an intelligence-based route choice model for pedestrian flow in a transportation station
TRANSCRIPT
-
8/11/2019 An intelligence-based route choice model for pedestrian flow in a transportation station
1/9
Applied Soft Computing 24 (2014) 3139
Contents lists available at ScienceDirect
Applied Soft Computing
j ournal homepage : www.elsevier .com/ locate /asoc
An intelligence-based route choice model for pedestrian flow
in a transportation station
J.K.K. Yuen a,, E.W.M. Lee a, W.W.H. Lam a,b
a Department of Civil and Architectural Engineering, City University of Hong Kong, Tat Chee Avenue, Kowloon Tong, Hong Kong (SAR),PR Chinab MTR CorporationLimited, MTRHeadquarter Building, Telford Plaza, 33WaiYip Street, Kowloon Bay, Hong Kong (SAR), PR China
a r t i c l e i n f o
Article history:
Received 13 December 2012
Received in revised form 10 June 2013
Accepted 29 May 2014
Available online 15 July 2014
Keywords:
Artificial neural networks
Intelligent system
Crowd movement
Evacuation
Human behaviour
Transportation
a b s t r a c t
This study proposes a method that uses an artificial neural network (ANN) to mimic human decision-
making about route choice in a crowded transportation station. Although ANN models have been
developed rapidly and widely adopted in various fields in the last three decades, their application to
predict human decision-making in pedestrian flows is limited, because the video clip technology used to
collect pedestrian movement data in crowded conditions is still primitive. Data collection must be carried
out manually or semi-manually, which requires extensive resources and is time consuming. This study
adopts a semi-manual approach to extract data from video clips to capture the route choice behaviour of
travellers, and then applies an ANN to mimic such decision-making. A prediction accuracy of86% (ANN
model with ensemble approach) is achieved, which demonstrates the feasibility of applying the ANN
approach to decision-making in pedestrian flows.
2014 Published by Elsevier B.V.
Introduction
To cope with rapid population growth, increasing numbers of railway systems
have been constructed acrossthe world, ranging from high-speed railways such as
the Eurostar in Europe to the more recent magnetic levitation trains such as the
Shanghai airport train in China. The projected trend is that railway systems will
continue to increase in size, and thus crowd movement and evacuation planning
will play a more importantrolein railway systems than ever before. TheHongKong
transportation system is characterised by high passengerflows,a short train head-
way andlimitedcapacityin thetransportation stations.The failure offacilitiesinside
a station may cause accidentsand threatenpassengerslives. A study of route choice
between escalators andstairways in subwaystationsin Hong Kong wascarried out
[1] to explore the optimisation of station facilities, which is critical for both safety
and operational efficiency.
In the past three decades, several dynamic evacuation and pedestrian mod-
els have been developed for modelling complex crowd movements. These models
include social force (SF) models [2], cellular automata (CA) models [3], lattice gas
(LG) models [4], fluid-dynamic model [5], agent-based (AB) models [6] and SGEM
model [7,8]. They provide important information about the spatial design of com-plexbuildings,undergroundstationsand other publicamenities.In additionto these
microscopic [24,68] and macroscopic models [5], network models [9,10] have
proved useful in the design of emergency evacuation plans, because they allow
the detailed modelling of human cognitive processes. However, most of the exist-
ing pedestrian flow models that simulate the dynamic movement of pedestrians
are based on mathematical models, which may not be able to sufficiently mimic
actual human behaviour. Indeed,the movementdecisionsthat areobtainedby these
Corresponding author. Tel.: +852 3442 2307; fax: +852 2788 7612.E-mail addresses:[email protected], [email protected]
(J.K.K. Yuen).
models are determined by either empirical equations (e.g., Bradley [5] employed
NavierStokes equations that govern fluid motion to describe crowd movementat high densities) or by assumptions (e.g., the SF models [2], CA models [3], and
LG models [4] treat individuals or groups as homogeneous, and do not consider
heterogeneous human behaviour such as herding). Even with AB models such as
Simulex [11] and buildingEXODUS [12], occupant behaviour is assignedby theoper-
ator accordingto hisor herpreference,and theresultant simulatedpedestrian flow
patterns may not reflect real-life situations.
One of the critical behavioural reactions of humans during evacuation and in
moving crowds is route choice.Route choice is influencedby manyfactors, including
personal experience, building geometry, interactions among occupants and envi-
ronmental factors [8,12,13]. In general, passengers will choose the path with the
shortest travel time, traveldistanceor a combination of both [14]. However, Proulx
[15] pointed outthat evacueestend to prefer familiar routes rather thanthe shortest
path to theexit, because they feel that unknown paths increase thethreat. Proulxs
pioneering work depicted the complexity of route choice in human movement.
Gwynne et al. [12] proposed an exit selection behaviour model that is based on
Queuing and Familiarity Behaviour. Lo et al. [8] introduced a game theory based
exit selection model for evacuation. Hoogendoorn and Bovy [16] proposed a newtheory of pedestrianroutechoice behaviourunderuncertainty basedon theconcept
ofutility maximisation.
These route choice models all usemathematics to simulatethe humandecision-
making.In contrast,thispaper proposesthe alternative approachof applyingan ANN
model to capture human decision-making behaviour by data collected from actual
passengers.
