Model-based design of optimal experiments for guaranteed ...€¦ · Model-based design of optimal...
Transcript of Model-based design of optimal experiments for guaranteed ...€¦ · Model-based design of optimal...
Model-based design of optimal experiments for
guaranteed parameter estimation of nonlinear
dynamic systems
A.R. Gottu Mukkula and R. Paulen
Technische Universitat Dortmund
September 29, 2016
SCAN ’16
1 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Motivation
Classical Parameter Estimation
p1
p2
P0
t
u(t)
Optimal Experiment Design
Guaranteed Parameter Estimation
p1
p2
Pe
P0
t
u(t)
?
Optimal Experiment Design
SCAN ’16
2 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Classical Parameter Estimation
model: y = F (p)
{
dynamic system: x = f (x ,u,p), x(0) = g(p)
output equation: y = h(x ,p)
SCAN ’16
3 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Classical Parameter Estimation
model: y = F (p)
{
dynamic system: x = f (x ,u,p), x(0) = g(p)
output equation: y = h(x ,p)
y
t
SCAN ’16
3 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Classical Parameter Estimation
model: y = F (p)
{
dynamic system: x = f (x ,u,p), x(0) = g(p)
output equation: y = h(x ,p)
y , y
t p1
p2
P0
SCAN ’16
3 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Classical Parameter Estimation
minp∈P0
‖y − y‖22
s.t. y = F (p)
y , y
t p1
p2
P0
SCAN ’16
3 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Classical Parameter Estimation
minp∈P0
‖y − y‖22
s.t. y = F (p)
y , y
t
y(e) ∈ y + [ e, e ]
p1
p2
P0
SCAN ’16
3 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Classical Parameter Estimation
e ∈ [e,e] : minp∈P0
‖y+e − y‖22
s.t. y = F (p)
y(e), y
t p1
p2
P0
SCAN ’16
3 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Classical Parameter Estimation
e ∈ [e,e] : minp∈P0
‖y+e − y‖22
s.t. y = F (p)
y(e), y
t p1
p2
P0
SCAN ’16
3 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Guaranteed Parameter Estimation (GPE)
∀e ∈ [e,e] : minp∈P0
‖y+e − y‖22
s.t. y = F (p)
y , y
t p1
p2
Pe
P0
SCAN ’16
4 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Guaranteed Parameter Estimation (GPE)
findp∈P0
all p
s.t. e ≤ y − y ≤ e
y = F (p)
y , y
t p1
p2
Pe
P0
SCAN ’16
4 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Guaranteed Parameter Estimation (GPE)
findp∈P0
all p
s.t. e ≤ y − y ≤ e
y = F (p)
y , y
tp1
p2
Pe
P0
SCAN ’16
4 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Model-based Optimal Experiment Design (OED)
Idea: Improve the reliability of parameter estimates by
optimizing some measure of Fisher Information Matrix (FIM)
minu(t)
φ (FIM) s.t. FIM =
Ne∑
k=1
∂y(k)
∂pTQ∂y(k)
∂p, y = F (p,u)
t
u(t)
p1
p2
P0
A design: φ(FIM) = tr(FIM), D design: φ(FIM) = det(FIM), . . .
SCAN ’16
5 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Optimal Experiment Design (OED) for GPE
The classical least-squares and guaranteed parameter
estimation yield different results.
y , y
tp1
p2
Pe
P0
How to perform OED in the context of GPE?
SCAN ’16
6 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Optimal Experiment Design (OED) for GPE
1 Find an over-approximation of GPE solution set
maxp/p
np∑
j=1
pj − pj
s.t. y(t) = F (π), ∀π ∈ {E(p),p/p}
ei ≤ y(ti ,E(p))− y(ti ,pj/pj) ≤ ei , ∀j ∈ {1, . . . ,np}
t
y , y
p1
p2
Pe
E(p)p
1
p2
p1
p2
SCAN ’16
7 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Optimal Experiment Design (OED) for GPE
2 Optimize over the over-approximation of GPE solution set
minu
maxp/p
np∑
j=1
pj − pj
(A design)
s.t. y(t) = F (π), ∀π ∈ {E(p),p/p}
ei ≤ y(ti ,E(p))− y(ti ,pj/pj) ≤ ei , ∀j ∈ {1, . . . ,np}
t
y , y
p1
p2
Pe
E(p)p
1
p2
p1
p2
SCAN ’16
7 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Optimal Experiment Design (OED) for GPE
2 Optimize over the over-approximation of GPE solution set
minu
maxp/p
np∏
j=1
pj − pj
(D design)
s.t. y(t) = F (π), ∀π ∈ {E(p),p/p}
ei ≤ y(ti ,E(p))− y(ti ,pj/pj) ≤ ei , ∀j ∈ {1, . . . ,np}
t
y , y
p1
p2
Pe
E(p)p
1
p2
p1
p2
SCAN ’16
7 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Optimal Experiment Design for GPE
Reformulation into Mathematical Program with Equilibrium
Constraints (MPEC)
minu,λ,λ
np∑
j=1
pUj − pL
j
s.t. y(t) = F (π,u),
0 = ∇πL(x(t),u(t), π, λπλπ)
ei ≤ y(ti ,E(p))− y(ti ,pj/pj) ≤ ei
0 = diagλij(y(ti ,pj/pj)− y(ti ,E(p)) + ei)
0 = diagλij(y(ti ,E(p))− y(ti ,pj/pj)− ei)
