1/ 73 Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks How to predict an epidemic of Zika virus? A challenge in nonlinear stochastic dynamics Americo Cunha Jr Universidade do Estado do Rio de Janeiro – UERJ NUMERICO – N´ ucleo de Modelagem e Experimenta¸c˜ ao Computacional http://numerico.ime.uerj.br In collaboration with: Eber Dantas (UERJ) Michel Tosin (UERJ) COLMEA 2019 - IM-UFRJ March 28, 2019 Rio de Janeiro, Brazil
103
Embed
How to predict an epidemic of Zika virus?coloquiomea/apresentacoes/cunha_2019.pdf · C. Manore and M. Hyman, Mathematical Models for Fighting Zika Virus, SIAM News, May 2016. 10
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
1/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
How to predict an epidemic of Zika virus?A challenge in nonlinear stochastic dynamics
Americo Cunha Jr
Universidade do Estado do Rio de Janeiro – UERJ
NUMERICO – Nucleo de Modelagem e Experimentacao Computacional
http://numerico.ime.uerj.br
In collaboration with: Eber Dantas (UERJ)Michel Tosin (UERJ)
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Calibrated model response
10 20 30 40 50time (weeks)
0
100
200
300
num
ber
of p
eopl
e
103
datamodel
cumulative number of infectious
0
6
13
19
25
num
ber
of p
eopl
e
103
10 20 30 40 50time (weeks)
datamodel
new infectious cases
30/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Calibrated model response
10 20 30 40 50time (weeks)
0
100
200
300
num
ber
of p
eopl
e
103
datamodel
cumulative number of infectious
0
6
13
19
25
num
ber
of p
eopl
e
103
10 20 30 40 50time (weeks)
datamodel
new infectious cases
Robust predictions demandssome kind of “certification”.
31/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Section 4
Sensitivity Analysis
32/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Parametric study for βh
10 20 30 40 50time (weeks)
0.0
100.0
200.0
300.0
400.0
num
ber
of p
eopl
e
104
85%90%95%nominal105%110%
cumulative number of infectious
0.0
25.0
50.0
75.0
100.0
num
ber
of p
eopl
e
103
10 20 30 40 50time (weeks)
85%90%95%nominal105%110%
new infectious cases
33/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Parametric study for αh
10 20 30 40 50time (weeks)
0.0
50.0
100.0
150.0
200.0
num
ber
of p
eopl
e
103
85%90%95%nominal105%110%
cumulative number of infectious
0.0
7.5
15.0
22.5
30.0
num
ber
of p
eopl
e
103
10 20 30 40 50time (weeks)
85%90%95%nominal105%110%
new infectious cases
34/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Parametric study for γ
10 20 30 40 50time (weeks)
0.0
50.0
100.0
150.0
200.0
num
ber
of p
eopl
e
103
85%90%95%nominal105%110%
cumulative number of infectious
0.0
6.3
12.5
18.8
25.0
num
ber
of p
eopl
e
103
10 20 30 40 50time (weeks)
85%90%95%nominal105%110%
new infectious cases
35/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Parametric study for βv
10 20 30 40 50time (weeks)
0.0
50.0
100.0
150.0
200.0
num
ber
of p
eopl
e
104
85%90%95%nominal105%110%
cumulative number of infectious
0.0
12.5
25.0
37.5
50.0
num
ber
of p
eopl
e
103
10 20 30 40 50time (weeks)
85%90%95%nominal105%110%
new infectious cases
36/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Parametric study for αv
10 20 30 40 50time (weeks)
0.0
75.0
150.0
225.0
300.0
num
ber
of p
eopl
e
103
85%90%95%nominal105%110%
cumulative number of infectious
0.0
7.5
15.0
22.5
30.0
num
ber
of p
eopl
e
103
10 20 30 40 50time (weeks)
85%90%95%nominal105%110%
new infectious cases
37/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Parametric study for δ
10 20 30 40 50time (weeks)
0.0
75.0
150.0
225.0
300.0
num
ber
of p
eopl
e
103
85%90%95%nominal105%110%
cumulative number of infectious
0.0
6.3
12.5
18.8
25.0
num
ber
of p
eopl
e
103
10 20 30 40 50time (weeks)
85%90%95%nominal105%110%
new infectious cases
38/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Variance-based sensitivity analysis
Mathematical model:
Y =M(X), Xi ∼ U(0, 1), (i.i.d.)