Route choice behaviour on escalators
In transportation stations, escalators, stairways and elevators serve as vertical
transport between theconcourse andthe platforms. Passengers tend to useescala-
tors rather than stairs or lifts. They avoid climbing stairs to save energy and avoid
spending time waiting for lifts. Escalators play an important role in transportation
http://dx.doi.org/10.1016/j.asoc.2014.05.031
1568-4946/ 2014 Published by Elsevier B.V.
http://localhost/var/www/apps/conversion/tmp/scratch_1/dx.doi.org/10.1016/j.asoc.2014.05.031http://www.sciencedirect.com/science/journal/15684946http://www.elsevier.com/locate/asocmailto:[email protected]:[email protected]://localhost/var/www/apps/conversion/tmp/scratch_1/dx.doi.org/10.1016/j.asoc.2014.05.031http://localhost/var/www/apps/conversion/tmp/scratch_1/dx.doi.org/10.1016/j.asoc.2014.05.031mailto:[email protected]:[email protected]://crossmark.crossref.org/dialog/?doi=10.1016/j.asoc.2014.05.031&domain=pdfhttp://www.elsevier.com/locate/asochttp://www.sciencedirect.com/science/journal/15684946http://localhost/var/www/apps/conversion/tmp/scratch_1/dx.doi.org/10.1016/j.asoc.2014.05.031 -
8/11/2019 An intelligence-based route choice model for pedestrian flow in a transportation station
2/9
32 J.K.K. Yuen et al. / Applied Soft Computing 24 (2014) 3139
Fig. 1. Example of routes choice decision-making.
around a station. Optimising the use of escalators is thus the first priority in the
internal transportation system of stations.
This study presents an example of route choice in a transportation station. Pas-
sengers approach two escalators and choose one of them to travel to the upper
floor. The decision-making involves the consideration of various factors, which are
explored in the following sections.
Routes choice behaviour
Initial research in thisarea [17,18] focused on the correlation between through-
put capacity and thewidth of exits. Fig. 1 shows the layout of a compartment with
two exits. In an emergency (e.g., the outbreak of fire), each occupant inside the
compartment will choose one of theexits to leave thecompartment. This is a typi-
cal example ofroute choice decision-making. In predictingthe evacuationtime from
the origin of an evacuee to the exit of the compartment, Togawa [18] established
empirically that the evacuation time Te can be expressed as a sum of the flow time
(To) and the travel time (Tf) of theevacuee as follows.
Te= To + Tf (1)
In the simplest form,To can be further expressed in Eq. (2), where ks is the dis-
tance from thepointof theevacueeto thedestination, andvrepresents the walking
speed of thecrowd.
To=ksv
(2)
The travel time Tf isestimated by Eq. (3), where Na is the total numberof evac-
uees, fp represents theflow rate perunit width of theexit andB isthe width of the
exit.
Tf=NafpB
(3)
Usually, pedestrians choose the route with the least travelling time. According
to Togawa [18], the factors that contribute to thetravelling time are as follows.
1. Thewalking velocity of thepedestrian.
2. Thedistance from the originof thepedestrian to the exit.
3. The maximum flow capacitiesof the exits.
Thesefactors formthe basisof thisstudyto investigatethe parametersin human
route choice decision-making for ANN model training.
The remainder of this paper is organised as follows. Section Artificial neural
network introduces the ANN for route choice in moving crowds. Section Data
collection discusses how the data was collected from the transportation station.
SectionDevelopmentof theMLP modeloutlinesthedevelopmentand architecture
ofthe ANN model. Sections Model training and Results and discussion, respec-
tively, present the model training process and evaluatethe performance of the ANN
model. Section Conclusion concludes the paper.
Artificial neural network
ANN models have developed rapidly in the last few decades,
to the extent that they are now able to mimic the correlation of
system parameters that are unknown or complex [19] and capture
the nonlinear behaviour of a system via a learning process (also
known as the training of the network function) from historical
system data. ANN models have also become a popular approach to
the prediction of non-linear functions in the past decade. Among
the various ANN models, the multi-layered perceptron (MLP) [20]
is one of the most widely used for forecasting due to its simple
and flexible nature. MLP has been successfully used to predict the
weather, flank wear in drills and thermal load predictions [2126].
However, researchershave seldom applied ANNmodels to simulate
pedestrian decision-making in moving crowds or evacuation plan-
ning due to the extensive resources required for the data collection
and pre-processing.
One of the merits of ANN models is that they do not require
highly specialised human expertise nor any assumptions. The
learning feature of ANNmodels is especially usefulfor human deci-
sion models, as the relationships between the input parameters
are less well known than in highly structured expert systems or
equation-base approaches [19]. In this study, the MLP is adopted to
predict route choice in pedestrian flows.
Data collection
The data was collected from a transportation station in HongKong. A bank of escalators inside the station, as shown in Fig. 2,
was offered by the transportation company for the study.There are
three escalators, one of which had been stopped to save energy.
The other twoescalators were moving upwardat an equal andcon-
stant speed. The passenger flow was unidirectional. All passengers
are required to ascend to the upper floor via escalator to leave the
station. Passengers approaching the escalators have to choose one
of the two escalators to ride to the upper floor. This is the decision
making that is investigated and mimicked by the MLP model.
It is normal practice in Hong Kong, as in some other countries,
for passengers not in a hurry to stand on one side of the escala-
tor and leave the other side for passengers in a hurry to walk up
to shorten the time spent riding the escalator. Both of the escala-
tors shown in Fig. 2 are 60m long. This long length prevents most
-
8/11/2019 An intelligence-based route choice model for pedestrian flow in a transportation station
3/9
J.K.K. Yuen et al. / Applied Soft Computing 24 (2014) 3139 33
Fig. 2. Route selection by passengers in the transportation station.
passengers from walking all the way up the escalators and
simplifiesthe decision-makingprocess. Accordingto our siteobser-
vations, the general behaviour of passengers is in line with our
assumption, with most passengers in the video footage standing
on the escalators and just a few walking up.
