Hopfield Neural Network for Microscopic Evacuation of Buildings
Boutheina Amina Aoun
1
, Zouhour Neji Ben Salem
2
and Hend Bouziri
1
1
LARODEC Laboratory, Higher Institute of Management University, Tunis, Tunisia
2
Artificial Intelligence Unit, National School of Computer Sciences, Tunis, Tunisia
Keywords:
Microscopic Evacuation, Evacuee Characteristics, Itinerary, Artificial Neural Networks, Hopfield Networks.
Abstract:
The problem of evacuation raised a lot of interest as its objective of saving lives is of an extreme importance. In
this context, many researches supplied solutions allowing to plan the process of evacuation in case of disaster.
Certain solutions took into account the behavior of the crowd, while others treated the evacuees in an inde-
pendent way. For that purpose, we dedicate our study to this last type of evacuation, namely the microscopic
evacuation. Our approach is based on the artificial neural networks which we considered capable of generating
a human behavior thanks to their neuronal aspect. We proposed a solution capable of planning a microscopic
evacuation of building by having recourse to Hopfield neural networks. We supplied an experimental study
on the real cases of two hospitals. This study also brought a comparison of our model with another neuronal
model for evacuation which is the self organizing map.
1 INTRODUCTION
Public areas such as schools, hospitals and shopping
malls are buildings where there is generally an accu-
mulation of persons. When everythingis normal, peo-
ple manage to reach the exit easily, once their need ac-
complished. However, when an unforeseen event like
fire arrives, the access to the same exit becomes com-
plicated. Indeed, the desperation, the speed of peo-
ple, the time which is urgent and several other factors,
make that the exit becomes harder and harder.
Naturally, if the building, the district or even the
city is designed in a way that makes the evacuation
process easier, more people could succeed to reach
the safety places.
Many researches supplied solutions allowing to
plan the process of evacuation in case of disaster. Cer-
tain solutions took into account the behavior of the
crowd (Macroscopic evacuation), while others treated
the evacuees in an independent way. For that pur-
pose, we dedicate our study to this last type of evacu-
ation, namely the microscopic evacuation. This kind
of evacuation tends to propose the adequate itinerary
of an evacuee according to the characteristics which
define him in a panic situation.
In this context, we found that the neuronal aspect
of artificial neural networks could be a good alterna-
tive to represent the human behavior in evacuation sit-
uations. However, only one type of artificial neural
networks, namely the Self Organizing Map (SOM),
treated the problem evacuation. For that purpose, we
decided to investiguate the track of Hopfield neural
networks.
2 MICROSCOPIC EVACUATION
OF BUILDINGS
Evacuation’s essential aim is moving a population out
of a dangerous place (an hypermarket, a work place
and even a district or a city). Each evacuee is pre-
sented by some parameters like age, gender, or walk-
ing speed (Klupfel et al., 2000). According to these
parameters, he would make his choice to move from
one position to another. Naturally, simulating the oc-
cupants of one building behaviors would help in plan-
ning evacuation and preventing the loss of lives.
Indeed, architects and persons in charge must take
the safety aspect into a high level of consideration and
apply security measures. To be able of this, people be-
havior in trouble cases must be predicted in advance.
To realize such simulations, technical tools are re-
quired. In this context, different approaches handeled
the problem. We essentially distinguished two big
sectors in this domain; a sector treating the problem
from a mathematical point of view, namely the cellu-
lar automata and the social forces, and another solv-
ing it using the artificial neural netwotks. We then
576
Amina Aoun B., Neji Ben Salem Z. and Bouziri H..
Hopfield Neural Network for Microscopic Evacuation of Buildings.
DOI: 10.5220/0004168905760581
In Proceedings of the 4th International Joint Conference on Computational Intelligence (NCTA-2012), pages 576-581
ISBN: 978-989-8565-33-4
Copyright
c
2012 SCITEPRESS (Science and Technology Publications, Lda.)
decided to investiguate this truck by focusing on Hop-
field neural networks and trying to apply it to the evac-
uation field.
3 HOPFIELD NETWORK
PRINCIPLE
Hopfield network is a very simple modeling of unsu-
pervised neural network. It is simply a set of neurons
fully connected with some strengths or weights; they
have two possible states, on and off (Heaton, 2005).
However, Hopfield neurons could be classified
into two logical layers.
• Input Layer: The layer containing the neurons
which values have to be initialized and provided
to the network.
• Output Layer: The layer of neurons which val-
ues are generated by the network.
