You are on page 1of 28

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/263857557

The validation of evacuation simulation models through the analysis of


behavioural uncertainty

Article  in  Reliability Engineering [?] System Safety · November 2014


DOI: 10.1016/j.ress.2014.07.007

CITATIONS READS

86 2,087

3 authors:

Ruggiero Lovreglio Enrico Ronchi


Massey University Lund University
103 PUBLICATIONS   1,882 CITATIONS    130 PUBLICATIONS   2,992 CITATIONS   

SEE PROFILE SEE PROFILE

Dino Borri
Politecnico di Bari
75 PUBLICATIONS   598 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Evacuation Study View project

Physiological Limitations of Oxygen Uptake (Kinetics) and Leg Muscle Activity (Fatigue) during Ascending Evacuation using Stairs View project

All content following this page was uploaded by Ruggiero Lovreglio on 11 October 2017.

The user has requested enhancement of the downloaded file.


Published as: Lovreglio R., Ronchi E., Borri D., 2014, The validation of evacuation simulation models through the
analysis of behavioural uncertainty, Reliability Engineering & System Safety, Vol. 131, pp. 166-174, DOI:
http://dx.doi.org/10.1016/j.ress.2014.07.007

The validation of evacuation simulation models through the analysis of


behavioural uncertainty

Ruggiero Lovreglioa, Enrico Ronchib, Dino Borric

a
Department of Civil, Environmental, Planning, Building and Chemistry, Technical University of
Bari, Via Edoardo Orabona, 4, 70126 Bari, Italy; r.lovreglio@poliba.it
b
Department of Fire Safety Engineering, Lund University, P.O. Box 118 SE-221 00 Lund, Sweden;
enrico.ronchi@brand.lth.se
c
Department of Civil, Environmental, Planning, Building and Chemistry, Technical University of
Bari, Via Edoardo Orabona, 4, 70126 Bari, Italy; d.borri@poliba.it

Address correspondence to:


E. Ronchi, e-mail: enrico.ronchi@brand.lth.se ; Phone: +460462227694, P.O. Box 118 SE-221 00
Lund, Sweden

Abstract
Both experimental and simulation data on fire evacuation are influenced by a component of
uncertainty caused by the impact of the unexplained variance in human behaviour, namely
behavioural uncertainty (BU). Evacuation model validation studies should include the study of this
type of uncertainty during the comparison of experiments and simulation results. An evacuation
model validation procedure is introduced in this paper to study the impact of BU. This methodology
is presented through a case study for the comparison between repeated experimental data and
simulation results produced by FDS+Evac, an evacuation model for the simulation of human
behaviour in fire, which makes use of distribution laws.

Keywords: evacuation modelling; model validation; behavioural uncertainty; evacuation


experiments; FDS+Evac
Abbreviations:
ASET: Available Safe Egress Time
BU: behavioural uncertainty
ERD: Euclidean Relative Difference
EPC: Euclidean Projection Coefficient
ExC: Experimental Curve
FED: Fractional Effective Dose
OETC: Occupant-Evacuation Time Curve
RD: Relative Difference
RSET: Required Safe Egress Time
SC: Secant Cosine
SD: Standard Deviation
SmC: Simulated Curve
𝑡̅: Average evacuation time
TET: Total Evacuation Time
TR: Threshold
TV: Threshold Values
1. Introduction

The achievement of an acceptable level of safety is generally the main objective of fire safety engineering
[1]. In the context of fire evacuation, the level of fire safety is studied by comparing the time required by the
occupants to evacuate buildings safely (i.e. the RSET, Required Safe Egress Time) and the time until the
conditions are not tenable (i.e. the ASET, Available Safe Egress Time) [2]. In order to estimate and compare
these two times, a variety of egress simulation models have been developed [3] [4] [5]. Validation studies are
needed to quantify the accuracy of the prediction of simulation models and their subsequent suitability for
fire safety studies [6].

Validation studies can be distinguished between qualitative versus quantitative and macroscopic versus
microscopic analyses [7] [8] [9]. Qualitative validation is the easiest procedure; it allows verifying if certain
collective phenomena are reproduced by the models (e.g., formation of congestions) without quantifying it
analytically. Quantitative validation provides analytic criteria to compare outcomes. Moreover, quantitative
studies can be divided in macroscopic or microscopic [8] [9]. Macroscopic validation focuses on observable
data averaged over time, space or agents, while microscopic validation tests investigate individual agents’
properties (e.g. walking speeds, travel paths, etc.).

To date, different validation approaches have been used. Kuligowski et al. [3] categorizes these approaches
according to the benchmark employed for the comparison: (1) against code requirements; (2) against fire
drills or other people movement experiments/trials; (3) against literature on past experiments and (4) against
other models. Kady and Tolk [10] provide information on the methods used to validate evacuation models
adopting those different approaches. The comparison against code or other simulation models does not allow
a direct comparison between experimental and simulated data because results of models are compared with
safety expected value established by the legislature on the base of previous studies. This is instead possible
in the case of dedicated fire drills or experiments/trials or past experiments.

This issue is reflected in the methods adopted for the study of uncertainty in experiments and simulation
results. Uncertainty in evacuation modelling is divided in four different components: model input uncertainty
(uncertainty of the parameters used as input), measurement uncertainty (uncertainty of data collection and its
measurement techniques), intrinsic uncertainty (uncertainty of the model formulation and its physical and
mathematical assumptions and methods), and behavioural uncertainty (BU) [11] [12]. BU is a peculiarity of
evacuation data if compared with fire simulations and it is closely connected to the representation of human
behaviour during evacuations situations [11]. Even if a behavioural experiment is repeated under identical
conditions, the outcome may vary to a certain degree [13].

To date, two approaches have been used in order to model human behaviour and its uncertainty during
evacuation: deterministic and stochastic [8]. The deterministic approach is easier to use but it has the
limitation of representing average behaviours only [14] [15]. Thus, a deterministic model or sub-model may
not be able to represent the unexplained variance in human behaviour associated with human behaviour
in an exhaustive way. The stochastic approach is a strategy to simulate BU because it allows modelling
different behaviours starting from the same conditions [11] [15] [16]. According to this approach, three
strategies can be used. The first one employs a deterministic model together with random input represented
by probabilistic distributions [11] [17] [18]. The second strategy is to use stochastic models (e.g., Random
Utility Models). These models directly reproduce BU by generating different results starting from the same
input [19]. The third strategy is the combination of a stochastic model and random input. Stochastic models
may not have closed formulation and may require numerical solutions that increase the computational
burden. Nevertheless, the advances in computer science gradually reduce the problems associated with
computational costs [20].

