UQL AB U SER M ANUAL

A CTIVE L EARNING R ELIABILITY

M. Moustapha, S. Marelli, B. Sudret

C HAIR OF R ISK , S AFETY AND U NCERTAINTY Q UANTIFICATION Risk, Safety &

S TEFANO -F RANSCINI -P LATZ 5 Uncertainty Quantification

CH-8093 Z ÜRICH
How to cite UQL AB

S. Marelli, and B. Sudret, UQLab: A framework for uncertainty quantification in Matlab, Proc. 2nd Int. Conf. on
Vulnerability, Risk Analysis and Management (ICVRAM2014), Liverpool, United Kingdom, 2014, 2554-2563.

How to cite this manual

M. Moustapha, S. Marelli, B. Sudret, UQLab user manual – Active learning reliability, Report UQLab-V2.0-117,
Chair of Risk, Safety and Uncertainty Quantification, ETH Zurich, Switzerland, 2022

BIBTEX entry

@TechReport{UQdoc_20_117,
author = {Moustapha, M. and Marelli, S. and Sudret, B.},
title = {{UQLab user manual -- Active learning reliability}},
institution = {Chair of Risk, Safety and Uncertainty Quantification, ETH Zurich,
Switzerland},
year = {2022},
note = {Report UQLab-V2.0-117}
}
Document Data Sheet

Document Ref. UQL AB-V2.0-117

Title: UQL AB user manual – Active learning reliability

Authors: M. Moustapha, S. Marelli, B. Sudret

Chair of Risk, Safety and Uncertainty Quantification, ETH Zurich,
Switzerland
Date: 01/02/2022

Doc. Version Date Comments

V2.0 01/02/2022 UQL AB V2.0 release
V1.4 01/02/2021 Initial release
 Abstract

Structural reliability methods aim at the assessment of the probability of failure of com-
plex systems due to uncertainties associated to their design, manifacturing, environmental
and operating conditions. The name structural reliability comes from the emergence of such
computational methods back in the mid 70’s to evaluate the reliability of civil engineering
structures. As these probabilities are usually small (e.g. 10−2 − 10−8 ), this type of problems
is also known as rare events estimation in the recent statistics literature. A variety of meth-
ods have been developed to estimate the reliability of structures. However they are often
time-consuming as they require repeated evaluations of the computational model describing
the behavior of the system. Alternative and more affordable methods have been developed
thanks to the introduction of surrogate models as cheap proxies of the computational model.
This manual presents the UQL AB framework for active learning, a process by which the
surrogate models are sequentially enriched so as to accurately estimate the failure probability
with the smallest possible computational cost. The framework is made of four different
components (namely surrogate model, estimation method, learning function and stopping
criterion) which can be combined to build custom active learning schemes.
The active learning reliability user manual is divided in three parts:

• A short introduction to the main concepts and techniques used to solve structural re-
liability problems using active learning, with a selection of references to the relevant
literature

• A detailed example-based guide, with the explanation of most of the available options
and methods

• A comprehensive reference list detailing all the available functionalities in the UQL AB
active learning reliability module.

Keywords: Active Learning, Structural Reliability, Surrogate models, Importance Sampling,

Monte Carlo Simulation, Subset Simulation, AK-MCS, UQL AB, rare event estimation.
Contents

1 Theory 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Limit-state function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.2 Failure Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Strategies for the estimation of Pf . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Surrogate-model-based strategies . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4.1 Global framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4.2 Surrogate modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4.3 Reliability analysis algorithm . . . . . . . . . . . . . . . . . . . . . . . 6
1.4.4 Learning function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4.5 Convergence criterion . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Multiple-point enrichment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2 Usage 11
2.1 Reference problem: R-S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Problem set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Basic active learning reliability . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Accessing the results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5 Advanced usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.5.1 Individually configure each component . . . . . . . . . . . . . . . . . . 14
2.5.2 Other general options . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5.3 Asynchronous learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.6 Advanced limit-state function options . . . . . . . . . . . . . . . . . . . . . . . 18
2.6.1 Specify failure threshold and failure criterion . . . . . . . . . . . . . . 18
2.6.2 Vector Outputs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.7 Excluding parameters from the analysis . . . . . . . . . . . . . . . . . . . . . 19

3 Reference List 21
3.1 Create a reliability analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 Accessing the results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 Printing/Visualizing of the results . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.3.1 Printing the results: uq print . . . . . . . . . . . . . . . . . . . . . . . 32
3.3.2 Graphically display the results: uq display . . . . . . . . . . . . . . . . 33
Chapter 1

Theory

1.1 Introduction

An engineering system is defined as a system required to provide specific functionality under

well-defined safety constraints. Such constraints need to be taken into account during its
design phase in view of the expected environmental/operating loads it will be subject to.
In the presence of uncertainties in the physical properties of the system (e.g. due to tolerances
in the manufacturing), in the environmental loads (e.g. due to weather conditions), or in the
operating conditions (e.g. traffic), it can occur that the system operates outside of its nominal
range. In such cases, it can encounter failure.
Reliability analysis deals with the quantitative assessment of the probability of occurrence of
such failures (a.k.a. probability of failure), given a probabilistic model of the uncertainty in
the structural, environmental and load parameters.
The UQL AB reliability module offers a wide range of state-of-the-art techniques for the so-
lution of structural reliability problems, summarized in detail in the UQL AB User Manual –
Structural Reliability. Consistently with the overall design philosophy of UQL AB, all these
algorithms follow a black-box approach, i.e. they rely on the point-by-point evaluation of a
computational model, without knowledge about its inner structure. In some cases, e.g. for
Monte-Carlo based methods, the associated computational cost can be extremely high when
the limit-state function is expensive-to-evaluate (see Section 1.2.1 for proper definitions).
This document focuses on a specific class of methods, namely active learning-based reliability
methods, whose aim is to reduce the overall computational costs of the analysis through the
use of surrogate models. This is achieved by iteratively building locally accurate surrogates
of the limit state function, which will then serve as a tool to explore the random variable
space.
In this chapter, we first briefly recall the formalism introduced in UQL AB User Manual –
Structural Reliability, based on Sudret (2007). We then introduce a modular framework for
active learning reliability, that allows one to easily reproduce a wide array of methods that
have recently been proposed in the reliability analysis literature. This framework is made of
independent blocks, which are themselves based on existing UQL AB modules. The latter are
interconnected non-intrusively to build custom solution strategies.

1
UQL AB user manual

Figure 1: Schematic representation of the safe and failure domains Ds and Df and the cor-
responding limit-state surface g(x) = 0.

1.2 Problem statement

1.2.1 Limit-state function

A limit state can be defined as a state beyond which a system no longer satisfies some perfor-
mance measure (ISO Norm 2394). Regardless of the choice of the specific criterion, a state
beyond the limit state is classified as a failure of the system.
Consider a system whose state is represented by a random vector of variables X ∈ DX ⊂ RM .
One can define two domains Ds , Df ⊂ DX that correspond to the safe and failure regions of
the state space DX , respectively. In other words, the system is failing if the current state
x ∈ Df and it is operating safely if x ∈ Ds . This classification makes it possible to construct
a limit-state function g(X) (sometimes also referred to as performance function) that assumes
positive values in the safe domain and negative values in the failure domain:

x ∈ Ds ⇐⇒ g(x) > 0
(1.1)
x ∈ Df ⇐⇒ g(x) ≤ 0

The hypersurface in M dimensions defined by g(x) = 0 is known as the limit-state surface,

and it represents the boundary between safe and failure domains. A graphical representation
of Ds , Df and the corresponding limit-state surface g(x) = 0 is given in Figure 1.

