4 Reliability Based Optimization

Application of optimization methods and algorithms in reliability-based design of engineering or

economic systems has mainly concentrated on models where the potential failure modes (failure
elements) are modelled as elements in a series system. In reliability-based optimization some of
the quantities defining the model used are modelled as stochastic variables. The reliability can
generally effectively be measured using first order reliability methods (FORM).
In this paper systems are considered where the reliability model consists of a series system of
parallel systems. Such a model is used for example when failure of a system is defined by the
event that more than one failure element have to fail simultaneously. In section 4.1 the
reliability-based optimization problem is formulated. The objective function can for example
model the total expected cost of an engineering system, and the constraints signify that the
reliability of the system is adequate during the expected lifetime of the system.
In section 4.2 a sensitivity analysis of the optimization problems considered is presented, and in
section 4.3 an optimization procedure is described.

4.1 Problem Formulation

The optimization problem considered is written
min W  z  (1)
s.t .  S  z    min (2)
zimin  zi  zimax , i  1,..., N (3)
where z   z1 ,..., z N  are the optimization variables which are assumed to be real-valued. N is the
number of optimization variables. As a measure of the reliability of the system the so-called
systems reliability  S , see below, is used.  min is the minimum acceptable reliability index.
zimin and zimax are simple lower and upper bounds of zi. W  z  is the objective function. W  z 
can for example model the total costs of the system (for structural systems the structural weight
is often used as an objective function). The optimization problem (1) - (3) is generally non-linear
and non-convex,
In system modelling some of the uncertain quantities are modelled by stochastic variables
X   X 1 ,..., X n  , i.e. the number of stochastic variables is n.
The potential failure modes (failure elements) are described by failure functions
gi  x   0 , i  1,..., m (4)
Realizations x of X where gi  x   0 correspond to failure states, and realizations where
gi  x   0 correspond to safe states. The failure elements can e.g. in structural systems model
such failure modes as yielding of cross-sections, stability failure, fatigue failure, etc.
In first order reliability methods a transformation T of the correlated and non-normally
distributed stochastic variables X into standardized and normally distributed variables U =

6
(U1,...,Un) is defined. Let X = T ( U ). An element reliability index  i corresponding to failure
element  i is then determined from the following optimization problem
 i  min uTu (5)
gi  T  u    0

In some applications a system is not considered to be in a state of failure if only one local
failure mode fails. If failure is defined by the event that M failure elements have to be in a state
of failure then the whole system can be modelled by a series system of parallel systems. A rough
estimate of the reliability index  p of one parallel system can be determined by linearizing the
 
safety margins M i  gi T  U  of the failure elements in the parallel system, see figure 1,
M i   iT U  i (6)
where
gi
i  (7)
g i
and using

 p  11  M    ,    (8)

Figure 1. Illustration of parallel system (n = 2 and M = 2) and

approximations of failure regions (BAC and GEF).

where  M is the M-dimensional normal distribution function    1 ,...,  M  and the elements
in the correlation coefficient matrix  are determined by
ij   iT  j (9)
This approximation is illustrated in figure 1 for the case where M = n = 2. The failure
region is approximated by the region defined by  BAC.

7
 Usually a much better approximation of the reliability of a parallel system is obtained by
using the so-called joint design point E (see figure 1) as linearization point. The joint design
point is determined from the following optimization problem
1
min   u T u
u 2
s.t. g1  u   0 (10)

gM  u   0
Generally this optimization problem (10) can be solved using standard mathematical
programming algorithms. For example the following methods can be used:
a) Penalty function methods
b) Augmented Lagrangian multiplier methods
c) Feasible direction method
d) Generalized reduced gradient (GRG)
e) Sequential quadratic problems (SQP)
f) Sequential linear problems (SLP)
Also combinations of these methods can be powerful which use elements from Augmented
Lagrangian multiplier methods and from SQP methods.
In the following an optimality criteria method to solve the optimization problem is
presented. The Lagrangian function corresponding to (10) is
M
1
L  u ,    u T u   i gi  u  (11)
2 i 1

where  is the Lagrangian multipliers. If u  is the joint design point and   the corresponding
Lagrange multipliers then the Kuhn-Tucker conditions must be satisfied at this point:
M dg j
ui    j  0 , i  1,..., n (12)
j 1 dui
gi  u    0 , i  1,..., M (13)

i gi  u    0 , i  0 , i  1,..., M (14)

