Roadmap to Membership of RICS - Pathways and Routes
A_R_Gottu_Mukkula_Escape_26.pptx
1. Department of Biochemical and Chemical Engineering
Process Dynamics and Operations Group (DYN)
D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
13.06.2016
Optimal dynamic experiment design for
guaranteed parameter estimation
A.R. Gottu Mukkula and R. Paulen
Technische Universität Dortmund
3. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
2
Mathematical modelling
Mathematical models play an important role in:
Process design
Control system design
Process optimization
Steps for model development:
Specification of model structure
Design and realization of experiments
Estimation of uncertain model parameters
4. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
2
Mathematical modelling
Mathematical models play an important role in:
Process design
Control system design
Process optimization
Steps for model development:
Specification of model structure
Design and realization of experiments
Estimation of uncertain model parameters
5. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
2
Mathematical modelling
Mathematical models play an important role in:
Process design
Control system design
Process optimization
Steps for model development:
Specification of model structure
Design and realization of experiments
Estimation of uncertain model parameters
6. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
2
Mathematical modelling
Mathematical models play an important role in:
Process design
Control system design
Process optimization
Steps for model development:
Specification of model structure
Design and realization of experiments
Estimation of uncertain model parameters
8. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
3
What experiment(s) give the best data
Control policy, initial conditions, sampling time
Criteria for ‘the best experiment’
Most informative data
Precise parameter estimation
Efficient model discrimination
Optimal experiment design
9. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
3
What experiment(s) give the best data
Control policy, initial conditions, sampling time
Criteria for ‘the best experiment’
Most informative data
Precise parameter estimation
Efficient model discrimination
Assumption:
Usually: normal distribution of measurement error
In this work: unknown-but-bounded measurement error
Optimal experiment design
10. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
3
What experiment(s) give the best data
Control policy, initial conditions, sampling time
Criteria for ‘the best experiment’
Most informative data
Precise parameter estimation
Efficient model discrimination
Assumption:
Usually: normal distribution of measurement error
In this work: unknown-but-bounded measurement error
Optimal experiment design
11. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
3
What experiment(s) give the best data
Control policy, initial conditions, sampling time
Criteria for ‘the best experiment’
Most informative data
Precise parameter estimation
Efficient model discrimination
Assumption:
Usually: normal distribution of measurement error
In this work: unknown-but-bounded measurement error
Optimal experiment design
12. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
3
What experiment(s) give the best data
Control policy, initial conditions, sampling time
Criteria for ‘the best experiment’
Most informative data
Precise parameter estimation
Efficient model discrimination
Assumption:
Usually: normal distribution of measurement error
In this work: unknown-but-bounded measurement error
Optimal experiment design
13. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
4
Approximate the joint-confidence region of parameters in the
parametric space using confidence ellipsoid
Assumption: normally distributed measurement error
Classical parameter estimation
14. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
4
Approximate the joint-confidence region of parameters in the
parametric space using confidence ellipsoid
Assumption: normally distributed measurement error
Classical parameter estimation
15. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
4
Approximate the joint-confidence region of parameters in the
parametric space using confidence ellipsoid
Assumption: normally distributed measurement error
Classical parameter estimation
16. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
4
Approximate the joint-confidence region of parameters in the
parametric space using confidence ellipsoid
Assumption: normally distributed measurement error
Classical parameter estimation
17. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
4
Approximate the joint-confidence region of parameters in the
parametric space using confidence ellipsoid
Assumption: normally distributed measurement error
Classical parameter estimation
18. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
4
Approximate the joint-confidence region of parameters in the
parametric space using confidence ellipsoid
Assumption: normally distributed measurement error
Classical parameter estimation
19. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
4
Approximate the joint-confidence region of parameters in the
parametric space using confidence ellipsoid
Assumption: normally distributed measurement error
Classical parameter estimation
22. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
6
Classical OED : Design criteria
Minimize or maximize a certain measure of covariance matrix
Experimental design criteria
A design:
Minimize the trace of the covariance matrix
Franceschini, G., Macchietto, S., 2008.Model-based design of experiments for
parameter precision: State of the art. Chem. Eng. Sci. 63 (19), 4846–4872.
23. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
6
Classical OED : Design criteria
Minimize or maximize a certain measure of covariance matrix
Experimental design criteria
A design:
Minimize the trace of the covariance matrix
D design:
Minimize the determinant of the covariance
matrix
Franceschini, G., Macchietto, S., 2008.Model-based design of experiments for
parameter precision: State of the art. Chem. Eng. Sci. 63 (19), 4846–4872.
24. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
6
Classical OED : Design criteria
Minimize or maximize a certain measure of covariance matrix
Experimental design criteria
A design:
Minimize the trace of the covariance matrix
D design:
Minimize the determinant of the covariance
matrix
Optimization problem:
Franceschini, G., Macchietto, S., 2008.Model-based design of experiments for
parameter precision: State of the art. Chem. Eng. Sci. 63 (19), 4846–4872.
33. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
9
Step 1: over-approximate the GPE solution using an orthotope (box)
Methodology : Optimal Experiment Design for GPE
Enclose the GPE solution set by finding the upper and lower limits of
each uncertain parameter
Note: Similar to classical optimal experiment design, is assumed to be
given
40. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
11
Optimal experiment design for GPE: Design criteria
Minimize or maximize a function of the over-approximation box
Experimental design criteria
A design:
Minimize the perimeter of the box
D design:
Minimize the volume of the box
41. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
12
Objective: identify the control input which yields the minimum
value for the design criterion over the over-approximation box
OED for GPE problem:
Step 2 : Optimize over the over-approximation box
42. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
12
Objective: identify the control input which yields the minimum
value for the design criterion over the over-approximation box
OED for GPE problem:
Step 2 : Optimize over the over-approximation box
43. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
12
Objective: identify the control input which yields the minimum
value for the design criterion over the over-approximation box
OED for GPE problem:
Step 2 : Optimize over the over-approximation box
Hatz, K., Leyffer, S., Schlöder, J. P., Bock, H. G., 2013. Regularizing bilevel nonlinear programs by lifting. Preprint ANL/MCS-P4076-0613.
44. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
12
Objective: identify the control input which yields the minimum
value for the design criterion over the over-approximation box
OED for GPE problem:
Step 2 : Optimize over the over-approximation box
Hatz, K., Leyffer, S., Schlöder, J. P., Bock, H. G., 2013. Regularizing bilevel nonlinear programs by lifting. Preprint ANL/MCS-P4076-0613.
45. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
13
Case Study: Lotka-Volterra system
Lotka-Volterra (predator-prey) model modified to control the prey
population
Model :
Measurements:
Measurement error bounds: [-0.1,0.1]
Fixed sampling time: 0.1
Control input:
Uncertain parameters: (True parameters = (0.9,1.5))
46. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
14
Case Study: Lotka-Volterra system
Optimal experiment design for GPE – A criterion
Solved using CasADi + Ipopt using sequential approach with CVODE
integrator
GPE solution sets obtained from MC++ (Chachuat et al.)
47. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
14
Case Study: Lotka-Volterra system
Optimal experiment design for GPE – A criterion
Solved using CasADi + Ipopt using sequential approach with CVODE
integrator
GPE solution sets obtained from MC++ (Chachuat et al.)
0.8 0.84 0.88 0.92 0.96 1
1.42
1.46
1.5
1.54
1.58
k
d
99% confidence ellipsoid for GPE A-design
GPE solution set for GPE A-design
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
t
u
scaled
control input for GPE A-design
48. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
14
Case Study: Lotka-Volterra system
Optimal A design for GPE (magenta) compared to classical A design
(green)
Solved using CasADi + Ipopt using sequential approach with CVODE
integrator
GPE solution sets obtained from MC++ (Chachuat et al.)
0.8 0.84 0.88 0.92 0.96 1
1.42
1.46
1.5
1.54
1.58
k
d
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
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
t
u
scaled
control input for classical A-design
control input for GPE A-design
51. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
16
Conclusions
Problem of optimal experiment design for guaranteed parameter
estimation formulated as a bi-level program
The approximation of the GPE set is realized as a box (this might
include parts of parametric space not in the GPE set)
52. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
16
Conclusions
Problem of optimal experiment design for guaranteed parameter
estimation formulated as a bi-level program
The approximation of the GPE set is realized as a box (this might
include parts of parametric space not in the GPE set)
This methodology for optimal experiment design for GPE always yields
a non-convex optimization problem (computationally intensive
problem; challenging to solve for global optimality)
53. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
16
Conclusions
Problem of optimal experiment design for guaranteed parameter
estimation formulated as a bi-level program
The approximation of the GPE set is realized as a box (this might
include parts of parametric space not in the GPE set)
This methodology for optimal experiment design for GPE always yields
a non-convex optimization problem (computationally intensive
problem; challenging to solve for global optimality)
Results of optimal experiment design for guaranteed parameter
estimation differ from the classical experiment design
54. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
17
Future work
Look for alternative definite shapes (ellipsoid) to over-approximate the
GPE solution set
Exploring the possibility of using other solvers to achieve the global
optimum
Apply this method on robust control and design procedures
57. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
58
Lotka-Volterra system
Model:
Optimal D design for GPE compared to classical D design
Sager, S., 2013. Sampling decisions in optimum experimental design in the light of pontryagin’s maximum principle. SIAM Journal on Control and Optimization 51 (4), 3181–3207.