1.2.2 Failure Probability

If the random vector of state variables X is described by a joint probability density function
(PDF) X ∼ fX (x), then one can define the failure probability Pf as:

Pf = P (g(X) ≤ 0) . (1.2)

UQL AB-V2.0-117 -2-

 Active learning reliability

This is the probability that the system is in a failed state given the uncertainties of the state
parameters. The failure probability Pf is then calculated as follows:
Z Z
Pf = fX (x)dx = fX (x)dx. (1.3)
Df {x: g(x)≤0}

Note that the integration domain in Eq. (1.3) is only implicitly defined by Eq. (1.1), hence
making its direct estimation practically impossible in the general case. This limitation can be
circumvented by introducing the indicator function of the failure domain, a simple classifier
given by: (
1 if g(x) ≤ 0
1Df (x) = , x ∈ DX .
0 if g(x) > 0
In other words, 1Df (x) = 1 when the input parameters x cause the system to fail and
1Df (x) = 0 otherwise. This function allows one to cast Eq. (1.3) as follows:
Z
 
Pf = 1Df (x)fX (x)dx = E 1Df (X) , (1.4)
DX

where E [·] is the expectation operator with respect to the PDF fX (x). This reduces the
calculation of Pf to the estimation of the expectation value of 1Df (X).

1.3 Strategies for the estimation of Pf

From the definition of 1Df (x) in Section 1.2.2 it is clear that determining whether a certain
state vector x ∈ DX belongs to Ds or Df requires the evaluation of the limit-state function
g(x). In the general case this operation can be computationally expensive, e.g. when it
entails the evaluation of a computational model on the vector x. For a detailed overview
of standard structural reliability methods and applications, see e.g. Ditlevsen and Madsen
(1996); Melchers (1999); Lemaire (2009).
In the following, three strategies are discussed for the evaluation of Pf , namely approxima-
tion, simulation and adaptive surrogate-modelling-based methods.

Approximation methods
Approximation methods are based on approximating the limit-state function locally at a
reference point (e.g. with a linear or quadratic Taylor expansion). This class of methods
can be very efficient (in that only a relatively small number of model evaluations is needed
to calculate Pf ), but it tends to become unreliable in the presence of complex, non-linear
limit-state functions. Two approximation methods are currently available in UQL AB:

• FORM (First Order Reliability Method) – it is based on the combination of an iterative

gradient-based search of the so-called design point and a local linear approximation of
the limit-state function in a suitably transformed probabilistic space.

• SORM (Second Order Reliability Method) – it is a second-order refinement of the solution

UQL AB-V2.0-117 -3-

UQL AB user manual

of FORM. The computational costs associated to this refinement increase rapidly with
the number of input random variables M .

Simulation methods
Simulation methods are based on sampling the joint distribution of the state variables in
X and using sample-based estimates of the integral in Eq. (1.4). At the cost of being com-
putationally very expensive, they generally have a well-characterized convergence behaviour
that can be exploited to calculate confidence bounds on the resulting Pf estimates. Three
sampling-based algorithms are available in UQL AB:

• Monte Carlo simulation – it is based on the direct sample-based estimation of the expec-
tation value in Eq. (1.4). The total costs increase very rapidly with decreasing values
of the probability Pf to be computed.

• Importance Sampling – it is based on improving the efficiency of Monte Carlo simulation

by changing the sampling density so as to favour points in the failure domain Df . The
choice of the importance sampling (a.k.a. instrumental) density generally uses FORM
results.

• Subset Simulation – it is based on iteratively solving and combining a sequence of condi-

tional reliability analyses by means of Markov Chain Monte Carlo (MCMC) simulation.

Surrogate model-based adaptive methods

Surrogate model-based adaptive methods rely on iteratively building models that approxi-
mate the limit-state function in the direct vicinity of the limit-state surface. These surrogates
are adaptively refined by adding limit-state function evaluations to their experimental de-
signs until a suitable convergence criterion related to the accuracy of Pf is satisfied. This
manual specializes in such approaches. They are thus described in details in the remainder
of this chapter.

Note: This user manual focuses only on surrogate model-based adaptive methods. For
more details on the other two classes of methods, please refer to the UQL AB User
Manual – Structural Reliability

1.4 Surrogate-model-based strategies

1.4.1 Global framework

This module is based on the global framework for active learning reliability presented in
Moustapha et al. (2021). The framework aims at building a variety of active learning schemes
by non-intrusively combining methods from four different core components. These compo-
nents, illustrated by a few examples, are presented in Figure 2. The first two modules, i.e.

UQL AB-V2.0-117 -4-

 Active learning reliability

Surrogate model Reliability method Learning function Stopping criterion

Kriging Monte Carlo U LF-based

PC-Kriging Subset simulation EFF Stability of β
PCE Importance sampling FBR Stability of Pf
SVR ... CMM Bounds on β
LRA ... Bounds on Pf
... ...

Figure 2: Active learning reliability framework with methods available in UQL AB.

surrogate modelling and reliability method capitalize on readily available tools within UQL AB.
They will be briefly described in the next section and the user is referred to the corresponding
manual for a more detailed description. The other two, i.e. learning function and conver-
gence criteria will be detailed in this manual instead.
Using these four components, an algorithm for active learning reliability can be devised and
summarized as follows:

1. I NITIALIZATION : Generate a small initial experimental design X = x(1) , . . . , x(N0 )



and evaluate the corresponding limit-state function responses Y = y (1) , . . . , y (N0 ) =



g(x(1) ), . . . , g(x(N0 ) ) ;


2. S URROGATE M ODEL : Train a surrogate model gb (column 1 of Figure 2) on the current

experimental design {X , Y};

3. R ELIABILITY METHOD : Using the chosen reliability method (column 2 of Figure 2),
estimate a failure probability Pbf with the current surrogate model;

4. L EARNING FUNCTION : based on an appropriate learning function (column 3 of Fig-

ure 2), choose the best next sample(s) xnext to be added to the current experimental
design X ;

5. S TOPPING CRITERION : check whether a chosen stopping criterion (column 4 of Fig-

ure 2) is met. If it is, skip to Step 7, otherwise continue with Step 6;

6. E NRICHMENT: add xnext and the corresponding limit-state function response(s) y next =
g(xnext ) to the experimental design of the surrogate model. Return to Step 2;

7. C ONCLUSION : Return the estimated failure probability Pbf .

1.4.2 Surrogate modelling

Surrogate models are used in active learning reliability as a tool to explore the random space
in a much affordable cost. They are built adaptively in an attempt to finely approximate the
limit-state surface in areas of high probability mass.

UQL AB-V2.0-117 -5-

UQL AB user manual

There are various surrogate modelling approaches currently offered by UQL AB. The follow-
ing will be used within the active learning reliability framework:

• Gaussian process a.k.a. Kriging (UQL AB User Manual – Kriging (Gaussian process mod-
elling));

• Polynomial chaos expansions (UQL AB User Manual – Polynomial Chaos Expansions);

• Polynomial chaos-Kriging (UQL AB User Manual – PC-Kriging);

• Canonical low-rank tensor approximations (UQL AB User Manual – Canonical low-rank

approximations);

• Support vector machines for regression (UQL AB User Manual – Support vector ma-
chines regression).

The user is referred to the respective manuals for more details about each of the surrogate
models. Throughout this document, they will be treated as black-boxes, i.e. we will only
be interested in their input-output structures. More specifically, for a given random variable
x, we denote by gb (x) the corresponding surrogate model prediction. Note that in some
cases, such as Kriging and polynomial chaos-Kriging, a second output, namely the prediction
variance, is also available. This can be quite useful in active learning and will be denoted by
sb2 (x).

1.4.3 Reliability analysis algorithm

Reliability algorithms lie at the core of active learning methods, as they are primarily used
to estimate the failure probability. As explained in previous sections, simulation methods
are preferred to approximation ones as they often lead to more accurate results. This comes
however at the expense of generally increased computational costs. When used in conjunc-
tion with surrogate-based methods, however, they can be effectively exploited, due to the
extreme efficiency of surrogate model predictors, which are typically inexpensive to evaluate
(often O(105−6 s−1 )).
This module is also considered black-box and directly calls the following methods from the
UQL AB reliability module:

• Monte Carlo simulation, UQL AB User Manual – Structural Reliability, Section 1.5.1

• Importance sampling, UQL AB User Manual – Structural Reliability, Section 1.5.2

• Subset simulation, UQL AB User Manual – Structural Reliability, Section 1.5.3

1.4.4 Learning function

A learning function is used to identify the points that, once added to the experimental de-
sign, would best reduce the surrogate-induced error in the estimated failure probability. The
candidate point is normally chosen from a suitably defined candidate set S, often a large

UQL AB-V2.0-117 -6-

 Active learning reliability

Monte-Carlo sampling of the input space, or, in this framework, the full set of limit state eval-
uations performed by the reliability algorithm of choice at the current algorithm iteration. A
wide variety of learning functions have been proposed in the recent literature.
In the current version, UQL AB offers the following four:

Deviation number
Also known as U learning function, the deviation number was developed by Echard et al.
(2011) based on the concept of misclassification probability, and reads:

|b
g (x)|
U (x) = . (1.5)
sb(x)

This quantity is directly related to the probability of misclassification in case of Gaussian

distributed variables (as is the case in Kriging-based predictors) as:

Pm (x) = Φ (−U (x)) , (1.6)

where Φ is the standard Gaussian CDF.

The next sample candidate is chosen as the one from the candidate pool S that maximizes
the probability of misclassficiation or equivalently minimizes U :

xnext = arg min U (s) = arg max Pm (s) . (1.7)

s∈S s∈S

Efficient feasibility function

Another popular learning function was introduced by Bichon et al. (2008) and reads:
      
−b g (x) − − gb(x)  − gb(x)
EF F (x) = gb(x) 2Φ −Φ −Φ
sb(x) sb(x) sb(x)
      
−b g (x) − − gb(x)  − gb(x)
− sb(x) 2ϕ −ϕ −ϕ
sb(x) sb(x) sb(x)
    
 − gb(x) − − gb(x)
+ Φ −Φ , (1.8)
sb(x) sb(x)

where  = 2b
s(x) and ϕ is the PDF value of a standard normal Gaussian variable. The EFF
function provides an indication of how well the true model response is expected to satisfy
g(x) = 0 and provides a good exploration/exploitation balance for enrichment. It takes large
values when a point is close to the limit-state threshold, i.e. gb(x) ≈ 0 and/or when there
is large uncertainty in the prediction, i.e. sb(x) is large. The next candidate sample is then
chosen as follows:
xnext = arg max EF F (s) . (1.9)
s∈S

Fraction of bootstrap replicates

This learning function, introduced in Marelli and Sudret (2018), uses B bootstrap repli-
cates of the experimental design to build multiple surrogates and then eventually assess their

UQL AB-V2.0-117 -7-

UQL AB user manual

consistency in predicting the failed/safe state of a candidate sample. Because it is a direct

measure of the misclassification probability of the current surrogate, it is closely related to
the deviation number introduced earlier, but it does not require the predictions to be normally
distributed.
It is defined as:
|Bs (x) − Bf (x)|
UFBR (x) = , (1.10)
B
where Bs (x) and Bf (x) ∈ [0, . . . , B] are respectively the number of safe and failed boot-
strap replicate predictions at the point x. Samples whose prediction is uncertain tend to have
close values of Bs and Bf , hence the next sample is chosen as:

xnext = arg min UFBR (s) . (1.11)

s∈S

Constrained min-max
This learning function was introduced in Moustapha and Sudret (2019) as an adaptation
of Basudhar and Missoum (2008). It aims at finding in the vicinity of the limit-state approx-
imation samples that are the furthest from existing training points. The next sample to add
then reads:
xnext = max0 min ks − x(i) k, (1.12)
s∈S i=1,...,N

where S 0 = {s ∈ S : |b
g (s)| ≤ q} with q being an γ-quantile of |b
g (s)|. The default value
γ = 0.01 is used in UQL AB.

1.4.5 Convergence criterion

Various convergence criteria can be used to terminate the enrichment process. The stopping
criteria implemented in UQL AB can be split into three categories:

Variance-based criteria
These criteria relate to the surrogate-induced uncertainty in the estimates of the failure
probability or in the associated reliability index β.
b This uncertainty is estimated through the
built-in error measure provided by some surrogates. As such, they apply only to Kriging and
PC-Kriging.
The first criterion is based on bounds of the estimated failure probability and is computed as
follows:
Pbf+ − Pbf−
≤ bound
Pbf
, (1.13)
Pb 0
f

where the three failure probabilities are defined as:

Pbf0 = P (b
g (x) ≤ 0) , (1.14)

Pbf± = P (b
g (x) ∓ kb
s(x) ≤ 0) , (1.15)

where k = Φ−1 (1 − α/2) sets the confidence level (1 − α), typically k = Φ−1 (97.5%) = 1.96.

UQL AB-V2.0-117 -8-

 Active learning reliability

The second criterion is similarly defined, however using the reliability index instead:

βb+ − βb−
≤ bound
βb
, (1.16)
βb0

where the three reliability indices correspond to the aforementioned failure probabilities:

βb0 = −Φ−1 (Pbf0 ), (1.17)

βb± = −Φ−1 (Pbf∓ ). (1.18)

Stability based criteria

The second family of convergence criteria applies to all surrogate types and relates to the
stability of the failure probability or reliability index estimates. They respectively read:
(j) (j−1)
Pbf − Pbf
(j)
≤ stab
Pb
(1.19)
Pbf f

and
βb(j) − βb(j−1)
≤ stab
βb
(1.20)
βb(j)
(j)
where Pbf and βb(j) respectively represent the estimated failure probability and reliability
index at the j-th iteration.

