1 Stochastic Structural Dynamics Lecture-31 Dr C S Manohar Department of Civil Engineering Professor of Structural Engineering Indian Institute of Science Bangalore 560 012 India [email protected] Monte Carlo simulation approach-7
111
Stochastic Structural Dynamics
Lecture-31
Dr C S ManoharDepartment of Civil Engineering
Professor of Structural EngineeringIndian Institute of ScienceBangalore 560 012 India
Monte Carlo simulation approach-7
22
( ) 0
1
21
( )
1
(1 )1Var (1 )
( ) 0( ) ; ( ) .
0
f X Xg x
n
ii
nF F
F Fi
XF V F h
V
XV
F
P p x dx I g x p x dx I g X
I g Xn
P PP Pnn
I g x p xP F x h x dx F x P F X
h x
I g v p vh v
P
Probability of failure
Variance reduction
3
(a) Variance reduction can be viewed as a means to use known information about the problem.
(b) If nothing is known about the problem, variance reductionis not achievable.
(c) At the o
Variance reduction
ther extreme, that is, when everything about the problem is known, variance reduces to zero but then simulation itself is not needed.
(d) How do we get information about the problem? - Perform a few cycles of brute force simulations and learn something about the problem.
4
Sub‐set simulations using Markov Chain Monte Carlo (MCMC)
• S K Au and J L Beck, 2001, Estimation of small failure probabilities in high dimension by subset simulation, Probabilistic Engineering Mechanics, 16, 263-277
• J S Liu, 2001, Monte Carlo strategies in scientific computing, Springer, NY.
5
Small failure probability can be expressed as a product of larger conditional failure probabilities.These larger conditional failure probabilities can be
estimated with lesser computation
Basic idea
al effort.The method is applicable to a wide class of problems
6
0
1
2
, , ; 0 , 0 specified
: zero mean, stationary Gaussian random process.
cos sin
where , ~ 0, , , ,&
, 1, ;
N
n n n nn
n n n n k n k
n k
my cy ky f y y t q t y y
q t
q t a t b t
a b N a a n k b b n k
a b n k N
Subset simulation : motivation
12
*
2
Let , , a metric of system performance.
We are interested in estimating P 0, .
: The system parameters could also be random .
n
n
qq nS d
z t h y t y t t
z t z t T
Note
7
0
*
*
0,
*
0,
*
*1
1 0,
max
0
0
max
, , ,
0
F
t T
m
m t T
m
Nn n n
F X
P P z t z t T
P z t z
P Z X z
P g X
Z X z t
g X z Z X
X a b z
P I g x p x dx
8
1
2ˆ
0
1ˆ 0
ˆ is an unbiased and consistent estimator of P with minimum variance. The optimal variance is given by
(1 ) .F
F X
Ni
Fi
F F
F FP
P I g x p x dx
P I g XN
P
P Pn
Remark
9
1 2
1
1
1 1
1 1
1
11
1
1 11
0 Failure event
Definesuch that
, 1, 2, ,
|
|
|
mk
k ii
m
F m ii
m m
m i ii i
m
m m ii
m
i ii
F g X
F F F F
F F k m
P P F P F
P F F P F
P F F P F
P F P F F
Subset simulations
10
1
1 11
1 1
6
|
If -s are configured such that | and are much larger than , then we will be able to estimate
in terms of product of "large" probabilities.
Suppose, ~ 10 ,
m
F i ii
i i i
F
F
F
P P F P F F
F P F F P FP
P
P
Remarks
6 1 1 1 1 1 1
then we could obtain an estimate of
as 10 ~ 10 10 10 10 10 10 .
Estimation of probability of failure of the order of 0.1 can be easily done using MCS because the failure events here are m
FP
orefrequent.
11
1
1 11
1
1
|
can be estimated using a "brute force" Monte Carlo.
| , 1, 2, , 1 can be estimated using MCMC.
m
F i ii
i i
P P F P F F
P F
P F F i m
Remarks (continued)
12
th
1. Run a brute force Monte Carlo using, say, 200 samples. Evaluate the realization of the performance function at these 200 points. Rank order the these realizations and pick the 20 ranke
Steps
1
* *1 1 1
1 1
1
d member and denote the performance function as . Define a new performance function .
Define 0ˆ Clearly, Estimate of 0 0.1.
2. Store 20 members of which lieF
g g X g X g
F g X
P P g X
X
1
1
in the failure region of .3. Run 20 episodes of MCMC with each episode commencing from one of the 20 points in faiure region of . In each run continue with the simulaitons till 9 points are
g X
g X
1
obtained in failure region of .g X
13
1
th
*2
4. This leads to 200 points in failure region of . Rank order the value
of ( ) at these 200 points and identify the 20 ranked member and denoteit by . Define a new performance
g X
g Xg
Steps (Continued)
2
*2 2
2 2
2 1
1 1
function .
Define 0ˆ Clearly, Estimate of 0 | 0 0.1.