Thus, similarly to a single experiment, also a single model run (obtained with a stochastic model or by a
random input) may not be representative of a full range of occupant behaviours [11]. However, to date, the
majority of validation studies are based on the comparison of a single experimental data-set with simulation
results [3] [21] [22] [23] [24].

The definition of the appropriate number of runs for the simulation of a single evacuation scenario is a key
issue in egress modelling. To address this issue, Ronchi et al. [11] proposed a methodology based on
functional analysis operators. This method is based on a set of criteria used to study the convergence of the
simulated occupant-evacuation time curves (OETCs) representing data using vectors. This method allows
estimating an average OETC with a pre-defined acceptable threshold of BU. This approach is based on the
central limit theorem, i.e. the impact of BU on average behaviours tends to decrease with the increase of the
number of runs.

To date, functional analysis operators have been used to compare both fire model and evacuation model
results with experimental data [1] [25] [26]. However, functional analysis operators have not been used to
perform a quantitative comparison between evacuation experimental and simulation data taking into account
BU.

This paper introduces a methodology to perform macroscopic quantitative validation studies which include
BU. This method is based on the concept that BU and the convergence towards the OETCs produced by
evacuation models and observed during repeated evacuation experiments can be studied using functional
analysis operators [11]. The scope is the performance of a quantitative comparison between BUs from
experimental and simulated data. An explanatory case study is presented in this paper to introduce the
validation procedure. Simulation results using an open source computational model for fire evacuation (i.e.,
the evacuation model FDS+Evac [27]) are produced for the study of a set of building evacuation experiments
conducted by Guo et al. [28]. A discussion on the advantages and limitations of the proposed methodology
are presented.

2. Methodology

In order to describe the proposed method, the convergence measures presented by Ronchi et al. [11] are
briefly presented. Those measures are used as benchmark to perform statistical tests for the comparison
between evacuation model predictions and experimental data.

2.1 Convergence measures

The convergence measures used for the study of BU are based on functional analysis [11]. Functional
analysis is here intended as the branch of mathematics used for the study of vector spaces through operators.
Functional analysis operators are here used to investigate the OETCs, which are studied as vectors where the
components are ordered times. Three concepts of functional analysis can be used to analyses these curves,
namely the Euclidean Relative Difference (ERD), the Euclidean Projection Coefficient (EPC) and the Secant
Cosine (SC) [11] [25] [26] (see the Appendix of this paper). The mathematical formulations for these
functional operators are described in detail by Ronchi et al. [11] and Galea et al. [24].

Let x = {x1; x2; …; xk;…; xn} and y = {y1; y2; …; xk; …; yn} be n dimensional vectors which describe two
different OETCs, then ERD quantifies the normalized difference between the two curves x and y. Therefore,
a small value of ERD indicates small differences between the two curves. On the other hand, EPC quantifies
the normalized scalar product of the two curves x and y. Thus, if EPC is multiplied by x or y, then the
distance between the vectors is reduced to its achievable minimum. SC is a factor that measures the
differences of the shapes of two curves by analysing an approximation of the first derivative of both curves.
SC incorporates the skin parameter, s, which is a factor that represents the period of noise due to microscopic
behaviours [24]. Thus, this parameter is useful for smoothing the noise of the data. The s parameter should
be neither too small nor too large. if it is too small than the noise effects may dominate the analysis whileas
if the value is too large that will lead to loss of the nature of the data [24]. Moreover it is chosen according to
the size of x and y (n). Therefore, it may be preferable to use the ratio s/n. However a preliminary sensitivity
analysis can be useful to understand how this parameter influences the SC.

Ronchi et al. [11] also introduce five convergence measures for the study of BU, which are presented in
Equations 1-5. They are based on the central limit theorem, i.e., they investigate the convergence towards the
average of consecutive aggregate occupant-evacuation curves. Equation 1 and 2 investigate the variation of
Total Evacuation Time (TETi), namely the highest time of an OETC. In fact, Equation 1 and 2 investigates
̅̅̅̅̅̅𝑖 ) and standard deviations (SDi) of total evacuation times from
respectively the consecutive average (𝑇𝐸𝑇
different curves. Similarly, convergence measures ERDconv i, EPCconv i and SCconv i are produced for the
functional analysis operators ERDi, EPCi and SCi.
̅̅̅̅̅̅𝑖 − 𝑇𝐸𝑇
𝑇𝐸𝑇 ̅̅̅̅̅̅𝑖−1
𝑇𝐸𝑇𝑐𝑜𝑛𝑣 𝑖 = | | Eq 1; [11]
̅̅̅̅̅̅𝑖
𝑇𝐸𝑇
𝑆𝐷𝑖 − 𝑆𝐷𝑖−1
𝑆𝐷𝑐𝑜𝑛𝑣 𝑖 = | | Eq. 2; [11]
𝑆𝐷𝑖

𝐸𝑅𝐷𝑐𝑜𝑛𝑣 𝑖 = |𝐸𝑅𝐷𝑖 − 𝐸𝑅𝐷𝑖−1 | Eq. 3; [11]

𝐸𝑃𝐶𝑐𝑜𝑛𝑣 𝑖 = |𝐸𝑃𝐶𝑖 − 𝐸𝑃𝐶𝑖−1 | Eq. 4; [11]

𝑆𝐶𝑐𝑜𝑛𝑣 𝑖 = |𝑆𝐶𝑖 − 𝑆𝐶𝑖−1 | Eq. 5;[11]

2.2 Validation procedure

Three different steps are proposed for the validation procedure. The first step deals with the
assessment of BU within experimental data. To address this issue, repeated experimental data of the
same scenario are necessary to define the experimental OETCs (see Step 1).

Figure 1 - Schematic flow chart of the validation procedure


Step 1

Step 1 analyses the experimental data. The main goal of this step is to investigate BU in
experimental data and its variation around the average value with the increase of the number of
trials. In order to make this estimation, the convergence measures shown in Section 2.1 are used.
Thus, Equations 1-5 are applied to the m experimental OETCs (see Equation 6).

𝐸 𝐸 𝐸 𝐸 𝐸
(𝑇𝐸𝑇𝑐𝑜𝑛𝑣 𝑗 ; 𝑆𝐷𝑐𝑜𝑛𝑣 𝑗 ; 𝐸𝑅𝐷𝑐𝑜𝑛𝑣 𝑗 ; 𝐸𝑃𝐶𝑐𝑜𝑛𝑣 𝑗 ; 𝑆𝐶𝑐𝑜𝑛𝑣 𝑗 ) j = 1,…, m-2 Eq. 6

Starting from these measures, the assessment of threshold values for the five convergence measures
is performed, namely TRTET, TRSD, TRERD, TREPC and TRSC. These thresholds will be used during
the study of simulation results in Step 2 in order to perform the convergence criteria proposed.

Step 2

Step 2 is focused on the analysis of evacuation simulation results. The number of repeated runs of a
single scenario is defined in accordance with a set of acceptance criteria on the convergence of
evacuation model predictions. This criterion is based on the comparison of the convergence
measures used to investigate the experimental and simulated BUs. In fact, even if the simulated BU
around the average performance is unknown a priori, the impact of a single run on average tends to
decrease with the increase of the number of runs due to the law of large numbers. Thus, the
simulated BU becomes lower than the experimental BU after an a priori undefined number of runs.

In this step, an evacuation model is used to simulate the evacuation scenario. A finite set of n runs
of the same evacuation scenario are simulated. Thus, n OETCs are produced and the convergence
measures can be calculated for the simulation results (see Equation 7).

𝑆 𝑆 𝑆 𝑆 𝑆
(𝑇𝐸𝑇𝑐𝑜𝑛𝑣 𝑗 ; 𝑆𝐷𝑐𝑜𝑛𝑣 𝑗 ; 𝐸𝑅𝐷𝑐𝑜𝑛𝑣 𝑗 ; 𝐸𝑃𝐶𝑐𝑜𝑛𝑣 𝑗 ; 𝑆𝐶𝑐𝑜𝑛𝑣 𝑗 ) j = 1,…, n-2 Eq. 7

A set of acceptance criteria on the convergence of simulation results can be defined in accordance
with Equations 8- 12.

𝑆
𝑇𝐸𝑇𝑐𝑜𝑛𝑣 𝑗 < TR TET for b consecutive number of runs Eq. 8
𝑆
𝑆𝐷𝑐𝑜𝑛𝑣 𝑗 < TR SD for b consecutive number of runs Eq. 9

𝑆
𝐸𝑅𝐷𝑐𝑜𝑛𝑣 𝑗 < TR ERD for b consecutive number of runs Eq.10

𝑆
𝐸𝑃𝐶𝑐𝑜𝑛𝑣 𝑗 < TR EPC for b consecutive number of runs Eq.11

𝑆
𝑆𝐶𝑐𝑜𝑛𝑣 𝑗 < TR SC for b consecutive number of runs Eq.12

where j varies between n (total runs of simulation) and n-b. Thus, 5b inequities are studied. The b
parameter is necessary to verify that convergence measures are stable under certain thresholds.

Depending on the acceptance criteria (which may also include a sub-set of the inequalities), an
average OETC is produced in which the impact of simulated BU on the average behaviour is lower
than the experimental. This allows a comparison between experimental data and simulation results.

Step 3

Step 3 deals with the comparison between experimental data and simulation results. This
comparison includes four different criteria.

The first criterion is based on functional analysis operators and deals with the comparison between
experimental and simulated average OETCs according to a generalization of the criteria proposed
by Galea et al. [24]. Thus, a quantitative comparison can be performed by using the measures
introduced in Section 2.1, namely, the relative difference of experimental and simulated average
TETs (RDTET), the ERD, the EPC and the SC. The comparison between experimental and simulated
curves can be made according to different approaches: comparison between average experimental
and simulated curves; comparison between average experimental and the best or worst TET
simulated curves (i.e. the simulated curves that show the TET closest or farthest to the experimental
average TET) and equally for the other parameters: ERD, EPC and SC; etc..

Equations 13 - 16 show the application of this first criterion for experimental (ExC) and simulated
(SmC) curves.
|𝑇𝐸𝑇(𝑬𝒙𝑪) − 𝑇𝐸𝑇(𝑺𝒎𝑪)|
𝑅𝐷𝑇𝐸𝑇 (𝑬𝒙𝑪, 𝑺𝒎𝑪) = ≤ 𝑇𝑉𝑅𝐷 Eq.13
𝑇𝐸𝑇(𝑬𝒙𝑪)

𝐸𝑅𝐷(𝑬𝒙𝑪, 𝑺𝒎𝑪) ≤ 𝑇𝑉𝐸𝑅𝐷 Eq.14

𝑇𝑉𝐸𝑃𝐶 ′ ≤ 𝐸𝑃𝐶(𝑬𝒙𝑪, 𝑺𝒎𝑪) ≤ 𝑇𝑉𝐸𝑃𝐶 ′′ Eq.15

𝑆𝐶(𝑬𝒙𝑪, 𝑺𝒎𝑪) ≥ 𝑇𝑉𝑆𝐶 Eq.16

These criteria allow model users to quantify the differences between two curves. In fact, those
criteria compare the distance (𝑅𝐷𝑇𝐸𝑇 ; ERD; EPC) and the shape (SC) between the two curves. If
two curves are identical, then 𝑅𝐷𝑇𝐸𝑇 and ERD are equal to zero and EPC and SC are equal to 1.
EPC and SC may return value close to 1 although the average experimental and simulated curves
are not identical. The thresholds (𝑇𝑉𝑅𝐷 , 𝑇𝑉𝐸𝑅𝐷 , 𝑇𝑉𝐸𝑃𝐶 ′, 𝑇𝑉𝐸𝑃𝐶 ′′ and 𝑇𝑉𝑆𝐶 ) shown in Equation 13 -
16 are chosen by the tester in line with the required acceptance criteria of the accuracy of the model
predictions.

The second criterion investigates if the experimental and simulated BUs belong to the same
population in terms of the results of the above described functional analysis operators (RDTET, ERD,
EPC and SC). These operators can be estimated comparing both experimental and simulated
𝐸/𝑆
OETCs with the relative average curves. Thus, six couples of series are available ( 𝑅𝐷𝑇𝐸𝑇 𝑖 ,

𝐸𝑅𝐷𝐸/𝑆 𝑖 , 𝐸𝑃𝐶𝐸/𝑆 𝑖 and 𝑆𝐶𝐸/𝑆 𝑖 ). In order to perform the analysis, standard statistical tests can be
used to verify if these couple of series belong to the same distributions.

The third criterion verifies if the average time of experimental data and simulation results belong to
the same population applying statistical hypothesis testing. This criterion is a physical indirect
measure of the average evacuation times.

Let t = {t1; t2; …; tk;…; tn} an n dimensional vector, which represents an OETC, and p = {p1; p2;
…; pk;…; pn} an n dimensional vector, which represents the evacuated people at the time of t
(Figure 2), then the area subtended by the curve is shown in Equation 17.
Figure 2 – Example of occupant- evacuation time curve.

𝑛
𝑥𝑖 + 𝑥𝑖−1
𝐴𝑠𝑢𝑏 = ∑ (𝑝𝑖 − 𝑝𝑖−1 ) Eq.17
2
𝑖=2

If pi are a series of consecutive natural numbers then the area can be approximated according to
Equation 18.

𝑛 𝑛
1
𝐴𝑠𝑢𝑏 = ∑ 𝑥𝑖 − (𝑥1 + 𝑥𝑛 ) ≈ ∑ 𝑥𝑖 Eq.18
2
𝑖=1 𝑖=1

Figure 2 shows the amount of surplus of area (dark grey) which derives from the approximation
proposed in Equation 17. Thus, the area subtended by an experimental or simulated OETC can be
considered as an indirect measure of the average evacuation time (𝑡̅) of the same curve (Equation
19).

𝑛
1 1
𝑡̅ = ∑ 𝑥𝑖 = 𝐴𝑠𝑢𝑏 Eq.19
𝑛 𝑛
𝑖=1

Therefore, this criterion allows the study of the average simulated evacuation times (𝑡𝑖̅ 𝑆 ), i.e. if they
belong to the same population of average experimental evacuation times (𝑡𝑖𝐸̅ ).

In order to perform this analysis, different statistical tests can be used in relation to the data under
consideration (e.g., number of occupants in the building, etc.).
3. Case study

An explanatory case study is presented in this section as an application of the methodology for the
study of BU during validation studies. An experimental data-set available in the literature on
building evacuation has been employed [28]. This evacuation data-set includes repeated
experimental OETCs for the same scenario. The same scenario is simulated using the evacuation
model FDS+Evac [27]. FDS+Evac is an agent-based model which uses the social force model by
Helbing and Molnár [29] to simulate people movement in the case of fire. Then, this model allows
simulating the interaction between fire/smoke and agents through the concept of Fractional
Effective Dose (FED) [30] and a sub-model on the influence of visibility on agents speed based on
the experiments by Frantzich and Nilsson [31]. This model was chosen because both the runs of
simulations and the analysis of the outcomes can be easily managed by external code (e.g. VB.Net).

A tool is developed in VB.Net language to perform the validation procedure. This tool implements
the convergence measures and the validation procedure presented in Section 2 for the evacuation
model FDS+Evac. The output of the tool is the OETCs simulated by FDS+Evac and the
corresponding convergence measures produced for each consecutive run.

3.1 Step 1: Experimental data and threshold values

Guo et al. [28] conducted an evacuation experiment with 30 participants who repeatedly had to
evacuate simultaneously from a classroom. In total 12 repeated trials were realized in this
experiment. Two independent variables were manipulated. First, participants were randomly
assigned to one of six different seating positions in each trial. Second, in 50% of the trials
participants were blindfolded to simulate zero visibility conditions. That is, each participants
experienced each seating position once blindfolded and once not with full visibility. In each trial
participants started to evacuate from the classroom after they were given an acoustic signal.
Detailed information on the geometric layout, experimental conditions and data collection
procedures of the evacuation scenario are available in Guo et al. [28].

Six curves from experiments with full visibility are considered in order to estimate the experimental
BU from experiments with same conditions. The experimental OETCs of the six experiments are
shown in Figure 3.
30
Exp 1
25
Exp 2

20 Exp 3

Exp 4
Time [s]

15
Exp 5

10 Exp 6

Average
5 Curve

0
0 5 10 15 20 25 30
Evacuees [n]

Figure 3 - Experimental occupant - evacuation time curves

The convergence measures for these experimental curves are calculated. The calculation of the
SCconv is done setting s equal to 1, 2, 3 and 5. Table 1 and Figure 4 show all measures calculated
while Figure 5 shows the results depending on the value for skip s. In Table 1 the first two lines are
blank since ERDconv, EPCconv, SCconv can be estimated only with at least three repeated experiments
(see Equations A.1-A.5 and Appendix 1).

Table 1 - Experimental convergence measures (TETconv, SDconv, ERDconv, EPCconv, SCconv)


according to Equation 1-5.
Experiment TETconv SDconv ERDconv EPCconv SCconv (s/n=1) SCconv (s=2) SCconv (s=3) SCconv (s=5)
Experiment 1-2 / / / / / / / /
Experiment 2-2 / / / / / / / /
Experiment 3-2 0.021 0.230 0.062 0.106 0.0452 0.0259 0.0143 0.0090
Experiment 4-2 0.020 0.189 0.012 0.035 0.0036 0.0010 0.0013 0.0007
Experiment 5-2 0.006 0.113 0.004 0.015 0.0040 0.0006 0.0002 0.0001
Experiment 6-2 0.014 0.140 0.015 0.021 0.0010 0.0010 0.0008 0.0003
0,25
TETconv

TETconv; SDconv; ERDconv; EPCconv


SDconv
0,20
ERDconv
EPCconv
0,15

0,10

0,05

0,00
0 1 2 3 4 5 6 7
N° of experiments

Figure 4 – Experimental convergence measures versus number of runs.

0,050
SCconv (s=1)
0,045
SCconv (s=2)
0,040
SCconv (s=3)
0,035
SCconv (s=5)
0,030
SCconv

0,025
0,020
0,015
0,010
0,005
0,000
0 1 2 3 4 5 6 7
N° of experiment

Figure 5 – Variation of SCconv for different value the skip s

Figures 4 and 5 are used to define the convergence measures and experimental BU (Equation 20).

TR TET = 0.05; TR SD = 0.15; TR ERD = 0.05; TR EPC = 0.05; TR SC = 0.002 Eq.20

Figure 5 shows the impact of the skip s on the results. A large number for s parameter may
significantly modify the shape of the curves. Moreover, Galea et al. [24] suggest that reasonable
values of the s/n ratio are 0.03 - 0.05. According to this case study (n = 30), this means that s should
be equal to 0.9 - 1.5. However, a value equal to 2 is chosen because s must be an integer and greater
than 1 in order to smooth the noise of the data during the comparison.

An estimation of BU is performed for the experimental data. BU associated with simulation results
should be in the same order of magnitude to the experimental uncertainty. Due to the central limit
theorem, an increase in the number of runs for a same scenario leads to a lower impact of BU in
average simulation results. Thus, the threshold values were obtained by reducing the values shown
in Equation 20 by an order of magnitude (Equation 21).

TR TET = 0.005; TR SD = 0.015; TR ERD = 0.005; TR EPC = 0.005; TR SC = 0.0002 Eq.21


The number of consecutive runs b is set equal to 10. The choice is based on the study of the impact
of b on evacuation simulation results, i.e., the average OETCs for b equal to 10, 15 and 20 have all
a maximum error equal to 2% (Figure 6).

80

70

60

50
b=
Time [s]

40 20
b=
30 15
b=
20 10

10

0
0 5 10 15 20 25 30
Evacuees [n]

Figure 6 – Variation of average OETC for different values of b

3.2 Step 2: FDS+Evac simulations

A series of evacuation simulations are run with the evacuation model FDS+Evac [27]. The
evacuation scenario is represented within the evacuation model. For each run, 30 agents are
randomly assigned to the rows of seats used in the experiments. The remaining variables which
characterize the agents are set using the default values provided by the model developers (i.e. body
size distributions, walking speed distributions). The agents’ characteristics are based on the “Adult”
category [32]. This category was chosen because the students in the experiments are able-bodied.
Walking speeds of the agents are assumed to follow a normal distribution (average = 2.1 m/s and
standard deviation = 0.25 m/s) in line with data from similar experiments [33]. Pre-evacuation times
are set equal to 0 s for all agents in order to represent the experimental conditions, i.e., participants
began to evacuate as soon as a signal is given to them.

The aforementioned tool is used to run the simulations and calculate the convergence measures. The
tool automatically stops the simulations once the convergence criteria are met (in this example, 52
runs were necessary). The simulated consecutive convergence measures are plotted in Figures 7-11.
Figures 7-11 show that the most restrictive parameters are TETconv and SDconv while ERDconv is the
least restrictive. Moreover, those figures compare the convergence measures of the simulation
results with both their threshold values and their corresponding experimental measures.

Figures 7-11 show that simulated TETconv, SDconv, ERDconv, EPCconv and SCconv are in the same of
order of magnitude of the corresponding experimental measures for a number of runs lower than 7.
The experimental and simulation curves intersect each other for the first 6 runs, for which
experimental data are available. However, the low number of experimental curves does not allow
studying if this tendency is maintained even for large number of runs.

0,100
Sim TETconv
0,090
Exp TETconv
0,080
Treshold TETconv
0,070
0,060
TETconv

0,050
0,040
0,030
0,020
0,010
0,000
0 10 20 30 40 50
N° of runs

Figure 7 – Variation of experimental and simulated TETconv with increasing of number of runs. The
threshold criterion is plotted through a horizontal straight line. TETconv meets convergence threshold
after 42 runs.
0,250
Sim SDconv

0,200 Exp SDconv

Treshold SDconv

0,150
SDconv

0,100

0,050

0,000
0 10 20 30 40 50
N° of runs

Figure 8 - Variation of experimental and simulated SDconv with increasing of number of runs. The
threshold criterion is plotted through a horizontal straight line. SDconv meets convergence threshold
after 42 runs.

0,070
Sim ERDconv
0,060
Exp ERDconv
0,050
Treshold ERDconv
ERDconv

0,040

0,030

0,020

0,010

0,000
0 10 20 30 40 50
N° of runs

Figure 9 - Variation of experimental and simulated ERDconv with increasing of number of runs. The
threshold criterion is plotted through a horizontal straight line. ERDconv meets convergence
threshold after 20 runs.
0,120
Sim EPCconv
0,100
Exp EPCconv
Treshold EPCconv
0,080
EPCconv

0,060

0,040

0,020

0,000
0 10 20 30 40 50
N° of runs

Figure 10 - Variation of experimental and simulated EPCconv with increasing of number of runs. The
threshold criterion is plotted through a horizontal straight line. EPCconv meets convergence threshold
after 41 runs.

0,016
Sim SCconv
0,014
Exp SCconv
0,012
Treshold SCconv
0,010
SCconv

0,008

0,006

0,004

0,002

0,000
0 10 20 30 40 50
N° of runs

Figure 11 - Variation of experimental and simulated SCconv with increasing of number of runs. The
threshold criterion is plotted through a horizontal straight line. SCconv meets convergence threshold
after 29 runs.

3.2 Step 3: Data comparison

Once the convergence criteria are met, a comparison between the experimental data and simulation
results is performed in accordance with different and independent criteria described in Section 2.2.
According to the first criteria, different experimental and simulated curves can be compared. In this
work the comparison is made for the simulated average, best and worst (TET , ERD, EPC and SC)
curves with the experimental average one. These curves are plotted in Figure 12 - 14, while Table 2
shows values of 4 functional measures for the different simulated curves and the experimental
average curve.

50
Exp Average Curve
45
Sim Average Curve
40
Sim best TET Curve
35
Sim worst TET Curve
30
Time [s]

25
20
15
10
5
0
0 5 10 15 20 25 30
Evacuees [n]

Figure 12 – Experimental and simulated averages OETCs and the best or worst TET simulated
OETCs.

50
Exp Average Curve
45
Sim Average Curve
40
Sim best ERD and EPC Curve
35
Sim worst ERD and EPC Curve
30
Time [s]

25
20
15
10
5
0
0 5 10 15 20 25 30
Evacuees [n]

Figure 13 - Experimental and simulated averages OETCs and the best or worst TET and EPC
simulated OETCs
50
Exp Average Curve
45
Sim Average Curve
40
Sim best SC Curve
35
Sim worst SC Curve
30
Time [s]

25
20
15
10
5
0
0 5 10 15 20 25 30
Evacuees [n]

Figure 14 - Experimental and simulated averages OETCs and the best or worst SC simulated
OETCs

Table 2 - RDTET, ERD, EPC and SC between experimental average OETCs and simulated average
and the best and worst OETCs. The font style changes when the parameters do not satisfy the less
restrictive threshold values by Galea et al. [24].
Parameters average best TET worst TET best ERD worst ERD best EPC worst EPC best SC worst SC
N° curve / 40 28 7 17 7 17 44 28
RDTET 0.34 0.05 0.86 0.17 0.64 0.17 0.64 0.26 0.64
ERD 0.23 0.08 0.31 0.08 0.36 0.08 0.37 0.21 0.37
EPC 0.77 0.92 0.69 0.93 0.65 0.93 0.65 0.80 0.65
SC 1.00 0.96 0.88 0.97 0.85 0.97 0.85 0.99 0.85

Some considerations can be made analysing the values shown in Table 2 and those suggested in
Section 2.2. Table 2 shows that the worst TET and SC curves are the same while the best and worst
ERD and EPC curves are the same. The three measures - which estimate distances between the
curves (RDTET; ERD; EPC) - show that the distances between the simulated average, the best TET,
ERD, EPC and SC curves and experimental average curve are apparently close to zero while the
shape (according to the first derivative) seems to be very similar (SC ≈ 1). Differently, the worst
TET, ERD, EPC and SC curve shows a greater distance from experimental average curve and
different shapes in accordance to Figure 12 - 14. This outcome is also confirmed by the comparison
between the value shown in Table 2 and the threshold values (less restrictive: 𝑇𝑉𝑅𝐷 = 0.45 ,
𝑇𝑉𝐸𝑅𝐷 = 0.45 , 𝑇𝑉𝐸𝑃𝐶 ′ = 0.6 , 𝑇𝑉𝐸𝑃𝐶 ′′ = 1.4 and 𝑇𝑉𝑆𝐶 = 0.6 ; restrictive: 𝑇𝑉𝑅𝐷 = 0.15 , 𝑇𝑉𝐸𝑅𝐷 =
0.25, 𝑇𝑉𝐸𝑃𝐶 ′ = 0.8, 𝑇𝑉𝐸𝑃𝐶 ′′ = 1.2 and 𝑇𝑉𝑆𝐶 = 0.8) provided by Galea et al. [24]. In fact, all worst
curves show a RDTET greater than the less restrictive threshold value, thus failing this test.
The second criterion deals with a statistical comparison (based on the functional operators: RDTET,
ERD, EPC and SC) between average OETCs with all the curves both for experimental and
𝐸/𝑆
simulated data. A normality test is used in order to know if the six couple of series (𝑅𝐷𝑇𝐸𝑇 𝑖 ,

𝐸𝑅𝐷𝐸/𝑆 𝑖 , 𝐸𝑃𝐶𝐸/𝑆 𝑖 and 𝑆𝐶𝐸/𝑆 𝑖 ) belong to a normal distributions. A Shapiro-Wilk Test is used for
experimental data while a Kolmogorov-Smirnov test for simulated data because of the different
sample sizes (Table 3).

𝐸/𝑆
Table 3 – Normality test for 𝑅𝐷𝑇𝐸𝑇 𝑖 , 𝐸𝑅𝐷𝐸/𝑆 𝑖 , 𝐸𝑃𝐶𝐸/𝑆 𝑖 and 𝑆𝐶𝐸/𝑆 𝑖
Sample Statistic df P-value
𝐸
𝑅𝐷𝑇𝐸𝑇 0.949 6 0.730
𝑆
𝑅𝐷𝑇𝐸𝑇 0.159 52 0.002
𝐸𝑅𝐷𝐸 0.881 6 0.273
𝐸𝑅𝐷 𝑆 0.120 52 0.059
𝐸𝑃𝐶 𝐸 0.980 6 0.951
𝐸𝑃𝐶 𝑆 0.068 52 0.200
𝑆𝐶 𝐸 0.885 6 0.292
𝑆𝐶 𝑆 0.142 52 0.011

𝐸
Table 3 show that all the experimental series ( 𝑅𝐷𝑇𝐸𝑇 𝑖
, 𝐸𝑅𝐷𝐸 𝑖 , 𝐸𝑃𝐶 𝐸 𝑖 and 𝑆𝐶 𝐸 𝑖 ) are from a
𝑆
normally distributed population. The simulated series 𝑅𝐷𝑇𝐸𝑇 and 𝑆𝐶 𝑆 are not from a normally
distributed population while simulated series 𝐸𝑅𝐷 𝑆 𝑖 and 𝐸𝑃𝐶 𝑆 𝑖 are normally distributed. Thus, 2
sample unpooled t-test is used to verify if the two series 𝐸𝑅𝐷𝐸/𝑆 and 𝐸𝑃𝐶𝐸/𝑆 belong to the same
normal distribution while U test is used to verify if the remaining series belong respectively to the
same distributions (Table 4).

𝐸/𝑆
Table 4 – T-test (𝐸𝑃𝐶𝐸/𝑆 ) and U-test (𝑅𝐷𝑇𝐸𝑇 , 𝐸𝑅𝐷 𝐸/𝑆 , 𝑆𝐶𝐸/𝑆 )
Sample Statistic P-value
𝐸/𝑆 -1.123 0.261
𝑅𝐷𝑇𝐸𝑇
𝐸𝑅𝐷 𝐸/𝑆 -0.831 0.436
𝐸𝑃𝐶𝐸/𝑆 0.003 0.998
𝑆𝐶𝐸/𝑆 -2.836 0.004
𝐸/𝑆
Table 4 shows that 𝑅𝐷𝑇𝐸𝑇 , 𝐸𝑅𝐷𝐸/𝑆 and 𝐸𝑃𝐶𝐸/𝑆 come respectively from the same distributions
while 𝑆𝐶𝐸/𝑆 come from different distribution.

The third criterion deals with a statistical comparison between the average experimental evacuation
times (𝑡𝑖𝐸̅ ) and the average simulated evacuation times (𝑡𝑖̅𝑆 ).

A preliminary analysis is made in order to verify if the two samples are from normally-distributed
populations. A Shapiro-Wilk Test is used for experimental data while a Kolmogorov-Smirnov test
for simulated data because of the different sample sizes.

The experimental and simulated samples are from a normally distributed population (Exp: P-value
= 0.979; Sim: P-value = 0.200). Thus, a t-test is used to verify if the two samples are from the same
normal distribution. The results show that P-value is equal to zero. Thus, the two samples are from
different normally distributed populations. That means that the areas subtended by experimental and
simulated OETCs belong to different populations and we can assume that these areas are
statistically different.

4. Discussion

The proposed methodology is based on functional analysis measures introduces by Peacock et al.
[25] and Galea et al. [24] for the study of fire and evacuation curves and the study of BU according
to approach suggested by Ronchi et al. [11].

To date, the majority of validation studies are based on the comparison of individual experimental
and simulated total evacuation times and/or qualitative comparison between experimental and
simulated occupant - evacuation time curves [3] [11]. Such an approach is not able to compare
experimental and simulated data of the entire evacuation process in a quantitative way. Conversely,
a validation approach based on functional analysis measures allows the comparison of the full
OETCs. This approach has been used for the first time for validation studies of evacuation models
by Galea et al. [24]. However, the innovative aspect of the present study is the integration of BU
measures in validation studies. In fact, to date, BU has generally been treated only in a qualitative
manner while in the present work, it has a key role in the assessment of the impact of the number of
runs on the comparison. The simulated number of runs derives from a comparison between the
experimental and simulated BU (Step 1 and 2). This number is generally left to a qualitative
judgement [11] or it corresponds to a fixed number of runs [34]. However, these two approaches do
not investigate the level of uncertainty associated with the number of runs in validation studies.
Conversely, the use of measures for BU as a quantitative criterion for the comparison allows the
comparison of experimental and simulated uncertainty as well as the compliance with pre-defined
acceptance criteria [11]. Moreover, this paper presents a tool to calculate these measures in real
time (and subsequently deciding the optimal number of runs) in order to optimize computational
cost. In the present paper, the new tool permitted a reduction of computational time during the
FDS+Evac simulation work.

The assessment of BU permits a detailed comparison between experiments and models. Two
antithetical approaches are currently used to compare experimental and simulated data [11]. In the
first approach, the experimental data are compared with best model estimation for the OETC (based
on TET and ERD) [24]. The second approach employs average OETC [11]. The proposed method
increases the usability of both approaches. In fact, the first criterion in Step 3 provides a quantitative
comparison between the experimental and simulated occupant - evacuation time curves according
the both criteria. This criterion shows that the best and worst model estimated OETC can be chosen
also according to EPC and SC measures. Moreover, results show that the worst model estimated
OETC are very different from the experimental results. In fact, worst TET, ERD, EOC and SC
curves fail the RDTET test based on the less restrictive threshold values by Galea et al. [24] (Table
3).

Then, the second criterion deals with a direct comparison of experimental and simulated BU. In this
criterion the BU is seen as the differentiation of experimental and simulated curves from the
respective average curves. In order to implement this comparison statistical parametric test are used
and multiple experimental data are required. The results show that experimental and simulated BU
associated to the shape of OETCs is different because 𝑆𝐶𝐸/𝑆 come from different distributions. This
could mean that the experimental and simulated OETCs are distributed around the respective
average curve differently. The cause may be associated with congestions which cause steps in the
curves as that shows in Figure 12 and 13.

The third criterion in Step 3 allows modellers to statistically compare the differences between
experimental and simulated data by using average evacuation times. Statistical testing was used to
compare data according to a macroscopic validation approach. Also in this case, multiple
experimental data are necessary to perform the comparison. The results show that 𝑡𝑖𝐸̅ and 𝑡𝑖̅𝑆 come
from different distributions. This may be associated with the difference in the shape of the OETCs
shown in the second criterion.

5. Limitations

An intrinsic limitation of the method proposed in this paper is the availability of multiple
experimental occupant evacuation time curves of the same evacuation scenario for the performance
of the model validation study. This paper highlights that a single experimental OETC may not be
representative of a full range of the behaviours of the occupants because of BU.
The main goal of this paper is to propose a validation methodology for evacuation models rather
than the analysis of the result of the evacuation model under consideration. In fact, the case study
has been selected only to show the validation procedure. The assessment of the validity of a model
should be based instead on the analysis of a full range of scenarios similar to the experimental data-
sets employed in this paper as a case study.

The current lack of multiple experimental data-sets makes it difficult to apply the method for
several validation cases. However, the need to fill this gap with data collection efforts is known to
the evacuation research community and dedicated research efforts have been carried out [11] [13]
[35] [36] [37].

Another issue is associated with the evaluation of the required degree of accuracy during the
comparison of data including BU. This may depend on the scope of the analysis and the type of
validation study under consideration.

A study on the effect of BU in relation to experimental data is not made in this work because of the
small number of repeated experimental data. However, this paper raises a fundamental question
when performing validation studies: “is the impact of simulated and experimental BU the same if
the number of repeated experiments and simulated runs is equal?” The answer to this question may
be useful to perform a direct estimation of the accuracy of evacuation model predictions.

6. Conclusion

A methodology to implement quantitative validation studies based on the concept of BU is


presented. The method includes the use of functional analysis operators as well as statistical testing
for the study of the convergence of the evacuation simulation and experimental data. A new tool is
developed in order to implement the convergence criteria for the evacuation model FDS+Evac.
Three general criteria are proposed for the comparison between experimental and simulated data.
The methodology is presented by applying the tool for a set of building evacuation experiments
conducted by Guo et al. [28].
This paper highlights the need for multiple experimental data-set for the assessment of the
differences between evacuation simulation and experimental data in order to evaluate the impact of
BU.

7. Appendix 1. Definition of the ERD, the EPC and the SC

Let x = {x1; x2; …; xk;…; xn} and y = {y1; y2; …; xk; …; yn} be n dimensional vectors which
represent two different occupant-evacuation time curves, then the ERD, the EPC and the SC of
these two vector are shown in Equations A.1, A.2 and A.3.

∑𝑛𝑖=1(𝑥𝑖 − 𝑦𝑖 )2 [Eq. A.1]


𝐸𝑅𝐷(𝒙, 𝒚) = √
∑𝑛𝑖=1(𝑦𝑖 )2

∑𝑛𝑖=1(𝑥𝑖 𝑦𝑖 ) [Eq. A.2]


𝐸𝑃𝐶(𝒙, 𝒚) =
∑𝑛𝑖=1 𝑦𝑖 2

(𝑥𝑖 −𝑥𝑖−𝑠 )(𝑦𝑖 − 𝑦𝑖−𝑠 )


∑𝑛𝑖=𝑠+1
𝑠2 (𝑝𝑖 −𝑝𝑖−1 ) [Eq. A.3]
𝑆𝐶(𝒙, 𝒚) =
(𝑥 −𝑥 )2 (𝑦𝑖 −𝑦𝑖−𝑠 )2
√∑𝑛𝑖=𝑠+1 𝑠2 (𝑝
𝑖 𝑖−𝑠
∑𝑛𝑖=𝑠+1
𝑖−𝑝 𝑖−1 ) 𝑠2 (𝑝𝑖 −𝑝𝑖−1 )

The pi parameter is the index of the xi component if pi is referred to x while s is the skip parameter
(see Section 2.1 for more detail).
References:
[1] International Standards Organization. ISO/FDIS 16730. Fire safety engineering - assessment,
verification and validation of calculation methods. 2008; ISO/TC92/SC4
[2] Purser DA. ASET And RSET: Addressing Some Issues In Relation To Occupant Behaviour
And Tenability. Fire Safety Science 2003; 7: 91-102. doi:10.3801/IAFSS.FSS.7-91.
[3] Kuligowski ED, Peacock RD, Hoskins BL. A review of building evacuation models, 2nd edn.
Fire Research Division, National Institute of Standards and Technology 2010; Technical Note 1680.
[4] Liu Y, Wang W, Huang HZ, Li Y, Yang Y. A new simulation model for assessing aircraft
emergency evacuation considering passenger physical characteristics. Reliab Eng Syst Safe 2014;
121: 187-197.
[5] Ronchi E, Kinsey MJ. Evacuation models of the future: insights from an online survey of user's
experiences and needs. In: Advanced Research Workshop: Evacuation and Human Behavior in
Emergency Situations, 21 Oct 2011, Santander, Spain.
[6] Borg A, Paulsen Husted B, Njå O. The concept of validation of numerical models for
consequence analysis. Reliab Eng Syst Safe 2013.
[7] Klüpfel H. Ship Evacuation - Guidelines, Simulation, Validation, and Acceptance Criteria. In:
Klingsch WWF, Rogsch C, Schadschneider A, Schreckenberg M, Pedestrian and Evacuation
Dynamics, Springer; 2008, p. 257-262.
[8] Schadschneider A, Seyfried A. Empirical Results for Pedestrian Dynamics and their
Implications for Cellular Automata Models, In Timmermans H, Pedestrian Behavior: models, data
collection and applications, Emerald Group Publishing Limited; 2009, p. 27-43.
[9] Schadschneider A, Eilhardt C, Nowak S, Will R. Towards a Calibration of the Floor Field
Cellular Automaton. In Peacock·RD, Kuligowski·ED, Averill JD, Pedestrian and Evacuation
Dynamics, Springer; 2010, p. 557-566.
[10] Kady RA, Tolk A. Emergency Evacuation Modeling: A Novel Approach to Layout Designs
and Evacuation Procedures. In Peacock·RD, Kuligowski·ED, Averill JD, Pedestrian and Evacuation
Dynamics, Springer; 2010, p 471-480.
[11] Ronchi E, Reneke PA, and Peacock RD. A Method for the Analysis of Behavioural
Uncertainty in Evacuation Modelling, Fire Technol 2013. doi: 10.1007/s10694-013-0352-7
[12] Hamins A, McGrattan K. Verification and validation of selected fire models for nuclear power
plant applications. Final Report NUREG-1824. U.S. Nuclear Regulatory Commission 2007.
[13] Averill JD. Five grand challenges in pedestrian and evacuation dynamics. In: Proceedings of
the 5th international pedestrian and evacuation dynamics conference; 2011, p. 1-11.
[14] Gwynne SMV, Galea ED, Owen M, Lawrence PJ, Filippidis L. A Review of the
Methodologies Used in Evacuation Modelling, Fire Mater 1999; 23: 383-388
[15] Capote JA, Alvear D, Abreu O, Lázaro M, Cuesta A. An Evacuation Model for High Speed
Trains, In Peacock·RD, Kuligowski·ED, Averill JD, Pedestrian and Evacuation Dynamics,
Springer; 2010, p 421 - 431.
[16] Averill JD, Reneke P, Peacock RD. Required Safe Egress Time: Data and Modeling. 7th
Conference on Performance-Based and Fire Safety Design Methods. Auckland, New Zealand,
2008.
[17] Kuligowski E. Predicting Human Behavior During Fire. Fire Technol 2013, 49:101 - 120.
[18] Ronchi E, Kuligowski ED, Peacock RD, Reneke PA. A probabilistic approach for the analysis
of evacuation movement data. Fire Safety Journal 2014, 63, 69-78.
[19] Lovreglio R, Borri D, dell’Olio L, Ibeas A. A discrete choice model based on random utilities
for exit choice in emergency evacuations. Safety Sci 2014; 62: 418 - 426.
[20] Mollick E. Establishing Moore's Law, Ann Hist Comput 2006; 28 – 3: 62-75
[21] Johnson, P., Beck, D., Horasan, P. Use of Egress Modeling In Performance Based Fire
Engineering Design - A Fire Safety Study At The National Gallery of Victoria. In T. Kashiwagi
(Ed.), Fire Safety Science -- Proceedings of the 4th International Symposium 1994, p. 669-680.
[22] Olsson PA, Regan MA. A comparison between actual and predicted evacuation times, Safety
Sci 2001; 38- 2: 139–145.
[23] Schneider V, Konnecke R. Simulating Evacuation Processes with ASERI. In Tagungsband
International Conference on Pedestrian Evacuation Dynamics (PED), Duisburg, 2001.
[24] Galea ER., Filippidis L., Deere S., Brown R., Nicholls I, Hifi Y and Besnard N. IMO INF
Paper Summary - The SAFEGUARD validation data-set and recommendations to IMO to update
MSC/Circ.1238. In: SAFEGUARD Passenger Evacuation Seminar. 30 November 2012. London,
UK. Papers. The Royal Institution of Naval Architects, London, UK, p. 97. ISBN 978-1-909024-
08-3.
[25] Peacock RD, Reneke PA, Davis WD, Jones WW. Quantifying fire model evaluation using
functional analysis. Fire Saf J 1999; 33:167- 184.. doi:10.1016/S0379-7112(99)00029-6
[26] Galea ER, Deere S, Brown R, Filippidis L. An Evacuation validation data set for large
passenger ships. In: Proceedings of the Pedestrian and Evacuation Dynamics 2012.
[27] Korhonen T, Hostikka S. Fire Dynamics Simulator with Evacuation: FDS+Evac. Technical
Reference and User’s Guide,VTT Technical Research Centre of Finland, VTT Working Papers 119,
2010.
[28] Guo RY, Huang HJ, Wong SC. Route choice in pedestrian evacuation under conditions of
good and zero visibility: Experimental and simulation results, Transport Res B-Meth 2012; 46: 669
- 686.
[29] Helbing D, Molnár P. Social force model for pedestrian dynamics, Phys Rev E 1999; 51: 4282
- 4286
[30] Purser DA. Toxicity Assessment of Combustion Products, in SFPE Handbook of Fire
Protection Engineering, 2nd ed, National Fire Protection Association, Quincy, MA., 1995, p. 2/28 -
2/146.
[31] Frantzich H, Nilsson D. Utrymning genom tät rök: beteende och förflyttning, 75 p., Report
3126, Department of Fire Safety Engineering, Lund University, Sweden, 2003.
[32] Thompson P, Lindstrom H, Ohlsson P, and Thompson S. Simulex: Analysis and Changes for
IMO Compliance, Proceedings of 2nd International Conference: Pedestrian and Evacuation
Dynamics 2003, p. 173 - 184.
[33] Rinne T, Tillander K, Grönberg P. Data collection and analysis of evacuation situations. VTT,
Technical Research Center of Finland. Research note 2562, 2010.
[34] IMO 2007. Guidelines for a Simplified Evacuation Analysis for New and Existing Passenger
Ships” covered by MSC Circ 1238
[35] Gwynne SMV. Improving the collection and use of human egress data. Fire Technol 2013; 41-
1: 83 – 99.
[36] Gwynne SMV, Kuligowski ED, Nilsson D. Representing evacuation behavior in engineering
terms. J Fire Prot Eng 2012; 22-2:133–150
[37] Kuligowski ED. The Process of Human Behaviour in Fire, Fire Research Division, National
Institute of Standards and Technology, Technical Note 1632, 2009.

View publication stats

You might also like