Advanced Process Operations

1. Introduction to Course
1.1. Dynamic Optimization Formulations

1.1.1.
t Time tf Final time
x Differential state variables u Control variables
y Algebraic state variables v Time-invariant parameters
1.2. Batch Process Optimization

1.2.1. Determine the batch duration tf and the feed rate profile F(t) to maximize the
final amount of C, while ensuring the final purity of C (w.r.t. B, C and D) over
70%, and maintaining the cooling jacket temperature Tj(t) above 15°C at all
times
1.2.2. Model-Based Optimization Approach
1.2.2.1. Construct a mathematical model relating the control actions (flow rate)
with the states/outputs variables (concentrations, temperature)
1.2.2.2. Determine the optimal control actions based on this model
 1.2.3. Trajectory

1.2.3.1. Optimal feed trajectory happens to be discontinuous

1.2.3.2. Purity constraints require the feed to be turned off, otherwise it would
not reach the desired purity
1.3. Continuous Transition Optimization

1.3.1. Determine the feed rate profile F(t) to minimize the transition time between
ref ref
reference concentrations C 1 and C 2 , while keeping the cooling jacket
temperature Tj(t) above 15°C at all times
 1.3.2. Trajectory

1.3.2.1. Minimize the gap between the response and the desired set-point over
the prediction horizon, while penalizing control moves, and repeat on a
rolling horizon [tk ,tk + T]
1.4. Dynamic Model Building

1.4.1. Estimate (a subset of) the unknown parameters for the candidate model to
best fit the experimental data
1.4.2. Determine the best experimental conditions to maximize parameter precision
during parameter estimation
1.4.3. This is an optimization problem based on different states
2. Dynamic Optimization Formulations
2.1. Dynamic Optimization Problem Formulations
2.1.1. Expected to determine the time-varying and/or time-invariant inputs of a
dynamic system in order to minimize (or maximize) some performance
criterion while satisfying given physical and operational constraints
2.1.2. Main Steps of Problem Formulations
2.1.2.1. Specification/modelling of the controlled systems
2.1.2.1.1. Manipulated variables: controls u(t )∈ R n and decisions v ∈ R n
u v
 2.1.2.1.2. Response and dynamical model: states x ( t ) ∈ Rn and y ( t ) ∈ R n
x y

2.1.2.2. Specification of a performance criterion

2.1.2.2.1. point-wise and path objectives
2.1.2.3. Specification of physical constraints to be satisfied
2.1.2.3.1. point-wise and path constraints
2.2. Systems of Ordinary Differential Equations (ODEs)
2.2.1. ẋ ( t )=f ( x ( t ) )
2.2.2. Equations: n x
2.2.2.1. Differential: f : R n → Rn
x x

2.2.3. Variables: n x
2.2.3.1. Differential: x ∈ Rn x

2.2.4. Initial Conditions for ODEs

2.2.4.1. Conditions: n x
2.2.4.1.1. ẋ ( 0 )=f ( x ( 0 ))
2.2.4.2. Variables: 2 n x
2.2.4.2.1. x ( 0 ) , ẋ (0)
2.2.4.3. Consistent initialization: specify exactly n x initial conditions
2.2.5. Existence of Solutions to ODEs
2.2.5.1. A unique solution to ODEs exists locally if f is locally Lipschitz-
continuous
2.2.5.1.1. Locally refers to specified time horizon such as from t 0 → t 1 that
the problem can be solved
2.2.5.2. For global existence, f needs to be globally Lipschitz-continuous
2.2.5.2.1. The function must be finite if not it cannot be Lipschitz-
continuous
2.2.5.2.2. Setting a bound can make it Lipschitz-continuous
2.2.6. Lipschitz continuous
2.2.6.1. Must be defined at all areas
2.2.6.2. Gradient (slope) is always bounded
2.2.6.3. Not every continuous function is Lipchitz Continuous
2.2.6.3.1. Example x → √|x| is not Lipschitz continuous because at 0 it is
undefined (infinite)
2.2.7. A Simple ODE with Multiple Solutions
2.2.7.1. Consider the following ODE: ẋ ( t )= √ x (t ) with x ( 0 )=0
2.2.7.2. Check that all of the following trajectories are solutions of this initial
value problem
 {
0 if t ≤θ
x ( t )= ( t−θ )2
otherwise
4

2.2.7.2.1.

