Professional Documents
Culture Documents
DOI 10.1007/s40430-017-0815-8
TECHNICAL PAPER
13
J Braz. Soc. Mech. Sci. Eng.
with respect to design variables is not a trivial task when quadrature and building of the approximation functions
dealing with sampling approaches [21, 35]. as employed in this work.
Although the literature of approaches to reduce the com-
putational burden of RBDO problems is vast [1, 7, 8, 11–
13, 18, 20, 23, 28, 41–44, 47, 49, 51, 54], these efficient 2 Problem statement
RBDO approaches cannot be directly applied or adapted to
reduce the computational burden of RO problems for the RO and RBDO problems can be stated in the following
following reasons: general form:
13
J Braz. Soc. Mech. Sci. Eng.
comprised by the developments presented in [21, 35, Note that two cases are of particular interest. First, if
37, 40]. First, assuming that g < 0 indicates failure of dk affects fX but does not directly affect g (e.g., dk is the
the system under study, the probability of failure can be expected value or the variance of some random variable),
defined as follows: we have ∂g/∂dk = 0 and only the first part of Eq. (11)
remains. This case is discussed in detail in [35–37], and it
Pf = I(g(d, x))fX (x) = E[I(g(d, X))], (6) is the case typically addressed in the literature. On the other
� hand, if fX does not depend on dk (i.e., the design variable
where fX is the joint probability density function (PDF) dk is not a parameter of the random vector), then we have
of the random vector X with support , E[·] represents sk = 0 and only the second part of Eq. (11) remains. This
the expected value and I is the indicator function case was more recently addressed and it is discussed in
detail in [19, 21, 40]. It is worth to point out that in this
0, t ≥ 0
I(t) = . (7) paper, we aim to address both situations described above,
1, t < 0
i.e., we address problems in which dk affects and does not
We then wish to evaluate the sensitivity of Pf according affect fX .
to design variables dk , i.e., the derivatives ∂Pf /∂dk . The Here, we propose the use of polynomial expansions to
main difficulty is that I is discontinuous, and thus, its approximate Pf and ∂Pf /∂dk to alleviate the computational
derivative must be evaluated in the sense of distributions. burden required for this purpose. Hence, the details of this
One approach to obtain the desired result is to approxi- approximation are given in the next section.
mate it by a continuous function that, in the limit, con-
verges to I [21]. Here, we take the approximation
4 Probabilistic polynomial expansion
Ih (t) = 1 − �h (t), (8)
where �h (t) is the Gaussian cumulative distribution func- 4.1 Least squares polynomial expansion
tion (CDF) with mean 0 and standard deviation h. Note
that the approximation Ih converges to the indicator func- In polynomial expansion techniques, the original perfor-
tion I (in the sense of distributions) as h → 0. The failure mance function g(d, X) is substituted by a polynomial
probability can then be approximated by approximation g̃ of the form
q
Pf ≈ Ih (g(d, x))fX (x) = E[Ih (g(d, X))] (9)
� g̃(d, X) = ci (d, X)ψi (X), (13)
i=1
with arbitrary accuracy as h → 0. Application of the
chain rule to the above expression gives where ψi are basis functions from a polynomial set and ci
are coefficients to be determined. Once the approximation
∂Pf ∂fX (x) ∂g(d, x)
≈ Ih (g(d, x)) − φh (g(d, x)) fX (x), g̃ is built, probabilistic measures can be evaluated approxi-
∂dk ∂dk ∂dk
� �
(10) mately with little computational effort. The basis functions
employed in this work are described in Appendix 1.
where φh (t) is the Gaussian PDF with mean 0 and stand- Here, the polynomial representation is obtained with a
ard deviation h (known to be the derivative of the CDF). least squares (LS) approximation [6, 29]. In LS representa-
The expression above can be rewritten as tion, the coefficients can be found from the system of linear
equations [3]
∂Pf ∂g(d, X)
≈ E[Ih (g(d, X))sk (d, X)] − E φh (g(d, X)) q
∂dk ∂dk
(11) Aij cj = bi , i = 1, 2, . . . , q (14)
j=1
with
where cj are the coefficients of the expansion,
1 ∂fX (x)
sk (d, x) = . (12) Aij = �ψi , ψj �,
fX (x) ∂dk (15)
The function sk is known in literature as score function.