Learning function-based criteria

The final category of convergence criteria are directly based on the learning functions.
For the deviation number, the criterion reads:

min U (s) > U . (1.21)

s∈S

This is the original AK-MCS stopping criterion and basically means that the probability of
misclassifying any point in the candidate pool is smaller than Φ (−U ).
In a similar fashion, the original EFF stopping criterion is implemented and reads:

max EF F (s) < EFF . (1.22)

s∈S

For the fraction of bootstrap replicates, the stopping criterion reads:

min UFBR (s) < FBR . (1.23)

s∈S

Finally, for the min-max criterion, the stopping criterion refers to the minimal distance to the
other experimental design points

max0 min ks − x(i) k ≤ 10−2 · d0 (1.24)

s∈S i=1,...,N

UQL AB-V2.0-117 -9-

UQL AB user manual

where d0 is the minimum distance between the experimental design points at the first itera-
tion.

1.5 Multiple-point enrichment

It is possible to add K > 1 samples within a single enrichment iteration. This is achieved
in UQL AB using weighed K-means clustering (Zaki and Meira, 2014). First, a subset of the
candidate pool S is defined according to the learning function:

• For U and EF F , the subset is defined as the following margin of uncertainty:

S 00 = {x ∈ S : −2 sb (x) ≤ gb (x) ≤ 2 sb (x)} (1.25)

• For the fraction of bootstrap replicates, the subset is defined as:

S 00 = {x ∈ S : UFBR (x) < 0.5} (1.26)

The set S 00 is then reduced into K cluster centers obtained by weighted K-means clustering
using weights defined according to the learning function. They read as follows:

• Deviation number
w (x) = ϕ (−U (x)) ; (1.27)

• Expected feasibility
w (x) = ϕ (EF F (x)) ; (1.28)

• Fraction of bootstrap replicates

w (x) = ϕ (−UFBR (x)) ; (1.29)

where ϕ is the standard Gaussian PDF.

For the constrained min-max criterion, the enrichment is made sequentially one sample at a
time, with the min-max distance updated after the addition of each sample.

UQL AB-V2.0-117 - 10 -
Chapter 2

Usage

In this section, a reference problem will be set up to showcase how each of the techniques in
Chapter 1 can be deployed in UQL AB.

2.1 Reference problem: R-S

The benchmark of choice to showcase the methods described in Section 1.3 is a basic problem
in structural reliability, namely the R-S case. It is one of the simplest possible abstract setting
consisting of only two input state variables: a resistance R and a stress S. The system fails
when the stress is higher than the resistance, leading to the following limit-state function:

X = {R, S} g(X) = R − S; (2.1)

The two-dimensional probabilistic input model consists of independent variables distributed

according to Table 1.

Table 1: Distributions of the input parameters of the R − S model in Eq. (2.1).

Name Distributions Parameters Description

R Gaussian [5, 0.8] Resistance of the system
S Gaussian [2, 0.6] Stress applied to the system

2.2 Problem set-up

Solving a structural reliability problem in UQL AB requires the definition of three basic com-
ponents:

• a MODEL object that describes the limit-state function

• an INPUT object that describes the probabilistic model of the random vector X

• a reliability ANALYSIS object.

The UQL AB framework is first initialized with the following command:

11
UQL AB user manual

uqlab

The model in Eq. (2.1) can be added as a MODEL object directly with a M ATLAB vectorized
string as follows:
MOpts.mString = 'X(:,1) - X(:,2)'; % R−S
MOpts.isVectorized = 1;
myModel = uq_createModel(MOpts);

For more details on the available options to create a model object in UQL AB, please refer to
the UQL AB User Manual – the MODEL module.
Correspondingly, an INPUT object with independent Gaussian variables as specified in Table 1
can be created as:
IOpts.Marginals(1).Name = 'R'; % Resistance
IOpts.Marginals(1).Type = 'Gaussian';
IOpts.Marginals(1).Moments = [5, 0.8];
IOpts.Marginals(2).Name = 'S'; % Stress
IOpts.Marginals(2).Type = 'Gaussian';
IOpts.Marginals(2).Moments = [2, 0.6];

myInput = uq_createInput(IOpts);

For more details about the configuration options available for an INPUT object, please refer
to the UQL AB User Manual – the INPUT module.

2.3 Basic active learning reliability

Running an active learning reliability analysis on the specified UQL AB MODEL and INPUT ob-
jects does not require any specific configuration. The following minimum syntax is required:
ALROpts.Type = 'Reliability';
ALROpts.Method = 'ALR';
ALRAnalysis = uq_createAnalysis(ALROpts);

Once the analysis is performed, a report with the ALR results can be printed on screen by:
uq_print(ALRAnalysis)

which produces the following:

-------------------------------------------
Active learning reliability
-------------------------------------------
Pf 1.4002e-03
BetaHL 2.9888
CoV 0.0189
ModelEvaluations 11
PfCI [1.3484e-03 1.4120e-03]
BetaCI [3.0003e+00 2.9777e+00]
PfMinus/PfPlus [1.4002e+03 1.4002e-03]
-------------------------------------------

UQL AB-V2.0-117 - 12 -
 Active learning reliability

The results can be visualized graphically as follows:

uq_display(ALRAnalysis)

which produces the images in Figure 3. Note that the graphical representation of the experi-
mental design (right panel of Figure 3) is only produced for the 2-dimensional case.

Figure 3: Graphical visualization of the results of the basic ALR analysis Section 1.4.

Note: In the preceding example, no specifications are provided. If not further specified,
the ALR runs the following defaults:

– Surrogate model: PC-Kriging;

– Reliability algorithm: Subset simulation;
– Learning function : Deviation number (U -function);
– Stopping criterion: Bounds on the β estimate.

These defaults were chosen following an extensive benchmark carried out in

Moustapha et al. (2021).

2.4 Accessing the results

The results from the active learning reliability analysis are stored in the ALRAnalysis.Results
structure:
ALRAnalysis.Results
ans =
Pf: 0.0015
CoV: 0.0588
Beta: 2.9771
ModelEvaluations: 11
PfCI: [2x1 double]
BetaCI: [2x1 double]
Metamodel: [1x1 uq_model]

UQL AB-V2.0-117 - 13 -
UQL AB user manual

Reliability: [1x1 uq_analysis]

History: [1x1 struct]

The Results structure contains the following fields: Pf, the estimated Pbf ; Beta, the
corresponding generalized reliability index; CoV, the calculated coefficient of variation;
ModelEvaluations, the total number of limit-state function evaluations; PfCI, the confi-
dence intervals; BetaCI, the corresponding confidence intervals on βHL ; History, a struc-
ture containing the convergence of Pf , CoV , the stopping criteria values, upper and lower
bounds on the failure probability (when this applies). The content of the History structure
is used to produce the convergence plot shown in Figure 3. Note that the actual calculation
of each of these results is dependent on the reliability algorithm used. The user should refer
to the corresponding sections in the UQL AB User Manual – Structural Reliability for further
details.

