th

SEC18: Proceedings of the 11 Structural Engineering Convention - 2018

Jadavpur University, Kolkata, India, December 19 - 21, 2018
Paper No. 20180048_2

AN IMPROVED ROBUST DESIGN OPTIMIZATION OF

STRUCTURES WITH UNCERTAIN BUT BOUNDED PARAMETERS
Sujit Das1, Gaurav Datta2 and Soumya Bhattacharjya3*
1,2,3
Department of Civil Engineering, Indian Institute of Engineering Science and Technology, Shibpur,
West Bengal, India
E-mail: [email protected], [email protected], *[email protected]

ABSTRACT
This paper deals with Robust Design Optimization (RDO) of structures when uncertainty information
about the parameters is limited. Most often, only range of variation of the parameters due to
uncertainty is available, by which parameters can be treated uncertain-but-bounded (UBB) type. In
recent years, convex programming (CP) approaches are used to solve RDO problem with UBB
parameters. However, these approaches do not consider sensitivity information while formulating the
RDO problem. But, it is expected that due consideration of sensitivity measure in the RDO
formulation will improve robustness of RDO solution. Thus, concept of a new sensitivity measure is
proposed in this paper, which is minimized along with the conventional objective function of CP. The
improvement by the proposed approach is investigated by optimizing a three bar truss. The results
indicate that the proposed approach yields more robust solutions in comparison to the conventional
approach.
Keywords: Robust design optimization, Uncertain but bounded type parameters, Convex
modeling, Sensitivity Index

sensitive to input parameter uncertainty. Most

NOMENCLATURE often, sufficient realistic statistical data of
involved system parameters like load,
f Objective Ui UBB temperature difference, etc. are not available to
function parameters characterize an uncertain system and construct
g Constraint β Permissible probability density function to carry out a
function Sensitivity reliability analysis. Rather, only range of
level variation of the parameters due to uncertainty is
X Design η Safety available, by which parameters can be treated
variable index uncertain-but-bounded (UBB) type. RBDO
cannot be applied with UBB parameters, and the
1. INTRODUCTION RDO becomes only viable alternative. Most
conveniently, RDO with UBB parameters are
It is well established now that disregarding executed using first order perturbation approach
influence of uncertainty in optimization may [1] which is only valid for low level of
lead to unsafe design leading to catastrophic uncertainty. Also, this approach often leads to
failure consequences. Reliability based design uneconomic and very conservative design since
optimization (RBDO) brings specified it is based on worst case propagation uncertainty
reliability of structure, but the system may be [2]. In this regard, convex programming (CP)
still sensitive to input parameter variation due to approaches [2] are gaining increasing attention,
uncertainty, leading to large undesirable where the UBB parameters constitute a convex
deviations. On the other hand, the robust design set. In the CP, the design parameters are so
optimization (RDO) is more viable in this selected that provides a so-called „reliability
regard, which yields design which is least measure‟ more than the target one [2]. It may be