In some simple cases, the score function can be evaluated
bi = �ψi , g�, (16)
analytically [35–37]. If analytical evaluation is difficult and the internal product is taken as
or impossible, the score function can be evaluated numer-
ically with a little computational effort, since no evalua- �u, v� = u(x)v(x)w(x), (17)
tion of the performance function is involved. �
13
J Braz. Soc. Mech. Sci. Eng.
where
∂g(d, X)
bik = E ψi (X) . (30)
∂ci ∂dk
cik = . (21)
∂dk
From Eqs. (20) and (29), we conclude that the derivatives
The coefficients cik can be obtained by differentiation of of the expansion can be approximated by the expansions
Eq. (14) according to dk and give of the derivatives. We also observe that this approxima-
q q
tion converges to the exact value as g̃ converges to g.
∂cj ∂bi ∂Aij Besides, the approximation is exact if the score function
Aij = − cj , i = 1, 2, . . . , q. (22)
j=1
∂dk ∂dk ∂dk
j=1
sk = 0 in Eq. (28), i.e., if fX does not depend on dk .
We then obtain the derivatives of the expansion
Using the dot product from Eq. (18), differentiation of employing polynomial representations to the derivatives
Eqs. (15) and (16) according to dk gives of the performance function, as previously done in [40].
∂Aij Simulation can then be made using large samples with
= E[ψi (X)ψj (X)sk (d, x)] (23) low computational effort, since it only requires evalu-
∂dk
ation of closed form polynomials. Finally, the accuracy
and of the expansions can be improved by increasing the
13
J Braz. Soc. Mech. Sci. Eng.
13
J Braz. Soc. Mech. Sci. Eng.
are presented in Table 1. In this table, PfMCS stands for the is, the value in parenthesis is the actual objective function
probability of failure of the design d evaluated by a crude value at d. Finally, in this table, ε is the relative difference
MCS with 107 samples, and pt is the target probability of between PfPCE and PfMCS .
failure of the RBDO problem. Finally, ε is the relative dif- The first-order approximation leads the optimization
ference between PfMCS and pt of each approach evaluated algorithm to a non-optimal solution due to the inaccuracy
by Eq. (31). in the estimation of the failure probability. That is, we
Regarding the probability of failure of the optimum see that the value of f evaluated using PCE and MCS is
designs, it is easily seen that the Pf of the optima provided quite different (1.4301 and 1.5301, respectively). As the
by the RBDO approaches based on FORM (RIA, PMA, approximation order increases, these values get closer,
and ADA) is 39% higher than the target failure probability. since PfPCE converges to the reference value (PfMCS).
This occurs, because FORM assumptions are inaccurate Indeed, it is worth to note that the third-order approxi-
in this case, since the limit sate function is highly nonlin- mation is able to achieve accurate results regarding the
ear. Indeed, this is exactly one of the issues that we want to approximation of the probability of failure. In addition,
address in this paper: FORM-based algorithms may lead to the computational cost required for the solution is much
very inaccurate results from the probabilistic point of view, lower than a single run of any given simulation scheme.
and we want to avoid them for the solution of RO problems. The main conclusion, from both RBDO and RO ver-
The proposed approach, on the other hand, allows one sions of this problem, is that the proposed PCE-based
to obtain more accurate results by increasing the order of approach is able to provide accurate solutions requiring a
the polynomial expansion employed. In this example, we reasonable computational effort.