Hoeffding-Sobol’ decomposition:
Y =M0 +∑
1≤i≤nMi (Xi ) +
∑1≤i<j≤n
Mij(Xi ,Xj) + · · ·+M1···n(X1 · · ·Xn)
An orthogonal decomposition in terms of conditional expectations:
M0 = E {Y }Mi (Xi ) = E
{Y |Xi
}−M0
Mij(Xi ,Xj) = E{Y |Xi ,Xj
}−M0 −Mi −Mj
etc
39/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Sobol’ indices
Total variance:
D = Var[M(X )
]=∑
u⊂{1,··· ,k}
Var[Mu(Xu)
]First order Sobol’ indices:
Si = Var[Mi (Xi )
]/D
(quantify the additive effect of each input separately)
Second order Sobol’ indices:
Sij = Var[Mij(Xi ,Xj)
]/D
(quantify interaction effect of inputs Xi and Xj)
40/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Metamodelling via Polynomial Chaos
Assuming Y =M(X) has finite variance, then it admits aPolynomial Chaos expansion
Y =∑α∈Nk
yα Φα(X)
where
Φα(X): multivariate orthonormal polynomials
yα: real-valued coeficients to be determined
D. Xiu, and G. Karniadakis, The Wiener-Askey Polynomial Chaos for Stochastic Differential Equa-
tions. SIAM Journal on Scientific Computing, 24: 619-644, 2002.
41/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
PC-based Sobol’ indices
For computational purposes, a truncated PCE is employed
Y ≈∑α∈A
yα Φα(X)
Thus, Sobol’ indices are given by
Su = Du/D =∑α∈Au
y2α
/∑α∈A\0
y2α
Au ={α ∈ A : i ∈ u⇐⇒ αi 6= 0}
Sobol’ indices of any order can be obtained, analytically, from thecoefficients of the PC expansion!
B. Sudret, Global sensitivity analysis using polynomial chaos expansions. Reliability Engineering &
System Safety, 2016, 93(7): 964–979, 2008.
42/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Global sensitivity analysis: first order
43/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Global sensitivity analysis: second order
44/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Global sensitivity analysis: total order
45/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Global sensitivity analysis: general overview
→ Two most relevant: δ and βH (75% variance around 7th EW )
→ Third most, γ, mainly by nonlinear interactions with δ and βH
→ Parameters limited to nonlinear interactions have, in general,delayed effects (significant for EW > 15)
→ (sparsity-of-effects principle) Higher order interactions haveminor effect: 1st and 2nd are 99.8–96.7% variance on 5–10th EW
Around 7th EW → uncertainty propagation of {βh,δ}
46/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Section 5
Uncertainty Quantification
47/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Uncertainty Quantification (UQ) framework
Mathematical model:Y =M(X)
General steps for UQ:
1 Stochastic modeling−→ characterization of inputs uncertainties
(MaxEnt Principle)
2 Uncertainty propagation−→characterization of output uncertainties
(Monte Carlo Method)
3 Response certification−→ specification of reliability levels for predictions
(Nonparametric Statistical Inference)
C. Soize, Uncertainty Quantification: An Accelerated Course with Advanced Applications in Com-
putational Engineering, Springer, 2017.
48/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Maximum Entropy Principle (MaxEnt)
Among all the probability distributions, consistent with the knowninformation about a random parameter, choose the one whichcorresponds to the maximum of entropy (MaxEnt).
MaxEnt distribution = most unbiased distribution
Entropy of the random variable X is defined as
S (pX ) = −∫RpX (x) ln pX (x) dx ,
“measure for the level of uncertainty”
49/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
MaxEnt optimization problem
Maximize
S (pX ) = −∫RpX (x) ln pX (x) dx ,
respecting N + 1 constraints (known information) given by∫Rgk (X ) pX (x) dx = mk , k = 0, · · · ,N,
where the gk are known real functions, with g0(x) = 1.