The video technology used to collect pedestrian movement
data in crowded conditions is still primitive, and existing image
processing techniques are unable to fully overcome the occlusion
problem [36]. Therefore, data collection was carried out manually
or semi-manually, which requires extensive resources and is time
consuming.In thispaper, the locationsof the passengersat different
time steps were extracted from the video clips by using a semi-
manual approach with a Passenger Trap, which was presented
by Teknomo [36]. The Passenger Trap is an imaginary rectan-
gle marked as a boundary for counting passengers on the video
clips. Only the passengers who pass the trap are considered. The
trap width is perpendicular to the path of the passenger, and the
trap length is parallel to the walkway. Therefore, the area and the
coordinates of this rectangle can be obtained more easily.
In this study, the data were collected from video clips captured
by a closed-circuit television (CCTV), as shown in Fig. 3a. Fig. 3b
shows a typical snapshot (front view) of the train station. The loca-
tions of the passengers at different time steps were extracted from
the video clips using a semi-manual approach. A computer pro-
gram was developed to play the video on the computer screen. At
each time step, the computer mouse was moved over the screen
to click on the feet of passengers shown on the screen. The data on
Input1
Input 2
Input 3
Input Layer Hidden Layer(s) Output Layer
Output 1
Output 2
Fig. 4. Typical architecture of the multilayer perception (MLP) model.
the passengers locations were then converted to geometrical data
by the computer program. These data were further processed to
obtain the input and output data for the training of the ANN model.
Details of the input and output parameters of the ANN model are
discussed in Section Development of the MLP model.
Development of the MLP model
TheMLP architectural model consists of several layers: the input
layer, hidden layer(s), and output layer. Fig. 4 shows the archi-
tecture of a typical MLP model. The neurons in each layer are
interconnected with the neurons in the adjacent layers. There may
be more than one hidden layer in a model, but it has been proven
that an MLP with a single hidden layer is a universal function
approximator [27], subject to the provision of a sufficient number
of hidden neurons. A three-layered (input layer, hidden layer and
output layer) MLP model was thus adopted in this study.
Input layer
The number of neurons in the input layer corresponds to the
number of input parameters in the model. As mentioned in Section
Routes choice behaviour, the main factors affecting the evacua-
tion time according to Togawa [18] are (1) the walking velocity of
the evacuee, (2) the distance from the origin of the evacuee to the
exit and (3) the flow rate of the exits. However, non-linear human
behaviour such as queuing and familiarity behaviour is obviously
not included in Togawas mathematical model, and further mod-
ification based on empirical equations is necessary. We used the
following six parameters as inputs for the ANN model to decide
which escalator would be chosen by each passenger.
Fig. 3. (a) Snapshot from the CCTV footage and (b)Snapshotof theescalators (frontview).
-
8/11/2019 An intelligence-based route choice model for pedestrian flow in a transportation station
4/9
34 J.K.K. Yuen et al. / Applied Soft Computing 24 (2014) 3139
Fig. 5. Initial position of the passengers.
Input 1 (walking velocity of passengers)
Normally,the walking speedof a passengerreflectshow much of
a hurry he or sheis in.Passengers in a hurry will walk faster. Passen-
gersin a hurry may also be impatientto board thenearest escalator,
andmay overtake otherpassengers [11,28]. Alternatively,they may
choose the escalator that is not the nearest one but has less peo-
ple using it. Passengers walking slowly are less likely to be in a
hurry,and may choose a less dense escalator to avoid collisionwith
hurrying passengers.
Input 2 (initial position of passengers)
Pedestrians dislike taking detours or moving contrary to their
desired walking direction, even in crowded conditions [29]. How-
ever, due to the high population density in Hong Kong, passengers
usually move as a crowd due to the limited capacity of stations. The
initial position of passengers entering the crowd will affect their
route choice, in that such decisions a prevented when the densityreaches a certain level. In this situation, a passenger is most likely
to follow the crowd or follow the path of the passenger in front
to avoid collisions with nearby passengers [29]. Sudden changes of
direction are seldom found in normal situations, as this action may
cause collisions and block the routes of passengers nearby. The ini-
tial position of the passenger was thus also selected as one of the
inputs for the ANN model training.
According to our onsite observations, the initial position of the
passengers entering the crowd resulted in different probabilities of
passengers choosing the escalators. In our data capturing process,
we evenly divided the initial position of the passengers into six
locations, as shown in Fig. 5.
Input 36 (initial and final densities of the escalators)Humans are required to make decisions when they face two or
more options. Similarly, the passengers in this study were required
to make a decision as they walked to a position where they could
see the two escalators (i.e., where the two escalators fell into their
line-of-sight). We denote this location as the initial position, which
is chosen as one of the parameters influencing the decision of pas-
sengers and an input in the ANN training model. Fig. 6 illustrates
the initial decision point of passengers according to the concept of
the minimum line-of-sight.
The densities at the platform and on the steps of escalators 1
and 2 are captured when the passengers approach the initial deci-
sion point as input parameters 3 and 4, respectively. The idea of
level of service (LOS) proposed by Fruin [30] reveals that the walk-
ing velocity of a pedestrian and the density of the surrounding area
Fig. 6. Initial decision point of passengers with a minimum line-of-sight.
are highly correlated. Passengers approaching the two escalators
and deciding which escalator to use may also consider the densi-
ties at the entrances of the escalators to avoid being blocked in by
otherpassengers. Passengers typicallymake their initial decision at
the initial decision point based on their walking velocity, position
and the densities at the escalator entrance. If an alternative route
of the same length is available, then according to [29] passengers
will make the route choice decision as late as possible, meaning
that a change of mind is still possible at the last minute. Our model
accounts for this observation and considers the scenario at a final
decision point, as shown in Fig. 7. The densities at the landing plat-
forms of the escalators 1 and 2 when the passengers approach the
final decision point are respectively recruited as input variables 5
and 6.