5. Repeat this exercise till is reached.6. Obtain the final probability of failure by using
F
m
F i
g X g X g
F g X
P P g X g X
F F
P P F P F
1
1
|m
ii
F
14
RemarksThe definition of -s (as in the present illustrative explanation)
ensures that -s are all equal to 0.1.
Estimates for sampling variance can be deduced. Choice of proposal density functio
i
i
F
FP
n: In standard normal space, typically shifted normal pdf.
15
1,10
10
1
2
1 2 4 5 6 10
Let max
: zero mean Gaussian random variables with
covariance matrix given by
1 1,10
0 , 1,10 excepting
0.3; 0.4; 0.2
Estimate 5 using subset m
m ii
i i
i
i j
X
X X
X
X i
X X i j
X X X X X X
P
Example
Questionsimulations.
16
Number of samples: 200 at each subset
Proposal pdf | ~N ,i iq X x x I
17
0 1 2 3 4 5 610
-5
10-4
10-3
10-2
10-1
100
level
Fai
lure
pro
babi
lity
Level
1 FP
5
Blue line:Simulation with 10 samples110
210
310
410
18
Run PF
1 2.5388 1.5394 0.8291 0.1154 0.0 6.95E‐05
2 2.4819 1.6062 0.8591 0.1662 0.0 5.75E‐05
3 2.4454 1.4920 0.6616 0.0 ‐ 1.00E‐04
4 2.2659 1.2125 0.4420 0.0 ‐ 2.65E‐04
*1g *
2g *3g *
4g *5g
19
25
1
0 10
cos sin
1 1~ iid N 0, ; ~ iid N 0,2 2
, 1, 252
max
What is P 8 ?
n n n nn
n n
n k
n
m t
m
X t a t b t
a b
a b n kn
X X t
X
Example
Question
20
5
Number of samples: 200 per subset
Proposal pdf | ~N ,0.4
Brute force Monte Carlo with 10 samplesi iq X x x I
21
0 1 2 3 4 5 6 7 8 9 1010
-5
10-4
10-3
10-2
10-1
100
level
Fai
lure
pro
babi
lity
4*
12.9187, 1.1528, 0.4760, 0i i
g
22
Preliminaries
Let be a deterministic function defined over .2
Let us assume that ( ) is well behaved in a suitable sen
Tf t t
f t
Series representation for random processes :Karhunen - Loeve expans
revisitedion
1
2
2
se.
Consider a sequence of functions which satisfy
completeness requirements and the orthogonality conditionsn n
T
n k nkT
t
t t dt
23
1
22
12
can be expressed in terms of the convergent series
with a measure of error of representation given bytotal meansquare error given by
= .
The constants can b
n nn
T
n nnT
n
f t
f t b t
f t b t dt
b
e determined using the conditions
0; 1, 2, ,k
kb
24
2
2
0; 1, 2, ,
; 1, 2, ,
QuestionCan similar formulation be developed for representingrandom process ( )?
H K Van Trees, 2001, Detection, estimation, and modulationthe
kT
k kT
kb
b t f t dt k
x t
Reference :
ory, Vol. I, John Wiley, NY pp. 178-198.
2525
01
Let be a zero mean, stationary, Gaussian random process defined as
cos sin ;
~ 0, , ~ 0, , 0 , 0 ,
n n n n nn
n n n n n k n k
X t
X t a t b t n
a N b N a a n k b b n k
a
RecallFourier representation of a Gaussian random process
1
2
1 1
2
0 , 1,2, ,
cos sin 0
cos ;
2
n k
n n n nn
XX n n XX n n nn n
n nn
b n k
X t a t b t
R S S
S
26
01
If ( ) is mean square periodic we can use the Fourierrepresentation with uncorrelated coefficients.
cos sin ;
Can we obtain series representations with uncorrelatedcoeffcients w
n n n n nn
X t
X t a t b t n
hen ( ) is not mean square periodic?Or, more generally, when ( ) is not even stationary?How can we proceed if ( ) is non-Gaussian?
X tX t
X t
27
1
1
Consider ( ) to be a zero mean Gaussian random process-not necessarily stationary-not necessarily mean square periodicConsider the series
; 2
Here are a set of random variables a
n nn
n n
x t
Tx t a t t
a
1
2 2
2 2
1
1
nd
are a set of deterministic functions such that
We would like to select such that .
0 0
n n
T T
n m nm k kT T
n n k n nkn
n n nn
t
t t dt a t x t dt
t a a
a x t a t
28
1
1 11
1 11
2
1 1
2
2
1 1
2
; 2
If we impose the requirement we get
n nn
n nn
k k n nn
k n k nk
T
k k kT
T
k k kT
Tx t a t t
x t a t
a x t a a t
a a
x t t x t dt t
t x t x t dt t
29
2
2
, ;2
This is an integral eigenvalue problem.The kernel , is nonnegative definite.
