1. Automatic Control Laboratory, ETH Zürich
www.control.ethz.ch
Application of Set Membership
Identification to Controller Design
Marko Tanaskovic
Doctoral Exam 24.11.2015
www.control.ee.ethz.ch
2. Controller design
• Create mathematical control law for physical system mathematical model required
u y
ydes
f(y; ydes)
3. Controller design
• Create mathematical control law for physical system mathematical model required
• Derive mathematical model from measurements (system identification)
?
u y
u y
ydes
f(y; ydes)
4. System identification
• Measurements corrupted by noise and disturbance identified model uncertain
Probabilistic identification Set membership identification
Assumption on Statistical properties Bound
? +
u
e
y
e
5. System identification
• Measurements corrupted by noise and disturbance identified model uncertain
Assumption on Statistical properties Bound
Parameter estimate
Probability density function
Optimal estimate
maximizes likelihood
Set of possible parameters
Optimal estimate minimizes
worst case estimation error
? +
u y
e
e
Probabilistic identification Set membership identification
6. System identification
• Measurements corrupted by noise and disturbance identified model uncertain
Assumption on Statistical properties Bound
Parameter estimate
Probability density function
Optimal estimate
maximizes likelihood
Set of possible parameters
Optimal estimate minimizes
worst case estimation error
? +
u y
Application to
controller design
Probabilistic objectives Non-probabilistic objectives
e
e
Probabilistic identification Set membership identification
7. System identification
• Measurements corrupted by noise and disturbance identified model uncertain
Assumption on Statistical properties Bound
Parameter estimate
Probability density function
Optimal estimate
maximizes likelihood
Set of possible parameters
Optimal estimate minimizes
worst case estimation error
? +
u y
e
Application to
controller design
SM identification topic of the thesise
Probabilistic objectives Non-probabilistic objectives
Probabilistic identification Set membership identification
8. Controller design classification
Identification Design
Indirect Direct
Off-lineOn-line
ModelData Controller
Controller identification
Data Controller
Identification Design
ModelData Controller
Closed loop operation
Controller identification
Data Controller
Closed loop operation
9. Controller design classification
Identification Design
Indirect Direct
Off-lineOn-line
• SM useful for robust design (mature theory)
• Only few results on experiment design
• Experiment design for constrained MISO
systems
ModelData Controller
Controller identification
Data Controller
Identification Design
ModelData Controller
Closed loop operation
Controller identification
Data Controller
Closed loop operation
10. Controller design classification
Identification Design
Indirect Direct
Off-lineOn-line
ModelData Controller
Controller identification
Data Controller
Identification Design
ModelData Controller
• SM useful for handling constraints
• No computationally tractable algorithms
• Adaptive model predictive control (MPC)
for constrained MIMO systems
Closed loop operation
Controller identification
Data Controller
Closed loop operation
• SM useful for robust design (mature theory)
• Only few results on experiment design
• Experiment design for constrained MISO
systems
11. Controller design classification
Identification Design
Indirect Direct
Off-lineOn-line
• SM leads to formal stability and
performance guarantees
• Efficient algorithms based on SM
exist
ModelData Controller
Controller identification
Data Controller
Identification Design
ModelData Controller
Closed loop operation
Controller identification
Data Controller
Closed loop operation
• SM useful for robust design (mature theory)
• Only few results on experiment design
• Experiment design for constrained MISO
systems
• SM useful for handling constraints
• No computationally tractable algorithms
• Adaptive model predictive control (MPC)
for constrained MIMO systems
12. Controller design classification
Identification Design
Indirect Direct
Off-lineOn-line
ModelData Controller
Controller identification
Data Controller
Identification Design
ModelData Controller
Closed loop operation
Controller identification
Data Controller
Closed loop operation
• SM could lead to formal stability and
performance guarantees
• No algorithm based on SM exists
• On-line direct controller design for
nonlinear systems
• SM leads to formal stability and
performance guarantees
• Efficient algorithms based on SM
exist
• SM useful for robust design (mature theory)
• Only few results on experiment design
• Experiment design for constrained MISO
systems
• SM useful for handling constraints
• No computationally tractable algorithms
• Adaptive model predictive control (MPC)
for constrained MIMO systems
13. Thesis contributions
• Worst-case experiment design for constrained linear systems
• Adaptive model predictive control based on set membership identification
• On-line direct data driven control design based on set membership identification