2.2.7.2.2. At x(0)=0 the problem does not work as there are multiple
solutions possible, changing the initial condition to a real
number would fix this like x(0)=0.0001
2.2.7.3. The height of liquid in a gravity drained tank is described by:
2.2.7.3.1. ḣ ( t )=−C √h ( t )
2.2.7.4. Given that the tank is empty at T, i.e. h(T) = 0, determine the time t < T
at which the level was h(t)= H > 0
2.2.7.4.1. Hint: Consider reversing time as s = T – t
2.2.8. A Simple ODE with Finite Escape Time
2.2.8.1. Workshop: Consider the following ODE: ẋ ( t )=x ( t )2, with x ( 0 )=1
2.2.8.2. Does a solution exist locally?
2.2.8.3. Find an expression of the ODE solutions using the method of separation
of variables, and determine the escape time

x (t ) t

∫ dx 2 ∫
= dt
x(0) x 0

[ ]
x(t)
−1
=t
x x(0)
1 1
− =t
x ( 0) x ( t )
x (0)
x ( t )=
1−x ( 0 ) t
2.3. Systems of Differential-Algebraic Equations (DAEs)
2.3.1. f ( x ( t ) , ẋ ( t ) , y ( t ) )=0
2.3.2. g ( x ( t ) , y ( t ) )=0
2.3.3. Initial conditions for DAEs:
 2.3.3.1. Conditions n x + n y
2.3.3.1.1. f ( x ( 0 ) , ẋ ( 0 ) , y ( 0 ) )=0
2.3.3.1.2. g ( x ( 0 ) , y ( 0 ) )=0
2.3.3.2. Variables: 2 n x +n y
2.3.3.2.1. x ( 0 ) , ẋ ( 0 ) , y ( 0 )
2.3.3.3. Consistent initialization: specify at most n x initial conditions
2.3.3.4. Test: Can we specify arbitrary values for x ( 0 )and determine the
corresponding values of ẋ ( 0 ) and y ( 0 ) ?
2.3.3.4.1. If yes specify exactly n x initial conditions
2.3.3.4.2. If no specify fewer than n x initial conditions
2.4. Test Yourself
''
2.4.1. x 1 =−λ x 1
''
2.4.2. x 2 =−λ x 2−¿
2.4.3. x 21+ x22 =1
2.4.4. Differential states
'
2.4.4.1. dx 1=x 1
'
2.4.4.2. dx 2=x 2
2.4.4.3. dx 1=− λ x1
2.4.4.4. dx 2=− λ x2 −g
2.4.5. Algebraic State
2 2
2.4.5.1. x 1+ x2 =1
2.4.5.1.1. Differentiate to time
' '
2.4.5.1.2. 2 x1 x 1+2 x 2 x 2=0
2.4.5.1.3. x 1 dx 1 +x 2 dx 2=0
2.4.5.1.3.1. Differentiate to time
2.4.5.1.3.2. dx 21 + x 1 dx '1 +dx 22 + x 2 dx '2=0
2 2 2 2
2.4.5.1.3.3. dx 1 +dx 2−λ x 1− λ x 2−x 2 g=0
2.4.5.1.3.4. Divide by λ
2 2
2.4.5.1.3.5. dx 1 +dx 2−x 2 g−λ=0
2.4.6. Count number of equations and variables
2.4.6.1. 7 equations
2.4.6.2. 9 variables
2.4.6.3. Can supply 2 initial conditions
2.4.6.3.1. Specify x 1 and dx 1
2.4.6.3.1.1. Can use all relationships including hidden relationships
2.4.7. Index of DAE system is 3 as you need one more time differentiation to make an
ODE system
2.4.8. Dummy derivative method
2.4.8.1. Need to introduce the same number of dummy variables as hidden
relations
'
2.4.8.2. z 1=dx 1
'
2.4.8.3. z 2=dx 2
2.4.8.4. Equations
2.4.8.4.1. dx 1=x '1
 '
2.4.8.4.2. dx 2=x 2
2.4.8.4.3. z 1=−λ x 1
2.4.8.4.4. z 2=−λ x 2−g
2 2
2.4.8.4.5. x 1+ x2 =1
2.4.8.4.6. x 1 dx 1 +x 2 dx 2=0
2 2
2.4.8.4.7. dx 1 +dx 2−x 2 g−λ=0
2.4.8.5. Now there is an index 1 DAE system
2.5. Case Study of a Simple Car Control Problem
2.5.1. The motion of a car is described by the following ODEs
2.5.1.1. ṗ ( t ) =v ( t )
2.5.1.2. v̇ ( t )=u ( t )
2.5.2. Where states p(t) and v(t) denote position and speed and the control u(t) is the
force due to acceleration (≥0) or deceleration (≤0)
2.5.3. The optimization problem consists of driving the car from the position p 0
where it is initially parked, to its final destination p f where it is also parked,
using the least amount of fuel. The maximum speed limit is v U, and the car’s
maximal acceleration and decelaration are, respectively, uL and uU. For
simplicity, assume that the car’s instantaneous full consumption is
proportional to its squared velocity.
2.5.4. Formulate problem
2.5.4.1. Minimize
tf