MaxEnt general solution
pX (x) = 1K(x) exp (−λ0) exp
− N∑k=1
λk gk(x)
49/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
MaxEnt optimization problem
Maximize
S (pX ) = −∫RpX (x) ln pX (x) dx ,
respecting N + 1 constraints (known information) given by∫Rgk (X ) pX (x) dx = mk , k = 0, · · · ,N,
where the gk are known real functions, with g0(x) = 1.
MaxEnt general solution
pX (x) = 1K(x) exp (−λ0) exp
− N∑k=1
λk gk(x)
50/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Philosophy of MaxEnt Principle
The parameter of interesthas a unknown distribution
Distributions arbitrarilychosen can be coarse andbiased
A conservative strategy isto use the most unbiased(MaxEnt) distribution
real
50/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Philosophy of MaxEnt Principle
The parameter of interesthas a unknown distribution
Distributions arbitrarilychosen can be coarse andbiased
A conservative strategy isto use the most unbiased(MaxEnt) distribution
realbiased
50/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Philosophy of MaxEnt Principle
The parameter of interesthas a unknown distribution
Distributions arbitrarilychosen can be coarse andbiased
A conservative strategy isto use the most unbiased(MaxEnt) distribution
realbiased
unbiased
51/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Uncertainty propagation through the model
Monte Carlo Method
pre-processing
generationof scenarios
X1
...
XM
known FX
generator ofrandom vector X
processing
solution ofmodel equations
U = h(X)
computationalmodel
deterministic solverof u = h(x)
post-processing
computationof statistics
U1 = h(X1)
...
UM = h(XM )
estimated FU
statistical inferenceto estimate convergenceand distribution of U
52/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Probabilistic model 1
Random variables: βh and δ
Available information: support and mean (nominal) value
MaxEnt distribution
pX (x) = 1[a,b](x) exp (−λ0 − λ1 x)
“truncated exponential (2 parameters)”
53/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Confidence band for the QoIs
10 20 30 40 50
time (weeks)
0.0
1.5
3.0
4.5
6.0
num
ber
of p
eopl
e
105
cumulative number of infectious
10 20 30 40 50
time (weeks)
0.0
0.8
1.5
2.3
3.0
num
ber
of p
eopl
e
104
95% prob. mean nominal data
new infectious cases
54/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Probabilistic model 2
Random variables: βh and δ
Available information: support, mean (nominal) value and dispersion
MaxEnt distribution
pX (x) = 1[a,b](x) exp(−λ0 − λ1 x − λ2 x
2)
“truncated exponential (3 parameters)”
55/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Confidence band for the QoIs
βh dispersion = 5% , δ dispersion = 5%
10 20 30 40 50
time (weeks)
0.0
1.5
3.0
4.5
6.0
num
ber
of p
eopl
e
105
cumulative number of infectious
10 20 30 40 50
time (weeks)
0.0
0.8
1.5
2.3
3.0
num
ber
of p
eopl
e
104
95% prob. mean nominal data
new infectious cases
55/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Confidence band for the QoIs
