi:10.1016/S0019-9958(65)90241-X
21. Zadeh LA (1973) Outline of new approaches to analysis of complex systems and deci-
sion process. IEEE Trans Syst Man Cybern 3(1):28–44
Pre-evacuation Human Response in Domestic Building Fire 449

22. Zadeh LA (1987) A fuzzy-algorithmic approach to the definition of complex or impre-


cise concepts. Fuzzy sets and application: selected papers by L. A. Zadeh. John Wiley
& Sons, Inc., Philadelphia, PA, pp 147–192
23. Takagi T, Sugeno M (1985) Fuzzy identification of systems and its applications to mod-
eling and control. IEEE Trans Syst Man Cybern 15:116–132
24. Jang RJS (1993) ANFIS: Adaptive-Network-Based Fuzzy Inference System. IEEE
Trans Syst Man Cybern 23(3):665–685. doi:10.1109/21.256541
25. Rosnow RL (1993) Rosenthal R beginning behavioral research, Maxwell Macmillian
Canada, Inc, Toronto
26. Tong D, Canter D (1985) The decision to evacuate: a study of the motivations which
contribute to evacuation in the event of fire. Fire Safe J 9:257–265. doi:10.1016/0379-
7112(85)90036-0
27. Sekizawa A, Ebihara M, Notake H, Kubota K, Nakano M, Ohmiya Y et al (1999)
Occupants: behaviour in response to the high-rise apartments fire in Hiroshima City.
Fire Mater 23:297–303. doi:10.1002/(SICI)1099-1018(199911/12)23:6<297::AID-FAM
702>3.0.CO;2-2

You might also like