note that second- and third-order expansions are enough to
obtain accurate results from the probabilistic point of view. 5.2 Example 2: multiple limit states
We also note that the design obtained with a third-order
expansion is significantly different from those obtained This section investigates the effectiveness of the proposed
with FORM-based approaches. These results indicate approach in the case of multiple nonlinear limit state
that FORM-based approaches may not be appropriate in functions. It is originally an RBDO problem and it was
some cases. Finally, we observe that the increase in accu- studied by several authors, among them Yi et al. [50]. It
racy comes at the cost of an increase in the computational has two design variables that are also the mean values of
effort required. However, the computational effort remains two independent Uniform variables with√standard devia- √
much smaller than what would be required by simulation tion equal to 0.6, i.e., Xi ∼ U(di − 0.6 3, di + 0.6 3).
schemes (e.g., crude or even importance sampling MCS). The initial cost function is
5.1.2 Risk optimization
C0 (d) = d1 + d2 , (35)
and three limit state functions are
In the RO version of this problem, we set the cost of failure
X12 X2
as a constant and equal to Cf = 10. The starting point of the g1 (d, X) = − 1, (36)
search as well as the bounds on the design variables are the 20
same of the previous section. Solving the RO problem as
(X1 + X2 − 5)2 (X1 − X2 − 12)2
stated in Sect. 2, we obtain the results presented in Table 2 g2 (d, X) = + − 1, (37)
for first-, second-, and third-order polynomial approxima- 30 120
tions. This table also presents the probability of failure at d 80
evaluated using the PCE approach (PfPCE). Moreover, in the g3 (d, X) = − 1. (38)
X12 + 8X2 + 5
column that presents the optimal objective function f, the
first value is Eq. (2) evaluated using PfPCE , while the value In the next subsections, we first solve it as an RBDO and
in parenthesis employs PfMCS for this computation. That then as an RO problem.
13
J Braz. Soc. Mech. Sci. Eng.
Here, we employ the parameters proposed by Yi et al. To construct the RO version of this problem, we set the cost
[50], i.e., the allowable probability of failure is taken of failure as a constant and equal to Cf = 10, while we keep
as pti = 0.0228, the bound constraints 0 ≤ di ≤ 10 are the same starting point and bounds on the design variables
imposed, and the initial design vector is taken as (5.0, 5.0). of the previous section. The results obtained employing
The results obtained with the proposed approach (employ- first- and second-order expansions are presented in Table 4.
ing first- and second-order expansions) and results avail- Again, only the maximum probability of failure among the
able in literature [50] are presented in Table 3, where only three limit states is presented.
the maximum probability of failure among the three limit These results show that once again, the proposed PCE-
states is presented. based RO approach was able to provide an accurate opti-
Again, we note that the results obtained with FORM- mum solution regarding probability of failure approxi-
based approaches are very inaccurate from the probabil- mation. Furthermore, it was accomplished requiring a
istic point of view. In this case, the design obtained with reasonable computational cost, e.g., the second needed only
PMA is very conservative, while the design obtained with 648 constraint evaluations to reach the optimum design.
PMA-SAP does not meet the required reliability level. The
results obtained with the second-order polynomial expan- 5.3 Example 3: several design variables RO problem
sion, on the other hand, are very accurate and satisfy the
reliability constraint up to simulation errors. Finally, we This example is intended to evaluate the proposed approach
observe again that the proposed approach required more in the case of increasing number of design variables. In
computational effort when compared to FORM-based addition, since this RO benchmark problem is taken from
approaches, i.e., this is the price to pay to obtain more the literature, we can also show that the proposed approach
accurate results. It is worth to note that even though the is able to drastically reduce the computational effort in
NCE of the proposed approach was higher (1485 calls of comparison to simulation-based techniques. As commented
the constraint functions), it still is much lower than a sin- in the introduction of this paper, FORM-based approaches
gle evaluation of Pf by any simulation method, e.g., crude are, in general, not employed for RO, because the cor-
MCS, importance sampling MCS, among others. In other rect representation of the objective function requires an
words, we are able to solve an entire optimization under accurate estimation of the probabilities of failure. Hence,
uncertainty problem requiring a computational cost lower FORM results are not presented in this section.