OPT7 | P a g e
 SEC18: Paper No. 20180048_2

noted here that the term „reliability measure‟ is can be traced in direction finding problem while
not same as „reliability index‟ used for calculating reliability index [9]. The authors of
probabilistic parameters. Rather, it is a term the present paper have also investigated a new
used to denote the largest variation of the UBB formulation hinged on sensitivity based
parameters that the system can withstand. importance factors when addressing RDO with
Design x1 will have greater „reliability measure‟ probabilistic uncertainty [10] and with worst
than design x2, if x1 can withstand higher case propagation approach of bounded
deviation of UBB parameters. Thus, higher the uncertainty [11]. However, application of this
„reliability measure‟, higher will be the concept towards CP was not simple, since
robustness of the system and so higher is the importance factors were earlier used to modify
safety of the system. standard deviation or dispersion of objective
and constraint functions. But, in the CP, the
The state-of-the-art on mathematical
RDO is not achieved by simply minimizing the
formulation of RDO with probabilistic
standard deviation or dispersion, rather the
parameters is quite advanced [3]. Available
approach is non-probabilistic in nature. Hence, a
techniques enable one to formulate RDO
new sensitivity measure is defined in this paper
problem in presence of higher uncertainty level,
to cope with the CP problem, which is distinctly
mixed probability distribution and correlation.
different than the importance factors used in
In this regard, there is enough scope to explore
probabilistic formulations. Then, this sensitivity
RDO formulation with UBB parameters. Ben-
measure is imposed as a constraint in the
Tal and Nemirovski [4] detailed the concept of
proposed RDO formulation, which must be kept
convex modelling to treat UBB uncertainty.
at significantly lesser level to guarantee
Since then, researchers have tried various
„robustness‟ or „insensitiveness‟ of the system.
extension of the convex modelling to effectively
solve RDO problem. Au et al. [5] proposed an Thus, the unique contribution of this
unsatisfactory degree function following the paper is to propose a new RDO formulation
concept of Physical Programming to express the with UBB parameters which is hinged on the
feasibility of constraints under undesirable concept of imposing a sensitivity constraint. The
variation of UBB parameters. Wang et al. [6] improvement by the proposed approach is
applied convex modelling for optimization of a elucidated by optimizing a three-bar truss
simple semiactive tuned mass damper system. problem, and the results are compared with the
Chen et al. [2] presented hyper-ellipsoidal CP existing formulations of CP.
approach for RDO of structural system, where
the authors proposed a sub-optimization scheme 2. CONVENTIONAL CP BASED RDO
to efficiently yield RDO solution. Meng et al.
[7] proposed a super-parametric approach of The concept of CP is first described with help of
convex modeling which is a general case of Fig. 1. Let U1 and U2 are two UBB parameters.
ellipsoidal model. In this regard, the theoretical These can be represented by a general ellipse
review of Jiang et al. [8] on probability-interval defined by their ranges Ū1 and Ū2 as shown in
hybrid uncertainty analysis for structures with Fig. 1. Since, these have dimensions, it will be
both aleatory and epistemic uncertainties is of quite advantageous to transform the system in
worth mentioning. non-dimensional space. It is done by changing
the scale of UiL (lower bound) to UiU (upper
It may be noted here that the above bound) to (-1) (lower bound) to (+1) (upper
formulations ensure robustness by obtaining a bound). Thereby, the ellipse becomes a circle as
design solution which can withstand a specified shown in Fig. 2. Thus, by this normalization, the
level of undesirable deviation of UBB UBB parameters are transformed from U space
parameters. However, system sensitivity is not to u space. It can be written that
focussed in such formulation, which may be
even at high values, though the requirement of , where and
CP is made. But, unless the sensitivity of
structure is brought to significantly less value,
 
U i  max U iU  U , U  U iL . By this operation,

there is a chance of large deviation in system the variation of the interval variables will lie in
behaviour, and the essence of „robustness‟ or the interval [-1, 1].
„insensitiveness‟ will be lost. In fact, the use of
sensitivity measure in probabilistic formulations

OPT8 | P a g e
 SEC18: Paper No. 20180048_2

U2

Ū2

U1

Ū1

Fig. 1 Convex ellipsoidal space

u2 Fig. 3 Concept of η
Chen et al. [2] proposed a two-step optimization
process for the RDO. The main Optimization
problem is to find the DVs X as,

find X such as to,

minimize f ( X , U )
ū2 u1 subjected to, g j ( X , U )  0 , where, j=1,2,…M
X [ X L , X U ] (1)
A sub-Optimization problem is defined at each
update of X to find out the maximum deviation of
ū1 UBB parameters U that the system can tolerate.
This optimization formulation finds U, such as to,
Fig. 2 Unit radius circle case
minimize g j ( X , U ) , where, j=1,2,3,….,n
Now, the structure is said to be in reliable state
if the constraint function g(u) > 0. On the other subjected to, η =max (|U1 |,|U 2 |,|U3 |,.....,|U m |)  t arg et
hand, g(u)<0 indicates the failure of the structure.
g(u)=0 defines the critical condition or limiting (2)
state of the structures. This conditions are depicted The constraint function is minimized in the sub-
in Fig. 3. In Fig. 3, η defines the safety of the problem to find the most probable failure point or
structures. Higher the values of η, more will be the the set-up of U for which g(U) is just critical or
safety margin of the structures, and hence, more becomes an active constraint.
will be the robustness. Chen et al. [2] formulated
robust optimization problem such as to make η Apart from the proposition of [2], another
more than a target value ηtarget (shown in Fig. 3). In popular formulation regarding the CP, i.e. the
this proposition, the original unit circle is allowed formulation of Au et al. [5] has been also
to increase in size whose radius depicts the investigated in this study. As per this formulation,
uncertainty. η =1 presents the case of highest if the structure contains N UBB parameters, a set
variation of UBB parameters i.e. U i . ηtarget is is defined as follows:
generally taken as 1.0. But, by proper selection of  U1  U 1 U U2 
| | 1 ,| 2 | 2 ,
design variables (DVs), if it becomes possible to  U1 U2 
U(  )    (3)
achieve η > 1, the design is said to be more robust  UN  U N 
since it can endure substantial amount of  .......,| |  N 
 U 
undesirable variations which is unprecedented. N

But for η<1, the curve touches or becomes inside In the above, 𝛿i denotes the variations of the UBB
the circle which implies the unreliability of the parameters. Now, in this case, the system is said to
system. At η=1, the curve cuts the unit circle be robust, if the structure is safe for large values of
tangentially that implies the critical state of the 𝛿i, but if the structure fails for smaller values of 𝛿i
structures. then the structure is considered as unreliable. The