2.5.4.1.1. ∫ α v ( t )2 dt
0
2.5.4.1.2. For u ( t ) at t 0 → t f
2.5.4.2. Subject to
2.5.4.2.1. . ṗ ( t ) =v ( t )
2.5.4.2.2. v̇ ( t )=u ( t )
2.5.4.2.3. p ( 0 )= p 0
2.5.4.2.4. v ( 0 )=0
2.5.4.2.5. p ( t f ) =p f
2.5.4.2.6. v ( t f )=0
v (t )≤ v , ∀ t ∈[0 , tf ]
U
2.5.4.2.7.
u ≤u ( t ) ≤ u , ∀ t ∈ [ 0 , t f ]
L U
2.5.4.2.8.
2.5.5. What if your minimizing travel time?
2.5.5.1. Minimize
tf

2.5.5.1.1. t f =∫ 1dt
0
2.5.5.1.2. Takes it from an integral term to a point term
2.6. Performance Criteria in Dynamic Optimization
2.6.1. Functional used for quantitative evaluation of the system’s performance
2.6.2. Main types of Criteria
2.6.2.1. Lagrange form
tf

2.6.2.1.1. ∂ ( u , v )=∫ l ( x ( t ) , y ( t ) , u ( t ) , v ) dt
0
 2.6.2.1.2. Minimize average power consumption
2.6.2.2. Mayer form
N
2.6.2.2.1. ∂ ( u , v )=∑ φ(x ( t k ) , y ( t k ) , v)
k=1
2.6.2.2.2. maximize final batch product purity
2.6.2.2.3. minimise prediction error compared with time-series
measurements
2.6.2.3. Point Constraints
End point: ψ ( x ( t f ) , v ) ≤ψ
U
2.6.2.3.1.
Interior: ψ ( x ( t k ) , v ) ≤ ψ ,t k ∈ [ 0 , t f )
U
2.6.2.3.2.
2.6.2.3.3. batch process must reach at least 30% conversion after 2 hours:
x(7200) ≥ 0.3
2.6.2.4. Path Constraints
2.6.2.4.1. κ ( x ( t ) , y ( t ) , u ( t ) , v ) ≤ κU
2.6.2.4.2. product purity must never fall below 97%: x(t) ≥ 0.97, for all t ∈
[0,tf ] - temperature must always stay between 323 K and 500 K:
323 ≤ T(t) ≤ 500, for all t ∈ [0,tf ]
2.6.3. General Dynamic Optimization Formulation