than a single evaluation of the probability of failure by This RO example was taken from [15]. The initial cost
simulation approaches. is defined as
13
J Braz. Soc. Mech. Sci. Eng.
where n is the number of design variables. The random This paper presents a gradient-based PCE approach for
variables X1 , X2 , X3 are independent √Gaussian variables the solution RO and RBDO problems. It focuses on the
with mean 1.0 and standard deviation 0.2. Moreover, the utilization of gradient-based algorithms for the solution
cost of failure is c1 = 20. Since this is a RO problem, the of these problems. For the required probability of failure
probability of failure is multiplied by the cost of failure and estimation and its sensitivity analysis, PCE techniques
included in the objective function to compute the expected are employed, alleviating the computational burden and
cost of failure (see Eq. (2)). keeping the accuracy of the reliability analysis step. The
To assess the performance of the proposed approach, the proposed approach does not have any limitation regard-
problem was solved for different number of design vari- ing the type of design variable. That is, the design vari-
ables n = 5, 10, 20 as in the work [15]. Computational effi- ables can be some statistic of the random variables of the
ciency was compared considering the number of constraint problem (e.g., mean value) or not.
function evaluations (NCE) need to obtain the optimum Several benchmark problems were analyzed to assess
solution. In this example, we employed only first-order the effectiveness of the proposed approach in different
expansions, because it was enough to give accurate results. situations. The results show that the proposed gradient-
The results are presented in Table 5. A comparison to the based PCE approach is able to obtain accurate solutions
results obtained in [15] is also presented, but only a num- from the probabilistic point of view with tractable com-
ber of limit state function evaluations were available in that putational effort. That is, in all the analyzed cases, third-
work. order polynomial expansions were able to provide an
From these results, we note that the proposed approach accuracy compatible with a crude 107 sample MCS, while
allowed a drastic reduction of the computational effort the number of constraint evaluations was kept much
required for obtaining the optimum design. Besides, the lower than a single probability of failure evaluation by
results obtained are accurate enough from the probabilis- simulation techniques. This conclusion was true even for
tic point of view, since the Pf obtained with the first-order the cases where the number of design variables was high.
expansion agrees to the one obtained with MCS up to simu- Consequently, the proposed approach, in its present form,
lation errors. This indicates that the proposed approach is is recommended for problems with several design vari-
computationally efficient in comparison to existing RO ables and a few random variables.
13
J Braz. Soc. Mech. Sci. Eng.
A potential limitation of the proposed approach is linked In this work, quadrature rules for the n-dimen-
with one of the PCE: the stochastic dimension of the prob- sional case are obtained by full-tensor product between
lem. It happens, because the number of terms of the expan- one-dimensional rules. Suppose X1 , X2 , . . . , Xn and
sion increases exponentially with the number of random vari- W1 , W2 , . . . , Wn are sets of quadrature nodes and
ables of the problem. However, this issue may be alleviated weights according to variables x1 , x2 , . . . , xn , respec-
with the employment of sparse integration schemes [6, 25, tively. Then, the set of quadrature nodes and weights for
34] and will be the matter of future studies. x = {x1 , x2 , . . . , xn } can be obtained by
13
J Braz. Soc. Mech. Sci. Eng.
11. Cheng GD, Xu L, Jiang L (2006) Sequential approximate pro- 31. Lopez RH, Torii AJ, Miguel LFF, Cursi JESD (2015) An
gramming strategy for reliability-based optimization. Comput approach for the global reliability based optimization of
Struct 84(21):1353–1367 the size and shape of truss structures. Mech Ind 16(6):603.