Hidden layer
The number of neurons in the input and output layers is equal
to the number of input parameters and output parameters. This
is defined by the system itself. The required number of hidden
neurons is crucial to the performance of the model. Currently,
there is no analytical approach to determine the number of hidden
Fig. 7. Final decision point of passengers.
-
8/11/2019 An intelligence-based route choice model for pedestrian flow in a transportation station
5/9
J.K.K. Yuen et al. / Applied Soft Computing 24 (2014) 3139 35
Table 1
Input and outputparameters of theMLP model.
Parameter Unit
Input Walking velocity of passengers [m/s]
Initial position of passengers
(i.e., range from 1 to 6)
Initial density of escalator 1 [person/m2]
Initial density of escalator 2 [person/m2]
Final density of escalator 1 [person/m2]
Final density of escalator 2 [person/m2
]
Output Passengers choosing escalator
1 o r 2 (i .e. , 0 o r 1)
neurons, but different rules of thumb are available to provide
heuristic hints to the user. The rule of thumb adopted in this study
is therule developedby Ward [31] as describedin Eq. (4), whereNh,
Ninand Noutare respectively the number of neurons in the hidden
layer, the input layer and the output layer, and Ns is the number of
training samples.
Nh= (Nin +Nout)/2 +Ns (4)Before the adoption of the rule-of-thumb, the sensitivity of the
performance of the model to the number of hidden neurons was
studied. However, it was found that a change in the number of hid-
den neurons did not significantly affect the model performance in
terms of themean of theaccuracyof theactual outputs andthe out-
puts predicted by the model obtained from the training (described
in Section Model training). The number of hidden neurons exam-
ined was set at Nh5. For each number of hidden neurons, 5000trials were performed to determine the model performance with a
95% confidence level. If the 95% confidence intervals of the models
with differentnumbers of hidden neurons overlap, then from a sta-
tisticalpointof view itcan be concludedthatthe performanceof the
models with different numbers of hidden neurons is comparable.
It was thus concluded that the model performance is insensitive
to the number of hidden neurons, and the rule of thumb could beadopted.
Output layer
The output of the MLP model is whether passengers choose
escalator 1 or 2, which is represented by 0 or 1, respectively.
The inputs and output of the MLP model are listed in Table 1. The
architecture of the MLP model is illustrated in Fig. 8.
Passengerschoosingescalator 1 or 2
Input layer
Walking velocity of passengers
Initial position of passengers
Initial density of escalator 1
Initial density of escalator 2
Final density of escalator 1
Final density of escalator 2
layerHidden Output layer
Fig. 8. Architecture of theMLP model.
Prediction
Error
Minimumvalidation error
Network training stopshere, as there is nofurther improvement inthe validation error
TrainingEpoch
Trainingerror
Validationerror
Pre-defined number of epochs(1,000 epochs in this study)
Fig. 9. Early-stop validation approach.
Model training
Back-propagation (BP) [32] is the traditional training algorithm
usedfor the MLP model.It feeds back theprediction errorsfrom the
output layer to the input layer and adjusts the weights of the links
between the neurons. Upon completion of the weight adjustment,
a new prediction is carried out to evaluate a new prediction error
for the next epoch of weight adjustments. These procedures are
repeatednumeroustimes untila satisfactory predictionis achieved.A total of 621 samples were collected in August 2010 for the
network training and testing. Fifty per cent of the samples were
used for network training, 25% were hidden during the network
training phase and the remaining 25% served as a testing set to
evaluate the performance of the trained network. The training set
was used to train the model with the BP algorithm, whereas the
validation set was used to monitor and stop the BP training using
the early-stop validation approach. The testing set did not play a
role in thetraining ofthe MLPmodel. Uponcompletion of the model
training,the testing setwas used toevaluatethe performanceof the
trained model.
To prevent overfitted training, the intermediate-state trained
model in every training epoch was applied to the validation set to
evaluate the prediction error (i.e., the validation error). The net-work training was stopped when the validation error reached the
minimum value. As we had no prior knowledge of the trend of the
validation error, early-stop training was adopted.This records the
statusof themodel continuouslyin thecourse of thetraining.When
there is no reductionin thevalidation error over a predefinednum-
ber of epochs (the number of epochs selected here was 150), the
model state with the minimum validation error is taken to be the
trained model. Fig. 9 illustrates the early-stop training process.
Upon completion of the network training, the trained MLP was
applied to the testing set to evaluate the performance indices by
comparing the target values of the testing set and the values pre-
dicted by the trained model. The performance index used here was
the fraction of correct predictions (a), as defined in Eq. (5), whereN
is the total number of samples and {ti=pi}Ni=1are the target values
and the predicted values, respectively. A fraction of correct pre-
dictions of 1 indicates that 100% of the test samples are correctly
predicted.
a =Ni=1
(ai)/N where ai=
1 if ti=pi0 if ti /=pi
(5)
It shouldbe noted that a randomprocess is normally involved in
the network training of an MLP model, especially when the avail-
able samples are divided into training and validation sets. It is thus
possible for the random process to result in fortuitous samples
that show favourable evaluated performance indices. Instead of
reporting only the best simulation result, a less-prejudice statisti-
cal approach was adopted to minimise the effect of randomisation.