2.7. Useful Reformulation Tricks

2.7.1. Handling variable final time tf
2.7.1.1. Consider tf time-invariant parameter and define scaled time τ ∈ [ 0,1 ]
2.7.1.2. Replace t with tfτ and dt with tfdτ throughout
t t ∈ [ 0 ,t f ]
2.7.1.2.1. t→τ=
t f τ ∈ [ 0,1 ]

2.7.1.2.2.
v̇ ( t ) =u ( t ) }
ṗ (t )=v ( t ) ∀ t ∈ 0 ,t

tf
[ f]

2.7.2. Converting integral cost ∫ l ( x ( t ) ,u ( t ) ) dt into terminal costs

0

2.7.2.1. Append extra differential equation ẋ n +1 ( t )=l ( x ( t ) , u ( t ) )

x

2.7.2.2. Define ẋ n +1 ( 0 )=0 and replace integral cost with ẋ n +1 ( t f )

x x

tf tf

2.7.2.3. ∫ ẋ n +1 ( t ) dt=∫ l ( x ( t ) , u ( t ) ) dt
x
0 0

2.7.2.4. ẋ n +1 ( t f ) −x 0 (0)
x

Converting a path constraint κ ( x ( t ) , u ( t )) ≤ κ into a terminal constraint

U
2.7.3.
2.7.3.1. Measure infeasibility using integrals
 tf
2
∫ max {0 , κ ( x ( t ) , u (t ) ) −κU } dt ≤ 0
0

2.7.3.2.

2.7.3.2.1. Red area measures violation

2.7.3.3. Append extra differential equation
U 2
2.7.3.4. ẋ n +1 ( t )=max {0 , κ ( x ( t ) , u ( t ) )−κ }
x

2.7.3.5. Set initial condition ẋ n +1 ( 0 )=0

x

2.7.3.6. Replace path constraint with ẋ n +1 ( t f ) ≤ ϵ ≪ 1

x

2.7.3.6.1. 0 is typically not liked and can be infeasibile

2.8. Case Study of a Batch Reactor

2.8.1. Formulate a dynamic optimization problem to maximize the amount of C at a

given final time tf, by manipulating the reactor temperature, and subject to
minimum and maximum temperatures Tmin = 293 K and Tmax = 323 K,
respectively.
3. Direct Solution Methods
3.1. Workshop. Formulate a finite-dimensional approximation of the following optimal
control problem by using a piecewise-constant control parameterization on ns ≥ 1
stages:
1
min 1 [ u ( t ) ]2 dt
3.2. ∫
u (T ) 0 2

3.3. s . t . ẋ ( t ) =u (t )−x ( t )
x ( 0 ) =1. x (1 )=0
3.4. The parameterized control trajectories are given by

3.4.1. U k0 ( t ,ω k ) ≔ωk0 , ∀ t ∈
[( ) ( ))
k−1
ns
,
k
ns
, k=1 … . ns
 3.4.2.
3.5. Direct Sequential Approach: Single Shooting
3.5.1. Visual Depiction of the Approach

3.5.2. Compute the cost and constraint values using numerical integration
3.5.3. Compute the cost and constraint values using numerical sensitivity
3.5.4. Gradient free methods
3.5.4.1. Good if you only have an objective function
3.5.4.2. Don’t like constraints or lots of decision variables
3.5.5. Gradient based methods
3.5.5.1. Need to supply gradients to the optimization model
3.5.6. Pros and Cons
3.5.6.1. Relatively small-scale NLP problems in the variables p=( ω1 , … , ωn , v )
s

that is nu ( N +1 ) ns +n v variables
3.5.6.1.1. Stage times can be added to the decision vector p too
3.5.6.2. Accuracy of the state variables enforced via the error-control
mechanism of the ODE/DAE solver
3.5.6.3. Computing the gradients using numerical sensitivity analysis is often
the dominant cost
3.5.6.4. Feasible path method
3.5.6.4.1. The ODEs/DAEs are satisfied at each NLP iteration
3.5.6.4.2. This is computationally demanding
3.5.6.4.3. Only mildly unstable systems can be handled this way
3.5.6.5. Strategy is one implemented in gPROMS
3.6. Direct Sequential Approach: Multiple Shooting
3.6.1. Discretization as a finite dimensional NLP through parameterization of the
control variables, with the ODEs/DAEs embedded in the NLP problem but with
state discontinuities allowed at stage times
 3.6.2. Lifting of the optimization problem

3.7. Computing Gradients of State-Dependent Functions

3.7.1. State Dependent function in the Mayer form

3.8. Gradients: Forward and Adjoint Sensitivity Analysis

3.8.1. Forward Sensitivity Approach

3.8.1.1. Size of state and state sensitivity equations (np+1) × nx

3.8.1.2. But independent of the number of functionals n F
3.8.2. Adjoint (or reverse) Sensitivity Approach
Direct Initial Dynamic

3.8.2.1. Size of state & adjoint equations: (n F + 1) × nx

3.8.2.2. But, independent of the number of parameters, np
3.8.2.3. λ is the adjoint co-state
3.9. Procedure for Forward Sensitivity Analysis
3.9.1. State and Sensitivity Numerical Integration 0→T
 3.9.2. Function and Gradient Evaluation

3.9.3. Limitation is that more parameters creates more integration which gets more
complicated
3.10. Procedure for Adjoint Sensitivity Analysis
3.10.1. State Numerical Integration 0→T

3.10.2. Store state values x(t) at mesh points,

3.10.3. Adjoint Numerical Integration: T→0

3.10.4. Need to interpolate state values x(t) between mesh points

3.10.5. Function and Gradient Evaluation

3.11. The Collocation Principle

3.11.1. Consider the ordinary differential equation

3.11.2. With initial condition

3.11.3. Approximate the solution of this initial value problem (IVP) via a trial function,
most commonly a polynomial:
3.11.4. Substitution of trial function into the IVP gives residuals:

3.11.5. For an exact solution, and

3.11.6. In practice, require trial function to satisfy and at chosen
collocation points
 3.11.7. Collocation points are taken at the zeroes of orthogonal polynomials

4. Methodology and Tools for Dynamic Model Building

4.1. Mathematical Models and Modelling Goals
4.1.1. General Mathematical Model Formulation

4.1.2. Equation Forms of Process Models

4.1.3. Typical Modelling Goals

4.1.3.1. Output Prediction
4.1.3.2. Process Control
4.1.3.3. Process Design
4.2. Parameter Estimation as a Special Case
4.3. Comparison to “Typical” Optimization Problem

4.4. What do Dynamic Experiments Comprise?

4.4.1. Number of experiments, NE

4.4.2. Each experiment i = 1, . . . . , NE characterized by
4.4.2.1. Operating policy
4.4.2.1.1. Overall duration
4.4.2.1.2. Initial conditions
4.4.2.1.2.1. Initial concentrations, volumes
4.4.2.1.3. Other time invariant parameters
4.4.2.1.3.1. Reactor diameter, stirrer power input
4.4.2.1.4. Control variable inputs
4.4.2.1.4.1. Feed flowrate, heat input
4.4.2.2. Measurements
4.4.2.2.1. For each measured output variable j = 1, . . . . , NV(i):
4.4.2.2.1.1. Set of measurement times
4.4.2.2.1.2. Values of corresponding measurements
4.4.2.2.1.3. Uncertainty to which measurements are known
4.5. Sensors and Variance Models
 4.5.1. A property of the measurement technique associated with the sensor
4.5.2. Often assumed to be statistically independent, but not necessarily identically
distributed
4.5.2.1. Normally distributed with zero mean (Gaussian white noise)

4.5.2.2. Constant Variance x-axis is measured output of y, y-axis is true value

(predicted) of y
4.5.2.3. Heteroscedastic variance has a constant ω, y is the magnitude of the
measurement, has exponent γ
4.5.2.3.1. γ<0
4.5.2.4. A variance model may be estimated directly from the measurements