Like all artificial neural networks, Hopfield neu-
rons follow the different steps of figure 1 to provide
an output. This output is retrieved by passing the
Weighted Sum of Input (WSI) signals to the transfer
function (Abraham, 2005).
Figure 1: Artificial neuron.
This WSI value is the sum of the multiplication of
each synapses’ weight with its corresponding signal.
In fact, WSI value is calculated basing on the neurons
connected to the considered node following equation
(1) where w
i
is the weight between current neuron and
neuron i, x
i
is the input coming from neuron i.
WSI =
∑
i
w
i
x
i
(1)
This sum is passed to a transfer function to decide
whether to activate the neuron or not. Like a biologi-
cal neuron, the artificial neuron is inactive until a WSI
value is reached .
Hopfield networks are based on an auto-
associative memory concept. Like the brain, they are
able to store some memories that can be later recov-
ered when the network is provided with only a part of
the information. For example, a smell, a colour or a
situation makes us remember special persons or some
childhood memories. Also, Hopfield networks work
can complete information from corrupted or incom-
plete data.
Hopfield networks were widely used in optimiza-
tion problems like the traveller salesman (Letchford
et al., 2011) and the n-queens problem (Letavec
and Ruggiero, 2002) and in pattern recognition (Yu,
2003),(Pandey et al., 2010). However, it was never
used in the field of microscopic evacuation.
4 A HOPFIELD MODELING FOR
EVACUATION PROBLEM
In the microscopic evacuation problem we have to
present two essential information; the input which is
the evacuee’s parameters as well as the output de-
scribed by the itinerary he would follow.
We then opted for an architecture which allows the
distinction between the various data, those of input
and those of output; we subdivided the network into
two layers, an input layer which values are provided
to the network and an output layer that the network
generates. Both layers where combined to form one
vector as shown in figure 2.
Figure 2: The vector modeling.
The input part of the vector have to be initialized
and the output part is provided by the network.
4.1 Input Layer
The input layer is a description of the evacuee’s char-
acteristics that would make us guess the path he
would follow. Each parameter is defined by a neuron
having a state that takes the value 1 if the condition is
true and -1 if it is false.
After a prerequisite study of different researches
in the evacuation field, we decided to retain some pa-
rameters that almost all the studies showed effective
HopfieldNeuralNetworkforMicroscopicEvacuationofBuildings
577
in egress process. Later, these parameters will be val-
idated in the experimental phase, where we could de-
cide of their usefulness in our model.
• Age
The category of age to which the evacuee belongs.
This factor would influence the evacuee behav-
ior as experience and advanced years of life make
him more able to choose the best road.
• Gender
Men and women generally don’t think in the same
way and have different psychological characteris-
tics, especially in disaster situations.
• Fire Experience (FE)
The evacuee behavior would be influenced by his
experience in confronting egress situation as it
would help him in having this quality to react and
to keep control of the situation.
• Familiarity with the Building (FB)
When the evacuee knows the building well, the
egress process would be easier to him as he al-
ready knows the direction to exit.
• Starting Position (SP)
We also have to specify the starting position of the
evacuee that must be kept firing in every time step
of the execution.
As Hopfield model doesn’t allow to provide data
having integer values, each parameter could only be
on true, having the value 1 or -1 otherwise.
To train the network, we choose a set of in-
stances of input neurons to which we assigned the
corresponding and logical itineraries. The number of
training patterns should not exceed 14% the number
of nodes in the network. Each pattern represents a
specific person, having some characteristics and the
itinerary he took to exit the building.
4.2 Output Layer
This layer of the network represents the different
compartments of the building. If the area is visited
in the road the evacuee chooses, the corresponding
neuron will be activated. All the neurons of the out-
put layer are initialized to 0 except the starting point
which is initialized to 1 and kept in this state during
all the iterations of the process.
Figure 3 shows the resulting networks with its
both layers.
Figure 3: The resulting network.
5 CASE STUDIES OF BUILDING
EVACUATION IN THE CASE OF
FIRE
In order to apply the problem of evacuation on a real
situation, we will test the network behavior in the case
of two hospital buildings.
5.1 Case of the Second Floor of the
Tunisian Children Hospital
This floor of the building constituted the neurons of
the output layer. We supposed that the fire is located
in the transfer room. In table 1 we find the legend of
the different sectors of the building.
Table 1: Legend of the building areas.
Neuron name Description
Sri Septic room number i
Laui Laundry number i
offi Office number i
SR Room of Rest
Row Row
CM/W Cloak rooms Men/ Women
Rt Transfer room
TP Technical premise i
STE Sterilization
Was Washing
ELE Elevator
Gri Guard room number i
CiPi Corridor number i part i
exti Exit number i
The itinerary of each person will appear as red col-
ored neurons while inactive neurons will remain blue.