βh dispersion = 10% , δ dispersion = 5%
10 20 30 40 50
time (weeks)
0.0
1.5
3.0
4.5
6.0
num
ber
of p
eopl
e
105
cumulative number of infectious
10 20 30 40 50
time (weeks)
0.0
0.8
1.5
2.3
3.0
num
ber
of p
eopl
e
104
95% prob. mean nominal data
new infectious cases
55/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Confidence band for the QoIs
βh dispersion = 10% , δ dispersion = 10%
10 20 30 40 50
time (weeks)
0.0
1.5
3.0
4.5
6.0
num
ber
of p
eopl
e
105
cumulative number of infectious
10 20 30 40 50
time (weeks)
0.0
0.8
1.5
2.3
3.0
num
ber
of p
eopl
e
104
95% prob. mean nominal data
new infectious cases
56/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Probabilistic model 3
Random variables: βh, δ and σ
Available information for βh and δ: support, mean (nominal) value
Distribution for βh and βv
pX (x) = 1[a,b](x) exp(−λ0 − λ1 x − λ2 x
2)
Available information for σ: support
MaxEnt distribution for σ
pX (x) = 1[a,b](x)1
b − a
“uniform”
57/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Confidence band for the QoIs
random dispersion ∼ U(5%, 10%)
10 20 30 40 50
time (weeks)
0.0
1.5
3.0
4.5
6.0
num
ber
of p
eopl
e
105
cumulative number of infectious
10 20 30 40 50
time (weeks)
0.0
0.8
1.5
2.3
3.0
num
ber
of p
eopl
e
104
95% prob. mean nominal data
new infectious cases
58/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Confidence band for the QoIs
no dispersion
10 20 30 40 50
time (weeks)
0.0
1.5
3.0
4.5
6.0
num
ber
of p
eopl
e
105
10 20 30 40 50
time (weeks)
0.0
0.8
1.5
2.3
3.0
num
ber
of p
eopl
e
104
95% prob. mean nominal data
σ = {5%,5%}
10 20 30 40 50
time (weeks)
0.0
1.5
3.0
4.5
6.0
num
ber
of p
eopl
e
105
10 20 30 40 50
time (weeks)
0.0
0.8
1.5
2.3
3.0
num
ber
of p
eopl
e
104
95% prob. mean nominal data
σ = {10%,5%}
10 20 30 40 50
time (weeks)
0.0
1.5
3.0
4.5
6.0
num
ber
of p
eopl
e
105
10 20 30 40 50
time (weeks)
0.0
0.8
1.5
2.3
3.0nu
mbe
r of
peo
ple
104
95% prob. mean nominal data
σ = {10%,10%}
10 20 30 40 50
time (weeks)
0.0
1.5
3.0
4.5
6.0
num
ber
of p
eopl
e
105
10 20 30 40 50
time (weeks)
0.0
0.8
1.5
2.3
3.0
num
ber
of p
eopl
e
104
95% prob. mean nominal data
σ ∼ U(5%, 10%)
10 20 30 40 50
time (weeks)
0.0
1.5
3.0
4.5
6.0
num
ber
of p
eopl
e
105
10 20 30 40 50
time (weeks)
0.0
0.8
1.5
2.3
3.0
num
ber
of p
eopl
e
104
95% prob. mean nominal data
59/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Evolution of QoIs PDFs
random dispersion ∼ U(5%, 10%)
0.0 1.5 3.0 4.5 6.0
number of people
0.0
0.4
0.7
1.1
1.5
prob
abili
ty
10-4
105
57101520
2530354552
cumulative number of infectious
0.0 0.8 1.5 2.3 3.0
number of people
0.0
0.8
1.5
2.3
3.0
prob
abili
ty
10-3
104
57101520
2530354552
new infectious cases
60/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Time-averaged cumulative infectious
10-5
1.0 1.8 2.5 3.3 4.0 number of people
0.0
0.5
1.0
1.5
2.0
pro
babi
lity
dens
ity fu
nctio
n
105
PDFmeanmean std 95% prob.
10-5
1.0 1.8 2.5 3.3 4.0 number of people
0.0
0.5
1.0
1.5
2.0
pro
babi
lity
dens
ity fu
nctio
n
105
PDFmeanmean std 95% prob.