4.5.2.5. Varaince parameters may also be estimated with other model

parameters
4.6. Objective Functions in Parameter Estimation
4.6.1. Parameter estimation primarily aims at minimizing the discrepancy between
experimental data and model predictions
4.6.2. Weighted Least-Square Errors (LSE) Regression

4.6.2.1. Trouble happens as outliers are given heavy weight

4.6.3. Other parameter estimation methods exist
4.7. Log-Likelihood Estimators

4.7.1. Consists of a constant term, variance term, and residual term

4.7.2. Implies weighted least-squares regression by assuming a known variance and

choosing
4.7.3. The variance and residual terms are pulling in different directions
4.8. Goodness of fit: Main Idea

4.8.1. Two sources of model prediction errors

4.8.1.1. Pure error due to measurement noise
4.8.1.2. Lack of fit error due to model inaccuracy
4.8.2. Are the model prediction error statistically consistent with the pure error
4.9. Goodness of fit: Statistical Assessment
 4.9.1. The χ2-distribution with k degrees of freedom is the distribution of a sum of the
squares of k independent standard normal random variables

4.9.2. Goodness of Fit Test

4.9.2.1. Is the behaviour of the model prediction error in agreement with the
(expected) measurement noise error property to a desired confidence
level (1-α)?

4.9.2.2. 1-α is the confidence

4.9.3. Interpretation
4.9.3.1. Failure to pass the above rest suggests a poor fit; the model needs
revising
4.9.3.2. Beware! An incorrect sensor model or a local optimum could falsify the
test results
4.10. What is Covariance?
4.10.1. Covariance measures the joint variability of two (random) variables

4.10.1.1. Cov(x,y)

4.10.2. Covariance of a single measurement

4.10.2.1. Taking the expected realization of a variable minus the expected value
squared δyi
4.10.3. Covariance of any two measurements

4.10.4. Covariance of a vector of measurements

n×1 n×n
1×n

4.10.5. Case of model parameters

4.10.5.1. Covariance of a single parameter
4.10.5.2. Covariance between any two parameters
 4.10.5.3. Covariance of a vector of parameters

4.10.5.3.1. Taken on the middle of the ellipsoid, is a measurement between

the ratio of one axis to the other

4.11. Precision of Parameter Estimates: Statistical Inference

4.11.1. ~y → ~θ ∈arg min Φ ( θ|~

y ) , Φθ ( θ^ |~y )=0
4.11.2. ~y+ ∂ y → ~θ+∂ θ ∈arg min Φ ( θ|~ ^ θ|~
y +∂ y ) , Φθ ( θ+∂ y +∂ y ) =0
4.11.3. Φθ ¿
∂ θ ≈ [ Φθθ Φ θy ]θ^ , ~y ∂ y
−1
4.11.4.
C θ=E ( δθ ∙ δθ ) ≈ [ Φθθ Φθy ]θ^ , ~y C y [ Φ θθ Φ θy ]θ^ , ~y
T −1 −1 T
4.11.5.
4.11.6. Noise covariance matric C y =E ¿
4.11.7. Only exact for linear models

4.11.8.
4.12. Obtaining Confidence Intervals from Covariance Matrices
4.12.1. Linearized 100(1-α)% Confidence Intervals
4.12.1.1.

4.12.1.2. t N −n 1−
θ ( α2 ) is the z(1-α) distribution which gives larger range of
Gausian
4.12.2. 100(1-α) t-value test
 4.12.2.1.
4.12.2.1.1. If 0 is within this confidence interval the parameter value is
garbage
4.12.3. Linearized 100(1-α)% Confidence Region
4.12.3.1. Large sample size-known measurement variance
4.12.3.1.1.
4.12.3.1.2. Ellipsoid gets bigger from the increase in confidence
4.12.3.2. Small sample size- uncertain measurement variance

4.12.3.2.1.
4.12.3.2.2. Using an F test is better because it is more conservative and is
better for low confidence
4.13. Precision of Parameter Estimates: Practical Aspects
4.13.1. Failure to pass a t-test for one/several parameters suggests a flat optimum in
certain directions
4.13.1.1. Cause: (i) low sensitivity; (ii) large correlations
4.13.1.2. Remedy: (i) focus on a parameters subset – apply
estimability/sensitivity analysis (ii) design experiments—apply
optimal experiment design
4.13.2. A model presenting a systematic bias (e.g. χ2-test not satisfied) or a highly
nonlinear model may invalidate the statistical analysis
4.13.2.1. Always be cautious with your conclusions

4.13.3. Likelihood-ratio Confidence Regions

4.13.3.1. Provide a better approximation than linearized regions for nonlinear
models and/or limited measurement points – but computationally
demanding
4.14. What does an experiment consist of?

4.14.1. Formulate as an optimization problem

4.15. General idea of experiment design
4.15.1. What experiments give the best data
4.15.1.1. Control policy, initial conditions, equipment design
 4.15.1.2. Duration, location, and number of sampling times
4.15.2. Can this be established a priori based on current knowledge
4.15.2.1. Model structures and a priori parameters
4.15.2.2. Data and senor properties
4.15.3. Criteria for best data
4.15.3.1. Precise parameter estimation
4.15.3.2. Efficient model discrimination
4.15.4. Constraints
4.15.4.1. Control variables
4.15.4.2. Equipment conditions
4.15.5. Iterative Procedure

4.15.5.1.
4.16. Model based design of experiments (MBDoE)
Design conditions, initial
condition, run time

4.17. Experiment Design Criteria

4.17.1. Parameter Covariance Matrix

4.17.2.
4.17.3. D-optimality
4.17.3.1. Minimize the volume of confidence ellipsoid
 1
min det C n
4.17.3.2.
u,v
[ θ] θ

4.17.4. A optimality
4.17.4.1. Minimize the average variance of the individual parameter estimate
min 1 trace [C ]
4.17.4.2. θ
u , v nθ
4.17.5. E optimality
4.17.5.1. Minimizes the longest axis of confidence ellipsoids
min λ [C ]
4.17.5.2.
u , v max θ
4.17.6. Modified-E optimality
4.17.6.1. Minimizes the ratio of longest to shortest axis of confidence ellipsoids
4.17.7. Differences in the resulting designs can be significant
4.18. Experiment Design Criteria: Covariance Matrix

4.18.1. Parameter Covariance Matrix

4.18.2. Alternative (simplified) covariance matrix

4.19. Sensitivity and Uncertainty Analysis

4.19.1. Uncertainty analysis

4.19.1.1. Propagate factor uncertainty to quantify output uncertainty
4.19.2. Sensitivity Analysis
4.19.2.1. Apportion the output uncertainty back into factor contributions
4.20. Derivative-based Local Sensitivity Analysis
4.20.1. Local sensitivity coefficients at θ :

4.20.1.1.
 4.20.2. Average sensitivity coefficients over factor domain:

4.20.2.1.

4.20.3. Could be used for ranking the factors in order of importance

4.20.3.1. Needs input factor normalizations, e.g. in [0,1]d
4.20.4. But does not capture factor interactions
4.20.4.1. Could be extended to second- or higher-order derivatives
4.21. Variance-based Global Sensitivity Analysis
4.21.1. Any model M may be decomposed as follows (HDMR) (anova)

4.21.1.1.
4.21.1.2. Impose extra conditions, e.g. nased on orthogonality for f0, fi, fi1, i2, …
4.21.2. The terms in this decomposition can be calculated recursively

4.21.3. The variance of any model M may be decomposed likewise:

4.21.4. The variance terms may be calculated recursively:

4.21.5. Sobol’ Sensitivity Indices:

4.21.6. First Order (a.k.a Main-Effect) Sensitivity Indices

4.21.6.1.
4.21.6.2. Effect of varying θi alone, averaged over variations in other parameters
4.21.7. Total-Order Sensitivity Indices:

4.21.7.1.
 4.21.7.2. Effect of varying θi , including all variance caused by its interactions, of
any order, with any other parameters ST,i = Si + (S1,i + · · · + Si−1,i + Si,i+1 + ·
· · + Si,d) + · · · + S1,2,...,d
4.22. Sampling Strategies in Multiple Dimensions
4.22.1. Monte-Carlo integration

4.22.2. Sampling Strategy

4.22.2.1. Low-discrepancy is more random than a regular grid but utilizes the
knowledge of previous points to create a more uniform coverage
without being too uniform
4.23. How to use the first and total order sensitivity indices

5. Optimal Control Theory

5.1. Scope and Motivation
5.1.1. Constrained NLP Problem: NCOs: KKT Conditions

Dual Feasibility

5.1.2. Constrained DO Problem NCOs

Complimentary
 5.1.3. NCOs are useful to:
5.1.3.1. Single out candidate optimal controls
5.1.3.2. Determine the structure (arcs) of optimal controls
5.1.3.3. Derive optimal control laws along certain arc
5.2. Basic Optimal Control Formulation

5.3. Admissible and Feasible Controls

5.3.1. Which linear space U for the time-varying controls
5.3.1.1. Continuous controls ---- not big enough
5.3.1.2. More generally, piecewise continuous controls:

5.4. Global and Local Optimality in Optimal Control

5.4.1. (Globally) Optimal Control (Minimize Case)
5.4.1.1. A feasible control u* is a global minimum of the functional

5.4.1.2. This global condition does not require a distance (or a norm) on U: the
control u* is compared to every other feasible control
5.4.2. Locally Optimal Control (Minimize Case)
5.4.2.1. A feasible control u ∗ is a local optimum of the functional
5.5. Weak Neighborhoods in Optimal Control

5.6. Strong Neighborhoods in Optimal Control

5.7. Necessary Conditions for Optimality

 5.7.1. Euler-Lagrange Optimality Conditions
stationarity Hamiltonian function
integrand
terminal

feasibility
5.8. Remarks about Euler-Lagrange Conditions
5.8.1. Complete set of necessary conditions
5.8.1.1. 2 x nx ODEs in the state variable x(t) and adjoint variables λ(t), with
initial/terminal conditions
5.8.1.2. Nu AEs in the variables u(t)
5.8.2. But need to solve a two-point boundary value problem
5.8.3. Same set of conditions for minimize and maximize problems (in the absence of
terminal conditions
5.8.4. Second-order (Legendre-Clebsch) Condition:

5.8.5. For autonomous problems (i.e. no explicit time dependence), the Hamiltonian
is invariant along an optimal trajectory:

5.9. Applying the Euler-Lagrange Conditions

5.10. Necessary Conditions for Optimality

5.10.1. Rest of terms remain the same

5.11. Remarks about Pontryagin Maximum Principle

5.11.1. Handles control bounds in a very natural way: boils down to solving an NLP
problem at each time along [t0, tf]
5.11.2. Control bounds give rise to a solution structure comprised of interior and

boundary arcs
5.12. Applying the Pontryagin Maximum Principle

5.12.1.
5.13. Singularity

5.13.1. H is not dependent on u

5.14. Singular Optimal Control: Definition

5.14.1. In a scalar singular control problem the stationarity condition

does not depend explicitly on the control – it cannot be

used directly to provide candidate optimal controls!
5.14.2. The system becomes a high index DAE system
5.15. Singular Optimal Control: Solution Strategy
5.15.1. The singular arc (τ1, τ2) is such that

5.15.1.1.
5.15.2. Apply successive time differentiations
 5.15.3. From previous example

5.15.4. Further rules

5.16. Optimal Control with State Path Constraints

5.16.1. Notoriously difficult to solve
5.16.2. Two Main types
5.16.2.1. Mixed state-control constraint

5.16.2.1.1.
5.16.2.2. Pure-state constraints

5.16.2.2.1.
5.16.2.3.