14. Thesis contributions
• Worst-case experiment design for constrained linear systems
• Adaptive model predictive control based on set membership identification
• On-line direct data driven control design based on set membership identification
15. Set membership identification
je(t)j · ²; 8t
• System model linear in parameters, disturbance magnitude bounded
• Collection of input-output data
y(t) = '(t)Tµ + e(t)
yN = fy(t)gN
t=1 ; 'N = f'(t)gN
t=1
16. Set membership identification
je(t)j · ²; 8t
• System model linear in parameters, disturbance magnitude bounded
• Feasible parameter set: all parameters consistent
with assumptions and collected data
• Collection of input-output data
y(t) = '(t)Tµ + e(t)
yN = fy(t)gN
t=1 ; 'N = f'(t)gN
t=1
FPS =
n
µ : jy(t) ¡ '(t)T µj · ²; t = 1; : : : ; N
o
17. Set membership identification
je(t)j · ²; 8t
• System model linear in parameters, disturbance magnitude bounded
• Feasible parameter set: all parameters consistent
with assumptions and collected data
• Optimal estimate
• Collection of input-output data
• Worst case estimation error (radius of information) measure of uncertainty
y(t) = '(t)Tµ + e(t)
yN = fy(t)gN
t=1 ; 'N = f'(t)gN
t=1
FPS =
n
µ : jy(t) ¡ '(t)T µj · ²; t = 1; : : : ; N
o
^µ¤
p = arg inf
v
sup
µ2FPS
jjv ¡ µjjp
Rp(FPS) = sup
µ2FPS
jjµ¤
p ¡ µjjp
Rp(FPS)
^µ¤
p
18. • Worst case radius of information upper bound on
• Experiment design: find input sequence that minimizes
• Existing results for SISO systems, FIR model and input magnitude bound
Worst case experiment design
Depends on the
applied inputs only
Rp(FPS)
Rp('N) = sup
µ;je(t)j·²;t=1;:::;N
Rp(FPS)
Rp('N) R
¤
p
19. • Worst case radius of information upper bound on
• Experiment design: find input sequence that minimizes
• Existing results for SISO systems, FIR model and input magnitude bound
• We consider MISO systems subject to polytopic constraints
Worst case experiment design
Depends on the
applied inputs only
'(t) = A'(t ¡ 1) + Bu(t ¡ 1)
Constraints form a convex compact set that contains origin
controllable, stableA; B A
Rp(FPS)
Rp('N) = sup
µ;je(t)j·²;t=1;:::;N
Rp(FPS)
y(t) = '(t)Tµ + e(t)
Cu(t) · o
L¢u(t) · q
Rp('N) R
¤
p
22. Designing the input sequence
• Design input sequences that creates all regressors of interest
• Ideally shortest possible input sequence hard combinatorial problem
• Our solution input sequence that visits all points computationally tractable
23. Designing the input sequence
• Design input sequences that creates all regressors of interest
• Ideally shortest possible input sequence hard combinatorial problem
• Our solution input sequence that visits all points computationally tractable
• A sequence of inputs that brings to in steps has to satisfy'0 'T k
• Requires solution of the linear program (LP)
• Greedy algorithm always find the closest regressor vector
'(0) = '0
'(k) = 'T
'(t) = A'(t ¡ 1) + Bu(t)
Cu(t) · o
L¢u · q
9
>=
>;
8t = 1; : : : ; k
24. Numerical example
• Compare with random input sequence
• Large number of experiments with same disturbance for both input types
• In 13.7% of the experiments with random input actual radius > worst case bound
radius of information
Empirical probability density functionsRadius of information values
experiment realization
optimalinputrandominput
optimalinputrandominput
25. Thesis contributions
• Worst-case experiment design for constrained linear systems
• Adaptive model predictive control based on set membership identification
• On-line direct data driven control design based on set membership identification
26. Adaptive control
• Adaptive control automatically improve controller performance over time
• Useful for time varying or uncertain time invariant systems
Controller Plant
System
identification
27. Standard adaptive control
• Certainty equivalence adaptive control
• Cannot guarantee output constraint satisfaction
Provide a point estimate of the system model
Set the controller as if there is no uncertainty
in the identified model
Plant estimate - point
Controller Plant
System
identification
28. • MPC is method of choice for constrained systems
• Combine MPC with recursive set membership identification
Adaptive control with constraints
Use MPC to satisfy output constraints
for all the models in the set
Provide a set of models consistent with initial
assumptions and measured data
Plant estimate - set
Plant estimate - point
Controller Plant
System
identification
29. Problem definition
• Discrete, open-loop stable, LTI,
SISO system, FIR model
Can generalize to MIMO systems, basis
transfer function parameterization
'(t) = [u(t ¡ 1); : : : ; u(t ¡ m)]T
Length of the impulse response
y(t) =
mX
k=1
#(k)u(t ¡ k) + e(t)
= '(t)T µ + e(t)
µ = [#(1); : : : ; #(m)]T
m
30. Problem definition
• Discrete, open-loop stable, LTI,
SISO system, FIR model
Can generalize to MIMO systems, basis
transfer function parameterization
'(t) = [u(t ¡ 1); : : : ; u(t ¡ m)]T
Length of the impulse response
y(t) =
mX
k=1
#(k)u(t ¡ k) + e(t)
= '(t)T µ + e(t)
µ = [#(1); : : : ; #(m)]T
• