RELY

Structural Reliability powered by Strurel®

SOFiSTiK | 2020
 RELY
Structural Reliability powered by Strurel®

RELY Manual, Service Pack 2020-6 Build 289

Copyright © 2020 by SOFiSTiK AG, Oberschleissheim, Germany.

SOFiSTiK AG

HQ Oberschleissheim Office Nuremberg

Bruckmannring 38 Flataustraße 14
85764 Oberschleissheim 90411 Nuremberg
Germany Germany

T +49 (0)89 315878-0 T +49 (0)911 39901-0

F +49 (0)89 315878-23 F +49(0)911 397904
info@[Link]
[Link]

This manual is protected by copyright laws. No part of it may be translated, copied or reproduced, in any form or by
any means, without written permission from SOFiSTiK AG. SOFiSTiK reserves the right to modify or to release
new editions of this manual.

The manual and the program have been thoroughly checked for errors. However, SOFiSTiK does not claim that
either one is completely error free. Errors and omissions are corrected as soon as they are detected.

The user of the program is solely responsible for the applications. We strongly encourage the user to test the
correctness of all calculations at least by random sampling.

Front Cover
Project: Queensferry Crossing | Photo: Bastian Kratzke
 Contents | RELY

Contents

Contents i

1 Task Description 1-1

2 Theoretical Principles 2-1

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-1
2.2 The reliability problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-2
2.3 Monte Carlo simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-3
2.4 FORM - First-Order Reliability Method . . . . . . . . . . . . . . . . . . . . . . 2-4
2.5 SORM - Second-Order Reliability Method . . . . . . . . . . . . . . . . . . . . 2-7
2.6 Importance sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-7
2.7 Line sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-7
2.8 Directional sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-9
2.9 Adaptive sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-9
2.10 Subset simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-10

Literature 2-13

3 Input Description 3-1

3.1 Input Language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1
3.2 Input Records . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1
3.3 CTRL – Control switches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
3.4 OUTP – Output control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-4
3.5 VAR – Random Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-5
3.6 CORR – Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-8
3.7 PROJ – Define the SOFiSTiK project that evaluates the limit-state func-
tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-9
3.8 FUNC – Define a limit-state function . . . . . . . . . . . . . . . . . . . . . . . 3-10
3.9 SLSF – Define the system limit-state function . . . . . . . . . . . . . . . . . 3-11
3.10 MC – Monte Carlo simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-12
3.11 IS – Importance sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-14
3.12 LS – Line sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-15
3.13 DS – Directional sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-16
3.14 ADS – Adaptive sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-17
3.15 FORM – First order reliability method . . . . . . . . . . . . . . . . . . . . . . . 3-18
3.15.1 SPNT: Starting solution . . . . . . . . . . . . . . . . . . . . . . . . . 3-19
3.15.2 Convergence criteria . . . . . . . . . . . . . . . . . . . . . . . . . . 3-19
3.15.3 OPTA: Optimization algorithms . . . . . . . . . . . . . . . . . . . . 3-19
3.15.4 Finite difference scheme . . . . . . . . . . . . . . . . . . . . . . . . 3-21
3.16 SORM – Second order reliability method . . . . . . . . . . . . . . . . . . . . 3-22
3.17 MVFO – Mean value first order . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-23

SOFiSTiK 2020 i
RELY | Contents

3.18 SUBS – Subset simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-24

3.19 EXDS – Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-25
3.20 RSF – Response Surface Function . . . . . . . . . . . . . . . . . . . . . . . . 3-27

4 Output Description 4-1

5 Examples 5-1
5.1 Simple Beam: Explicit Limit-State Function . . . . . . . . . . . . . . . . . . . 5-2
5.2 Simple Beam: Finite Element Limit-State Function . . . . . . . . . . . . . . 5-4

6 Appendix 6-1

Literature 6-19

ii SOFiSTiK 2020
 Task Description | RELY

1 Task Description

Traditionally, compliance of an engineering system is demonstrated by means of safety fac-

tors. Structural reliability offers an alternative approach to assess the safety of an engineering
system. In structural reliability, uncertainties in loading, material properties and geometry are
taken into account explicitly; i.e., the uncertainties are described in a probabilistic manner. Our
aim is to compute the probability that the engineering system will fail. The concept of structural
reliability is also included in Eurocode 0 and DIN 1055-100.

Rely is an add-on to the SOFiSTiK program that performs reliability analysis, where the engi-
neering system of interest is modeled using one of the SOFiSTiK finite element modules. The
kernel of Rely is powered by the stand-alone software package Strurel. Different structural
reliability methods are provided:

• Monte Carlo simulation (Section 2.3)

• FORM (Section 2.4)
• SORM (Section 2.5)
• Importance sampling (Section 2.6)
• Line sampling (Section 2.7)
• Directional sampling (Section 2.8)
• Adaptive sampling (Section 2.9)
• Subset simulation (Section 2.10)

SOFiSTiK 2020 1-1

RELY | Task Description

1-2 SOFiSTiK 2020

 Theoretical Principles | RELY

2 Theoretical Principles

2.1 Introduction
In engineering, the system of interest is approximated by a model. The behavior of the model
depends on its input parameters; for a finite element model of a structure, this can be material
parameters like the Young’s modulus, or the applied loading. The state of the input parameters
is usually not known with certainty: e.g., the snow that lies on a roof, the wind-load that acts
on a building, or the future Young’s modulus of a concrete slab to be designed are uncertain.
In reliability analysis, all parameters whose uncertainty influences the behavior of the system
are modeled as random variables.

For some outcomes of the random variables, the behavior of the system will be undesirable:
e.g. parts of the structure will collapse, the state of the system can lead to fatalities, or the
serviceability of the system is no longer maintained. Such undesired events are called failure
F. Reliability analysis aims at evaluating the probability that a failure event occurs, pf = Pr(F).
The smaller the probability of failure, the larger is the reliability of the system. A measure
for the degree of safety of the system is given by the reliability index β. The reliability index
is a decreasing function of the probability of failure and is defined through the inverse of the
standard normal CDF at the probability of failure: β = −−1 (pf ). This relation is illustrated in
Fig. 2.6.

In structural reliability, the failure event F is defined in terms of a so-called limit-state function
g(X), where X is the vector of random variables of the problem. Failure occurs if and only
if g(X) ≤ 0. This concept of the limit-state function defining failure is illustrated in Fig. 2.1.
For some problems the limit-state function can be defined in terms of the demand S and the
capacity R as g(X) = R − S (Fig. 2.2). If demand and capacity cannot be separated easily,
the limit-state function is often defined as the difference between some threshold value and
the corresponding model output; e.g. the displacement at the tip of a cantilever beam versus
the maximum allowed displacement for that system. For a detailed introduction to structural
reliability, the reader is referred to Ditlevsen and Madsen [2], Melchers [3], Papaioannou [4].

x2
0.06 fX(x)
15 safe domain
Joint distribution
g(x)>0
fX(x)
0.04

10 g(x)=0
failure domain
0.02
g(x)<0

5 0
15
g(x)=0
10
safe domain
fX2(x2) x2 failure domain 15
5 10
0 x1
0 5 10 15 5 x1
0 0
Marginal fX1(x1)
distributions

Figure 2.1: Illustration of the general reliability problem, for the case of two random variables.
(taken from Straub [1])

SOFiSTiK 2020 2-1

RELY | Theoretical Principles

fS(s), fR(r) Pr(F) = Pr(R < S)

Capacity R

Demand S

r, s

Figure 2.2: The basic structural reliability problem.

(taken from Straub [1])

2.2 The reliability problem

The probability of failure pf can be expressed as
Z
pf = ƒX (x) dx (2.1)
g(X)≤0

where ƒX (x) denotes the joint probability density function of the random vector X, and x is an
outcome of X. In Eq. (2.1) the probability of failure is expressed as the integral over the failure
domain. However, the failure domain is usually only known implicitly through the definition of
the limit-state function g(X). Eq. (2.1) can equivalently be expressed as:
Z
pf = g (x) ƒX (x) dx (2.2)
X

where g (x) is the indicator function of g(X) defined as


1, if g(x) ≤ 0
g (x) = (2.3)
0, if g(x) > 0


For some reliability methods it is convenient to work in the so-called standard normal space.
Let us assume that the vector of random variables X can be expressed in terms of a vector of
independent standard normal random variables U and a transformation T : U → X. Outcomes
x of X can consequently be expressed as functions of underlying independent standard normal
random variables U with outcome . The reliability problem in Eq. (2.1) and Eq. (2.2) can then
be written as:
Z
pf = g (T()) φU () d (2.4)
U

where φU () is the joint probability density function of the vector of independent standard
normal random variables U. The transformation is illustrated in Fig. 2.3.

For the transformation T, different transformation methods are available in the literature. Three
well-known methods are:

• Rosenblatt transformation: requires the uncertainty vector to be represented as a sequence

2-2 SOFiSTiK 2020

 Theoretical Principles | RELY

Original space Standard Normal space

100 4
g‘(uR ,uS ) = 0
3
80
2

1
60

s g(r,s) = 0 uS 0

40
-1

-2
20
-3

0 -4
0 20 40 60 80 100 120 -4 -3 -2 -1 0 1 2 3 4
r uR

Figure 2.3: Example of a transformation to standard normal space with two random variables R
and S: The probability mass in the failure domain (colored areas) is identical in the original and in
the standard normal space.
(taken from Straub [1])

of conditional distribution functions (Hohenbichler and Rackwitz [5]).

• Marginal transformation based on the Nataf distribution: Represents the joint distribution in
terms of the marginal distributions and the correlation matrix Liu and Der Kiureghian [6].
• Hermite transformation

The integral in Eq. (2.2) or Eq. (2.4) can usually not be solved analytically, and, therefore, has
to be handled numerically. The various reliability methods differ in their treatment of the integral
in Eq. (2.4). Most methods aim at minimizing the number of limit-state function calls. This is
because in structural reliability the limit-state function is commonly expressed as a function that
depends on the outcome of a finite element analysis. Consequently, every time the limit-state
function is evaluated, a finite element analysis must be performed - which renders the reliability
analysis of large finite element systems computationally expensiive.

2.3 Monte Carlo simulation

The problem stated in Eq. (2.4) can also be expressed by means of the following expectation

pf = EU g (T(U))
 
(2.5)

where EU [·] denotes the expectation with respect to the vector of uncorrelated standard normal
random variables U. Drawing N random samples 1 , 2 , . . . , N from U, this expectation can
be approximated as:
N
1X
p̂f = g (T( )) (2.6)
N =1

where p̂f is an estimate of pf .

Please note that the value of g (T( )) follows a Bernoulli distribution with mean value pf .
Consequently, p̂f · N has a Binomial distribution with mean pf · N. In plain terms, p̂f equals the

SOFiSTiK 2020 2-3

RELY | Theoretical Principles

100
Samples in the failure domain Samples

80

60

s g(r,s) = 0
40

20

0
0 20 40 60 80 100 120
r

Figure 2.4: 1000 samples obtained with Monte Carlo simulation. The problem has two random
variables: capacity R and demand S. The failure domain is illustrated in standard normal space.
(taken from Straub [1])

number of failed samples Nƒ , i.e. the number of samples for which g (T( )) ≤ 0, divided by
N
the total number of samples, i.e. p̂f = Nƒ . Therefore, the estimate p̂f has a mean value of pf
(i.e. the estimate is unbiased), and the standard deviation σp̂f of p̂f is
v
t pf (1 − pf ) t pf
u v
σp̂f = ≈ (2.7)
N N

where the approximation is valid if pf is small such that pf  pf 2 . The coefficient of variation
δp̂f of p̂f is
v
f
u1 − p 1
δp̂f = (2.8)
t
≈p
N · pf N · pf

The variance of the estimate p̂f depends on the probability of failure pf and on the number of
samples N used in Eq. (2.6). The dependency between pf , N and δp̂f is illustrated in Fig. 2.5.
It is shown that the required number of samples for a target coefficient of variation increases
fast with a decrease of pf . Usually, the true value of pf is not known, since it is the target of the
analysis. Therefore, equation Eq. (2.8) is approximated using the estimate p̂f as:
1
δp̂f ≈ Æ (2.9)
N · p̂f

2.4 FORM - First-Order Reliability Method

FORM linearizes the limit-state function g(T()) with respect to  at the so-called design point
∗ , and computes the probability of failure based on the linearized limit-state function G0 ().
The design point ∗ is defined as the point on the failure surface g(T()) = 0 that is closest
to the origin. It is obtained by solving the following constraint optimization problem:

T  subjected to g(T()) = 0
p
∗ = rg min (2.10)

2-4 SOFiSTiK 2020

 Theoretical Principles | RELY

1e+14

1e+12

1e+10
Number of samples N

1e+08 1%
20% 50%
5% 10%
1e+06

10000

100

1
1 0.1 0.01 0.001 0.0001 1e-05 1e-06 1e-07 1e-08 1e-09
probability of failure pf

Figure 2.5: Number of samples N required for different coefficients of variation δp̂f .

0.1

0.01
probability of failure pf

0.001

0.0001

1e-05

1e-06

1e-07

1e-08

1e-09
-2 -1 0 1 2 3 4 5 6
reliability index β

Figure 2.6: Relation between the reliability index β and the probability of failure pf .

SOFiSTiK 2020 2-5

RELY | Theoretical Principles

u* φ(u*)

Original variable space Standard Normal space Pr(F)

βFORM

16 5 G´(u)=0
fX(x)
φU(u)
12

0 g(T(u))=0

U2
X2

u*
4
−5
g(x)=0
0
40 50 60 70 80 90 0 5 10
X1 U1

Figure 2.7: Design point and linear approximation of the limit state surface. Left side: original
random variable space; right side: standard normal space.
(taken from Klüppelberg, Straub, and Welpe [7])

The design point ∗ is the most likely point of failure in the investigated problem and, thus,
also referred to as Most Likely ∗
p Failure Point. The distance of  to the origin, i.e. the length of
∗ , is denoted as βFORM = ∗ T ∗ .

If the design point of the problem is known, a first order probability estimate of pf can be
computed as:

p̂f = (−βFORM ) (2.11)

where (·) denotes the CDF of the standard normal distribution. The concept of FORM is
illustrated in Fig. 2.7.

The main problem in FORM is to solve the optimization problem that gives the design point ∗ .
Since an optimization problem has to be solved, the limit-state function is usually required to
be differentiable. In Rely , different optimization algorithms are available.

A well-known problem in optimization is the presence of multiple optima. Note that a single run
of an optimization algorithm can only find a single (local) minimum. To obtain a good approx-
imation of the probability of failure, all dominating minima must be detected and considered.
Repeated runs with random starting points will at least indicate the presence of multiple min-
ima. If there are multiple local design points, an upper bound of the probability estimate can be
obtained by adding up the failure probabilities of all detected (the global and all local) design
points.

2-6 SOFiSTiK 2020

 Theoretical Principles | RELY

2.5 SORM - Second-Order Reliability Method

FORM uses a first-order approximation of the limit-state function around the design point.
If instead of a first-order approximation a second-order approximation is used, the reliability
method is referred to as SORM. Additional to the FORM approximation, SORM takes the cur-
vature of the limit-state surface around the design point into account. This means that the
Hessian matrix of the limit-state function at the design point has to be computed. Usually, this
is only feasible for problems with a small number of random variables.

For a description of SORM, the reader is referred to Gollwitzer [8].

2.6 Importance sampling

The problem with Monte Carlo simulation for problems with small probabilities of failure is that
many samples are required for a stable estimate of the probability of failure. The reason is that
on average only pf · N out of N samples fall in the failure domain. Importance sampling aims at
artificially increasing the number of samples in important regions of the failure domain. For this
purpose, the general reliability problem is rewritten as (here, for the standard normal space):

φU ()
Z Z
pf = g (T()) φU () d = g (T()) ψ() d (2.12)
U X ψ()

where ψ() is called importance sampling density and is a joint PDF over . In importance
sampling the second integral in Eq. (2.12) is approximated: Instead of sampling from φU (),
samples are drawn from ψ(). Let z with  = 1, . . . , N denote samples from ψ. The probability
of failure pf is estimated as:
N
1X φU (z )
p̂f = g (T(z )) (2.13)
N =1 ψ(z )

In the theoretical optimal case, i.e. for ψopt () = p1f g (T()) φU (), only a single sample is
needed to determine the true probability of failure. However, ψopt itself requires knowledge
about the true probability of failure and the explicit shape of the failure domain, and, thus, the
use of ψopt is not practical. In general, importance sampling methods aim at selecting the
importance sampling density ψ such that it approximates ψopt .

In practice, the importance sampling density is often centered around the design point (or
an approximated design point). Usually, the importance sampling density is selected as
ψ(1 , . . . , N ) = N
Q

=1
φ   −  , cen , where φ denotes the standard normal PDF and ,cen is
the th coordinate of the (approximated) design point.

2.7 Line sampling

This method (Hohenbichler and Rackwitz [9]) generates samples on the tangent hyperplane to
the failure surface centered at the design point (or an approximate design point). The sampling
space is thus rotated and reduced by one dimension.

The method is illustrated in Fig. 2.9. Here, the coordinate system ν is selected in such a way
that the hyperplane through the design point is perpendicular to the first Cartesian coordinate
ν1 . The origin of the coordinate system ν is the same as the origin of . Samples of ν are

SOFiSTiK 2020 2-7

RELY | Theoretical Principles

Importance sampling (nS = 100) Monte Carlo simulation (nS = 1000)

100 100

80 80

60 60

s s g(r,s) = 0
40 g(r,s) = 0 40

ψ(r,s) f (r,s)
20 20

0 0
0 20 40 60 80 100 120 0 20 40 60 80 100 120
r r

Figure 2.8: Samples of capacity R and demand S. Left side: samples according to the importance
sampling density ψ centered around the design point in original space. Right side: samples from
the original PDF ƒX (x); i.e., samples from Monte Carlo simulation. It can be seen that importance
sampling produces more samples in the failure domain than Monte Carlo sampling - even if the
total number of samples used in importance sampling is smaller than the one used in Monte Carlo
sampling.
(taken from Straub [1])

ν1

GHIS(ν)
design point

FORM hyperplane

d(νa) d(νb)
βFORM

ν2-n
νa νb
samples from ψ(ν)
Figure 2.9: Illustration of line sampling
(taken from Straub [1])

2-8 SOFiSTiK 2020

 Theoretical Principles | RELY

generated on the hyperplane that is parallel to the FORM hyperplane and goes through the
origin. Let ψ(ν) denote the importance sampling density that is used to generate this samples.
Typically, ψ(ν) is selected as the standard normal distribution for ν2 , . . . , νn , where n is the
total number of random variables. Since the samples should lie on the hyperplane, ν1 = 0 for
all samples. Next, for each sample ν  from ψ(ν), a line search is performed to find the distance
of the limit-state surface perpendicular to the hyperplane. This distance is denoted d(ν  ). In
this case, the probability of failure pf is estimated from the samples as:
N
1X
p̂f = (−d(ν  )) (2.14)
N =1

where N is the total number of samples used.

2.8 Directional sampling

This method expresses a random sample  as a point on a hypersphere with radius r . Any
sample  can be expressed uniquely by a radius r and n − 1 direction angles j , where n is the
total number of random variables and j = 1, . . . , n− 1. Let r and the j be realizations of random
variables denoted R and Aj , respectively. To generate samples that follow the multivariate
standard normal distribution, the Aj must be uniformly distributed random variables and R2 has
to follow the chi-squared distribution with n degrees of freedom.

The idea of directional sampling is to first generate a random direction; i.e. generate realiza-
tions j with j = 1, . . . , n − 1. Next, a gradient-free line search in the generated direction is
applied to find the point for which the limit-state function becomes zero. The distance of that
point to the origin is denoted r . Based on a total number of N samples, the probability of failure
pf can be approximated as:

N ”
1X € Š—
2 2
p̂f = 1 − χn r (2.15)
N =1

where χn2 is the CDF of the chi-squard distribution with n degrees of freedom.

This method can cope with problems that have multiple (local) design points, or if failure is
defined as the union of different failure modes. However, directional sampling is not suitable
for problems with many random variables.

2.9 Adaptive sampling

Adaptive sampling is in-principle importance sampling - with one addition: If the limit-state
function of a sample becomes negative and if that sample is closer to the origin (in standard
normal space) than the center of the importance sampling density, the importance sampling
density will be centered around that point from then on.

Adaptive sampling does not require the limit-state function to be differentiable. If more than one
important regions are detected, a warning is issued. The efficiency of the method decreases
with the number of random variables of the problem.

SOFiSTiK 2020 2-9

RELY | Theoretical Principles

2.10 Subset simulation

The subset simulation method proposed by Au and Beck [10] is an adaptive Monte Carlo
method that is efficient for high dimensional problems. This method expresses the probability
of failure pf as a product of larger conditional probabilities, i.e.:
M
Y
pf = Pr(F |F−1 ) (2.16)
=1

where F are intermediate failure events that are nested, i.e. F0 ⊃ F1 ⊃ . . . ⊃ FM = F. The
failure events F are defined as F = {g(X) ≤ c }, where c are real coefficients with c0 = ∞ >
c1 > . . . > cM = 0. Commonly, the values of c are chosen adaptively such that the conditional
probabilities Pr(F |F−1 ) are equal to a predefined threshold probability p0 . The probability p0
is usually set to 10%. The algorithm stops as soon as c becomes smaller or equal than zero.
In this case c is set to zero; i.e., c = cM = 0.

The number of samples at each level is kept constant in Rely . Standard Monte Carlo simulation
is applied to compute Pr(F1 ) = Pr(F1 |F0 ). All other conditional probabilities are approximated
by means of Markov Chain Monte Carlo (MCMC) techniques. In Rely the adaptive MCMC with
optimal scaling proposed by Papaioannou et al. [11] is implemented: The proposal distribution
in the MCMC step is scaled such that the acceptance rate of a proposed sample is close to an
optimal acceptance rate.

The subset simulation method is illustrated in Fig. 2.10. The limit-state surface of the example-
problem is illustrated in sub-figure (l). As shown in sub-figure (a), the method starts with
Monte Carlo sampling. The samples that are contained in the region outlined in sub-figure
(b) represent the p0 -fraction of the samples in sub-figure (a) with the smallest value of the
limit-state function. The first intermediate failure domain (black line in sub-figure (b)) is chosen
such that only the p0 -fraction of the samples (depicted in sub-figure (b)) fall into it. This samples
are used as seed values for the MCMC sampling, as shown in sub-figure (c). The described
procedure is continued until at least a p0 -fraction of samples is in the true failure domain.

2-10 SOFiSTiK 2020

 Theoretical Principles | RELY

a) Samples of U b) 1st intermediate limit state surface and complying samples

5 5

4 4

3 3

2 2

1 1

u2 0 u2 0 G(U) - c1 = 0

−1 −1

−2 −2

−3 −3

−4 Pr(E1) = 0.1 −4

−5 −5
−5 −4 −3 −2 −1 0 1 2 3 4 5 −5 −4 −3 −2 −1 0 1 2 3 4 5

u1 u1

c) Samples of U conditional on {G(U) - c1 < 0} d) 2nd intermediate limit state surface and complying samples
5 5

4 4

3 3

2 2

1 1

u2 0 G(U) - c1 = 0 0 G(U) - c2 = 0

−1 −1

−2 −2

−3 −3

−4 Pr(E2|E1) = 0.1 −4

−5 −5
−5 −4 −3 −2 −1 0 1 2 3 4 5 −5 −4 −3 −2 −1 0 1 2 3 4 5

u1 u1

e) - j) repeat for intermediate limit states 2 - 5

k) Samples of U conditional on {G(U) - c5 < 0} l) Original limit state surface and complying samples
5 5

4 4

3 3

2 2

1 1

u2 0 G(U) - c5 = 0 0 G(U) = 0

−1 −1

−2 −2

−3 −3

−4 Pr(E6|E5) = 0.4 −4

−5 −5
−5 −4 −3 −2 −1 0 1 2 3 4 5 −5 −4 −3 −2 −1 0 1 2 3 4 5

u1 u1

Figure 2.10: Illustration of subset simulation.

(taken from Straub [1])

SOFiSTiK 2020 2-11

RELY | Theoretical Principles

2-12 SOFiSTiK 2020

 Literature | RELY

Literature

[1] Daniel Straub. Lecture Notes in Structural Reliability. Tech. rep. TU München, 2012.
[2] Ove Ditlevsen and Henrik O Madsen. Structural reliability methods. Vol. 178. Wiley New
York, 1996.
[3] Robert E Melchers. Structural reliability analysis and prediction. John Wiley New York,
1999.
[4] Iason Papaioannou. Non-intrusive Finite Element Reliability Analysis Methods. TU
München, 2012.
[5] Michael Hohenbichler and Rudiger Rackwitz. “Non-normal dependent vectors in struc-
tural safety”. In: Journal of the Engineering Mechanics Division 107.6 (1981), pp. 1227–
1238.
[6] Pei-Ling Liu and Armen Der Kiureghian. “Multivariate distribution models with pre-
scribed marginals and covariances”. In: Probabilistic Engineering Mechanics 1.2 (1986),
pp. 105–112.
[7] Claudia Klüppelberg, Daniel Straub, and Isabell M. Welpe. Risk - A Multidisciplinary In-
troduction. Springer, 2014.
[8] Stephan Gollwitzer. Strurel – Users Manual. Tech. rep. RCP Consult GmbH, 2014.
[9] Michael Hohenbichler and Ruediger Rackwitz. “Improvement of second-order reliability
estimates by importance sampling”. In: Journal of Engineering Mechanics 114.12 (1988),
pp. 2195–2199.
[10] Siu-Kui Au and James L Beck. “Estimation of small failure probabilities in high dimensions
by subset simulation”. In: Probabilistic Engineering Mechanics 16.4 (2001), pp. 263–277.
[11] Iason Papaioannou et al. “MCMC algorithms for subset simulation”. In: Manuscript, En-
gineering Risk Analysis Group, Technische Universität München (2014).
[12] A. Hasofer and N. Lind. “An Exact and Invariant First-order Reliability Format”. In: Journal
of the Engineering Mechanics Division, ASCE 100.EM1 (1974), pp. 111–121.
[13] Rüdiger Rackwitz and Bernd Fiessler. “Structural reliability under combined random load
sequences”. In: Computers & Structures 9.5 (1978), pp. 489–494.
[14] Armen Der Kiureghian and Mario De Stefano. “Efficient algorithm for second-order relia-
bility analysis”. In: Journal of engineering mechanics 117.12 (1991), pp. 2904–2923.
[15] T Abdo and R Rackwitz. “Reliability of uncertain structural systems”. In: Proc. Finite
Elements in Engineering Applications (1990), pp. 161–176.

SOFiSTiK 2020 2-13

RELY | Literature

2-14 SOFiSTiK 2020

 Input Description | RELY

3 Input Description

3.1 Input Language

Input is defined using the free format of the CADINP input language (see the general manual
’SOFiSTiK Basics’).

3.2 Input Records

The following records are defined:

Record Items
CTRL TRAN SEED SVAL CDBP CLSE ENTR
OUTP NPRI IMON LSFR PRUV PRXV
VAR NAME TYPE PTYP P1 V1 P2 V2
P3 V3 P4 V4 VAL UMIN UMAX
CHAV TXT TID
CORR VAR1 VAR2 C
PROJ DAT CDB
FUNC NAME EXPR
SLSF VARN EXPR
MC NSPL LSF SAVE SAPD FSAV FSAP SPLS
SENS
IS NSPL DSPT LSF
LS NSPL SPNT LSF EPSC GSTP
DS NSPL LSF EPSC GSTP
ADS NSPL SPNT LSF
FORM SPNT OPTA GRDM GSTP MIT1 MIT2 EPSC
SMU LSF
SORM SPNT OPTA GRDM GSTP MIT1 MIT2 EPSC
SMU LSF
MVFO SPNT GRDM GSTP RELA LSF
SUBS NSPL P0 LMDA NCS LSF
EXDS SDES CNTR SAIF SAVE SAPD LSF SPRD
RSF NAME ANSA LSF LOAD SAVE

The records HEAD, END and PAGE are described in the genearl manual ’SOFiSTiK Basics’.

SOFiSTiK 2020 3-1

RELY | Input Description

3.3 CTRL – Control switches

CTRL

Item Description Unit Default

TRAN Type of transformation LT NATF

ROSE Rosenblatt transformation
NATF Nataf transformation
HRMT Hermite transformation

SEED ID of a seed value to use NT 8

SVAL User-defined starting solution in U- or in X-space LT U

U U-space (standard normal)
X X-space (physical)

CDBP Clear points in the CDB at start? LT DEL

DEL Delete points stored in the CDB
KEEP Consider points stored in the CDB

CLSE Treat errors in Limit-State function call as fatal? LT NO

NO Utter only a warning.
YES Output an error.

ENTR Repeated trials after failed sample NT 2

TRAN specifies the method that is used to transform probability distributions from an underlying
standard normal space.

With the flag SEED, seed values can be specified. In Strurel, seven predefined seed values
are available, corresponding to SEED 1,. . . ,7. If SEED is set to 8, the seed is computed based
on the current time.

SVAL defines whether starting solutions are defined in U- or in X-space. Thus, the state of
SVAL influences the option VAL in VAR.

Some reliability methods require the design-point or an approximated design-point as input.

Consequently, for these methods the (approximated) design-point of the problem must be com-
puted beforehand. If such a point is obtained, it is stored in the CDB. Points from previous runs
of Strurel can be used, if CDBP is set to KEEP. If CDBP is set to DEL, points computed in
previous runs of Strurel that are stored in the CDB will never be used. Note that if the model,
the limit-state or the definition of the random variables is modified, CDBP should always be set
to DEL.

For some problems it might be required that errors in the limit-state function are not treated as

3-2 SOFiSTiK 2020

 Input Description | RELY

fatal; i.e. that the simulation is allowed to continue even if some of the limit-state function calls
failed. The flag CLSE controls how errors in the limit-state function call are treated.

A failed limit-state evaluation can have multiple reasons. One the one hand, there might be
a syntax error in the input file of the finite element model, or the finite element solver might
not achieve convergence for the specified parameter values. One the other hand, there could
be a temporary problem with the SOFiSTiK license (e.g., because the user starts accidentally
another finite element calculation without having enough licenses). The error mentioned in the
second case could possibly be resolved by restarting the last call to the finite element solver.
By means of ENTR, the number of repeated trials after an error in the limit-state function call
can be specified.

SOFiSTiK 2020 3-3

RELY | Input Description

3.4 OUTP – Output control

See also:
OUTP

Item Description Unit Default

NPRI Verbosity of prints to output files (results) LT MEDI

NONE no output
MIN minimum output
MEDI average output
MAX verbose output

IMON verbosity of output to monitoring file LT LRGE

NONE no output
SMAL minimum output
LRGE normal output
DBG for debugging
VDBG verbose debugging

LSFR output realizations of limit-state function LT NO

NO
YES

PRUV output coordinates of sample vector (in standard Nor- LT NO

mal space) for each realization
NO
YES

PRXV output coordinates of sample vector for each realiza- LT NO

tion
NO
YES

3-4 SOFiSTiK 2020

 Input Description | RELY

3.5 VAR – Random Variables

See also: CTRL, PROJ, FUNC, CORR

VAR

Item Description Unit Default

NAME Name of the random variable LT8 !

TYPE Type of the distribution of the random variable LT !

TID ID of the type (form Strurel) NT *

PTYP The distribution is defined in terms of its: LT PARA

MOME moments
PARA parameters

P1 mean or 1st parameter − -

P2 standard deviation or 2nd parameter − -
P3 3rd parameter − -
P4 4th parameter − -

Link value of another random variable as:

V1 mean or 1st parameter LT8 -
V2 standard deviation or 2nd parameter LT8 -
V3 3rd parameter LT8 -
V4 4th parameter LT8 -

VAL Value for starting solution − 0.0

UMIN smallest value allowed in U-space − *

UMAX largest value allowed in U-space − *

CHAV Characteristic value − -

TXT A string describing the random variable. LT60 -

This record defines the random variables of the problem. At least a single random variable has
to be defined in a project.

NAME defines the name that is associated with the random variable. The name of a random
variable can be up to eight characters long. Empty names are not allowed. No two random
variables with the same name must exist. No distinction is made between upper and lower
case letters.

The type of the distribution of the random variable is set by means of either the key-word TYPE
or the distribution id TID. Either TYPE or TID must be specified. An overview of the distribution

SOFiSTiK 2020 3-5

RELY | Input Description

types TYPE available in Strurel is given in Table 3.5. The specification of the distribution type by
means of the key-word TYPE is only available for the most common types. All other distribution
types must be defined in terms of their id through the key-word TID. The ids of all distributions
available is listed in the table in Appendix A.

PTYP specifies how the distribution of the random variable is defined. Available options are
MOME and PARA:

• PARA: the distribution is defined in terms of its parameters. The number of parameters
required to define a distribution depends on the type of the distribution (specified in TYPE).
• MOME: the distribution is defined in terms of its mean value and standard deviation. If the
distribution has more than two parameters, then the 3rd parameter and, optionally, the 4th
parameter have to be supplied as well.

The mean value of the distribution or its first parameter can either be defined as as a fixed value
in P1 or as a link to another random variable in V1. If P1 is set, V1 must not be defined - and
vice versa. P1 expects a floating-point number, and V1 expects the name of random variable
that is already defined. If P1 is used, the mean value or first parameter of the distribution is
constant and equals the value specified in P1. If V1 is used, the mean value or first parameter
of the distribution is not constant and equals the realization of the random variable specified
in V1. If neither P1 nor V1 is specified, P1 is set to zero by default. P2/V2, P3/V3 and P4/V4
behave accordingly.

If a starting solution for the random variable is requested, the value specified in VAL will be
used. Depending on whether CTRL SVAL U or CTRL SVAL X is used, VAL is defined in U- or
in X-space, respectively.

UMIN and UMAX define the range of admissible values in U-space. If their value is not explicitly
specified, the value is set based on the floating point precision of your system.

If CHAV is not set explicitly, the specified mean of the random variable will be used as its
characteristic value.

The following table lists the available distribution types for random variables:

Table 3.5: Distribution types for random variables available in Strurel

TYPE Name of distribution Parameters Available options

for PTYP
UNIF Uniform , b MOME, PARA
NORM Normal μ, σ MOME, PARA
LOGN Lognormal (Shifted) ξ, δ, τ MOME, PARA
EXPO Exponential τ, λ MOME, PARA
GAMMA Gamma (Shifted) k , λ, τ MOME, PARA
BETA Beta r , t , , b MOME, PARA
GUMBMAX Gumbel (max) , α MOME, PARA
FRECMAX Frechet (max) , k, τ PARA
WEIBMIN Weibull (min) , k , τ MOME, PARA

3-6 SOFiSTiK 2020

 Input Description | RELY

Table 3.5: (continued)

TYPE Name of distribution Parameters Available options

for PTYP
RAY Rayleigh α, τ MOME, PARA
STUD Student ν PARA
TRAP Trapezoid , b, c, d PARA
BIRN Birnbaum/Saunders α, β MOME, PARA
CAUCHY Cauchy θ, λ PARA
INVGAU Inverse Gauß μ, λ MOME, PARA
GUMBMIN Gumbel (min) , α MOME, PARA
FRECMIN Frechet (min) , k, ω PARA
WEIBMAX Weibull (max) , k , ω MOME, PARA
PARETO Pareto , k MOME, PARA
LAPL Laplace θ, ϕ MOME, PARA
LOGISTIC Logistic μ, σ MOME, PARA
HALFN Halfnormal m, σ MOME, PARA
CHIS Chi-square ν PARA
FISH Fisher’s F ν1 , ν2 PARA
STUDG non-standard Student m, σ , ν PARA
NEVILLE Neville r , k, τ PARA
HERM Hermite m, σ , γ, ε MOME, PARA

SOFiSTiK 2020 3-7

RELY | Input Description

3.6 CORR – Correlation

See also: VAR

CORR

Item Description Unit Default

VAR1 Random Variable 1 LT8 !

VAR2 Random Variable 2 LT8 !

C Correlation between VAR1 and VAR2 − !

C is the correlation coefficient between random variables VAR1 and VAR2. By default, all
random variables are considered uncorrelated.

3-8 SOFiSTiK 2020

 Input Description | RELY

3.7 PROJ – Define the SOFiSTiK project that evaluates the limit-state
function

See also: FUNC, VAR

PROJ

Item Description Unit Default

DAT Name of the DAT-file of the project LT !

CDB Name of the CDB-file of the project LT *

Defines the SOFiSTiK project that needs to be computed in order to evaluate the limit-state
function. For the special case that the limit-state function is defined explicitly by means of a
mathematical function, this record does not have to be specified.

If CDB is not specified, it is set equal to the value specified in DAT.

Within the dat-file specified in DAT, the current realization of the random variable VARNAME
can be addressed as #VARNAME.

SOFiSTiK 2020 3-9

RELY | Input Description

3.8 FUNC – Define a limit-state function

See also: PROJ, VAR, SLSF

FUNC

Item Description Unit Default

NAME Name of the limit-state function LT8 REL_FUN

EXPR mathematical definition of the limit-state function [∗] -

If multiple limit-state functions are defined, the individual functions can be addressed by their
NAME.

If no SOFiSTiK project was specified by means of PROJ, EXPR defines the limit-state function
as an explicit mathematical expression. The functional expression of the limit-state function
must start with an equal sign (i.e., with ’=’). An example of an explicit limit-state function
definition that does not involve the SOFiSTiK FEM-module is:

FUNC 'rel_fun' EXPR '=0.1-800000/2.1952*#rv_q/#rv_E'

Please do not forget the equal sign at the beginning of the functional expression.

If a SOFiSTiK project was specified through PROJ, EXPR is ignored. In this case, the value of
NAME is retrieved from the CDB specified in PROJ. Consequently, NAME must be defined in
the dat-file specified in PROJ; e.g., by means of #STO.

Note: If the limit-state function is expressed in terms of a finite element analysis, the SOFiSTiK
@KEY can be used in the DAT-file specified in PROJ to access the CDB of the finite element
solver. An introduction to the @KEY concept can be found in the SOFiSTiK Basics-manual.
The structure of the CDB is documented in the file [Link] that can be found in the
installation folder of SOFiSTiK.

3-10 SOFiSTiK 2020

 Input Description | RELY

3.9 SLSF – Define the system limit-state function

See also: PROJ, VAR, FUNC

SLSF

Item Description Unit Default

VARN Name of FEM-variable that stores the result of the LT8 -

system limit-state function.

EXPR explicit representation of the system limit-state func- [∗] -

tion

With SLSF, a user-defined limit-state function for the system reliability problem can be speci-
fied. By default, the system reliability problem is defined as follows: The limit-state functions
defined through FUNC are treated as belonging to components of a series system. Thus, if
failure occurs in at least one local limit-state function, the entire system is assumed to fail.

SLSF can be called only once. Ideally, the command should be called after all local limit-state
function were defined through FUNC. Either VARN or EXPR must be specified, but not both at
the same time.

If the local limit-state functions are defined explicitly using FUNC EXPR, then the system limit-
state function must be specified by means of EXPR. EXPR defines the limit-state function of
the system reliability problem as a explicit mathematical expression. Note that similar to FUNC
EXPR, the functional expression of the limit-state function must start with an equal sign (i.e.,
with ’=’). The names of the local component limit-state functions can be used as variables (i.e.,
’#’ followed by the name of the LSF) in the functional expression EXPR.

If the local limit-state functions are defined within a separate DAT-file specified in PROJ DAT,
then the system limit-state function must be specified also in that DAT-file. In this case, the
name of the variable used to store the value of the evaluated limit-state function in the CDB
must be given in VARN.

SOFiSTiK 2020 3-11

RELY | Input Description

3.10 MC – Monte Carlo simulation

See also: IS, LS, DS, ADS, FORM, SUBS, EXDS

MC

Item Description Unit Default

NSPL Number of samples − 2500

LSF Name of the limit-state function LT8 *

SAVE Name of a file to which the generated samples are LT ""
written.

SAPD Should the generated samples be appended to the LT NO

file specified in SAVE?
NO an already existing file is overwritten
YES generated samples are appended to an al-
ready existing file

FSAV Name of a file to which the failed samples are written. LT ""

FSAP Should the failed samples be appended to the file LT NO

specified in FSAV?
NO an already existing file is overwritten
YES generated samples are appended to an al-
ready existing file

SPLS Sampling technique to be used to generate samples LT PRAN

PRAN pseudo random sampling
LHS Latin Hypercube sampling is employed

SENS Evaluate sensitivities. LT YES

REST Restart simulation using samples from specified file LT ""

Performs a Monte Carlos simulation (see Section 2.3) with NSPL samples.

If multiple limit-state functions are defined, the limit-state function to use can be specified with
LSF. If the key LSF is not set explicitly, the first limit-state function that was defined is [Link]
LSF is a proper limit-state function (and not a response surface function), then the only effect
of LSF is on the ’analysis log’. The final results will be computed for all limit-state functions
defined.

All processed samples are written to the file specified in SAVE. The structure of the file is as
follows: Each generated sample is written to a separate row. The first K columns contain the
values of all defined limit-state functions evaluated for the sample at hand, where K denotes the
number of limit-state functions defined. The ordering of the limit-state functions corresponds

3-12 SOFiSTiK 2020

 Input Description | RELY

to the order of their definition. Thereafter comes a column that lists the values of the limit-state
function of the system reliability problem (see SLSF). If the specified limit-state function is a
resonse surface function, then only the the value of the evaluated response surface function
is listed in the first column; no other results of limit-state function are outputted. The next
M columns list the value of the specified random variables, where M denotes the number of
random variables defined. The last M columns list the value of the specified random variables
transformed to the underlying independent standard Normal space. Note that if a response
surface function is specified as limit-state function, only the value of the specified function is
written to the file (i.e., K = 1 in this case). The same holds for the faileed samples which are
written to the file specified in FSAV.

The values of all random variables as well as the values of all specified limit-state functions are
stored in memory for each evaluated sample. For how this data can be re-used, please have a
look at the documentation of RSF.

A Monte-Carlo simulation can be restarted, meaning the sampling performed in a previous

run can be continued. The file that contains the data from the previous run is specified in
REST. The expected format is the same as for the file specified in SAVE. If REST is used, the
number specified in NSPL is the number of samples that are to be generated within the current
simulation run. It is perfectly valid to specify the same file in REST and in SAVE. However, in
this case SAPD should be set to YES - so that the existing samples are not overwritten in the
current simulation. If the limit-state function specified in LSF is a response surface function
(RSF), the restart-feature cannot be used; i.e., the value of the option REST is ignored in case
of a RSF.

SOFiSTiK 2020 3-13

RELY | Input Description

3.11 IS – Importance sampling

See also: MC, LS, DS, ADS, FORM, SORM, SUBS

IS

Item Description Unit Default

NSPL Number of samples − 2500

DSPT Where to center the importance sampling density? LT USE

NO at the origin in standard normal space
(equivalent to Monte Carlo simulation)
USE get starting point from VAR VAL
COMP design point is computed if no design point
is in CDB
CDB load point stored in CDB

LSF Name of the limit-state function LT8 *

Performs an importance sampling (see Section 2.6) with NSPL samples.

DSPT controls where the importance sampling density is centered:

• NO: the importance sampling density is centered around the origin in standard normal
space. Consequently, for this option importance sampling is equivalent to Monte Carlo
simulation.
• USE: the coordinates specified in VAR VAL are used as the coordinates for the centering.
• COMP: If no (approximated) design point is stored in the CDB, the design point is computed
by means of FORM. The FORM analysis is performed with default settings.
• CDB: If an already computed design point of the problem or an approximated design point
is stored in the CDB, this point is used as center of the importance sampling density. Please
have a look at CTRL CDBP. An approximated design point stored in the CDB is used only
if no design point is stored in the CDB. If no point is stored in the CDB, a warning is issued
and DSPT is changed to USE.

If multiple limit-state functions are defined, the limit-state function to use can be specified with
LSF. If the key LSF is not set explicitly, the first limit-state function that was defined is used.

3-14 SOFiSTiK 2020

 Input Description | RELY

3.12 LS – Line sampling

See also: MC, IS, DS, ADS, FORM, SORM, SUBS

LS

Item Description Unit Default

NSPL Number of samples − 2500

SPNT Where to get the (approximate) design point from? LT USE
USE get starting point from VAR VAL
COMP design point is computed if no design point
is in CDB
CDB load point stored in CDB

LSF Name of the limit-state function LT8 *

GSTP Step size for the finite difference scheme in line- − 0.01
search algorithm.
EPSC Convergence criterion for line-search algorithm. − 0.001

Performs a line sampling (see Section 2.7) with NSPL samples.

SPNT controls the orientation of the hyperplane on which the samples are generated.

• USE: the coordinates specified in VAR VAL are used.

• COMP: If no (approximated) design point is stored in the CDB, the design point is computed
by means of FORM. The FORM analysis is performed with default settings.
• CDB: If an already computed design point of the problem or and approximated design
point is stored in the CDB, then this point is used. Please have a look at CTRL CDBP. An
approximated design point stored in the CDB is used only if no design point is stored in the
CDB. If no point is stored in the CDB, a warning is issued and SPNT is changed to USE.

If multiple limit-state functions are defined, the limit-state function to use can be specified with
LSF. If the key LSF is not set explicitly, the first limit-state function that was defined is used.

A description of the options GSTP and EPSC is given in the section of FORM.

SOFiSTiK 2020 3-15

RELY | Input Description

3.13 DS – Directional sampling

See also: MC, IS, LS, ADS, FORM, SORM, SUBS

DS

Item Description Unit Default

NSPL Number of samples − 2500

LSF Name of the limit-state function LT8 *

GSTP Step size for the finite difference scheme in line- − 0.01
search algorithm.
EPSC Convergence criterion for line-search algorithm. − 0.001

Performs a directional sampling (see Section 2.8) with NSPL samples.

If multiple limit-state functions are defined, the limit-state function to use can be specified with
LSF. If the key LSF is not set explicitly, the first limit-state function that was defined is used.

A description of the options GSTP and EPSC is given in the section of FORM.

3-16 SOFiSTiK 2020

 Input Description | RELY

3.14 ADS – Adaptive sampling

See also: MC, IS, LS, DS, FORM, SORM, SUBS

ADS

Item Description Unit Default

NSPL Number of samples − 2500

SPNT Where to initially center the samples? LT USE

NO no starting solution specified
USE get starting point from VAR VAL
COMP starting solution is computed if none is in
CDB
CDB load the point stored in CDB

LSF Name of the limit-state function LT8 *

Performs an adaptive sampling (see Section 2.9) with NSPL samples.

SPNT controls the orientation of the hyperplane on which the samples are generated.

• NO: the search for the design point starts at the origin in standard normal space.
• USE: the coordinates specified in VAR VAL are used.
• COMP: If no (approximated) design point is stored in the CDB, the design point is computed
by means of FORM. The FORM analysis is performed with default settings.
• CDB: if some prior analysis already computed the design point of the problem or an ap-
proximated design point, this point is used as starting point. Please have a look at CTRL
CDBP. An approximated design point stored in the CDB is used only if no design point is
stored in the CDB.

If multiple limit-state functions are defined, the limit-state function to use can be specified with
LSF. If the key LSF is not set explicitly, the first limit-state function that was defined is used.

SOFiSTiK 2020 3-17

RELY | Input Description

3.15 FORM – First order reliability method

See also: MC, IS, LS, DS, ADS, SORM, SUBS

FORM

Item Description Unit Default

SPNT Where to start the search for the design point? LT NO
NO no starting solution specified
USE get starting point from VAR VAL
RND random starting point
GRAD from a gradient evaluation
CDB load point stored in CDB

OPTA Optimization algorithm for the beta-point search LT HLRF

RFLS RFLS algorithm
NLPQ NLPQL algorithm
HLRF HLRF algorithm
COBY COBYLA

GRDM Finite difference scheme to evaluate gradients numer- LT FWD

ically
FWD forward differences
CTRD central differences
GSTP Step size for the finite difference scheme − 0.01

MIT1 for the RFLS algorithm: maximum number of itera- − 50

tions in the approximate line search step
for the NLPQL algorithm: maximum number of itera-
tions in the line search step
MIT2 for the RFLS algorithm: maximum number of itera- − 50
tions in the exact line search step
for the NLPQL algorithm: maximum number of limit-
state evaluations in one iteration of the line search

EPSC Convergence criterion for beta-point search. − 0.001

SMU Convergence criterion for Armijo line-search. − 0.001

LSF Name of the limit-state function LT8 *

Performs a first order reliability analysis (see Section 2.4).

If multiple limit-state functions are defined, the limit-state function to use can be specified with
LSF. If the key LSF is not set explicitly, the first limit-state function that was defined is used.

3-18 SOFiSTiK 2020

 Input Description | RELY

3.15.1 SPNT: Starting solution

If the search for the design point did not converge, specifying a starting point for the search
can help finding the design point of the problem. Using a random starting point and performing
the analysis multiple times can give information about the existence of multiple (local) design
points. Moreover, if multiple (local) design points exist, using different starting points might help
finding the local design points.

• NO: The search for the design point starts at the origin in standard Normal space.
• USE: The coordinates specified in VAR VAL are used as the coordinates of the starting
point.
• RND: The starting point will be selected randomly in standard normal space.
• GRAD: The starting point will be generated from a gradient evaluation at the origin in stan-
dard normal space.
• CDB: If some previous analysis already computed the design point of the problem or an
approximated design point, this point is used as starting point. Please have a look at CTRL
CDBP. An approximated design point stored in the CDB is used only if no design point is
stored in the CDB.

Note: after a design point was detected with FORM (or SORM) the design point is stored in
the CDB. If already a design point is in the CDB, the design point in the CDB is overwritten if
the detected point is ’better’ than the one already in the CDB.

3.15.2 Convergence criteria

The convergence of the search for the design point can be controled with EPSC and SMU.

The global convergence in the search for the design point is controlled by EPSC. The search
stops if the following condition fulfilled:

|k+1 − k | ≤ ϵ ∀  ∧ |g(T(k+1 ))| ≤ ϵ (3.1)

where ϵ is equal to EPSC. The recommended range for EPSC ranges from 10−4 to 10−2 .

SMU is the relative precision in the Armijo-type stop criterion in the line search. The value of
SMU must be in the range 0 < SMU < 0.5.

3.15.3 OPTA: Optimization algorithms

Different algorithms that search the design point are implemented in Rely : HLRF, RFLS,
NLPQL, and COBYLA. The first three mentioned algorithms require the evaluation of gradi-
ents of the limit state function, and, thus, the limit-state function must be differentiable. In Rely
the gradient of the limit-state function is evaluated numerically. In case of SORM the limit-state
function must be twice differentiable at the design point. The fourth algorithm (COBYLA) is a
gradient-free algorithm.

HLRF (Hasofer and Lind [12] Rackwitz and Fiessler [13]) is widely used due to its simplicity and
rapid convergence. The design point is recursively approximated by the following expression:

∇g(T(k )) ” —
2 ∇g(T(k )) k − g(T(k ))
T
k+1 =  (3.2)
∇g(T(k ))

SOFiSTiK 2020 3-19

RELY | Input Description

However, the algorithm might fail to converge under certain conditions (Der Kiureghian and
Stefano [14]).

Both RFLS and NLPQL are non-linear sequential quadratic programming algorithms; i.e., they
consist of two steps: first the direction is determined and then the step length. For a small
number of random variables, the implemented version of RFLS is slightly less efficient than
NLPQL. However, for problems with many random variables, the NLPQL algorithm might not
achieve convergence.

The RFLS algorithm is especially adjusted to the the objective T  (Abdo and Rackwitz
p

[15]). In Rely , two line search strategies to determine an optimal step size are implemented:
an approximate line search and an exact line search. The exact line search is only done if
the approximate line search did not converge. The approximate line search is performed by
means of a quadratic interpolation/extrapolation. In each iteration step only one additional limit-
state evaluation is required. The maximum number of iterations in the approximate line search
algorithm is set in MIT1. The exact line search is performed by means of the golden section
method. Consequently, the exact line search requires considerable more limit-state evaluations
than the approximate line search. The maximum number of iterations in the exact line search
algorithm is set in MIT2. If convergence problems occur, it is recommended to increase MIT2.
For most problems the approximate line search should converge fast and the exact line search
is not needed. However, in the presence of strong curvatures, the approximate line search
might be inefficient or might not converge. For such problems, MIT1 can be set to zero, such
that only an exact line search is performed. In Rely , no updating scheme for the Hessian matrix
is used, because all known updating schemes can result in singular Hessians. Although some
precautions to prevent singularity of the Hessian during the iteration can be made, the solution
can be strongly affected by ill-conditioned Hessians. Especially for problems with many random
variables, this can lead to oscillations and convergence problems.

The NLPQL algorithm was originally designed by Schittkowskip (1981,1983). In Rely , the algo-
rithm is adapted to the special form of the objective function ( T ). The algorithm is suited
for problems with a small number of random variables (up to 20). It is considered to be very
robust, reliable and efficient in finding the design point. An efficient numerical Hessian update
scheme and a nearly exact line search algorithm is used. If the so-called quadratic subprob-
lem cannot be solved, it creates a similar quadratic subproblem which can always be solved.
This algorithm is recommended for problems that have a complicated and computationally de-
manding limit-state function but not more than 20 random variables. The maximum number of
iterations in the line search step is set in MIT1. The maximum number of limit-state evaluations
in one iteration of the line search is set in MIT2.

The COBYLA algorithm is a gradient-free algorithm based on Nelder/Mead (1965) and modi-
fied by Powell (1994) to include inequality constraints. It is less efficient and (in some cases)
also less reliable than the other mentioned algorithms. However, if the limit-state function is
non-differentiable (but it must still be smooth), COBYLA can be applied. If the limit-state func-
tion is neither smooth nor differentiable, FORM (and, thus, SORM) cannot be applied. Note:
COBYLA should only be used to determine a first approximation of the design point that is
used as input for other reliability methods. The maximum number of limit-state function calls
is determined as MAX(MAX(MIT1,MIT2)*N,100), where N is the total number of random vari-
ables.

3-20 SOFiSTiK 2020

 Input Description | RELY

3.15.4 Finite difference scheme

The finite difference scheme is set in GRDM. It can be either a forward difference scheme
(FWD) or a central difference scheme (CTRD). In general, CTRD is a better estimate than
FWD; however, CTRD requires twice as much limit-state function calls than FWD. The step
size of the finite difference scheme is set in GSTP.

SOFiSTiK 2020 3-21

RELY | Input Description

3.16 SORM – Second order reliability method

See also: FORM

SORM

Item Description Unit Default

SPNT Where to start the search for the design point? LT NO
NO no starting solution specified
USE get starting point from VAR VAL
RND random starting point
GRAD from a gradient evaluation
CDB load point stored in CDB

OPTA Optimization algorithm for the beta-point search LT HLRF

RFLS RFLS algorithm
NLPQ NLPQL algorithm
HLRF HLRF algorithm
COBY COBYLA

GRDM Finite difference scheme to evaluate gradients numer- LT FWD

ically
FWD forward differences
CTRD central differences
GSTP Step size for the finite difference scheme − 0.01

MIT1 for the RFLS algorithm: maximum number of itera- − 50

tions in the approximate line search step
for the NLPQL algorithm: maximum number of itera-
tions in the line search step
MIT2 for the RFLS algorithm: maximum number of itera- − 50
tions in the exact line search step
for the NLPQL algorithm: maximum number of limit-
state evaluations in one iteration of the line search

EPSC Convergence criterion for beta-point search. − 0.001

SMU Convergence criterion for Armijo line-search. − 0.001

LSF Name of the limit-state function LT8 *

Performs a second order reliability analysis (see Section 2.5).

Please note: SORM has exaclty the same input parameters as FORM. For a detailed descrip-
tion of the input parameters please refer to FORM.

3-22 SOFiSTiK 2020

 Input Description | RELY

3.17 MVFO – Mean value first order

See also: MVFO

MVFO

Item Description Unit Default

SPNT Where to start the search for the design point? LT NO
NO no starting solution specified
USE get starting point from VAR VAL
RND random starting point
GRAD from a gradient evaluation
CDB load point stored in CDB

GRDM Finite difference scheme to evaluate gradients numer- LT FWD

ically
FWD forward differences
CTRD central differences
GSTP Step size for the finite difference scheme − 0.01

RELA Output estimated reliability LT NO

NO do not output estimated reliability
YES output estimated reliability

LSF Name of the limit-state function LT8 *

Evaluates the derivative of the specified limit-state function (LSF) at the indicated starting point
(SPNT). Based on this information, sensitivity factors are derived.

If RELA is activated, an estimate for the design point is computed, based on the evaluated
derivative. However, for non-linear limit-state functions, the so-obtained approximation is typi-
cally rather crude.

For a detailed description of the input parameters please refer to FORM.

SOFiSTiK 2020 3-23

RELY | Input Description

3.18 SUBS – Subset simulation

See also: MC, IS, LS, DS, ADS, FORM, SORM

SUBS

Item Description Unit Default

NSPL Number of samples per subset − 1000

P0 Threshold probability of the intermediate failure do- − 0.1

mains.
LMDA Initial scaling factor − 0.6
NCS Chain parameter for adaptation − 15

LSF Name of the limit-state function LT8 *

Performs a subset simulation (see Section 2.10) with NSPL samples. It is recommended to set
NSPL always greater or equal than 500.

PZERO is the threshold probability of the intermediate failure domains. It can be any value
between ]0; 1[ . Usually, this value should not be changed.

LMDA defines the initial scaling factor of the proposal distribution. The admissible range for
LMDA is ]0; 1] . During the simulation, LMDA can be modified adaptively such that the ac-
ceptance of a sample proposed in a MCMC step is close to an optimal acceptance rate. The
absolute value of the parameter NCS defines after how many seed values LMDA is modified.
If NCS is set to zero, LMDA is not modified. If NCS is not zero, it’s absolute value should be
larger or equal than 5. If NCS is smaller than zero, all directions are treated as equal and LMDA
corresponds to the standard deviations of the proposal distribution. If NCS is larger than zero,
the standard deviations of the proposal distribution are estimated from the seed values, and
LMDA acts as a multiplier to achieve a near-optimal acceptance rate.

If multiple limit-state functions are defined, the limit-state function to use can be specified with
LSF. If the key LSF is not set explicitly, the first limit-state function that was defined is used.

3-24 SOFiSTiK 2020

 Input Description | RELY

3.19 EXDS – Experimental Design

See also: MC, RSF

EXDS

Item Description Unit Default

SDES Sampling Design LT JOA

DAT points are read from an input file
JOA points located just on axis
CC central composite design
FF full factorial design
FFS full factorial design (spherical)
P1 points located just on axis (one point in each
direction)

CNTR Location of the center point of the design (relevant for LT NO
JOA, CC, FF, FFS, and P1).
NO no center point specified (origin is used)
USE get center point from VAR VAL
CDB load point stored in CDB

SAIF Name of a file from which samples are to be read. LT ""

SAVE Name of a file to which the generated samples are LT ""
written.

SAPD Should the generated samples be appended to the LT NO

file specified in SAVE?
NO an already existing file is overwritten
YES generated samples are appended to an al-
ready existing file

LSF Name of the limit-state function. This option is only LT8 *

used if CNTR is set to CDB.

SPRD Spread of the sampling design. − 1.0

Evaluates the underlying model for a specified set of points. The points can be selected based
on a specific design, or read from an input file. The sampling design is specified using keyword
SDES.

If SDES is set to DAT, the points are retrieved from an input file. The path and filename of the
input file must be specified using keyword SAIF. The format of this input file is as follows: One
sampling point to evaluate per line. The values of the uncertain model parameters (random
variables) are assigned based on the columns in this file; i.e., the first column is associated

SOFiSTiK 2020 3-25

RELY | Input Description

with the first parameter defined, and so on.

Alternatively, a predefined sampling design can be specified using keyword SDES. Available
options are (Note: M is used in the following to denote the number of random variables defined
for the problem at hand.):

• JOA: A total of 2 · M + 1 points are specified. The center point as well as two points on each
axis; one on the positive side and one on the negative side of the center point on each axis.
• CC: A central composite design is used, which means that 2 · M + 1 + 2M points are used.
• FF: A full factorial design is used, which means that 3M points are used. The points are
located in a hypercube.
• FFS: A full factorial design is used, which means that 3M points are used. The points are
located in a hypersphere.
• P1: Similar to JOA, however, only M + 1 points are specified. The center point as well as
one point on each axis; one located on the side of the center point that point towards the
origin.

The center point of the sampling design can be specified using the keyword CNTR. Available
options are:

• NO: The origin in standard Normal space is used as center point for the sampling design.
• USE: The coordinates specified in VAR VAL are used as the coordinates of the center point.
• CDB: If some previous analysis already computed the design point of the problem or an
approximated design point, this point is used as center point. Please have a look at CTRL
CDBP. An approximated design point stored in the CDB is used only if no design point is
stored in the CDB.
If the center point is to be retrieved from the CDB, the limit-state function of which the
design point is to be taken must be specified using the keyword LSF.

The spread of the employed sampling design is specified in SPRD.

All processed samples are written to the file specified in SAVE. The structure of the file is as
follows: Each generated sample is written to a separate row. The first K columns contain the
values of all defined limit-state functions evaluated for the sample at hand, where K denotes the
number of limit-state functions defined. The ordering of the limit-state functions corresponds
to the order of their definition. Thereafter comes a column that lists the values of the limit-state
function of the system reliability problem (see SLSF). If the specified limit-state function is a
resonse surface function, then only the the value of the evaluated response surface function
is listed in the first column; no other results of limit-state function are outputted. The next
M columns list the value of the specified random variables, where M denotes the number of
random variables defined. The last M columns list the value of the specified random variables
transformed to the underlying independent standard Normal space.

The values of all random variables as well as the values of all specified limit-state functions are
stored in memory for each evaluated sample. For how this data can be re-used, please have a
look at the documentation of RSF.

3-26 SOFiSTiK 2020

 Input Description | RELY

3.20 RSF – Response Surface Function

See also: MC, EXDS

RSF

Item Description Unit Default

NAME Name with which to associate the response surface LT !

function (RSF)

LSF Name of the limit-state function that is to be approxi- LT !

mated by the RSF

ANSA Ansatz Space LT P1

P1 polynomials of first order
P2NM polynomials up to second order (without
mixed terms)
P2MT polynomials up to second order (with mixed
terms)

LOAD Load the parameter vector from the specified file. LT ""

SAVE Save the computed parameter vector to the specified LT ""
file.

If no file is specified in LOAD, the parameters of the response surface function are computed
based on performed simulation runs. The data from the last performed Monte Carlo (MC)
simulation or the last performed experimental design (EXDS) are used.

The parameter vector can be written to a file, which can be specified through SAVE. Note: In
case the parameter vector is read from a file specified in LOAD, it will not be written to the file
specified in SAVE.

SOFiSTiK 2020 3-27

RELY | Input Description

3-28 SOFiSTiK 2020

 Output Description | RELY

4 Output Description

During a run of Rely the following output files are generated:

• the URSULA output and graphic representations (.PLB)

• the Rely -results file (.RTI)
• the Rely -monitoring file (.MTI)
• the Rely -post-processing file (.PTI)

If no errors occur, for most users, only the URSULA output will be of interest. In case of errors,
the monitoring and the results file can be helpful in tracking the errors down.

The verbosity of the output to the results file can be controlled with the command OUTP NPRI.
The verbosity of the output to the monitoring file can be controlled with the command OUTP
IMON.

For problems with a limit-state function that is computationally demanding to evaluate (i.e.,
already a single call of the limit-state function takes a considerable time), you might want to
output every single result of a limit-state function call. This can be done with the command
OUTP LSFR YES.

SOFiSTiK 2020 4-1

RELY | Output Description

4-2 SOFiSTiK 2020

 Examples | RELY

5 Examples

The input files which are explained here are to be found in the installation directory SOFiSTiK
in the subdirectory [Link]/english. Alternatively, you can find these examples in the TEDDY
menu HELP > EXAMPLES sorted by program name and language.

For additional help like tutorials, tutorial movies and practical examples please refer to the
SOFiSTiK Infoportal ([Link]/Infoportal).

SOFiSTiK 2020 5-1

RELY | Examples

5.1 Simple Beam: Explicit Limit-State Function

A single-span beam with a length of 20m is investigated. The beam is subjected to a uniform
distributed load q, and the displacement  at the center of the beam is measured (Fig. 5.1).
The Young’s modulus E as well as the value of the loading q are considered uncertain. The
Young’s modulus follows a Log-Normal distribution with a mean value of 45 · 103 MN/ m2 and a
coefficient of variation of 10%. The value of the loading has a Normal distribution with a mean
of 7kN/m and a standard deviation of 1.6kN/m. The height and the width of the beam are
regarded as deterministic with values h = 0.7m and b = 0.2m, respectively. The system fails if
the displacement  is larger or equal than 0.1m.

w
E
l=20m

Figure 5.1: Model of the investigated beam

The center displacement  of the mechanical model at hand has an analytical solution:

5q4 8 · 105 q
(E, q) = = (5.1)
32Ebh3 2.1952 E

The limit-state function of the problem can be expressed as:

(E, q) = 0.1m − (E, q) (5.2)

The input to perform the reliability analysis reads:

PROG RELY
HEAD

!Define the random variables

!Young's modulus
!Type: Log-Normal distribution
!Mean: 45e3 MN/m2 = 45e6 kN/m2
!C.o.V.: 10%
VAR 'rv_E' TYPE LOGN PTYP MOME P1 45.e+6 P2 4.5e+6

!Distributed loading
!Type: Normal distribution
!Mean: 7 kN/m

5-2 SOFiSTiK 2020

 Examples | RELY

![Link].: 1.6 kN/m

VAR 'rv_q' TYPE NORM PTYP MOME P1 7. P2 1.6 P3 0 P4 16

!Define the limit-state function

FUNC 'rel_fun' EXPR '=0.1-800000/2.1952*#rv_q/#rv_E'

!Perform a FORM analysis

FORM

!Perform line sampling

LS SPNT CDB

!Perform subset simulation

SUBS NSPL 2000

!Perform a Monte Carlo simulation

MC NSPL 1e5

END

First, the two random variables of the problem are defined: the variable modeling the Young’s
modulus E is called rv_E, the variable modeling the loading q is called rv_q. The limit-state
function is defined using the command FUNC. Note that the first character in the explicit ex-
pression of the limit-state function must be the equal sign (=).

Next, a FORM analysis with default settings is performed. Thereafter, the design point com-
puted with FORM is used to perform a Line Sampling (LS). The default number of samples
(2500) is used. Afterwards, subset simulation (SUBS) with 2000 samples per conditioning step
is done. The last reliability method run is Monte Carlo sampling (MC) with 1 · 105 samples.

The outputted summary of the results of the different reliability methods is:

LSF Pf Beta C.o.V.

FORM REL_FUN 3.53751E-03 2.69330E+00
Line Sampling REL_FUN 3.35353E-03 2.71106E+00 1.42193E-03
Subset Simulation REL_FUN 2.59500E-03 2.79501E+00
Monte Carlo Simulation REL_FUN 3.16000E-03 2.73071E+00 5.61657E-02

Note that by default CTRL SEED is set to 8, and, therefore, only FORM will give the same result
each time the analysis is repeated. A reference solution of the problem was obtained as pf =
3.349 · 10−3 . In this case, the best estimate is given by line sampling: the estimate is closest to
the reference solution and has the smallest coefficient of variation. The coefficient of variation
of the estimate obtained with Monte Carlo simulation is more than an order of magnitude larger
than the one obtained with line sampling; despite the fact that Monte Carlo simulation required
almost 10 times as many limit-state function calls than line sampling. The estimate computed
with subset simulation is not as good as the one obtained from line sampling; however, subset
simulation with 2000 samples per conditioning step needed in total only half the number of
limit-state function calls that line sampling with 2500 samples needed.

SOFiSTiK 2020 5-3

RELY | Examples

5.2 Simple Beam: Finite Element Limit-State Function

The single-span beam illustrated in Fig. 5.1 and discussed in Section 5.1 is now modelled by
means of finite elements (File: strurel_beam_FEM_model.dat):

+PROG AQUA
! define the cross-section of the beam
! define the material: the Young's modulus is a random variable
HEAD Plane Beam
NORM DIN 1045-1
MATE 1 E #rv_E
SREC 1 MNO 1 H .7 B .2
END

+PROG SOFIMSHA
! define the mesh of the beam with length 20m
HEAD Plane Beam
SYST FRAM
GRP 1
NODE NO X FIX
1 0 PY
21 20 PP
(2 20 1) (1 1)
BEAM (1 20 1) (1 1) (2 1)
END

+PROG SOFILOAD
! define the distributed loading -> the value of the loading is random!
HEAD Distributed load
LC 1
BEAM GRP 1 PA #rv_q
END

+PROG ASE
! solve the system
HEAD
LC 1
END

+PROG TEMPLATE
! Store the result in variable relfun
HEAD performance function
! max uy in m
LET#lf 1
@KEY N_DISP #lf
LET#uy @(11,UY)

5-4 SOFiSTiK 2020

 Examples | RELY

! store the result of the limit-state function in the CDB:

STO#rel_fun (10./100.-#uy)
END

Note: the random variables are used exactly like normal variables in the input file. The Young’s
modulus is called rv_E and the loading is referred to as rv_q. The result of the limit state
function is stored as variable with name rel_fun in the CDB.