-
8/11/2019 An intelligence-based route choice model for pedestrian flow in a transportation station
6/9
36 J.K.K. Yuen et al. / Applied Soft Computing 24 (2014) 3139
Beta distribution
95% of totalarea under
x
95% Confidencex
p
(x)
10
Fig. 10. Example of thestatistical performance evaluation of theANN model.
We carried out the network training and performance evaluation
process 5000 times. The 5000 results were statistically analysed by
evaluating the mean and limit of the 95% confidence intervals of
the results.
In the model training, the training, validation and testing sam-
ple sets were obtained by random sampling. The performance of
the trained model might not always be favourable, as it very muchdepends on the quality of the training sample. It is possible that
the sample batch favours the network training, and thus produces
a better model. To verify the performance of the trained model, it
is thus necessary to evaluate it statistically.
As thesample extraction process is random, we carried out5000
trials of model training and testing (i.e., the results converged at
5000 trials). In each trial, the training, validation and testing sam-
ples were randomly grouped. The model was trained and tested
with the samples and the performance index of that trial was
obtained. The 5000 trials gave 5000 performance indices. As the
performance indices (i.e., the fractions of correct predictions) were
well bounded between 0 and1, a beta distribution was used to rep-
resent the probability distribution of the accuracy of the model, as
shown in the example in Fig. 10. The limit of the one-sided 95%confidence level from the right (i.e.,x95) represents the fraction of
correct predictions of the model. This statistical approach alleviates
the effect of randomness in evaluating the model performance.
Results and discussion
MLPmodel
The number of hidden neurons was determined by the rule-
of-thumb in Eq. (6), in which the number of training samples
(Ns), is taken to be 50% of the available samples initially (i.e.,
0.5621= 311). In this case, the number of input (Nin) and out-put parameters (Nout) is 6 and 1, respectively. According to Eq. (6),
25 hidden neurons are needed.
Nh= (6 + 1)/2 +
311 = 21.14 21 (6)
A study of the sensitivity of the number of hidden neurons to
the performance of the model was carried out. The number of hid-
den neurons to be investigated ranged from 17 to 26 (i.e., 215).For each number of hidden neurons, 5000 trials of model training
andperformance evaluation were conductedto obtainthe 95% con-
fidence intervals of the models with different numbers of hidden
neurons. The results are shown in Fig. 11.
The 95% confidence intervals overlap, which indicates that the
performance of the models with different numbers of hidden
neurons is comparable, and thus the number of hidden neuron
is insensitive to the model performance. Thus, the number of
NO. OF HIDDEN NEU RONS
CORRELATION
COEFFICIENT
17
18
19
20
21
22
23
24
25
260.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
1.00
Fig. 11. Performance of models with different numbers of hidden neurons.
hidden neurons determined by the rule of thumb (25) was justi-
fiably adopted.
Back-propagation was applied to train the MLP model with the
training sample. As explainedin SectionModel training, thesame
process was repeated 5000 times to obtain the total accuracy of
the model. Upon the completion of model training, the model was
applied to the testing dataset to evaluate the performance. The
mean of accuracy of the prediction results is 0.7782.The results are
presented in histogram form in Fig. 12. As the accuracy was well
bounded between 0 and 1, the beta distribution shown in Eq. (7)
was used to describe the accuracy distribution.
f(x|,) = x1(1 x)1
1
0u1(1 u)1du
(7)
By applying the beta distribution to describe the distributionof the 5000 samples, the parameters of the distribution were esti-
mated to be = 116.4770 and =33.2015. The profile of the beta
distribution reasonably matches the profile of the histogram. The
minimum fraction of correct predictions obtained by the trained
MLP model of 0.7204with a 95% confidence level indicates that the
MLP model performs reasonably well.
Fig.12. Histogramshowingthe distribution of the fraction of correct predictions of
the model approximated by a beta distribution.
-
8/11/2019 An intelligence-based route choice model for pedestrian flow in a transportation station
7/9
J.K.K. Yuen et al. / Applied Soft Computing 24 (2014) 3139 37
A sample is randomlydrawn as test sample
For Training and Validation For Testing
Original data set
The order of thesamples is shuffled
Step 2
Step 1
Fig. 13. Iterative process of the leave-one-out approach.
Leave-one-out validation approach
As there were only 621 samples available for the MLP model
training and testing, we adopted the leave-one-out approach
[33,34] to evaluate theperformance of theMLP model. In theleave-
one-out approach, in every trial, only one sample is reserved as
the test sample to evaluate the performance of the model beingtrained by the other 620samples. Fig. 13 illustrates the mechanism
of the leave-one-out approach. It is, in general, a two-step process.
In the first step, the order of the samples is randomly shuffled. This
process ensures that the samples used for training, validation and
testing are different foreach trial.In thesecondstep,assuming that
there areNnumbers of available samples, one sample is randomly
taken out from the pool and kept back to evaluate the performance
of the trained model while the other samples (i.e., N1 samples)
are used as training and validation samples in the model training.
Upon completion of the training, the trained MLP model is applied
to the omitted sample to evaluate its performance. This process is
repeated 5000 times to obtain 5000 performance indices.
As mentioned in Section MLP model, the dataset sample size
ratio for training, validation and testing in the MLP model com-
prised 50%, 25% and 25%, respectively, of the total number of
samples. To evaluate the model performance of the MLP model by
the leave-one-out approach, the sample size ratio for training and
validation with the leave-one-out approach is 70% and 30% of the
total number of samples. A further 5000 trials were run to ensure
that all of the samples were tested thoroughly at each iteration.
In each trial, the predicted output was compared with the target
output of the test sample to determine whether the prediction was
correct. Among the 5000trials,3974 trials were correctlypredicted.