In order to estimate  k   k 1   2 given a point u0 close to u  and given an estimate of the

set I of active constraints (failure elements), consider the following approximation of (12) - (14)
M dg j  u0 
ui    j  0 , i  1,..., n (15)
j 1 dui
gi  u    g i  u0  u  u0   0 , i  1,..., M
T 
(16)


i gi  u0   g i  u0 
T
u 

 u0   0 , i  0 , i  1,..., M (17)
or if only active constraints are considered
u   G    0 (18)

8
 G0  GT  u   u0   0 ,   0 (19)
  is now assumed to contain the multipliers of the active constraints numbered 1 to
MA and
G  g1  u0  ,..., g M A  u0  

G0   g1  u0  ,..., g M A  u0  
T

(18)-(19) can then be written

 
1
    G T G G (21)

 
1
u   G  G T G G (22)

Based on these equations the following algorithm can be formulated (k is the iteration
number)
step 0: initialize: k = 0, uk  0 , k  0 ,  0  0
input the positive convergence parameters  , 1 ,  2 ,  3 , and  4
step 1: calculate gi  uk  , i = 1,..., M
select the set I of active constraints (failure elements) as the constraints with
gi  uk    or i  1
step 2: calculate G and G0
step 3: convergence check (if k > 0)
If  k   k 1   2 and
 dg j  uk  
max i  ui  k     j  k     3 and
 jI dui 
gi  uk    4 , i  1,..., M
Then stop
Else k = k+1
step 4: calculate k , uk and  k from (21), (22) and (10)
Go to step 1
The identification of the active set can be improved if second order estimates of the
Lagrangian multipliers are used.
When the joint design point E is determined, see figure 1, the MA active failure elements
T
are linearized and written M i  i   i U . An estimate of the reliability index  P of the parallel
system can be determined using (8). The failure region is now approximated by the region
defined by  GEF, see figure 1.
For the parallel system an equivalent linear failure element with a linear safety margin
T
MP   P  P U (23)
can be determined.

9
 Figure 2. Series system of parallel
systems.
When the significant
parallel systems (sequences of failure elements) have been identified then these are modelled as
elements in a series system see figure 2, and the system reliability index is determined from

 S   1 1   M P   P ,  P   (24)
where Mp is the number of significant parallel systems and  P and  P are the reliability index
vector and correlation matrix corresponding to the equivalent linear failure elements for the
parallel system.

4.2 Sensitivity Analysis

If optimization algorithms such as NLPQL or VMCWD are used then first order derivatives of
the objective function and the constraints with respect to the optimization variables are required.
If no probability estimates are used in the objective function W then W can usually be
determined without significant computer time consumptions (when compared with the time used
for estimating  S ).
Estimation of the derivatives of the system reliability index j1s with respect to the
optimization variables using finite differences will generally be too expensive with respect to
computer time. Also derivatives estimated using analytical expressions can give problems. One
of the reasons is that for M > 2 the M–dimensional normal distribution function  M can only be
evaluated approximately (except in some special cases). Another reason is that the analytical
expressions become very complicated because reliability indices and correlation coefficients in
both the parallel systems and the series system are dependent on the optimization variables z .
In the following a quasi-analytical method to estimate  S is proposed. The basic failure
functions are written
gi  u , z   0 , i  1,..., M (25)
When the joint design point or the individual reliability indices of the failure elements have
been determined for each parallel system then only the linearized failure functions are used to
evaluate the probability of failure of the parallel systems and the series systems reliability index.
The linearized failure functions are written, see (6)
gij  u , z    ij  z    ijT  z  u , i  1,..., M , j  1,..., M P (26)
where  ij  z  and  ij  z  are the reliability index and the  -vector, respectively, for the ith
failure element in the jth parallel system.

10
 The system reliability index   z  is written (the detailed relationship between  S and
S

 and  is described in section 2)

 S  z    S    z  ,  z   (27)

where,s =    11 ,  21 ,...,  M M P  and    11 , 21 ,..., M M P  . The derivative of  S with respect
T T T

to zk is then written
 S   S  ij  S  ijl 
      (28)
zk i 1, M   ij zk l 1,n  ijl zk 
j 1, M P

If the reliability indices of the elements in the parallel systems are determined without
determining the joint design point, i.e. (5) - (7) are used, then  / zk and  / zk for a typical
element with failure function g  u , z   0 can be estimated from (using a sensitivity analysis of
the Kuhn-Tucker conditions corresponding to the optimization problem (5).
 1 g
 (29)
zk u g zk

 1  u  
   (30)
zk   zk zk 
where
u  1  g  
 B 1 b    u g T 1
B b   u g  (31)
zk  u g T B 1u g  zk  
 u g
b (32)
u g zk

BI  u g (33)
u g
Using a sensitivity analysis of the Kuhn-Tucker conditions corresponding to (19) then
 / zk and  / zk can also be determined if the joint design point are used as linearization
point. The sensitivity coefficients for each parallel system are determined simultaneously.
 S /  ij and  S /  ijl can be determined using finite differences. If the failure functions
are continuously differentiable (with respect to z ) and the approximation used for  M (for
example the Hohenbichler approximation, is continuously differentiable (with respect to  and
 ) then application of (28) gives continuous derivatives of  S with respect to z . First order
optimization methods can then be applied.
When structural systems are considered then from a numerical point of view, estimation of
all the derivatives using (28) and the above method will generally not be significantly more
expensive than to estimate the systems reliability index. The reason is that the determination of
single reliability indices or joint design points is generally very computer time consuming

11
because in many practical cases evaluation of the failure functions requires a complete structural
analysis.

4.3 Optimization Procedure

Several procedures to solve series system reliability constrained optimization problems have
been suggested. At each step in this procedure an element reliability index based optimization
problem is solved and the lower bounds of the element reliability indices are adjusted such that
the system reliability index constraint is satisfied. The advantage of this procedure is that
element reliability index based optimization problems are generally much easier to solve than the
system reliability index based optimization problem.
The following optimization procedure is developed to solve problems where a very large
number of parallel systems is possible. Only the most significant parallel systems are assumed to
be identified and modelled as elements in a series system.
A general first order optimization method is assumed to be used, e.g. the NLPQL
algorithm. Initially M0 significant parallel systems are identified corresponding to the initial
value of z . During the first I1 iterations the set of significant parallel systems is fixed, see figure
3. Then the set is updated using the current value of z . New significant parallel systems are
added to the set of significant parallel systems. Next the optimization is continued until iteration
I2 where a new updating is performed. This procedure is continued until some convergence
criteria are satisfied. These criteria will be related to the change in the set of significant parallel
systems and to the Kuhn-Tucker conditions corresponding to (1) - (3).

Figure 3. Number of parallel systems in series system as

a function of number of iterations.

If in some applications convergence problems occur then move limits on the optimization
variables can be used until the set of significant parallel systems does not change considerably.

12
5 Conclusion
A parallel system with n components has :
i) n-1 redundant components.
ii) Fails if and only if all n components fail.
iii) All components start operating simultaneously.
iv) Failed components are not repaired or replaced.
v) Total system reliability is greater than reliability of any single component of system.
Systems modelled as a series system of parallel systems are considered and it is shown how the
reliability of such systems can be estimated using first order reliability methods (FORM). This
type of systems can be used to estimate the reliability of systems in a large number of disciplines,
for example structural, economic and electrical systems. In order to estimate the reliability of a
parallel system using FORM an optimization problem with quadratic objective function and non-
linear constraints has to be solved. An algorithm to solve this optimization problem is presented.
A reliability-based optimization problem is formulated where the expected costs of the
system is used as objective function and the reliability (estimated by FORM) of the series system
of parallel systems is used as constraints. The optimization variables are quantities defining the
model. A sensitivity analysis of this problem (especially the reliability constraint) is performed
and a general procedure to solve the optimization problem is presented.