The strurel input is (File: strurel_beam_FEM.dat):

PROG RELY
HEAD
!define the random variables
!Young's modulus
!type: Log-Normal distribution
!mean: 45e3 MN/m2
!C.o.V.: 10%
VAR 'rv_E' TYPE LOGN PTYP MOME P1 45.e+3 P2 4.5e+3

!Distributed loading
! type: Normal distribution
! mean: 7 kN/m
! [Link].: 1.6 kN/m
VAR 'rv_q' TYPE NORM PTYP MOME P1 7. P2 1.6

!define the limit-state function

!name of the variable to read from the CDB of
!the project 'strurel_beam_FEM_model'
FUNC rel_fun

!the project that needs to be computed to evaluate the limit-state function

PROJ DAT 'strurel_beam_FEM_model'

!output the result of each limi-state function call

OUTP LSFR YES

!perform a FORM analysis

FORM

!perform line sampling

LS NSPL 100 SPNT CDB
END

The definition of the two random variables is equivalent to the one of the explicit example
(Section 5.1). The limit-state function is defined using the commands FUNC and PROJ: PROJ
specifies which DAT file is to be executed, and FUNC defines under which name the returned
value of the limit-state function is stored in the CDB.

SOFiSTiK 2020 5-5

RELY | Examples

The reliability analysis is performed using FORM and LS. First, the design point is computed
with FORM; and second, the estimate of the linearized problem is improved by Line Sampling
with 100 samples. The outputted summary of the results is:

LSF Pf Beta C.o.V.

FORM REL_FUN 3.53752E-03 2.69330E+00
Line Sampling REL_FUN 3.37592E-03 2.70885E+00 7.14756E-03

5-6 SOFiSTiK 2020

 Appendix | RELY

6 Appendix

SOFiSTiK 2020 6-1

 APPENDIX A: DISTRIBUTION FUNCTIONS AND THEIR PARAMETERS

Parameters of Distribution Functions

(MOME: Moments, PARA: Parameters)

Type ID Name of Parameters on Type of parameter input
TID distribution P1,P2,P3,P4 (default is PARA)
0 Fixed v PARA
1 Rectangular. (uniform) a, b MOME or PARA
2 Normal (Gauß) m,  MOME or PARA
3 Lognormal , MOME or PARA
4 Exponential , MOME or PARA
5 Gamma k,  MOME or PARA
6 Beta r, t, a, b, MOME or PARA
7 Gumbel (max) u,  MOME or PARA
8 Frechet (max) v, k,  PARA
9 Weibull (min) w, k,  MOME or PARA
10 Shifted Lognormal ,, MOME or PARA
11 Rayleigh , MOME or PARA
12 Student (standard)  PARA
13 Trapezoid a, b, c, d PARA
14 Birnbaum/Saunders , MOME or PARA
15 Cauchy , PARA
16 Shifted Gamma k, , MOME or PARA
17 Inverse Gauß , MOME or PARA
18 Gumbel (min) u,  MOME or PARA
19 Frechet (min) v, k,  PARA
20 Weibull (max) w, k,  MOME or PARA
21 Pareto a, k MOME or PARA
22 Laplace , MOME or PARA
23 Logistic , MOME or PARA
24 Halfnormal m,  MOME or PARA
25 Chi-square  PARA
26 Fisher’s F  1,  2 PARA
27 Student m, , PARA
28 Neville r, k,  PARA
29 Hermite m, ,, MOME or PARA
30 [Link] (max) t, , n PARA
31 [Link] (min) t, k, n PARA
32 [Link] (max) t, k, n PARA
33 [Link] , h PARA
34 [Link] (m unknown) ,, n PARA
35 [Link] ( unknown) m, s,  PARA
36 [Link] (m  ) x , n, s,  PARA
37 [Link] (max) t, , n PARA
38 [Link] (min) t, k, n PARA
39 [Link] , h PARA
40 [Link] (m) ,  PARA
41 [Link] () s,  PARA
42 [Link] (m  ) x , , n PARA
43 [Link] (max) t, k, n PARA
44 4-Par. Lognormal ,,, PARA

SOFiSTiK 2014 Appendix A (taken from Strurel – Users Manual, Gollwitzer, 2014) Page A-1
 APPENDIX A: DISTRIBUTION FUNCTIONS AND THEIR PARAMETERS

Notes to Parameters of Distribution Functions:

 Greek characters and other symbols used in this table are converted to plain text in the stochastic model
window (for example =lamda, =nue and x =xq).
 The distribution types in the table above are ordered according to their ID, TID in command VAR.
 For input of PTP equal to MOME the mean and standard deviation are the first and second entry for each
variable. If there are more than two parameters then parameter no.3 and optionally parameter no.4 have to be
supplied. So also for input of moments - in case of a distribution with more than 2 distribution parameters -
the distribution parameters 3 and 4 have to be provided in addition to the mean and standard deviation.

SOFiSTiK 2014 Appendix A (taken from Strurel – Users Manual, Gollwitzer, 2014) Page A-2
 APPENDIX B: Table 1: Standard Continuous Distributions

APPENDIX B: Univariate Distribution Functions

1. Standard Distribution Functions

Type Name Density/Distribution Function Range Parameter Mean Standard deviation

Restriction
1 Rectangular 1
fX (x)  ba
(Uniform) ba ab
xa a  x b ab 2 (1 2 )
FX ( x ) 
ba
V,VI)
2 Normal  1  x  m 2 
1 1  xm
fX (x)  exp         
2   2            x    0 m 
xm
FX ( x )    
  
3 Lognormal V,VI)  1  ln  x /    2 
1
fX (x)  exp       0 
2

2

x 2  
2

modal value =  2   0 x   e 2

 e 2
e 1
  0
   M edian
 ln  x /   
exp   
2 1
   
x    M ean   ln  C ov  1 
2 2
[Link].  
 ln  x /    C ov =
M ean C ov  1
2

FX ( x )    
  
4 Exponential f X ( x )   exp     x     1 1
  x  0  
 
F X ( x )  1  exp     x    

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-1
 APPENDIX B: Table 1: Standard Continuous Distributions

Type Name Density/Distribution Function Range Parameter Mean Standard deviation

Restriction
5 Gamma I,II) 1 k 1
fX (x)    x  ex p    x 
 k  k  0 k k
0 x  0 
( x, k ) 
FX ( x ) 
(k )
III,IV) r 1 t 1
6 Beta  xa  xa
  1   a  b r
ba  ba a  b  a  b  a 
rt
fX (x)  a  x b r  0 rt
B  r, t  b  a   r  t   r  t  1
2

t  0
B ( y, r, t) xa
FX ( y )  ; y 
B (r, t) ba
7 Gumbel 
f X ( x )   exp    x  u   exp     x  u   
(max. Type I) C VIII) 1 
F X ( x )  exp   exp    ( x  u )     x   u =  = 
 0   6
   
F X ( x )  ex p   ex p    x     C 
   6 
8 Frechet k  v  
k 1
  v   k 
f X (x)  exp      x  k 0
(max. Type      1  2 2  1
II) I) v    x     x     v  (v   ) 1     v     1     1  
 k  k   k 
  v   k 
F X ( x )  ex p     k  1 k  2
 
  x    
9 Weibull k 1
  x   k 
k  x  
(min. Type fX (x)    exp       1
w    k 0  2 2  1
III) I)  w     w       x  w     1    w     1     1  
w   k  k   k 
  x   k 
F X ( x )  1  ex p     
  w    

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-2
 APPENDIX B: Table 1: Standard Continuous Distributions

Type Name Density/Distribution Function Range Parameter Mean Standard deviation

Restriction
10 Shifted Lognormal VI)  1  ln(  x    /  )  2 
1
fX (x)  exp       0 
2

2

 x    2  
2

 2     x   e 2
 e 2
e 1
  

 ln(( x   ) /  ) 
FX ( x )    
  
11 Rayleigh   x   2 
x 
fX (x)  exp     
 2
2 2
    x   0    2
2 2
 1 ( x   )2 
F X ( x )  1  ex p   

2
 2 
12 Standard Student t I)   1
   1
1  2   x 
2
 
fX (x)   1  
2
 x  0 0 ;   2 
       2
 
2
x
FX ( x )   
f X ( z ) dz

13 Trapezoid

ax d ab see X) see X)

a b c d
 cd
x
FX ( x )   a
f X ( z ) dz

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-3
 APPENDIX B: Table 1: Standard Continuous Distributions

Type Name Density/Distribution Function Range Parameter Mean Standard deviation

Restriction
14 Birnbaum/  1  x/ 
½ V)
1  1 / 2
 
x /    x /  
1/ 2
Saunders V,VI) fX (x)   
   5
2
  
2
2 x
3/2

0  x  0  1     1
 2  4
 1  1

1
  0
FX ( x )     x /  2  x /   2

  
15 Cauchy 1
fX (x) 
  1    x    /   
2

   x 0 non-existent non-existent

 1 1  x  
FX ( x )   tan  
2    
k 1
16 Shifted Gamma    x   
exp     x    
I,II)
fX (x)    x  0 k k
 k   
k 0  
 (  ( x   ), k )
FX ( x ) 
(k )
17 Inverse 1/ 2
   x   2 
  
Gauss VII) fX (x) 
 2 x 3 
exp     0 
3/2

2 x
2
    
0 x  0 
1   (x  )  1  2  
 erf    exp   
2  2  x  2    
FX ( x )  1   
   (x  )  
 erf  
 2  x  
  
18 Gumbel f X ( x )   exp   x  u   exp   x  u    C 1 
   x u
VIII)
(min, Type I)  0 
 6
F X ( x )  1  exp   exp  ( x  u )  

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-4
 APPENDIX B: Table 1: Standard Continuous Distributions

Type Name Density/Distribution Function Range Parameter Mean Standard deviation

Restriction
19 Frechet I) k 1
    v k 
k  v
(min. Type f X ( x)    exp       1
       x      v    1   2 2  1
II)     x      x k 0    v  1     1  
 k  k   k 
    v k   
F X ( x )  1  ex p     k  1 k  2
 
    x  
20 Weibull I) k 1
    x k 
k   x 
(max. Type fX (x)    exp       1
  w     w      w    1   2 2  1
III)     w      x k 0    w  1     1  
 k  k   k 
  w
    x k 
F X ( x )  ex p     
    w  
21 Pareto a k
a

fX (x)  a 1 a k
x 0  k  x a 0 ;  a  1 k a
a 1
; a  2
k 
a
k  0 a 1 a 2
FX ( x )  1   
 x
22 Laplace 1  x  
fX (x)  ex p   
2   
 x  0  2
 1  ex p    
1 x 
fo r x  
 2  
FX ( x )  
 2 ex p   
1 x  fo r x  

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-5
 APPENDIX B: Table 1: Standard Continuous Distributions

Type Name Density/Distribution Function Range Parameter Mean Standard deviation

Restriction
23 Logistic  x 
e x p   
   3 
fX (x)   x  0  

2
3   x   
 1  e x p   
   3 
1
FX ( x ) 
 x
1  exp    
  3 
24 Half- 2  1  x  m 2 
normal V,VI) fX (x)  exp       2 2
2   2     mx m0 m    1
 
2  xm  0
  
   
x
2  zm  xm
FX ( x )    
  
 dz  2  
  

m

25 Central  
1  1   x
f X ( x)  x
2 
Chi-Square I) exp   
  / 2 
 /2
2  2
0 x   0  2
x

f x  z  d z  F  x  | k
1
FX ( x )   
2
; 
2
0

F see Type 5;
x here corresponds to  in textbooks
2

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-6
 APPENDIX B: Table 1: Standard Continuous Distributions

Type Name Density/Distribution Function Range Parameter Mean Standard deviation

Restriction
1
26 Central
1   2   1  2
1
1
Fisher F I)    x 2 2
;  2  2   2  2
  2  2
;  2  4 
2 1
 2   2  0 x  1  0  2  2
fX (x) 
  2  2  2  4 
2
 1  2  / 2
     1  2 0 1

 1  2  1  x
 2   2   2 
x
FX ( x )   
f X ( z ) dz

27 Student I)   1   1
  
 0
 2  t  2 xm 
2
1
fX (x)  1   ;t   x m;   1   ;   2 
       0  2
   
2

x
FX ( x )   
f X ( z ) dz
k 1
28 Neville k  x x  k0
fX (x)  ;  x    x
k  1 k  2
IX ) IX )