10-5
1.0 1.8 2.5 3.3 4.0 number of people
0.0
0.5
1.0
1.5
2.0
pro
babi
lity
dens
ity fu
nctio
n
105
PDFmeanmean std 95% prob.
no dispersion σ = {10%,10%} σ ∼ U(5%, 10%)
61/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
(mean) Cumulative infectious CDF until EW 20
σ ∼ U(5%, 10%)
1.0 1.3 1.5 1.8 2.0
number of people
0.0
0.3
0.5
0.8
1.0
prob
abili
ty
100
105
Statistics of C
mean = 1,47× 105
std. dev. = 1,53× 104
skewness = 0.084kurtosis = 2.605
P(C ≥ c∗) = 87.10%
c∗ = 130,000
Half the maximum C (data)
62/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
(mean) New cases CDF until 10th EW
σ ∼ U(5%, 10%)
1.0 1.3 1.5 1.8 2.0
number of people
0.0
0.3
0.5
0.8
1.0
prob
abili
ty
100
104
Statistics of Nw
mean = 1,57× 104
std. dev. = 1,35× 103
skewness = −0.032kurtosis = 2.656
P(Nw ≥ NC ∗) = 83.40%
NC∗ = 14,440
average NC (data) until EW 10
63/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Section 6
Ongoing
64/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Investigation of control strategies
65/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
SVEIR-SEI model for Zika virus dynamics
human population
Sh
ν
Vh
(βhI )v /Nv
(σβhI )v /Nv
Ehαh Ih
γRh
vector population
Nδv Sv
δ
(βvI )h /Nh Ev
δ
αv Iv
δ
H. S. Rodrigues et al. Vaccination models and optimal control strategies to dengue. Mathematical
Biosciences, 247 (2014) 1–12.
66/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Associated dynamical system
dSh
dt= −
(βh
IvNv
+ ν
)Sh
dVh
dt= νSh − σβh
IvNv
Vh
dEh
dt= βh (Sh + σ Vh)
IvNv−αh Eh
dIhdt
= αh Eh − γ Ih
dRh
dt= γ Ih
dSv
dt= δ − βv Sv
IhNh− δ Sv
dEv
dt= βv Sv
IhNh− (δ + αv )Ev
dIvdt
= αv Ev − δ Iv
dC
dt= αh Eh
+ initial conditionsS - Population of susceptibleV - Population of vaccinatedE - Population of exposedI - Population of infectiousR - Population of recovered
C - Infected humans cumulativeN - Total populationα - Incubation ratioβ - Transmission rateγ - Recovery rate
δ - Vector lifespan ratioσ - Infection rate of vaccinatedν - Fraction of vaccinatedh - Human-relatedv - Vector-related
H. S. Rodrigues et al. Vaccination models and optimal control strategies to dengue. Mathematical
Biosciences, 247 (2014) 1–12.
67/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Time series for QoI’s (SVEIR-SEI model)
10 20 30 40 50time (weeks)
0
100
200
300
num
ber
of p
eopl
e
103
cumulative number of infectious
0
6
13
19
25
num
ber
of p
eopl
e
103
10 20 30 40 50time (weeks)
datamodel
new infectious cases
68/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Quantification of model discrepancy
69/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Calculation of model discrepancy
Conventional statistical calibration:
y︸︷︷︸truth
= f (x ,p)︸ ︷︷ ︸model
+ ε︸︷︷︸error
Novel approach:
y︸︷︷︸truth
≈ f (x ,pε)︸ ︷︷ ︸model
, pε =∑k
αk Φk(ξ)
Bayesian inversion to identify α
π(model | data)︸ ︷︷ ︸posterior
∝ π(data | model)︸ ︷︷ ︸likelihood
×π(model)︸ ︷︷ ︸prior
K. Sargsyan, H. N. Najm and R. Ghanem, On the statistical calibration of physical models.
International Journal of Chemical Kinetics, 47 (2015) 246-276.