Further, the field Metamodel contains the final metamodel used to estimate the failure prob-
ability. This metamodel can be reused within UQL AB for any other purpose.

2.5 Advanced usage

2.5.1 Individually configure each component

In this section, we will discuss how the options for each module of the framework can be set
independently.

2.5.1.1 Surrogate model

The surrogate model can be specified using the command .ALR.Metamodel. For instance,
the user may specify the following command to use Kriging:
ALROptions.ALR.Metamodel = 'Kriging';

Furthermore, the user can specify the options of the chosen surrogate by using the command
.ALR.(MetamodelName), e.g. for Kriging:
KrgOpts.EstimMethod = 'ML' ;
KrgOpts.Corr.Family = 'Gaussian' ;
KrgOtps.Corr.Nugget = 1e-12;

ALROptions.ALR.Kriging = KrgOpts ;

UQL AB-V2.0-117 - 14 -
 Active learning reliability

Note: When no surrogate model is specified, PC-Kriging is used with the following de-
faults:

– PCE degree: adaptively calibrated in the range 1 to 3

– Kriging nugget: 10−10
– Kriging estimation method: maximum likelihood

The remaining options are the defaults as defined in the UQL AB User Manual –
PC-Kriging.

All the options available for each surrogate model can be specified here without restrictions,
except for the experimental design, which is handled by the ALR algorithm.

Table 2: List of supported surrogate models and user manuals

Surrogate Model User Manual

Kriging UQL AB User Manual – Kriging (Gaussian process modelling)
PC-Kriging UQL AB User Manual – PC-Kriging
PCE UQL AB User Manual – Polynomial Chaos Expansions
LRA UQL AB User Manual – Canonical low-rank approximations
SVR UQL AB User Manual – Support vector machines regression

2.5.1.2 Reliability algorithm

To specify a reliability algorithm, the user must set the command .ALR.Reliability. For
instance, the following code selects subset simulation:
ALROptions.ALR.Reliability = 'Subset' ;

All the options relative to the reliability algorithm can be set as defined in the UQL AB User
Manual – Structural Reliability. For instance, the user can specify the conditional target
failure probability in subset as follows:
ALROptions.Subset.p0 = 0.1 ;

Similarly, simulation options can be defined as follows:

ALROptions.ALR.Simulation.BatchSize = 1e3 ;
ALROptions.ALR.Simulation.MaxSampleSize = 1e4 ;

UQL AB-V2.0-117 - 15 -
UQL AB user manual

Note: When no surrogate model is specified, subset simulation is used with the follow-
ing options:

– Batch sample size: 105

– Maximum sample size: 2 · 106
– Conditional probability: p0 = 0.15

The remaining options are the defaults as defined in the UQL AB User Manual –
Structural Reliability.

2.5.1.3 Learning function

The learning function may be specified using the command .ALR.LearningFunction. For
instance, the following line of code specifies the deviation number as learning function:
ALROptions.ALR.LearningFunction = 'U'

It is worth noting that some learning functions are available exclusively for a specific type of
surrogate model. Table 3 shows all the possible combinations.

Table 3: Available combinations of surrogate models and learning function in UQL AB.

U EFF FBR CMM

Kriging 3 3 7 3
PC-Kriging 3 3 7 3
PCE 7 7 3 3
LRA 7 7 7 3
SVR 7 7 7 3

Note: When not specified by the user, the default learning function depends on the
surrogate model and is chosen as follows:

– 'U' for PC-Kriging and Kriging;

– 'FBR' for PCE;
– 'CMM' for SVR and LRA.

2.5.1.4 Stopping criterion

The stopping criterion may be specified by the user using the command .ALR.Convergence.
The following line sets the β-bound criterion (See Eq. 1.16) as stopping criterion:
ALROptions.ALR.Convergence = 'StopBetaBounds' ;

To set the corresponding threshold, say bound = 0.05, the following command may be used:
bβ
ALROptions.ALR.ConvThres = 0.05 ;

UQL AB-V2.0-117 - 16 -
 Active learning reliability

2.5.2 Other general options

There are several overarching options that may be needed to configure an ALR analysis.
Some of the most common are summarized below:

• Specify the initial experimental design: Apart from specifying a number of points
and a sampling strategy, the initial experimental design can be specified by providing
X = x(1) , . . . , x(N0 ) in the matrix X and the corresponding limit-state function values


g(x(1) , . . . , x(N0 ) in G:


ALROptions.ALR.IExpDesign.X = X;
ALROptions.ALR.IExpDesign.G = G;

• Specify the maximum number of additional experimental design points: The max-
imum number of samples added to the initial experimental design can be specified to
e.g. 100 as:
ALROptions.ALR.MaxAddedED = 100;

Note that the total number of runs of the limit-state function then is at most the initial
ED size plus the above number.

• Specify multiple enrichment points: An arbitrary K > 1 number of points can be

added simultaneously to the ED at each iteration as described in Section 1.5. For
instance, K = 3 points can be added simultaneously using:
ALROptions.ALR.NumOfPoints = 3;

2.5.3 Asynchronous learning

In some contexts, it is desirable to evaluate the limit state function outside the main UQL AB
or even M ATLAB session. We identify this process as asynchronous learning, because the
learning algorithm and the limit state function evaluations are performed separately, possibly
at different locations and times.
The ALR method supports an asynchronous learning feature, which makes it possible to per-
form active-learning-based reliability analysis also without defining a computational model
within UQL AB. To enable this feature, the following option is needed:
ALROptions.Async.Enable = true ;

When this is option is enabled, instead of evaluating the limit state function, UQL AB will
interrupt the execution of the analysis and return the next sample that the user needs to eval-
uate to continue the algorithm. This sample is provided in the field .Results.NextSample.
The user can then evaluate this sample outside of UQL AB or M ATLAB, and provide the model
evaluations Ynext back to UQL AB with the uq_resumeAnalysis command:
myALRAnalysis = uq_resumeAnalysis(Ynext)

UQL AB-V2.0-117 - 17 -
UQL AB user manual

Note: Beside the next sample that is returned, UQL AB also saves in the field
.Results.Snapshot, a .mat file containing a snapshot of the analysis before
it was interrupted. This can be used to restart to the analysis if the UQL AB ses-
sion / M ATLAB workspace was cleared or if the M ATLAB session was closed. This
can also be used to further investigate the current state of the analysis, e.g. by
graphically displaying the evolution of the failure probability estimates.

2.6 Advanced limit-state function options

2.6.1 Specify failure threshold and failure criterion

While it is normally good practice to define the limit-state function directly as a UQL AB
MODEL object as in Section 2.2, in some cases it can be useful to be able to create one from
small modifications of existing MODEL objects. A typical scenario where this is apparent is
when the same objective function needs to be tested against a set of different failure thresh-
olds, e.g. for a parametric study. In this case, the limit-state specifications can be modified.
As an example, when g(x) ≤ T = 5 defines the failure criterion, one can use the following
syntax:
MCOpts.LimitState.Threshold = 5;
MCOpts.LimitState.CompOp = '<=';

UQL AB offers several possibilities to create simple (or arbitrarily complex) objective functions
from existing MODEL objects (see also UQL AB User Manual – Structural Reliability).
For an overview of the advanced options for the limit-state function, refer to Table 5, page
25.