1    x  
2
r k r
 0
1
FX ( x )  1 
(1   ( x ) )
k

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-7
 APPENDIX B: Table 1: Standard Continuous Distributions

Type Name Density/Distribution Function Range Parameter Mean Standard deviation

Restriction
29 Hermite V,VI)
  y  1  h 3  y  1   h 4  y  3 y  
1
fY ( y ) 
2 3
(modified)   
 0
FY ( y )    y   h 3 y  y   h 4 ( y  1)  y 
2

 x m 
0
xm  ( X  m ) k 
y  ; k  E   ;   3;   4;
     
  3
h3  ; h4  ;    ;
6 24
for  3  3 :
(1  36 h 4 ) 1
1/ 2
h3 1
h4  ; h3  ; 
1  6 h4 (1  2 h  6 h 4 )
2 2 1/ 2
18 3

for  3  3 :
h3 1
h 4  h 4  27 h 4 ; h 3  ; 
2

1  24 h 4 (1  10 h  42 h 4 )
2 2 1/ 2
3

44 4-Parameter   x  
Lognormal  ln  /     x    numerical numerical
1     x 
V,VI) f (x)   integration integration
 ( x   )(  x )    ,  0
 
 
  x  
 ln  /  
  x 
F (x)   
  
 
 

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-8
 APPENDIX B: Table 1: Standard Continuous Distributions

Notes:


I. Complete Gamma function:   a    t

a 1
e dt
t

II. Incomplete Gamma function:   x , a    t a 1 e  t d t ;

0

 a   b 

b 1
III. Complete Beta function: B  a , b    t a 1  1  t  dt 
0
 a  b

x
b 1
IV. Incomplete Beta function: B  x , a , b    t a 1  1  t  dt
0

1  x2 
V. Standard Gaussian density:  ( x )  ex p   
2  2 
x

VI. Standard Gaussian distribution function:  ( x )    u  du



VII. Error function: erf  x   2   x 2   1

 

VIII. Euler number: C=0.577216

IX. Can be expressed in terms of binomial coefficients or by numerical integration

(d  a )  ( c  a )   ( d  a )  ( c  a )   ( b  a )
3
2 2
( d  a )  ( d  a )( c  a )  ( c  a )  ( b  a )
2 2 2

E X   a V ar  X    (E  X   a)
2

3 (c  d  a  b) 6 (c  d  a  b)
X.

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-9
 APPENDIX B: Table 2: Continuous Distribution Functions for Bayesian Analysis

2. Bayesian Distribution Functions

Type no. Variable unknown Priori/posteriori Posterior parameters Predictive density/distribution

(predictive (density/distribution) Parameters density/distribution
posterior) of parameters
Exponential I,II,III,IV) Gamma Invers-beta
33 (pred.)  x     exp    x    
h 1  
h " "
h n h"
fX e
f     "  x fY  y 
 h 
i h " 1
i !  y   "
39 (post.)
   , h  B  y /  y   "  , 1, h " 
F     h "  h  n
FX  x    1  exp    x   h 
FY  y 
B  1, h " 

Gamma I,II,III,IV) Gamma Invers-beta

not implem.  x
k k 1
e
x
  
g g 1
e
 n
"
g
y
k 1

fX x ,k   f     "   x fY  y 

 k   g
i g "k
i ! B  k , g "   " y 
not implem.
   , g  B  y /  y   " , k , g "
F     g "  g  n k  y 
FX x  , k      x , k  /   k  1)  g
FY
B k , g "

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-10
 APPENDIX B: Table 2: Continuous Distribution Functions for Bayesian Analysis

Type no. Variable un- Priori/posteriori Posterior Predictive density/distribution

(predictive (density/distribution) known density/distribution parameters
posterior) Param. of parameters
Normal V,VI) Normal Normal
34 (pred.)  X  x m,  m
"
n    n x Y  y  
1  1m    2
n  n  2

40 (post.) 1  1  x  m 2   M m   exp      n"  
exp     
  2   2     n " 1 1  y   " 
2   2     1 ex p    

" 2  "  2
 " n " 1 

n  n  
  n"  
n "  n  n
x

x     t m ,  dt
m
X m , X
M  m  ,      t  ,  dt
 y  ",  ", n " 

 Y

Normal I,II,III,V,VI) Invers-Gamma-2 Student

35 (pred.)  X  x m,   
s "  "  s    s 
2 2 2
fY  t  
 2  1
  1  " 1
 s2    s2 
 1  x  m 2   x  m
2 
2 2 s 
2 2
41 (post.) 1 f     ex p   1 1 1  t 
2 2
exp     2  2  
i
    s 1 
 2        2   
i
2    s" 1 1  "   "
B ,  "
2 2 2 
x    s 2  
X x m ,     tX
m ,  dt    ,
 2   2

 y
ym

F    1
  FY y s ",  "    fY  t  dt ;t
  / 2  
s"
 "   

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-11
 APPENDIX B: Table 2: Continuous Distribution Functions for Bayesian Analysis

Type no. Variable un- Priori/posteriori Posterior parameters Predictive density/distribution

(predictive (density/distribution) known density/distribution
posterior) Param. of parameters
Normal I,II,III,V,VI) Normal-Invers-Gamma-2 Student
36 (pred.)  X  x m,  m, n  nm  x   2 n "  n  n fY  t  
f M ,  m ,   exp      n " x "  nx   nx  " 1
 1  x  m 2  2   2    


42 (post.) t 
2
1 1 1 1 2
exp      1 1 
 
2   2      2

x 
n
x i s" 1 1
B , 

"
"   "
 1 i
  2 2 
2 2  s 2
 2  s  2
 " s "    s   n  x  
2 2 2
x   exp   2 
  s   2 
2

  
  s  nx   n " x " y x"
2 2 2
n"
2 t 
1 s" n " 1
  xi  x
x 2
s 
2

X x m ,     tX
m ,  dt FM , ( m ,  )  F ( ) FM  
(m  ) n 1 i s"  s"
n " 1

v "  v     n 
 n"
   s 2   VII)
   , 
 2   2  v   n    n "
 
F    1
 ( / 2 )
y
 m  x  FY y x ", n ", s ", v "    fY  t  dt
FM   m     
 / n  

 

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-12
 APPENDIX B: Table 2: Continuous Distribution Functions for Bayesian Analysis

Type no. Variable un- Priori/posteriori Posterior Predictive density/distribution

(predictive (density/distribution) known density/distribution parameters
posterior) Param. of parameters
I,VIII)
Gumbel (max)  n " p   exp    y 
fU  u   t 
30 (pred.) fX x u,  u fY  y  ( n " 1 p )
n t"  exp    y  
exp   n  p   u  t exp  u    t
n p

37 (post.) 
 exp    x  u )  exp     x  u     exp    x  i 

1
t" 

   n  p 
i 1

u  n " p
 exp    y  
FX  x u ,    exp   exp    x  u    FU  u    fU  z  dz
n "  n  n FY  y   1  
 t"
t "  t  t  
I,II,IX)
Frechet (max)
1 q n
fY  y 
n k
32 (pred.)
 x 
k 1
k v  v  k
v
nk  q
exp   v t  kt
k k t   1 q 
fX x v, k     exp      v fV  v   i 1
i
 1 q 
k
y 
 n "
 k
1 
 k 1
43 (post.)
v  x   x     n  1  q  / k  n "  n  n  n "   1   k 1
 k  t"  t" y
t "  t  t
 1 q 
 n "
  v k    t v , n  1  p  / k 
k 
 y 
k
 k 
F X  x v , k   ex p      FV  v   FY  y   1  
  x     n  1  p  / k   t" 
I,II,X)
Weibull (min)
m 1 n
fY  y 
n
31 (pred.)
 x 
k 1
  x   k k
k  x  w
 nk  m
exp   w
k
t  kt k t   m 1 
fX x w, k     exp      w fW w  i 1
i
 m 1 y 
k
 n "
 k
1 
 k
  n   m  1 / k 
ww k 1
38 (post.)   w   n "  n  n  n "   1   y
 k  t"  t"
  x k   t w
k
, n   m  1 / k  t "  t  t
F X  x w , k   1  exp      FW w  1 
  w     n   m  1 / k  
 n "
m 1 

 y 
k
 k 
FY  y   1  1  
 t" 

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-13
 APPENDIX B: Table 2: Continuous Distribution Functions for Bayesian Analysis

Notes:


I. Complete Gamma function:   a    t

a 1 t
e dt
0

II. Incomplete Gamma function:   x , a    t a 1 e  t d t ;

0

 a   b 

b 1
III. Complete Beta function: B  a , b    t a 1  1  t  dt 
0
 a  b

x
b 1
IV. Incomplete Beta function: B  x , a , b    t a 1  1  t  dt
0

1  x2 
V. Standard Gaussian density  ( x )  ex p   
2  2 
x

VI. Standard Gaussian distribution function:  ( x )    u  du



VII.   x    1 for x  0
0 for x  0

VIII. p  0.46

IX. q  0.55 k for k  1.0

X. m  1 + 0.38 k for k  1.0

SOFiSTiK 2014 Appendix B (taken from Strurel – Users Manual, Gollwitzer, 2014) Page B-14
RELY | Appendix

6-18 SOFiSTiK 2020

 Literature | RELY

Literature

[1] Daniel Straub. Lecture Notes in Structural Reliability. Tech. rep. TU München, 2012.
[2] Ove Ditlevsen and Henrik O Madsen. Structural reliability methods. Vol. 178. Wiley New
York, 1996.
[3] Robert E Melchers. Structural reliability analysis and prediction. John Wiley New York,
1999.
[4] Iason Papaioannou. Non-intrusive Finite Element Reliability Analysis Methods. TU
München, 2012.
[5] Michael Hohenbichler and Rudiger Rackwitz. “Non-normal dependent vectors in struc-
tural safety”. In: Journal of the Engineering Mechanics Division 107.6 (1981), pp. 1227–
1238.
[6] Pei-Ling Liu and Armen Der Kiureghian. “Multivariate distribution models with pre-
scribed marginals and covariances”. In: Probabilistic Engineering Mechanics 1.2 (1986),
pp. 105–112.
[7] Claudia Klüppelberg, Daniel Straub, and Isabell M. Welpe. Risk - A Multidisciplinary In-
troduction. Springer, 2014.
[8] Stephan Gollwitzer. Strurel – Users Manual. Tech. rep. RCP Consult GmbH, 2014.
[9] Michael Hohenbichler and Ruediger Rackwitz. “Improvement of second-order reliability
estimates by importance sampling”. In: Journal of Engineering Mechanics 114.12 (1988),
pp. 2195–2199.
[10] Siu-Kui Au and James L Beck. “Estimation of small failure probabilities in high dimensions
by subset simulation”. In: Probabilistic Engineering Mechanics 16.4 (2001), pp. 263–277.
[11] Iason Papaioannou et al. “MCMC algorithms for subset simulation”. In: Manuscript, En-
gineering Risk Analysis Group, Technische Universität München (2014).
[12] A. Hasofer and N. Lind. “An Exact and Invariant First-order Reliability Format”. In: Journal
of the Engineering Mechanics Division, ASCE 100.EM1 (1974), pp. 111–121.
[13] Rüdiger Rackwitz and Bernd Fiessler. “Structural reliability under combined random load
sequences”. In: Computers & Structures 9.5 (1978), pp. 489–494.
[14] Armen Der Kiureghian and Mario De Stefano. “Efficient algorithm for second-order relia-
bility analysis”. In: Journal of engineering mechanics 117.12 (1991), pp. 2904–2923.
[15] T Abdo and R Rackwitz. “Reliability of uncertain structural systems”. In: Proc. Finite
Elements in Engineering Applications (1990), pp. 161–176.

SOFiSTiK 2020 6-19