The fraction of correct predictions was thus0.7948,which is slightly
higher than the MLP model (i.e. 0.7782) obtained in Section MLP
model.
Ensemble approach
The performance of an ANN model can be further enhanced
using the ensemble approach [35]. The general concept of theensemble approach is illustrated in Fig. 14. Assuming that we have
a total ofNsamples to evaluate the performance of the ensemble
approach,a total ofNtests need tobe carried out. Inthe first test,the
first sample is drawn out as the test sample. In the second test, the
secondsample is drawn as thetest sample, andso on.The ensemble
approach is a two-step process. In the first step, the jth sample is
drawn outas a test samplefromthe totalNnumbers of samples, and
the output of this test sample is denoted asT(j)o . In the first trial, the
Predicted Output
For testingShuffling the order of the trainingand validation samples
T1
T2
TM
T1
T2
TM
Predicted Outputs EnsembleOutput
TE
Final Result (byvoting)
Original data set
Trial M
Step 2
Trial 2
Trial 1
ModelTraining and
prediction
To
TargetOutput
Compare
P=1 if TE=To
P=0 if TETo
Step 1
To
Fig. 14. Iteration process of the ensemble approach.
-
8/11/2019 An intelligence-based route choice model for pedestrian flow in a transportation station
8/9
38 J.K.K. Yuen et al. / Applied Soft Computing 24 (2014) 3139
S pecific Ac curacy
ProbabilityDensity
0.6 0.65 0.7 0.75 0.8 0.85 0.90
50
100
150
2
0
0
250
300
350
400
450
Mean of Fraction of CorrectPredictions of the MLP =
0.7782
Mean of Fraction of CorrectPredictions of the MLP withthe Leave-one-out approach
= 0.7948
Mean of Fraction of CorrectPredictions of the MLP withthe Ensemble approach =
0.8647
Beta distribution of the specific accuracy
Fig.15. Comparison ofthe accuracyof theMLP, leave-one-outapproachand ensem-
ble approach.
order of the remainingN1 samples isshuffled andthesamplesaregrouped into training samples and validation samples to train theANN model. Upon completion of training, thetrained ANN model is
applied to predict the output of the test sample. The predicted out-
put is then denoted as T(j)1 , where the superscript (j) refers to the
test where the jth sample is taken as the test sample and the sub-
script 1 refers to the first trial of this test. In the second trial, the
N1 samples areshuffled andgrouped again.The grouped trainingand validation samples are used to train another ANN model and
areapplied to thetest sampleto generate thenext predictedoutput
(i.e., T(j)2
). The process is repeatedMtimes (30 times). Eventually,
Mnumbers of predictedoutputs (i.e., T(j)1 , T
(j)2 , . . ., T
(j)M
) are created
on which a plurality vote is applied to obtain the final prediction
result of the jthT(j), as shown in Eq. (8).
T(j) =Mi=1
T(j)iM
(8)
The performance of theprediction using theensembleapproach
was evaluated by Eq. (9).
P(j) =
1 if T(j) = T(j)o0 if T(j) /= T(j)o
(9)
The process was repeated N times. The overall performance
of the ensemble approach was evaluated by
Ni=1
P(i)
N
100%,
which represents the percentage of correct predictions obtained
with this approach. For the 621 samples used in this study, thefraction of correct predictions using the ensemble approach was
0.8647.
Fig. 15 shows a comparison of the means of the fraction of
correct predictions of the MLP,leave-one-out approach and ensem-
ble approach, which are 0.7782, 0.7948 and 0.8647, respectively.
This result indicates that applying the ensemble approach effec-
tively improves the performance of the ANNprediction, and further
demonstrates the feasibility of applying ANN model prediction in
our route choice study.
The performance of theleave-one-out approach andthe ensem-
ble approach was further investigated by presenting the results in
confusion matrices, as shown in Tables 2 and 3. Each is a two-
by-two matrix. The rows of the matrix represent the predicted
escalators chosen and the columns the actual escalators chosen.
Table 2
Prediction results with the leave-one-out approach.
Actual output
(escalator 1)
Actual output
(escalator 2)
Predicted output (escalator 1) 40.14% 10.64%
Predicted output (escalator 2) 9.88% 39.34%
Table 3
Prediction results with the ensemble approach.
Actual output
(escalator 1)
Actual output
(escalator 2)
Predicted output (escalator 1) 43.00% 6.60%
Predicted output (escalator 2) 6.92% 43.48%
The diagonal elements are the percentages of correct predictions
in all scenarios (i.e., the escalator used by a passenger is correctly
predicted)and the off-diagonal elements representthe percentages
of incorrect predictions in all scenarios. The tables show that the
percentage of correct predictions for escalator 1 is almost the same
as that for escalator 2, with a sample ratio of escalator 1 to esca-
lator 2 of 308 to 313. The prediction accuracy and the predictionerror of MLP model with leave-one-out approach is respectively,
79.48% (40.14% + 39.34%) and 20.52% (10.64% + 9.88%). By contrast,
theprediction accuracy andthe prediction error of MLP model with
ensemble approach is respectively, 86.48% (43.00% + 43.48%) and
13.52% (6.60% +6.92%). It can also be observed that every entry
of the confusion matrix is improved by applying the ensemble
approach.