2.6.2 Vector Outputs

In case the limit-state function g(x) results in a vector rather than a scalar value, the struc-
tural reliability module estimates the failure probability for each component independently.

Note: There is no system-type reasoning implemented to combine the failure probabil-

ities of each component.

However, the implemented methods make use of evaluations of the limit-state function if
available, as follows:

• at the surrogate level: The initial experimental design for the surrogate model of
output component i consists of the final experimental design of component i − 1.

• at the reliability level:

– Monte Carlo simulation: The enrichment of the sample size is increased until the
convergence criteria are fulfilled for all components.

UQL AB-V2.0-117 - 18 -
 Active learning reliability

– Subset Simulation: The first batch of samples (MCS) is reused for every output
component of the limit-state.

2.7 Excluding parameters from the analysis

In various usage scenarios (e.g. parametric studies) one or more input variables may be set
to fixed constant values. This can have important consequences for many of the methods
available in UQL AB whose costs increase significantly with the number of input variables.
Whenever applicable, UQL AB will appropriately account for the set of constant input param-
eters and exclude them from the analysis so as to avoid unnecessary costs. This process is
transparent to the user as the analysis results will still show the excluded variables, but they
will not be included in the calculations.
To set a parameter to constant, the following command can be used when the probabilistic
input is defined (See UQL AB User Manual – the INPUT module):
inputOpts.Marginals.Type = 'Constant' ;
inputOpts.Marginals.Parameters = value;

Furthermore, when the standard deviation of a parameter equals zero, UQL AB treats it as a
Constant. For example, the following uniformly distributed variable whose upper and lower
bounds are identical is automatically set to a constant with value 1:
inputOpts.Marginals.Type = 'Uniform' ;
inputOpts.Marginals.Parameters = [1 1];

UQL AB-V2.0-117 - 19 -
Chapter 3

Reference List

How to read the reference list

Structures play an important role throughout the UQL AB syntax. They offer a natural way
to semantically group configuration options and output quantities. Due to the complexity of
the algorithms implemented, it is not uncommon to employ nested structures to fine-tune
the inputs and outputs. Throughout this reference guide, a table-based description of the
configuration structures is adopted.
The simplest case is given when a field of the structure is a simple value or array of values:

Table X: Input
.Name String A description of the field is put here

which corresponds to the following syntax:

Input.Name = 'My Input';

The columns, from left to right, correspond to the name, the data type and a brief description
of each field. At the beginning of each row a symbol is given to inform as to whether the
corresponding field is mandatory, optional, mutually exclusive, etc. The comprehensive list
of symbols is given in the following table:

Mandatory
 Optional
⊕ Mandatory, mutually exclusive (only one of
the fields can be set)

Optional, mutually exclusive (one of them
can be set, if at least one of the group is set,
otherwise none is necessary)

When one of the fields of a structure is a nested structure, a link to a table that describes the
available options is provided, as in the case of the Options field in the following example:

21
UQL AB user manual

Table X: Input
.Name String Description
 .Options Table Y Description of the Options
structure

Table Y: Input.Options
.Field1 String Description of Field1
 .Field2 Double Description of Field2

In some cases, an option value gives the possibility to define further options related to that
value. The general syntax would be:
Input.Option1 = 'VALUE1' ;
Input.VALUE1.Val1Opt1 = ...;
Input.VALUE1.Val1Opt2 = ...;

This is illustrated as follows:

Table X: Input
.Option1 String Short description
'VALUE1' Description of 'VALUE1'
'VALUE2' Description of 'VALUE2'

.VALUE1 Table Y Options for 'VALUE1'

.VALUE2 Table Z Options for 'VALUE2'

Table Y: Input.VALUE1
 .Val1Opt1 String Description
 .Val1Opt2 Double Description

Table Z: Input.VALUE2
 .Val2Opt1 String Description
 .Val2Opt2 Double Description

Note: In the sequel, double and doubles mean a real number represented in double
precision and a set of such real numbers, respectively.

UQL AB-V2.0-117 - 22 -
 Active learning reliability

3.1 Create a reliability analysis

Syntax

myAnalysis = uq_createAnalysis(ALROpts)

Input

All the parameters required to determine the analysis are to be given as fields of the structure
ALROpts. Each method has its own options, that will be reviewed in different tables. The
options described in Table 4 are common to all methods.

Note: This manual focuses on active learning reliability methods but is generally part
of the UQL AB reliability manual. All options to any reliability method can be
defined in the fields of the structure ALROpts. However in this manual, only the
options directly related to active learning will be described. For the complete
set of options of the methods available in the UQL AB reliability module, the
user is referred to the corresponding manual (UQL AB User Manual – Structural
Reliability, Section 3).

Table 4: ALROpts
.Type 'uq_reliability' Identifier of the module. The options
corresponding to other types are in
the corresponding guides.
.Method String Type of structural reliability method.
The available options are listed
below:
'ALR' Active learning reliability.

'MCS' Monte Carlo simulation (without

active learning).

'FORM' First order reliability method

(without active learning).

'SORM' Second order reliability method

(without active learning).

'IS' Importance sampling (without active

learning).
'Subset' Subset simulation (without active
learning).
'AKMCS' Adaptive Kriging Monte Carlo
Simulation (AK-MCS).

UQL AB-V2.0-117 - 23 -
UQL AB user manual

 .Name String Name of the module. If not set by

the user, a unique string is
automatically assigned to it.

 .Input INPUT object INPUT object used in the analysis. If

not specified, the currently selected
one is used.

 .Model MODEL object MODEL object used in the analysis. If

not specified, the currently selected
one is used.

 .LimitState See Table 5 Specification of the limit-state

function.

 .Display String Level of information displayed by the

default: 'standard' methods.
'quiet' Minimum display level, displays
nothing or very few information.
'standard' Default display level, shows the most
important information.
'verbose' Maximum display level, shows all the
information on runtime, like updates
on iterations, etc.

 .ALR See Table 6 Options field for the active learning

reliability scheme.
 .Simulation See Table 7 Options field for the simulation
methods used within the active
learning scheme.

 .IS See Table 8 Options field for importance

sampling if used within the active
learning scheme.

 .Subset See Table 9 Options field for subset simulation if

used within the active learning
scheme.
 .SaveEvaluations Logical Storage or not of performed
default: true evaluations of the limit-state
function.
true Store the evaluations.
false Do not store the evaluations.

 .Async See Table 12 Options for asynchronous learning.

UQL AB-V2.0-117 - 24 -
 Active learning reliability

In order to perform a structural reliability analysis, the limit-state function g(x) is compared
to a threshold value T (by default T = 0). In analogy with Eq. (1.1), failure is defined as
g(x) ≤ T . Alternatively, failure can be specified as g(x) ≥ T by adjustment of the field
ALROpts.LimitState.CompOp to '>='. The relevant options are summarized in Table 5:

Table 5: ALROpts.LimitState
 .Threshold Double Threshold T , compared to the
default: 0 limit-state function g(x).

 .CompOp String Comparison operator for the

default: '<=' limit-state function.

'<', '<=' Failure is defined by g(x) < T .

'>', '>=' Failure is defined by g(x) > T .

The active learning scheme is built by choosing different methods within various modules of
UQL AB, as well as some overarching options such as the initial experimental design specifi-
cation. In Table 6, the available options are shown:

Table 6: ALROpts.ALR
 .Metamodel String Surrogate model type to use in the
default: 'PCK' active learning scheme (as described
in Section 1.4.2).
'PCK' Polynomial Chaos - Kriging
'Kriging' Kriging
'PCE' Polynomial chaos expansions
'SVR' Support vector machines for
regression
'LRA' Low rank approximation
 .Reliability String Reliability algorithm used in the
default: 'Subset' active learning scheme (as described
in Section 1.4.3).
'MCS' Monte Carlo simulation
'Subset' Subset simulation
'IS' Importance sampling

UQL AB-V2.0-117 - 25 -
UQL AB user manual

 String or cell of strings Learning function used in the active

.LearningFunction learning scheme (as described in
Section 1.4.4). The default depends
on the surrogate defined in
.ALR.Metamodel and is set as:
• 'U' when using 'PCK' or
'Kriging'
• 'FBR' when using 'PCE'
• 'CMM' when using 'SVR' or
'LRA'
'U' Deviation number as defined in
Eq. (1.5)
• Only valid when
.ALR.Metamodel is set to 'PCK'
or 'Kriging'
'EFF' Expected feasibility function as
defined in Eq. (1.8)
• Only valid when
.ALR.Metamodel is set to 'PCK'
or 'Kriging'
'FBR' Fraction of bootstrap replicates as
defined in Eq. (1.10)
• Only valid when
.ALR.Metamodel is set to 'PCE'
'CMM' Constrained min-max as defined in
Eq. (1.12)
 .Convergence String Convergence criterion used in the
default: active learning scheme (as described
'StopBetaBound' in Section 1.4.5)
when .ALR.Metamodel
is 'Kriging' or 'PCK'
and 'StopBetaStab'
otherwise
'StopBetaBound' Convergence criterion based on the
accuracy of the reliability index
estimate w.r.t. the surrogate error
(See Eq. (1.16))
• Only valid when
.ALR.Metamodel is set to 'PCK'
or 'Kriging'
'StopPfBound' Convergence criterion based on the
accuracy of the failure probability
estimate w.r.t. the surrogate error
(See Eq. (1.13))
• Only valid when
.ALR.Metamodel is set to 'PCK'
or 'Kriging'
'StopBetaStab' Convergence criterion based on the
stability of the reliability index
estimate (See Eq. (1.20))

UQL AB-V2.0-117 - 26 -
 Active learning reliability

'StopPfStab' Convergence criterion based on the

stability of the failure probability
estimate (See Eq. (1.19))
'StopLF' Convergence criterion based on the
learning function defined in
.ALR.Learning
• For 'U', see Eq. (1.21)
• For 'EFF', see Eq. (1.22)
• For 'FBR', see Eq. (1.23)
• For 'CMM', see Eq. (1.24)
 .ConvThres Double Threshold for the convergence
criterion defined in
.ALR.Convergence. Default is:
• 2 for U in Eq. (1.21)
• 10−3 for EFF in Eq. (1.22)
• 10−2 for the remaining stopping
criteria
• When multiple convergence criteria
are used, this can be either a scalar
(then the same threshold is used for
all criteria) or a vector of size
1 × Ncrit where Ncrit is the number
of convergence criteria defined.
 .NumOfPoints Integer Number of points to add within an
default: 1 enrichment step (See Section 1.5 for
multiple point enrichment).
 .ConvIter Integer Number of successive iterations the
default: 2 convergence criteria should be
satisfied for the algoritm to stop.
 .MaxAddedED Integer Maximum number of samples to add
default: 1000 to the experimental design of the
metamodel.
 .IExpDesign See Table 11 Specification of the initial
experimental design of the
metamodel.

The available methods to perform structural reliability analysis within an active learning
scheme are Monte Carlo simulation, importance sampling and subset simulation. They all
share the simulation options which are shown in Table 7:

UQL AB-V2.0-117 - 27 -
UQL AB user manual

Table 7: ALROpts.Simulation
 .Alpha Double Confidence level α. For the Monte
default: 0.05 Carlo estimators, a confidence
interval is constructed with
confidence level 1 − α.

 .MaxSampleSize Integer Maximum number of samples to be

default: 2 · 106 for 'SuS'; evaluated. If there is no target
105 for 'IS'; coefficient of variation (CoV), this is
107 for 'MCS' the total number of samples to be
evaluated. If the target CoV is
present, the method will run until
TargetCoV or MaxSampleSize is
reached. In this case, the default
value of MaxSampleSize, if not
specified in the options, is Inf,
i.e.the method will run until the
target CoV is achieved.

 .TargetCoV Double Target coefficient of variation. If

present, the method will run until
the estimate of the CoV (Eq. (1.40))
is below TargetCoV or until
MaxSampleSize function
evaluations are performed. The
value of the coefficient of variation of
the estimator is checked after each
BatchSize evaluations. By default
this option is disabled. Note: this
option has no effect in
Method = 'Subset' and
'AKMCS'.

 .BatchSize Integer Number of samples that will be

default: 106 for 'MCS'; evaluated at once.
105 for 'Subset';
104 for 'IS'

The options specifically set for the importance sampling are presented in Table 8. Note that
the options of .Simulation and .FORM (not shown in this manual) are also processed in the
case of importance sampling due to the nature of the MCS, FORM and IS.

Table 8: ALROpts.IS
 .Instrumental 1 × Nout INPUT object or Instrumental distribution defined as
Struct either a structure of input marginals
and copula or an INPUT object (refer
to UQL AB User Manual – the INPUT
module for details).

UQL AB-V2.0-117 - 28 -
 Active learning reliability

 .FORM FORM ANALYSIS object or FORM results computed previously.

FORMAnalysis.Results See Section 2.3.4.2 for details.
Struct

The options specifically set for subset simulation are presented in Table 9. Note that the
options of .Simulation are also processed in the case of subset simulation due to the similar
nature of Monte Carlo simulation and subset simulation.

Table 9: ALROpts.Subset
 .p0 Double Target conditional failure probability
default: 0.15 of auxiliary limit-states
(0 <p0≤ 0.5).

 .Proposal See Table 10 Description of the proposal

distribution in the Markov Chain.

 .MaxSubsets Integer Maximum number of subsets. In the

default: subset simulation algorithm, the
MaxSampleSize maximum number of subsets is set to
BatchSize·(1−p0)
the minimum of MaxSubsets and
MaxSampleSize
BatchSize·(1−p0) .

The settings of the Markov Chain Monte Carlo simulation in subset simulation are summa-
rized in Table 10. Note that the default values are taken from Au and Beck (2001).

Table 10: ALROpts.Subset.Proposal (Proposal distributions)

 .Type String Type of proposal distribution (in the
default: 'Uniform' standard normal space).

'Gaussian' Gaussian distribution.

'Uniform' Uniform distribution.

 .Parameters Double Parameter of the proposal

default: 1 distribution. Corresponds to the
standard deviation for a Gaussian
distribution and the half-width for
the uniform distribution.

The initial experimental design of surrogate model can either be given by a number of sam-
ples and a sampling method or by a matrix containing the set of input samples and the
corresponding values of the limit-state function.

UQL AB-V2.0-117 - 29 -
UQL AB user manual

Table 11: ALROpts.ALR.IExpDesign

 .N Integer Number of samples in the initial
default: max(10, 2 × M ) experimental design.

 .Sampling String Sampling techniques of the initial

default: 'LHS' experimental design. See UQL AB
User Manual – the INPUT modulefor
more sampling techniques.

 .X N × M Double Matrix containing the initial

experimental design.

 .G N × Nout Double Vector containing the responses of

the limit-state function
corresponding to the initial
experimental design, corrected by
the threshold k (see also Table 5):
(g(x) − T ) for Criterion = '<=',
(T − g(x)) for Criterion = '>='.

Asynchronous learning can be enabled when a computational model is not directly associated
to the analysis. In this case, the analysis is interrupted and the next sample to add to the
experimental design is returned.

Table 12: ALROpts.Async

 .Enable Logical Enabling or not the asynchronous
default: false learning option.
true Asyncrhonous learning is enabled
false Asynchronous learning is disabled.

 .InitED Logical Enabling or not the asynchronous

default: false learning option for the initial
experimental design.
true Asyncrhonous learning is enabled for
the initial experimental design
false Asynchronous learning is disabled
for the initial experimental design.

UQL AB-V2.0-117 - 30 -
 Active learning reliability

3.2 Accessing the results

Syntax

myAnalysis = uq_createAnalysis(ALROpts)

Output

The results are summarized in Table 13.

Table 13: myAnalysis.Results

.Pf Double Estimator of the failure probability, Pbf .

.Beta Double Associated reliability index, βb = −Φ−1 (Pbf ).

.CoV Double Coefficient of variation.

.ModelEvaluations Integer Total number of model evaluations performed

during the analysis.

.PfCI 1 × 2 Double Confidence interval of the failure probability.

.BetaCI 1 × 2 Double Confidence interval of the associated reliability

index.

.Metamodel uq_model Final surrogate model used in the analysis.

.Reliability uq_analysis Final reliability analysis carried out.

.History See Table 14 History of the active learning reliability analysis

.Xnext K × M Double Next sample(s) to evaluate to add to the

experimental design.
• Only available when .Async.Enable is
true.

.Snapshot String Path to the .mat file containing a snapshot of

the analysis
• Only available when .Async.Enable is
true.

The history of the active learning process is saved in the .History field. Its contents is
described in Table 14.

UQL AB-V2.0-117 - 31 -
UQL AB user manual

Table 14: myAnalysis.Results.History

.Pf Double Failure probability estimate before each
enrichment.
.PfLower Double History of the estimated lower bound of the
failure probability Pf− .
.PfUpper Double History of the estimated upper bound of the
failure probability Pf+ .
.NSamples Integer Number of samples added to the experimental
design at each iteration.
.NCurrent Integer History of the experimental design size
.NInit Integer Number of samples in the initial experimental
design.
.X N × M double Samples in the experimental design for the
given output
.G N × Nout Values of the limit-state function of the
double experimental design samples corrected by the
threshold T (see also Table 5):
(g(x) − T ) for Criterion = '<=',
(T − g(x)) for Criterion = '>='.
.ReliabilitySample Double Samples used in the latest reliability analysis
.Convergence Double History of the convergence criterion values

3.3 Printing/Visualizing of the results

UQL AB offers two commands to conveniently print reports containing contextually relevant
information for a given result object:

3.3.1 Printing the results: uq print

Syntax

uq_print(myAnalysis);
uq_print(myAnalysis, outidx);

Description

uq_print(myAnalysis) prints a report on the results of the reliability analysis stored in

the object myAnalysis. If the model has multiple outputs, only the results for the first
output variable are printed.

uq_print(myAnalysis, outidx) prints a report on the results of the reliability analysis

stored in the object myAnalysis for the output variables specified in the array outidx.

UQL AB-V2.0-117 - 32 -
 Active learning reliability

Examples:
uq_print(myAnalysis, [1 3]) prints the reliability analysis results for output variables 1
and 3.

3.3.2 Graphically display the results: uq display

Syntax
uq_display(myAnalysis);
uq_display(myAnalysis, outidx);

Description

uq_display(myAnalysis) creates a visualization of the results of the reliability analysis

stored in the object myAnalysis, if possible. If the model has multiple outputs, only
the results for the first output variable are visualized.

uq_display(myAnalysis, outidx) creates a visualization of the results of the reliability

analysis stored in the object myAnalysis for the output variables specified in the array
outidx.

Examples:
uq_display(myAnalysis, [1 3]) will display the reliability analysis results for output
variables 1 and 3.

UQL AB-V2.0-117 - 33 -
References

Au, S. K. and J. L. Beck (2001). Estimation of small failure probabilities in high dimensions
by subset simulation. Probabilistic Engineering Mechanics 16(4), 263–277. 29

Basudhar, A. and S. Missoum (2008). An improved adaptive sampling scheme for the con-
struction of explicit boundaries. Struct. Multidisc. Optim. 42(4), 517–529. 8

Bichon, B. J., M. S. Eldred, L. Swiler, S. Mahadevan, and J. McFarland (2008). Efficient global
reliability analysis for nonlinear implicit performance functions. AIAA Journal 46(10),
2459–2468. 7

Ditlevsen, O. and H. Madsen (1996). Structural reliability methods. J. Wiley and Sons,
Chichester. 3

Echard, B., N. Gayton, and M. Lemaire (2011). AK-MCS: an active learning reliability method
combining Kriging and Monte Carlo simulation. Struct. Saf. 33(2), 145–154. 7

Lemaire, M. (2009). Structural reliability. Wiley. 3

Marelli, S. and B. Sudret (2018). An active-learning algorithm that combines sparse poly-
nomial chaos expansions and bootstrap for structural reliability analysis. Struct. Saf. 75,
67–74. 7

Melchers, R.-E. (1999). Structural reliability analysis and prediction. John Wiley & Sons. 3

Moustapha, M., S. Marelli, and B. Sudret (2021). A generalized framework for active learning
reliability: survey and benchmark. Reliability Engineering & System Safety. (submitted). 4,
13

Moustapha, M. and B. Sudret (2019). Surrogate-assisted reliability-based design optimiza-

tion: a survey and a unified modular framework. Struct. Multidisc. Optim. 60, 1–20. 8

Sudret, B. (2007). Uncertainty propagation and sensitivity analysis in mechanical models -

Contributions to structural reliability and stochastic spectral methods. Habilitation thesis,
Université Blaise Pascal, Clermont-Ferrand, France. 1

Zaki, M. J. and W. J. Meira (2014). Data Mining and Analysis: fundamental Concepts and
Algorithms. Cambridge University Press. 10

35