70/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Section 7
Final Remarks
71/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Concluding remarks
Contributions:
Development of an epidemic model to describe Brazilianoutbreak of Zika virus
Calibration of this model with real epidemic data
Construction of a parametric probabilistic model ofuncertainties
Ongoing research:
Investigate the effectiveness of different control strategies
Quantify model discrepancy in a nonparametric way
Future directions:
Scenarios exploration with active subspace method
Data-driven identification of epidemiological models
72/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Acknowledgments
Invitation for the talk:
Profa. Maria Eulalia Vares
Prof. Leandro Pimentel
Financial support:
73/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Calibration of a SEIR–SEI epidemic model to describe Zika virus outbreak in Brazil,Applied Mathematics and Computation, 338: 249-259, 2018.https://doi.org/10.1016/j.amc.2018.06.024
E. Dantas, M. Tosin and A. Cunha Jr,
Uncertainty quantification in the nonlinear dynamics of Zika virus, 2019 (in preparation).https://hal.archives-ouvertes.fr/hal-02005320
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
nominal parameters
73/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Nominal parameters and initial conditions
α value unit
αh 1/5.9 days−1
αv 1/9.1 days−1
γ 1/7.9 days−1
δ 1/11 days−1
βh 1/11.3 days−1
βv 1/8.6 days−1
N 206× 106 people
S ih 205,953,959 people
E ih 8,201 people
I ih 8,201 people
R ih 29,639 people
S iv 0.99956 ——
E iv 2.2× 10−4 ——
I iv 2.2× 10−4 ——
73/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Model response with nominal parameters
10 20 30 40 50time (weeks)
205.8
205.9
206.0
num
ber
of p
eopl
e
106
Susceptible humans
10 20 30 40 50time (weeks)
20
80
140
200
num
ber
of p
eopl
e
103
Recovered humans
5 10 15 20 25 30 35 40 45 50time (weeks)
0.99955
0.99970
0.99985
1.00000
prop
ortio
n of
vec
tors
100
Susceptible vectors
10 20 30 40 50time (weeks)
0
7
13
20
num
ber
of p
eopl
e
103
Exposed humans
10 20 30 40 50time (weeks)
0
100
200
300
num
ber
of p
eopl
e
103
datamodel
Cumulative infectious
10 20 30 40 50time (weeks)
0.0
0.8
1.7
2.5
prop
ortio
n of
vec
tors
10-4
Exposed vectors
10 20 30 40 50time (weeks)
0
1
2
3
num
ber
of p
eopl
e
104
Infectious humans
0
6
13
19
25
num
ber
of p
eopl
e
103
10 20 30 40 50time (weeks)
datamodel
New cases
10 20 30 40 50time (weeks)
0.0
0.8
1.7
2.5
prop
ortio
n of
vec
tors
10-4
Infectious vectors
73/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Inverse Problem
73/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Well-posedness
Let the forward map φ : E → F associates to each parametervector x , restricted to be on the set of admissible values C in theparameter space E , an observable vector in the data space F . TheNLS problem is Quadratically (Q-) wellposed if, and only if, φ(C )possesses an open neighborhood ϑ such that
1 Existence and uniqueness: for every z ∈ ϑ, the inverseproblem has a unique solution x
2 Unimodality: for every z ∈ ϑ, the objective function x J(x)has no parasitic stationary point
3 Local stability: the mapping z x is locally Lipschitzcontinuous from (ϑ, || · ||F ) to (C , | · ||E ).
G. Chavent. Nonlinear Least Squares for Inverse Problems: Theoretical Foundations and
Step-by-Step Guide for Applications. Springer, 2010.
73/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks
Well-posedness
Theorem
Let the follow finite dimension minimum set of hypothesis hold:
E = finite dimensional vector space, with norm || · ||E ,C = closed, convex subset of E ,
Cη = convex open neighborhood of C in E ,
F = Hilbert space, with norm || · ||F ,z ∈ F ,
φ : Cη F is twice differentiable along segments of Cη,
and: V =∂
∂tφ((1− t)x0 + tx1),A =
∂2
∂t2φ((1− t)x0 + tx1)
are continuous functions of x0, x1 ∈ Cη and t ∈ [0, 1].
Then, if moreover C is small enough for the deflection condition θ ≤ π/2 to
hold, x is OLS-identifiable on C , or equivalently: the NLS problem is
Q-wellposed on C .
G. Chavent. Nonlinear Least Squares for Inverse Problems: Theoretical Foundations and
Step-by-Step Guide for Applications. Springer, 2010.
73/ 73
Introduction Dynamic Model Inverse Problem Sensitivity Analysis Uncertainty Quantification Ongoing Final Remarks