eigenfunction; =eigenvalueExact solutions are available for a few cases.Numeri
T
xxT
xx
TR t d t t
R t
t
Remarks
cal solutions can be obtained by usingGalerkin's method
30
2 2
2exp ;
exp
exp exp
Differentiate with respect to
exp
exp
xx xx
T
Tt T
T t
t
TT
t
PR P S
P t u u du t
P t u u du P u t u du t
t
P t u u du P
P u t u du P t
Example
31
2
2
exp exp
exp exp exp exp
Differentiate with respect to
exp exp exp exp
exp exp exp
t T
T tt T
T t
t
T
P t u u du P u t u du t
t P t u u du P t u u du
t
t P t u u du P t t t
P t u u du P t
2
2
exp
2 exp
2
T
tT
T
t t
P P t u u du
P t t
32
2
2
2
2 2
1 2
2
2
0
2
0 with
exp expIt can be shown that (Exercise) - s are roots of the equation
tan tan 0
t P t t
P
t t
P
t b t b
t c ibt c ibtb
bbT bTb
33
2 2
0.5
0.5
2 ; 1,2, ,
cos odd
sin 212
sin even ;
sin 212
The eigenfuncitons are sinusoids (as in Fourier series) but the frequencies are not uni
ii
ii
i
i
ii
i
i
P ib
b tt i
bTTbT
b tt i t T
bTTbT
Remark
formly spaced.
34
2
2
2 2 2
sin, for
sin
1 2 0; 1
; eigenvalue2
Eigenfunctions: angular prolate spheroidal functio
xx xx
T
T
t u PR t u P St u
t ut P u du
t u
t f t tf t c t f t t
Tc
Example : Bandlimited white noise process
ns
35
2
2 2 2
0 0
2 2 2 2
22
2 2
Consider to be the Brownian motion process defined over 0 .
0; , min ,
min ,
0
xx
T t T
tT T
t t
n
x tt T
x t R t u t u
t t u u du u u du t u du
t t t u du t t u du
t u
Tn
Example
12
2 2
2; sin 0.5 ;00.5
1,2, ,
ntt n t T
T T
n
36
Let ( ) be a random process whose first order pdf andthe ACF functions are available. No further
X t
Series represetation of partially specified non - Gaussianrandom processes using Nataf's transformation
informaiton about the process is available.
need not be stationary.How to represent ( ) in a series?X t
X t
37
2
Define so that
0 & 1.
Introduce a new random process Z(t) through the transformation
Here PDF of N 0,1 random variable.
is a zero mean Gaussian random process with a
X
X
Y
X t m tY t
t
Y t Y t
Z t P Y t
Z t
n unknowncovariance function.
38
1
1 1 *1 2 1 2 1 2 1 2
1 11
*1 22 1 2 1 2 1 2
*1 2
*1 2 1 2
, ;0,
, , ;0,
RHS is known and , is not known
, 1& ,
,
1
Y
Y
Y Y
XX Y Y
XX
Z t P Y t
Y t P Z t
Y t Y t P z P z z z dz dz
t t P z P z z z dz dz
t
t
t
t
t t
t
t
Remarks
*1 2 1 2, ;0, , 2 dimensional Gaussian pdfz z t t
39
2
2
1
1
1
Solve the eigenvalue problem
, ;2
by using numerical methods.
T
zzT
n nn
X X Y n nn
TR t d t t
Z t a t
X t m t t P a t
40
2 2 2 2
2 2 2 2
0
0
2 2
2 42 200
2 2
12 200
( ) ( ) ( , ) ,
,0
,0
;
;
x x lx x l
xx x l
y y y yEI x P t m x c x f x t x tx x x t t
y x y x
y x y x
y y y yEI x k EI x kx x x x
y yEI x k y EI xx x x x
2 0xk y
Monte Carlo simulation of response of systemswith spatially distributed random parameters
41
th4 order, 2-point stochastic boundary value problemEvolution of randomness in space and timeMarkovian properties in space is not possibleDiscretization of random fields is also essentialN
Remarks
atural frequencies, modeshapes, Green's functionsare all stochastic in nature.
( ), ( ), and ( ) cannot take negative values Gaussian models are not valid (especially when considering prob
EI x m x c x
lem of reliability evaluation)
42
Approach: employ KL expansions for ( ), ( ), and ( ).
Note: These processes are non-Gaussian in nature. Assume that they are independent.Discretization using KL-expansion and Nataf's trans
EI x m x c x
1
1
2
2
3
3
1
1
KL expansion
1
1
1
1
formationN
EI EI Y n nn
N
m m Y n nn
N
c c Y n nn
EI x m x x P a x
m x m x x P b x
c x m x x P d x
43
1
1
,
modeshapes of the system with
deterministic propertiesUse method of wieghted residues (e.g., Galerkin's method)
to getM along with associat
N
n nn
Nn n
y x t t x
x
C K F t
ed ics.
, , = random matrices (fully populated)Starting point for applicaiton of methods such as the subset
simulations
M C K
44
Summary
• Simulations of random variables and random processes
• Fourier and KL expansions• Introduction to statistical inference and estimation theory
• Introduction to calculus of Brownian motion and implications on numerical simulations
• Estimation of low probability of failure• Variance reduction: adaptive procedures• Discretization of spatially varying random quantities.