OPT9 | P a g e
 SEC18: Paper No. 20180048_2

formulation is composed of one main optimization elucidated by optimizing a three bar system
loop as below: system in the next section.

find X such as to,

minimize, f ( X , U ) 4. NUMERICAL STUDY
subjected to, g j ( X , U )  0 , where, j=1,2,….,M
A three bar structure problem presented in [2] is
X [ X , X ]
L U
(4) taken up (Fig. 4). The bars are of same material,
And a sub-optimization problem with the density ρ= 2.768 x10-3 kg/cm3 and the
Find U such as to, modulus of elasticity E=6.895x10-3 kN/cm2.
The distance L=50.8 cm. An external load P is
minimize, g j ( X , U ) , where, j=1,2,3,….,n
applied at the free node of the truss at an
Ui U i inclination of θ =45o with the horizontal. It is
subjected to, | |  i , i=1,2,3,….,N required to minimize weight of the truss
Ui subjected to displacement and stress constraints.
(5) The DVs are i) the cross-sectional area of the
inclined bars (A1), and ii) the cross sectional
3. PROPOSED RDO area of the vertical bar, A2. The load P,
maximum allowable tensile stress σ+ and
The primary requirement of a robust system is to compressive stress σ-, permissible limits of
obtain a set of DVs that minimizes or limits the horizontal deflection of free node (ux), and
sensitivity of the structural system without trying permissible limits of vertical deflection of free
to reduce input uncertainty levels. The concept of node (uy) are the UBB parameters. The nominal
importance factor is generally adopted to evaluate values of [P σ- σ+ ux uy] are [1779.2 kN 344.75
the gradient of the performance of structures with MN/m2 344.75 MN/m2 0.285 cm 0.217 cm],
respect to uncertain parameters involved in respectively.
constraint functions [9]. The importance factor of
gth constraint for ith parameter Iji is given by,

L L
, (6)

The RDO results yielded by equations (1) to (5)

may have high values of Iji which means the L
system is vulnerable to even small deviations of
UBB parameters. Thus, it has been felt that if a set
of DVs is possible to work out which restricts the
sensitivity measure at a level of 5% to 10%, the
θ
system will behave robustly. Hence, in this paper,
a modification to [2] and [5] has been proposed by
limiting a new sensitivity measure. The new P
sensitivity measure, SI is defined in this paper as:
SI  max I ji ;i  N;j  M  (7) Fig. 4: A three-bar structural system
Then, a constraint in the main optimization loop
(i.e. equation (1) and (4)) is imposed as : The deterministic optimization problem is
formulated as,
find X such as to, Find A1, A2 to
minimize f ( X , U )
minimize weight of the structure (W)
subjected to, g j ( X , U )  0 , where, j=1,2,…M
SI < β such that,
X [ X , X ]
L U
(8) Horizontal displacement of free node,
where, β is the limiting sensitivity based g1
importance factor, taken as 10% in the present
paper. The efficiency of the proposed procedure is Vertical displacement of free node,

OPT10 | P a g e
 SEC18: Paper No. 20180048_2

UBB parameters. However, this constant

g2
sensitivity alone is insufficient to ensure the true
Tensile stress acts on bar 1, robustness of the system, rather unless the
sensitivity measure is tuned to considerable low
g3
value (such as 10% by the proposed RDO
Tensile stress acts on bar 2, approach), the system cannot be said to provide
g4 guarantee of robustness. Thus, the proposed
RDO approach provides this guarantee of
Compressive stress acts on bar 3, robustness by ensuring the sensitivity of the
g5 system at significantly low values, which is not
available in the existing formulations of CP.
Chen et al. [2]
(9) Au et al. [5]
Proposed on [2]
100 Proposed on [5]
Then, the RDO is formulated by the existing 90 Worst Case Propagation [1]

formulations of equations (1-5) and the 80

Weight of truss (kg)

proposed formulations equations (6-8). The 70
optimization problem is solved by the 60
Sequential Quadratic Programming method in 50
MATLAB. 40