12. Croquet R, Lemosse D, de Cursi ES, El-Hami A (2013) Iterative doi:10.1051/meca/2015029
projection on critical states for reliability-based design optimiza- 32. Madsen H, Krenk S, Lind N (1986) Methods of structural
tion. Eng Optim 45(5):577–590 safety. Prentice Hall, Englewood Cliffs
13. Du X, Chen W (2004) Sequential optimization and reliability 33. Melchers R (1999) Structural reliability analysis and predic-
assessment method for efficient probabilistic design. ASME J tion. Wiley, New York
Mech Des 126(2):225–233 34. Nobile F, Tempone R, Webster CG (2008) A sparse grid
14. Ghanem R, Spanos P (1991) Stochastic finite elements: a spec- stochastic collocation method for partial differential
tral approach. Springer, New York equations with random input data. SIAM J Numer Anal
15. Gomes W, Beck A (2016) The design space root finding method 46(5):2309–2345
for efficient risk optimization by simulation. Probab Eng Mech 35. Rahman S (2009) Stochastic sensitivity analysis by dimen-
44:99–110. doi:10.1016/j.probengmech.2015.09.019 sional decomposition and score functions. Probab Eng Mech
16. Gomes WJS, Beck AT (2013) Global structural optimization
24(3):278–287
considering expected consequences of failure and using ann sur- 36. Rahman S (2016) Reliability-based design optimization by
rogates. Comput Struct 126:56–68 adaptive-sparse polynomial dimensional decomposition. Struct
17. Hildebrand F (1974) Introduction to numerical analysis, 2nd edn. Multidiscip Optim 53(3):425–452
McGraw-Hill, New York 37. Rubinstein R, Shapiro A (1993) Discrete event systems: Sen-
18. Huang ZL, Jiang C, Zhou YS, Luo Z, Zhang Z (2016) An incre- sitivity analysis and stochastic optimization by the score func-
mental shifting vector approach for reliability-based design opti- tion method. Wiley, New York
mization. Struct Multidiscip Optim 53(3):523–543. doi:10.1007/ 38. Rubinstein RY, Kroese DP (2007) Simulation and the Monte
s00158-015-1352-7 Carlo method, 2nd edn. Wiley, Hoboken
19. Keshavarzzadeh V, Meidani H, Tortorelli DA (2016) Gradi-
39. Suryawanshi A, Ghosh D (2016) Reliability based optimiza-
ent based design optimization under uncertainty via stochastic tion in aeroelastic stability problems using polynomial chaos
expansion methods. Comput Methods Appl Mech Eng 306:47– based metamodels. Struct Multidiscip Optim 53(5):1069–
76. doi:10.1016/j.cma.2016.03.046 1080. doi:10.1007/s00158-015-1322-0
20. Kuschel N, Rackwitz R (2000) A new approach for structural 40. Torii A, Lopez R, Miguel L (2017) Probability of failure sen-
optimization of series systems. Appl Stat Probab 2:987–994 sitivity analysis using polynomial expansion. Probab Eng
21. Lacaze S, Brevault L, Missoum S, Balesdent M (2015) Probabil- Mech (to appear)
ity of failure sensitivity with respect to decision variables. Struct 41. Torii AJ, Lopez R, Biondini F (2012) An approach to reliabil-
Multidiscip Optim 52(2):375–381 ity-based shape and topology optimization of truss structures.