For example, the evacuee of figure 4 took the itinerary
sr2-sas1-c1p3-off2-off3-off2-c1p3-rt2-c2p2-ext3.
IJCCI2012-InternationalJointConferenceonComputationalIntelligence
578
Figure 4: An example of output of the Hopfield network.
For the learning phase, we tried to choose profiles
as exhaustivly as possible because of the constraint
of limited number of patterns. To validate the train-
ing phase, we must be sure of the stability of all the
learned pattern. Meaning that if we provide the net-
work with an input layer of one learned pattern, the
generated output must be identical to the learned one.
We reached a threshold of 14 patterns, after which the
network showed a high level of error.
This restriction of number of patterns could be ex-
plained by the imminence of correlation between the
learned vectors every time we exceed a certain num-
ber of patterns.
We then choose 50 differentprofiles to test the net-
work. Each generated itinerary was compared to the
logical expected one in the validation set to retrieve
the error rate of the network. The network accom-
plished the tests with an error rate of 10% which is a
satisfying result. Furthermore, the network converged
after less than 50 epochs.
5.1.1 Experimental Results and Interpretations
One of the aims of our solution to evacuation problem
is to analyze the building architecture. We found a lot
of disparity in the tendency of the evacuees to visit
certain compartments.
Figure 5 shows the number of times each compart-
ment was visited in the evacuation of our sample of 50
persons. We notice for instance that the second part
of corridor2 and the third part of corridor 1 are very
frequently visited by evacuees.
Thus, it would be useful to widen them to rem-
edy the congestion, to put plans indicating the exit on
their walls and to place security cameras there to su-
pervise the situation. Another precaution to take is to
put obstacles in places of congestion to separate the
crowd.
Figure 6 shows the probability of taking the short-
Figure 5: Number of visits to each compartment.
est path according to each parameter.
Figure 6: Percentage of taking the shortest path.
People familiar with the building are more likely
to take the shortest path while old people are generally
not able to take the best alternative as only 15% out
of them guessed the best itinerary.
5.1.2 Sensitivity Analysis
In this phase of the work, we have to validate our
choice of input parameters by analyzing the effective-
ness of each of them in the network behavior.
For every given parameter, we made disturbances
on its value while keeping all the other parameters
fixed. We took 20 profiles for each parameter and
made the variation noticing the number of times out-
put changes.
The results of the tests gave the statistics illus-
trated in figure 7. This figure shows the percentage of
times where the output changes when we only change
the value of one of the parameters.
All of the model’s parameters were effective, as
the variation rate exceeded 50% for all of them. How-
ever the degree of effectiveness changed from one pa-
rameter to the other.
HopfieldNeuralNetworkforMicroscopicEvacuationofBuildings
579
Figure 7: Variation rate.
5.2 Case Study of the First Floor of
Cardiology Department of Charles
Nicole Hospital
The experimentalstudy will concern a building bigger
than the floor of the children hospital studied in the
previous section in order to test, in the same time, the
sensitivity of the Hopfield model to the dimensions of
the building.
Figure 8: Network output corresponding to the first floor of
cardiology department of Charles Nicole hospital.
The building of the first floor of the cardiology de-
partment of Charles Nicole hospital contains 63 areas.
The different areas of the building were considered as
neurons composing the network showed in figure 8.
The maximum number of patterns accepted in the
learning process was 18 patterns. After the learning
phase, we tested both networks with 30 different pro-
files.
5.2.1 Experimental Results and Interpretations
The network succeeded in acheiving 23 % of error
in less than 50 time steps. However, we noticed that
the model’s performance decreased comparing to the
previous case study as the number of compartments
increased.
When the number of compartments of the tested
building increased of 27 areas comparing to the build-
ing tested in previous section, the number of maxi-
mum learned patterns didn’t increased proportionally.
In fact, we were only allowed to add 4 patterns, reach-
ing the total number of 18 learned patterns.
In the example of 36 areas, the 14 allowed patterns
were able to give the network the sufficient knowl-
edge which permits it to generate intelligent and cor-
rect roads. However, in this example, the 18 learned
profiles weren’t able to cover the 63 neurons of the
network for different times. To be as exhaustive as
possible, we were obliged to place the evacuees of the
learning set in distant areas in order to visit the total-
ity of the neurons. The majority of the neurons were
then visited only one time in the learning process. As
a result, the outputs of the tests were very sensitive to
the starting position.