Telen, D., Logist, F., Derlinden, E. V., Tack, I., Impe, J. V., 2012. Optimal experiment design for dynamic bioprocesses: A multi-objective approach. Chemical Engineering Science 78, 82 – 97.
58. D
N
Y
D
D
N
N
Y
Y
Process Dynamics
and Operations
59
Semi-batch reactor system
Model:
Optimal A design for GPE
B. Srinivasan, S. Palanki, and D. Bonvin. Dynamic optimization of batch processes: I. Characterization of the nominal solution. Comput- ers & Chemical Engineering,
27(1):1–26, 2003.
0 0.2 0.4 0.6 0.8 1
0
0.02
0.04
0.06
0.08
0.1
t
u
GPE A design
0.052 0.054 0.056 0.058 0.06 0.062 0.064
-8
-7.8
-7.6
-7.4
-7.2
-7
-6.8
-6.6
-6.4
x 10
4
k
H
Guaranteed paramater estimation solution set
Optimistic over-approximation of GPE solution set
Editor's Notes
Mathematical models are now-a-days widely used in all fields of science and engineering.
In process industry, they play an important role in process design, control system design, process optimization and so on.
Model development involves three major steps
Identification of the model structure – That is the right hand side of the differrential equation
Design of experiments
3) Estimation of uncertain model parameters in the dynamic model
Throughout the presentation, I will represent the model in this form
In this work we concentrate on design of experiments
Optimal experiment design is important as it helps in identifying the best experimental conditions i.e. the set of control inputs, initial conditions & sampling time
One could use this best experimantal conditions to extract most inforamative data from the system and use it for both model discrimination and precise parameter estimation.
In general, it is assumed for optimal experiment design that the measurement noise follows normal distribution.
In this work, we assume that the measurement noise is bounded and can follows any arbitary distribution
In this work, we assume that the measurement noise is bounded and can follows any arbitary distribution
In this work, we assume that the measurement noise is bounded and can follows any arbitary distribution
Classical parameter estimation is an approach which over-approximates the joint-confidence region of the parameters in the parametric space using an ellipsoid with certain confidence level.
Let‘s say that we have some measurements from a system, and are corrupted with some measurement noise.
Using LSE, one could identify the best fit parameter for its model, such that the normed difference between the measurements and the model outputs is minimum
If in case, we have measurements with different measurement error, LSE identifies a different parameter for the model
Simillarly, measurements from n different experiments would lead us to n different parameters
The classical paramter estimation approach identifies an ellipsoid which encloses all these possibly identified parameters called the joint confidence region of the parameters in the paramteric space
Using this formulation, where \hat{p} here represents an expected value of parameter, which here is a LSE,
C denotes the covariance matrix
F represents the fisher distribution
And alpha is the probability level
The covariance matrix which is a function of control input defines the shape and orientation of the confidence ellipsoid
So, If a control policy like this yeilds the following confidence ellipsoid
The objective of classical OED is to identify the control policy which gives us the confidence ellipsoid whose shape and orientation satisfies certain criteria of our interest.
In the literature, several design criteria for classical OED are extensively studied. Here are few important ones:
A , D design
Optimization problem for classical OED has the following structure.
In the literature, several design criteria for classical OED are extensively studied. Here are few important ones:
A , D design
Optimization problem for classical OED has the following structure.
In the literature, several design criteria for classical OED are extensively studied. Here are few important ones:
A , D design
Optimization problem for classical OED has the following structure.
In contrast to the classical parameter estimation, we have the guaranteed parameter estimation, which assume that the measurement noise is bounded, and can take any arbitary distribution.
Let the blue dots be the measurements from a system for a given control input
P_0 represents the parametric space in within which the nominal parameter lies.