0 ≤ λ, λ
u ≤ u(t) ≤ u
SCAN ’16
8 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Optimal Experiment Design for GPE
Regularization of the MPEC (Hatz et al., 2013)
minu,λ,λ
w ,w
np∑
j=1
pUj − pL
j + ρ‖w‖1 + ρ‖w‖1
s.t. y(t) = F (π,u),
0 = ∇πL(x(t),u(t), π, λπλπ)
ei ≤ y(ti ,E(p))− y(ti ,pj/pj) ≤ ei
−w ij≥diagλij(y(ti ,pj/pj)− y(ti ,E(p)) + ei)
−w ij≥diagλij(y(ti ,E(p))− y(ti ,pj/pj)− ei)
0 ≤ λ, λ,w ,w
u ≤ u(t) ≤ u
SCAN ’16
8 / 16Process Dynamicsand Operations
DN
YDD
NNYY
OED for GPE: Case Study 1
Reaction A+Bk→C run in isothermal semi-batch reaction with the
feed of component B (Srinivasan et al., 2003)
dcA
dt= −kcAcB −
u
VcA
dcB
dt= −kcAcB +
u
V(cB,in − cB)
dV
dt= u, Tj = T +
−∆HkcAcBV
αA
u, cB,in
Tj,in
Tj
Outputs cA, cB ,Tj sampled at ten equidistant time instants,
corrupted with zero-mean noise with standard deviation of
0.005 mol/L and 0.05 K respectively
Estimated parameters (k ,∆H) = (0.05784 L/mol/h, −72 kJ/mol)
Feed u(t) ∈ [0,1]
SCAN ’16
9 / 16Process Dynamicsand Operations
DN
YDD
NNYY
OED for GPE: Case Study 1
Optimal A design for GPE (red) compared to classical A design
(black) confindence ellipsoid
0 0.2 0.4 0.6 0.8 1 1.2
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1Classical A design
GPE A design
time [h]
u(t)
0.05 0.055 0.06 0.065-8
-7.5
-7
-6.5
k [l/mol h]
∆H
[10
4J/m
ol]
Solved with CasADi + Ipopt + CVODE (sequential approach)
GPE solution sets obtained from CRONOS (Chachuat et al.)
SCAN ’16
10 / 16Process Dynamicsand Operations
DN
YDD
NNYY
OED for GPE: Case Study 2
Lotka-Volterra (predator-prey) model modified to control the prey
population
dx1
dt= −kx (x1 − u(t)), x1(0) = 10,
dx2
dt= (0.05x1 − d)x2, x2(0) = 10.
Output x2 sampled at equidistant time instants, corrupted with
zero-mean noise with standard deviation of 0.1
Estimated parameters (kx ,d) = (0.9,1.5)
Stimulus of growth u(t) ∈ [0,1]
SCAN ’16
11 / 16Process Dynamicsand Operations
DN
YDD
NNYY
OED for GPE: Case Study 2
Optimal A design for GPE (purple) compared to classical A
design (green); “classical” confidence ellipsoids included
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
control input for classical A−design
control input for GPE A−design
t
u(t)
0.82 0.86 0.9 0.94 0.981.42
1.46
1.5
1.54
1.58
99% confidence ellipsoid for classical A−design
99% confidence ellipsoid for GPE A−design
GPE solution set for classical A−design
GPE solution set for GPE A−design
kx
d
Solved with CasADi + Ipopt + CVODE (sequential approach)
GPE solution sets obtained from CRONOS (Chachuat et al.)
SCAN ’16
12 / 16Process Dynamicsand Operations
DN
YDD
NNYY
OED for GPE: Case Study 3
Lotka-Volterra (predator-prey) model modified to control the prey
population
dx1
dt= x1 − p1x1x2 − 0.4x1u, x1(0) = 0.5,
dx2
dt= −x2 + p2x1x2 − 0.2x2u, x2(0) = 0.7.
Outputs x1 and x2 sampled at optimized time instants (τ1, τ2) are
corrupted with zero-mean noise with standard deviation of 0.1
Estimated parameters (p1,p2) = (1,1)
Fishing (yes/no) u(t) ∈ {0,1} (binary)
SCAN ’16
13 / 16Process Dynamicsand Operations
DN
YDD
NNYY
OED for GPE: Case Study 3
Optimal D design for GPE (green) compared to classical D
design (purple); “classical” confidence ellipsoids included
0 2 4 6 8 10 12
0
1
time
u(t)
Classical D designGPE D design
0.85 0.9 0.95 1 1.05 1.1 1.15
0.85
0.9
0.95
1
1.05
1.1
1.15
p1
p2
GPE D designClassical D design
Solved with CasADi + Ipopt + CVODE (sequential approach)
GPE solution sets obtained from CRONOS (Chachuat et al.)
SCAN ’16
14 / 16Process Dynamicsand Operations
DN
YDD
NNYY
OED for GPE: Case Study 3
Interesting observation: Optimal D design for GPE is a local
optimum of classical D design
43
21
06
8
10
0
-400
-1200
-1000
-800
-600
-200
12
τ1τ2
−d
et(
FIM
)
SCAN ’16
15 / 16Process Dynamicsand Operations
DN
YDD
NNYY
Conclusions
Problem of optimal experiment design for guaranteed parameter
estimation formulated as a bilevel program
The approximation of the set of guaranteed parameter
estimates realized as a box
Computationally intensive problem but tractable for small-scale
cases
Classical optimal experiment design does not cope well with the
nonlinearity in the model → Results of optimal experiment for
guaranteed parameter estimation differ from the classical
experiment design
Acknowledgement:
SCAN ’16
16 / 16Process Dynamicsand Operations
DN
YDD
NNYY