Prediction by the shortest distance
Ortuzar and Willumsen [14] concluded that passengers will
choose the path with the shortest travel time, travel distance or a
combination of both. The simplest way to model route choice is by
stipulating minimum pathchoice and this approach was adopted in
Eindhoven [37]. Therefore, we further compared our trained MLP
models with the prediction by the shortest path model to deter-
mine their effectiveness. For prediction by the shortest path, we
simply assumed that all passengers will select the shortest esca-
lator (escalator 1) as their desired path. The means of the fraction
of correct predictions was 49.60% which is significantly lower than
the trained MLP models as shown in Table 4.
Classical route choicemodel
The multinomial Logit (MNL) model was proposed as general-
isation to the deterministic model. Borgers and Timmermans [38]
presenteda MNLmodel forpedestrian route choicewithin city cen-tres in the Netherlands. In 1998, Cheung and Lam [1] adopted the
MNL model to predict the passengers choice between escalators
Table 4
Comparison of different modelsadopted in this study.
Model Means of the
fraction of correct
predictions
MLP model 77.82%
MLP model with leave-one-out approach 79.48%
MLP model with ensemble approach 86.47%
Prediction by shortest distance 49.60%
MNLmodel 71.34%
-
8/11/2019 An intelligence-based route choice model for pedestrian flow in a transportation station
9/9
J.K.K. Yuen et al. / Applied Soft Computing 24 (2014) 3139 39
and stairs in a station. Therefore, a MNLmodel also developed to
evaluate the MLP models developed in this study as follows:
P1=1
(1 + exp X) (10)
where P1is the probability of passengers choosing escalator 1 and
X is the different of the perceived total travel time between using
escalator 1 and escalator 2, i.e.
X= a+ b(t1 t2) (11)where a and b are the parameters to be estimated, t1 and t2 is
respectively the perceived total travel time from the origin (i.e.,
position 1 to position 6) to the escalator 1 and escalator 2.
The result of passenger route choice by using the MNLmodel is
given as follow:
P1=1
1 + exp (5.7931 3.5096)t) (12)
The means of thefractionof correct predictions predictedby the
MNLmodel was 71.34%.
Table 4 shows a comparison of the performances of MLP model,
MLP model with leave-one-out approach, MLP model with ensem-
ble approach, prediction by the shortest distance, and MNLmodel.
The performances of the MLP models were better than those of theMNL model and the prediction by the shortest path in this case.
The mean specific accuracy of the MLP model with the ensemble
approach was 86.47%, which indicates that the performance of the
developed ANN model mimics route selection behaviour reason-
ably well in this study. It should be noted that even our trained
ANN models provided a better performance than the MNLmodel
and the prediction by the shortest distance in our case. This study
only provide an alternative-tool to predict route choiceby using the
ANN models, it cannot be concluded that our trained ANN model is
superior to the other models.
Conclusion
This study develops an intelligent approach to mimic the gen-eral behaviour of passengers making a route choice between two
escalators in a transportation station, based on the passengers
walking velocities and positions and the passenger densities at the
entrances of the escalators. This model is useful for both station
design and daily operation, as escalators are a critical transporta-
tion facility in transportation stations. The approach provides a
rapid method for engineers to estimate the loadings of escalators,
even for new stations, so that they can optimise their utilisation
to achieve maximum efficiency. This study successfully demon-
strates the feasibility of this approach. In future studies, we will
explore other facilities to improve overall crowd movement inside
transportation stations.
Acknowledgements
The authors wouldlike to acknowledge the MassTransitRailway
Corporation (MTRC) for their support for this research. The work
that is described in this paper was fully supported by a grant from
CityU (Project No. 7008028).
References
[1] C.Y. Cheung, H.K. Lam, Pedestrian routes choices between escalator and stair-way in MTRstations,J. Transp. Eng. 124 (1998) 277285.
[2] D. Helbing, P. Molnar, Social force model forpedestrian dynamics,Phys.Rev. E51 (5)(1995)42824286.
[3] S. Wolfram, Statistical mechanics of cellular automata, Rev. Mod. Phys. 55 (3)(1983) 601644.
[4] Y. Tajima, T. Nagatani, Scaling behavior of crowd flow outside a hall, Phys. A292 (2001) 545554.
[5] G.E. Bradley, A Proposed Mathematical Model for Computer Prediction ofCrowd Movements and their Associated Risks, Engineering for Crowd Safety,Elsevier, Amsterdam, 1993, pp. 303311.
[6] R.L. Goldstone, M.A. Janssen, Computational models of collective behavior,Trends Cogn. Sci. 9 (9) (2005) 424430.
[7] S.M. Lo,G.S. Zhi, An evacuation model:the SGEM package, Fire Saf. J. 39 (2004)169190.
[8] S.M. Lo, H.C. Huang, P. Wang, K.K. Yuen, A game theory based exit selectionmodel for evacuation,Fire Saf. J. 41 (2006) 364369.
[9] G.G.Lovas, Modelingand simulation of pedestriantrafficflow, Transp. Res.PartB 28B (1994) 429443.
[10] J.M. Smith, State-dependent queuing models in emergency evacuationnetworks, Transp. Res. Part B 25B (6) (1991) 373389.
[11] P. Thompson,E. Marchant,A computermodelfor theevacuationof largebuild-ing populations,Fire Saf. J. 24 (1995) 131148.
[12] S. Gwynne, E.R. Galea, P.J. Lawrence, L. Filippidis, Modelling occupant interac-tionwith fireconditions using the buildingEXODUSevacuation model, FireSaf.
J. 36 (2001) 327357.[13] H. Entamo, S. Helivaara, S. Hostikka, T. Korhonen, Modeling evacuees exit
selection with best response dynamics, in: 4th International Conference onPedestrian and Evacuation Dynamics, 2008.
[14] J. de D. Ortuzar, L.G. Willumsen, Modelling Transport, John Wiley & Sons, Inc.,New York, 1994.