30
The RDO results by the formulation of
20
CP [2], [5], worst case uncertainty propagation
10
principle [1], proposed RDO over formulation 10 20 30 40 50

of [2] and proposed RDO over formulation of [5] Dispersion (%)

are presented through Figs. 5 and 6 for varying Fig. 5 Optimal weight of truss for varying
range of uncertainty level of the UBB
uncertainty range
parameters. Fig. 5 presents variation of optimal
weight. It can be observed that the worst case
It has been observed that the proposed
uncertainty propagation approach based on first- RDO yields improved solution by almost same
order Taylor series expansion [1] requires computational time as required in the existing
substantially higher weight than the CP approaches. Hence, the computational efficiency
approaches. Thus, this method is uneconomical. of the existing approaches are preserved in the
Moreover, this approach is not applicable proposed RDO, as well
beyond 20% uncertainty level. The proposed
RDO approach requires marginally higher Chen et al. [2]
Au et al. [5]
0.40
optimal weight (1% to 12%) than the existing Proposed on [2]
Proposed on [5]
approaches. This higher weight requirement is 0.35
Worst Case Propagation [1]
due to imposition of sensitivity constraint or to 0.30

make the system insensitive to undesirable

Sensitivity

0.25

variation of UBB parameters. 0.20

Fig. 6 depicts the variation of sensitivity 0.15

measure as an indicator of robustness. More 0.10

robustness is implied for less values of 0.05

10 20 30 40 50
sensitivity measure. As the constraint of Dispersion (%)
maximum 10% sensitivity measure was added
in the proposed RDO formulation, the proposed Fig. 6 Sensitivity of design with varying
formulation consistently shows sensitivity uncertainty range
measure at 10% level; whereas, for other
conventional approaches this parameter vary 5. CONCLUSIONS
between 25% to 33%. It may be noted that the
variation of sensitivity is almost constant even An improved RDO formulation with UBB
with higher uncertainty ranges for all the CP parameter is proposed using CP, where the
approaches, which indicate that the system robustness of the system is improved by
becomes robust to undesirable deviation of the imposing a sensitivity constraint. The

OPT11 | P a g e
 SEC18: Paper No. 20180048_2

improvement by the approach is elucidated by application for non-probabilistic

optimizing a three-bar truss system. It has been reliability-based design optimization”,
observed that from the results that the proposed Applied Mathematical Modelling, 55,
approach yields more robust solution by 2018, pp. 354–370.
limiting the sensitivity by maximum 10% upon [8] Jiang, C., Zheng, J., and Han, X.,
sacrificing marginal increase in cost (1-12%). “Probability-interval hybrid uncertainty
On the other hand, the existing conventional analysis for structures with both aleatory
formulations are all safe, but have sensitivity and epistemic uncertainties: a review”,
measure even more than 25%. Thus, the Structural and Multidisciplinary
proposed RDO approach is felt to be promising Optimization, 57, 2018, pp. 2485–2502.
in the field of design of robust structures. Also, [9] Haldar, A., and Mahadevan, S.,
the proposed RDO approach yields the optimal Reliability Assessment Using Stochastic
weight with almost same computational time Finite Element Analysis, John Wiley &
than the existing CP formulations of RDO. Sons, US, 2000.
[10] Bhattacharjya, S., and Chakraborty, S.,
6. ACKNOWLEDGEMENTS “Robust optimization of structures
subjected to stochastic earthquake with
The authors gratefully acknowledge the funding limited information on system parameter
received from the CSIR to carry out this uncertainty”, Engineering Optimization,
research work (Scheme No. 22(0779)/18/EMR- 43(12), 2011, pp. 1311-1330.
II dated 02/05/2018). [11] Chakraborty, S., Bhattacharjya, S., and
Halder, A., “Sensitivity Importance-based
REFERENCES
Robust Optimization of Structures with
Incomplete Probabilistic Information”,
[1] Wu, J., Luo, Z., Zhang, N., and Zhang, Y.,
International Journal for Numerical
“A New Interval Uncertain Optimization
Methods in Engineering, 90(10), 2012,
Method for Structures Using Chebyshev
pp. 1207 - 1320.
Surrogate Models”, Computers and
Structures, Vol. 146, 2015, pp. 185-196.
[2] Chen, X., Fan, J., and Bian, X.,
“Structural robust optimization design
based on convex model”, Results in
Physics, 7, 2017, pp. 3068-3077.
[3] Saad, L., Chateauneuf, A., and Raphael,
W., “Robust formulation for Reliability-
based design optimization of structures”,
Structural and Multidisciplinary
Optimization, 57, 2018, pp. 2233–2248.
[4] Ben-Tal, A., and Nemirovski, A. “Robust
Convex Optimization”, Mathematics of
Operations Research, 23(4), 1998, pp.
769-1024.
[5] Au, F.T.K., Cheng, Y.S., Tham, L.G., and
Zeng, G.W., “Robust design of structures
using convex models”, Computers and
Structures, 81(28–29), 2003, pp. 2611-
2619S
[6] Wang, L., Wang, X.J., Wang, R.X., and
Chen, X., “Reliability-based design
optimization under mixture of random,
interval and convex uncertainties”,
Archive of Applied Mechanics, 86(7),
2016, pp. 1341–1367.
[7] Meng, Z., Hu, H., and Zhou, H., “Super
parametric convex model and its

OPT12 | P a g e