6 HOPFIELD VS. SELF
ORGANIZING MAP (SOM) FOR
EVACUATION
After adapting Hopfield networks to the problem of
microscopic evacuation of buildings, we thought of
pushing the study to a comparison with another neu-
ral network model. Both belonging to unsupervised
learning type of neural networks, Hopfield and SOM
for evacuation were the object of a comparative study
basing on the SOM model proposed in (Ben Salem
et al., 2011).
We begun by a theoretical comparison that
showed that Hopfield was simpler and easier in its
processing way. We then moved to a practical com-
parison. In tables 2 and 3 we find the results of the
comparison in the case of the two Hospitals quoted
previously.
Table 2: Comparison results in the case of the second floor
of children hospital.
Hopfield Som
Patterns number 14 50
Error rate 23 % 16%
Time steps 50 2500
IJCCI2012-InternationalJointConferenceonComputationalIntelligence
580
Table 3: Comparison results in the case of the first floor of
Charles Nicole hospital.
Hopfield Som
Patterns number 18 18
Error rate 23 % 36%
Time steps 50 350
The results of the tests were in the favor of Hop-
field as they showed respectable error rates and ra-
pidity in convergence time of Hopfield comparing to
SOM. Hopfield also provided more logical and intel-
ligent results.
In fact, we noticed that SOM tends to an illogical
generation of the shortest path even when it hasn’t to
do it.
7 CONCLUSIONS
The problem of evacuation is an important subject of
actuality which was the object of a lot of interest in the
research’s field. This paper dealt with the microscopic
type of evacuation which focuses on the behavior of
each evacuee during the egress process.
We carried our interest on Hopfield neural net-
works known by their simple architecture and their
effectiveness in different domains especially the do-
main of pattern recognition and combinatory opti-
mization. The model is based on an auto-associative
memory concept which permits to recoverany learned
information providing the network with a respectable
degree of intelligence.
We proposed a model able to intelligently draw
evacuee’s itineraries in a building, where there is
a panic situation, according to some characteristics.
The parameters we thought effective in getting the
evacuee behavior are his age category, his gender, his
familiarity with the building, his fire experience and
his starting position.
To lighten our model’s contribution in artificial
neural network approaches for evacuation, we sup-
plied a comparative study between our model and
SOM model proposed in (Ben Salem et al., 2011).
Hopfield gave largely better results than SOM. It con-
served its rapidity quality adding to it the accomplish-
ment of a better rate of error than SOM.
Even if Hopfield networks succeeded in simulat-
ing the evacuation process, the model could be im-
proved by including other factors as the time of evac-
uation and the speed of people.
It would be also interesting to improve the quality
of the learning set as we could deepen it by taking
the opinion of sociologists and specialists of human
behavior.
Finally, it would be promising to enlarge Hopfield
capacity of learning.
REFERENCES
Abraham, A. (2005). Artificial Neural Networks. Peter H.
Sydenham and Richard Thorn, John Wiley and Sons
Ltd.
Ben Salem, Z. N., Bouziri, H., and Jelassi, S. (2011). A
som based predictive model for fire building evacu-
ation. In Communications, Computing and Control
Applications (CCCA), 2011 International Conference,
pp.1-5.
Heaton, J. (2005). Introduction to Neural Networks with
Java. The Heaton Research Inc.
Klupfel, H., Konig, T., Wahle, J., and Schreckenberg, M.
(2000). Microscopic simulation of evacuation pro-
cesses on passenger ships. In Proceedings of the
Fourth International Conference on Celular Automata
for Research and Industry, Karlsruhe, Springer-
Verlag, pp. 63-71, London.
Letavec, C. and Ruggiero, J. (2002). The n-Queens Prob-
lem. The INFORMS Transactions on Education 2, pp.
101-103.
Letchford, A., Lodi, A., and Wenger, K. M. (2011). The
Travelling Salesman Problem. Mathematical Pro-
gramming Computation, vol. 3, no. 1, pp. 59-78.
Pandey, B., Ranjan, S., Shukla, A., and R., T. (2010). Sen-
tence recognition using hopield neural network. In
IJCSI International Journal of Computer Science Is-
sues, vol. 7, iss. 4, no. 6.
Yu, J. (2003). Artificial Neural Net-
work: Assignment. Retrieved from
http://users.cecs.anu.edu.au/jinyu/JinYules/preproj/A
NN.pdf.
HopfieldNeuralNetworkforMicroscopicEvacuationofBuildings
581