[15] G. Proulx, A stress model forpeople facinga fire, J. Environ. Psychol. 13 (1993)137147.
[16] S.P.Hoogendoorn, P.H.H. Bovy, Pedestrian route-choice andactivityscheduling
theoryand models, Transp. Res. Part B 38 (2004) 169190.[17] J. Pauls, Calculating evacuation times for tall buildings, Fire Saf. J., Amst. 12
(1987) 237245.[18] K . Togawa, St udy of fi re es capes bas ing on th e o bse rv at ion of multitude
currents, Report No. 14, Building Research Institute, Ministry of Construction,Tokyo, 1995.
[19] G.Q.Zhang, B. Patuwo, M.Y.Hu, Forecasting withartificialneural networks:thestate of art, Int. J. Forecast. 14 (1998) 3562.
[20] F. Rosenblatt, Principles of Neurodynamics, Spartan Books, New York, 1962.[21] S. Srinivasulu, A. Jain, A comparative analysis of training methods fr artifi-
cial neural network rainfall-runoff models, Appl. Soft Comput. 6 (3) (2006)295306.
[22] S.S. Panda, D. Chakraborty, S.K. Pal, Flank wear prediction in drilling usingback propagation neural network and radialbasis function network, Appl. SoftComput. 8 (2) (2008) 858871.
[23] W.M. Lee, K.K. Yuen, S.M. Lo, K.C. Lam, Prediction of sprinkler actuation timeusing the artificial neural network, J. Build. Surv. 2 (1)(2000)1013.
[24] E.W.M. Lee, K.K. Yuen, S.M. Lo, K.C. Lam, G.H. Yeoh, A novel artificial neu-
ral network fire model for prediction of thermal interface location in singlecompartment fire, Fire Saf. J. 39 (2004) 6787.[25] R.K.K. Yuen, E.W.M. Lee, C.P. Lim, G.W.Y. Cheng, Fusion of GRNN and FA for
online noisy data regression, Neural Process. Lett. 19 (2004) 227241.[26] R.K.K. Yuen,E.W.M. Lee,S.M.Lo, G.H.Yeoh,Predictionof temperature andveloc-
ity profiles in single compartment fire by the intelligent system, Fire Saf. J. 41(6) (2006) 478485.
[27] K. Hornik, M. Stinchcombe, H. White, Multilayerfeedforward network are uni-versal approximators, Neural Netw. 2 (1989) 359366.
[28] J.K.K. Yuen, E.W.M. Lee, The effect of overtaking behavior on unidirectionalpedestrian flow, Saf. Sci. 50 (8) (2012) 17041714.
[29] D. Helbing, L. Buzna, A. Johansson, T. Werner, Self-organized pedestriancrowddynamics: experiments,Simul. Des. Solut.39 (1) (2005) 124.
[30] J.J. Fruin, Pedestrian Planning and Design, Metropolitan Association of UrbanDesigners and Environmental Planners, New York, 1971.
[31] Ward, Neuroshell 2 Manual, Ward Systems Group Inc., Frederick, MA, 1996.[32] P.J. Werbos, Beyond Regression New Tools for Prediction and Analysis in the
Behavioural Sciences (PhD dissertation), Applied Mathematics, Harvard Uni-versity, Boston, 1974.
[33] L. Breiman, Bagging predictors, Mach. Learn.24 (1996) 123140.[34] B. Efron, Estimating the error rate of a prediction rule: improvement on cross
validation, J. Am.Stat. Assoc. 78 (382) (1983) 316330.[35] A.K. Jain, R.C. Dube, C.C. Chen, Bootstrap techniques for error estimation, IEEE
Trans.Pattern Anal. Mach. Intell., PAMI 9 (5) (1987) 628633.[36] K. Teknomo, Microscopic Pedestrian Flow Characteristics: Development of an
Image Processing Data Collection and Simulation Model (PhD thesis), Depart-ment of Human Social Information Sciences, Graduate School of InformationSciences, Tohoku University, Japan, 2002.
[37] P.H.L. Bovy, E. Stem, Route Choice Wayfinding in Transport Networks, Studiesin Operational Regional Science, No. 9, Kluwer Academic Publishers, Hingham,MA, U.S.A., 1990.
[38] A.W.J. Borgers,H.J.P.Timmermans,Cityucentreentrypoints,storelocationpat-terns and pedestrian route choice behaviour: a microlevel simulation model,SocioEcon.Plan. Sci. 20 (1986) 2531.
http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0005http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0005http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0005http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0005http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0005http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0010http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0010http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0010http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0010http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0015http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0015http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0015http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0020http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0020http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0020http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0025http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0025http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0025http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0025http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0030http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0030http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0030http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0035http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0035http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0035http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0040http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0040http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0040http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0045http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0045http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0045http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0045http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0050http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0050http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0050http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0055http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0055http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0055http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0060http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0060http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0060http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0060http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0060http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0065http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0065http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0065http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0070http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0070http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0075http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0075http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0075http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0100http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0165http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0165http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0190http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0185http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0180http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0175http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0170http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0165http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0165http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0165http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0165http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0165http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0165http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0165http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0165http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0165http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0165http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0160http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0155http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0150http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0145http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0140http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0135http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0130http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0125http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0120http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0115http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0110http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0105http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0100http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0100http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0100http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0100http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0100http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0100http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0100http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0100http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0100http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0100http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0095http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0090http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0085http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0080http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0075http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0075http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0075http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0075http://refhub.elsevier.com/S1568-4946(14)00258-0/sbref0075http://refhub.elsevier.com/S1568-4946(1