According to the assumption, the true measurements of the system should lie in this region
Lets assume that, plugging in a certain parameter from P_0 into the model yeilded us this trajectory which here is a possibly true or nominal output trajectory
It is equally likely that the model output trajectory using the parameter in red to be nominal
Using a set-inversion algorithm GPE finds all the parameters, whose corresponding measurement trajectories lie within this region.
Note that the GPE set varies in shape and size for each control policy.
Now, for such a control policy, if we end up in the following GPE set
The objecive of this work, OED for GPE is to identify the control policy which yeilds us the GPE set of our criteria.
Here I introduce the new methodology for OED for GPE.
The procedure is divided into steps.
First step is to over-approximate the GPE solution set using an orthtope
We have formulated threee distinct over-approximation techniques. Here, I will present one of them.
Note that, similar to classical optimal experiment design, an expected value of parameter is required here
In optimistic over-approximation we assume that model outputs using expected parameters are nominal(equal to the true outputs)
The black curve here represents the nominal output trajectory
Blue dots represent the model outputs using expected parameters.
We know that the GPE identifies all the parameters whose trajectories pass through these error bounds
Optmistic over approximation identifies the bounds on the GPE set using this formulation, where we try to maximize the box such that the difference between the model output using E(p) and PL/U lies between the error bounds
Note that the over-approximation box includes the paramteric space which does not belong to the GPE set.
For example, If we have a two parameter system, One has to identify the upper and lower bounds of each of these parameters to identify the joint-confidence region of parameters , for a given control input.
So, the problem for optimistic over-approximation in such a case will try to maximize the box in 2-dimensions such that the difference between the model outputs using the expected parameter and the model outputs using each of the parametric limits for lie between the error bounds.
Note that the error bounds can vary for each measured variable, for each instant of time.
In line with the classical-OED, we can formulate several criteria in the case of OED for GPE
A , D design
Having the box ready, Step 2 is to now identify the control policy which yeilds the minimum value for the design criterion
This leads to a bilevel optimzation problem, which in the lower level identifies the joint confidence region and optimizes in the upper level the geometry defined by the objective function.
A classical way to solve this problem would be to implicitly solve the lower level problem by expressing the necessary conditions for optimality(KKT) as constraints. It must be however assured that the KKT conditions yresult in a global optimum solution to the lower level problem.
An advantage of such transformation is that the optimization problem is no more bi-level.
However this reformulation yields a non-covex dynamic optimization problem.
This problem fails to satisfy common constraint qualifications i.e LICQ, therefore we regularize this problem.
In the regularization techniqe, we add slack variables to each of the inequality constraints.
This leads to a well posed optimization problem which satifies the LICQ
We check upon the convergence of solver, if the over-approximation of the GPE set is correct. If the test is positive, we claim that we found a local optimal solution
If the test is positive, we claim that we found a local optimal solution
We applied this methodology on the Lotka Volterra system which tries to control the population of the prey. It has two states x1-prey, x2-predator, where we measure the population of predator with an accuracy of +-0.1 every 0.1 time units. variable and 1 control input.
It has two uncertain parameters.
On the left is the control policy obtained by solving the classical A optimal experiment design problem using CASADI interface with IPOPT.
On the right is the GPE set and the confidence ellipsoid obtained using MC++
We see here that the classical ellipsoid captures the shape of the GPE set and also encloses the GPE set nicely. But this is bot the case always.
We solved the optimal experiment design problem for GPA for A desing criterion using Casadi library, ipopt and CVODE integrator. The GPE sets are obtained form MC++.
The control input here represents atleast the local optimal solution for the A optimal design for GPE whose objective is to minimize the perimeter of the over-approximation box.
On the right, we see the confidence ellipsoid which represents the joint confidence region assuming normal distribution for the control input.
The GPE set in magents represents th ejoint confidence region assuming uniform distribution for the same conrtol input.
We see that the joint-confidence region is not the same for both the assumptions on the error distribution. We can also see here that the confidence ellipsoid identifies the orientation of the GPE set closely, but doesnot enclose the GPE set.
The control input in green represents atleast the local solution for the A optimal experiment design for classical parameter estimation whose objective is to identify the the control input which minimzes the trace of the covariance matrix.
On the right hand side, we see the corresponding confidence ellipsoid and GPE set. In this case we see that the confidence ellipsoid not only find the orientation but also enclose it.
We can see that both classical OED and OED for GPE have bang bang control input and differ only in the swithcing time. We can also see that over-approximation box of the magenta set is smaller than the one for the green GPE set.
If the test is positive, we claim that we found a local optimal solution