6 References
1. Zuev K (2009) Advanced stochastic simulation methods for solving high-dimensional
reliability problems. Ph.D. thesis, The Hong Kong University of Science and
Technology
2.Wang H, Pham H (1997) Survey of reliability and availability evaluation of complex
networks using Monte Carlo techniques. Microelectron Reliab 37(2):187-209
3.Sultan KS, Al-Moisheer AS (2012) Approximate Bayes estimation of the parameters
and reliability function of a mixture of two inverse Weibull dis-tributions under type-2
censoring. J Stat Comput Simul, 1-15
4.Thierauf G, Cai J (1997) Parallel evolution strategy for solving structural optimization.
Eng Struct 19(4):318-324
5.Ye RD, Ma TF, Luo K (2012) Inferences on the reliability in balanced and unbalanced
one-way random models. J Stat Comput Simul 1-18
6.Yaun T, Kuo Y (2010) Bayesian analysis of hazard rate, change point, and cost-
optimal burn-in time for electronic devices. IEEE Trans Reliab 59(1):132-138
7.Levitin G, Lisnianski A, Ben-Haim H, Elmakis D (1998) Redundancy opti-mization for
series-parallel multi-state systems. IEEE Trans Reliab 47(2):165-172
8.Ram M (2010) Reliability measures of a three-state complex system: a cop-ula approach.
Appl Appl Math Int J 5(10):1483-1492

9.Schueller GI, Pradlwarter HJ, Beck J, Au S, Katafygiotis L, Ghanem R (2005)

13
Benchmark study on reliability estimation in higher dimensions of structural systems|an
overview. In: Soize C, Schueller GI (eds) Structural dynamics EURODYN 2005|
Proceedings of the 6th international conference on structural dynamics. Millpress,
Rotterdam, pp 717-722
10.Schueller G, Pradlwarter H, Koutsourelakis P (2004) A critical appraisal of reliability
estimation procedures for high dimensions. Probab Eng Mech 19(4):463-474
11.Umesha P, Venuraju M, Hartmann D, Leimbach K (2005) Optimal de-sign of truss
structures using parallel computing. Struct Multidisc Optim 29(4):285-297
12.Rahman S, Xu H (2004) A univariate dimension-reduction method for multi-
dimensional integration in stochastic mechanics. Probab Eng Mech 19(4):393-408
13.Marti K, Kaymaz I (2006) Reliability analysis for elastoplastic mechanical structures
under stochastic uncertainty. ZAMM 86(5):358-384
14.Pu Y, Das P, Faulkner D (1997) A strategy for reliability-based optimiza-tion. Eng
Struct 19(3):276-282
15.Petryna Y, Kratzig W (2005) Computational framework for long-term reliability
analysis of RC structures. Comput Methods Appl Mech Eng 194(12-16):1619-1639
16.Pellissetti MF, Schueller GI, Pradlwarter HJ, Calvi A, Fransen S, Klein M (2006)
Reliability analysis of spacecraft structures under static and dynamic loading. Comput
Struct 84(21):1313-1325
17.Pellissetti M (2009) Parallel processing in structural reliability. J Struct Eng Mech
32(1):95-126
18.Papadrakakis M, Lagaros N, Plevris V (2005) Design optimization of steel structures
considering uncertainties. Eng Struct 27(9): 1408-1418
19.Papadrakakis M, Lagaros N (2002) Reliability-based structural optimiza-tion using
neural networks and Monte Carlo simulation. Comput Methods Appl Mech Eng
191(32):3491-3507
20.Lee I, Choi K, Du L, Gorsich D (2008) Inverse analysis method using MPP-based
dimension reduction for reliability-based design optimization of nonlinear and multi-
dimensional systems. Comput Methods Appl Mech Eng 198(1):14-27
21.Kuschel N, Rackwitz R (1997) Two basic problems in reliability-based structural
optimization. Math Methods Oper Res 46(3):309-333
22.Kaymaz I, Marti K (2007) Reliability-based design optimization for elastro plastic
mechanical structures. Comput Struct 85(10): 615-625
23.Foschi R, Li H, Zhang J (2002) Reliability and performance-based design:
a computational approach and applications. Struct Saf 24(2-4):205-218
24.Au S, Beck J (2001) Estimation of small failure probabilities in high di-mensions by
subset simulation. Probab Eng Mech 16(4): 263-277
25.Au S (2006) Critical excitation of SDOF elasto-plastic systems. J Sound Vib 296(4-
5):714-733
26.Basudhar A, Missoum S (2008) Adaptive explicit decision functions for probabilistic
design and optimization using support vector machines. Com-put Struct 86(19-
20):1904-1917
27.Basudhar A, Missoum S, Sanchez A (2008) Limit state function identi ca-tion using
support vector machines for discontinuous responses and disjoint failure domains.