22. Lee T, Jung J (2008) A sampling technique enhancing accuracy Eng Optim 44(1):37–53
and efficiency of metamodel-based rbdo: constraint boundary 42. Torii AJ, Lopez RH, Miguel LFF (2016) A general rbdo
sampling. Comput Struct 86(13–14):1463–1476. doi:10.1016/j. decoupling approach for different reliability analysis meth-
compstruc.2007.05.023. http://www.sciencedirect.com/science/ ods. Struct Multidiscip Optim 54(2):317–332. doi:10.1007/
article/pii/S0045794907001836, structural Optimization s00158-016-1408-3
23. Lim J, Lee B (2016) A semi-single-loop method using approxi- 43. Torng T, Yang R (1993) An advanced reliability based optimiza-
mation of most probable point for reliability-based design opti- tion method for robust structural system design. In: Proceedings
mization. Struct Multidiscip Optim 53(4):745–757. doi:10.1007/ of the 34th AIAA/ASME/ASCE/AHS/ASC structures, structural
s00158-015-1351-8 dynamics, and material conference
24. Liu X, Wu Y, Wang B, Ding J, Jie H (2016) An adaptive local 44. Tu J, Choi KK, Park YH (1999) A new study on reliability-based
range sampling method for reliability-based design optimization design optimization. J Mech Des 121:557–564
using support vector machine and kriging model. Struct Multi- 45. Valdebenito MA, Schueller GI (2010) A survey on approaches
discip Optim 1–20. doi:10.1007/s00158-016-1641-9 for reliability-based optimization. Struct Multidiscip Optim
25. Long Q, Scavino M, Tempone R, Wang S (2013) Fast estimation 42:645–663
of expected information gains for bayesian experimental designs 46. Wiener N (1938) The homogeneous chaos. Am J Math
based on laplace approximations. Comput Method Appl Mech 60(23–26):897–936
259:24–39 47. Wu YT (2001) Safety-factor based approach for probability-
26. Lopez R, Torii A, Miguel L, Cursi JS (2015) Overcoming the based design optimization. In: Proceedings of the 42nd AIAA
drawbacks of the form using a full characterization method. SDM conference, Seattle
Struct Saf 54:57–63 48. Xiu D, Karniadakis GE (2002) The Wiener–Askey polynomial
27. Lopez RH, Beck AT (2012) Rbdo methods based on form: a chaos for stochastic differential equations. SIAM J Sci Comput
review. J Braz Soc Mech Sci 34(4):506–514 24(2):619–644
28. Lopez RH, Lemosse D, de Cursi ES, Jhojan R, El-Hami A
49. Yi P, Cheng GD (2008) Further study on efficiency of sequential
(2011) An approach for the reliability based design optimization approximate programming strategy for probabilistic structural
of laminated composites. Eng Optim 43(10):1079–1094 design optimization. Struct Multidiscip Optim 35:509–522
29. Lopez RH, Miguel LFF, de Cursi ES (2013) Uncertainty quan- 50. Yi P, Cheng G, Jiang L (2008) A sequential approximate pro-
tification for algebraic systems of equations. Comput Struct gramming strategy for performance-measure-based probabilistic
128:189–202 structural design optimization. Struct Saf 30(2):91–109
30. Lopez RH, Miguel LFF, Belo IM, de Cursi ES (2014) Advan- 51. Yi P, Zhu Z, Gong J (2016) An approximate sequential opti-
tages of employing a full characterization method over form in mization and reliability assessment method for reliability-
the reliability analysis of laminated composite plates. Compos based design optimization. Struct Multidiscip Optim 1–12.
Struct 107:635–642 doi:10.1007/s00158-016-1478-2
13
J Braz. Soc. Mech. Sci. Eng.
52. Youn BD, Choi KK, Du L (2005) Adaptive probability analysis compind.2011.05.008. http://www.sciencedirect.com/science/
using an enhanced hybrid mean value method. Struct Multidiscip article/pii/S016636151100073X
Optim 29:134–148 54. Zou T, Mahadevan S (2006) A direct decoupling approach for
53. Zhu P, Zhang Y, Chen G (2011) Metamodeling develop-
efficient reliability-based design optimization. Struct Multidiscip
ment for reliability-based design optimization of automotive Optim 31(3). doi:10.1007/s00158-005-0572-7
body structure. Comput Ind 62(7):729–741. doi:10.1016/j.
13