14
Probab Eng Mech 23(1):1-11
28.Bjerager P, Krenk S (1989) Parametric sensitivity in rst order reliability theory. J Eng
Mech 115(7):1577-1582
29.Royset J, Der Kiureghian A, Polak E (2006) Optimal design with proba-bilistic
objective and constraints. J Eng Mech 132(1):107-118
30.Winterstein S, Ude T, Cornell C, Bjerager P, Haver S (1994) Environmen-tal
parameters for extreme response: inverse FORM with omission factors. In: Schueller G,
Shinozuka M, Yao J (eds) Proceedings of the 6th interna-tional conference on structural
safety and reliability (ICOSSAR’93). A.A. Balkema, Rotterdam, pp 551-557
31.Xu H, Rahman S (2004) A generalized dimension-reduction method for multi-
dimensional integration in stochastic mechanics. Int J Numer Methods Eng 61(12):1992-
2019
32.Yang R, Gu L (2004) Experience with approximate reliability-based opti-mization
methods. Struct Multidisc Optim 26(1-2):152-159
33.Youn B, Choi K, Park Y (2003) Hybrid analysis method for reliability-based design
optimization. J Mech Des 125(2):221-232
34.Royset J, Polak E (2004a) Implementable algorithm for stochastic op-timization
using sample average approximations. J Optim Theory Appl 122(1):157-184
35.Thierauf G, Cai J (1997) Parallel evolution strategy for solving structural
optimization. Eng Struct 19(4):318-324
36.Ellingwood B (2001) Earthquake risk assessment of building structures. Reliab Eng
Syst Saf 74(3):251-262
37.Ditlevsen O (1978) Narrow reliability bounds for structural systems. Tech. Rep. 145,
DCAMM (The Danish Center for Applied Mathematics and Mechanics)
38.Der Kiureghian A, Zhang Y, Li CC (1994) Inverse reliability problem. J Eng Mech
120(5):1154-1159
39.Jin R, Du X, Chen W (2003) The use of metamodeling techniques for optimization
under uncertainty. Struct Multidisc Optim 25(2): 99-116
40.Johnson E, Proppe C, Spencer B Jr, Bergman L, Szekely G, Schueller G (2003)
Parallel processing in computational stochastic dynamics. Probab Eng Mech 18(1):37-
60
41.Katafygiotis L, Zuev K (2007) Estimation of small failure probabilities in high
dimensions by adaptive linked importance sampling. In: Papadrakakis M, Charmpis D,
Lagaros N, Tsompanakis Y (eds) ECCOMAS thematic con-ference on computational
methods in structural dynamics and earthquake engineering (COMPDYN). Rethymno,
Crete
42.Katafygiotis L, Zuev K (2008) Geometric insight into the challenges of solving high-
dimensional reliability problems. Probab Eng Mech 23(2-3): 208-218

15
PRESENTED BY
Sai Somyadarshan Bhuyan 2017B4A10720P

Rahul Ranjan 2017B4A80824P

Abhishek Pratik 2017B4A30864P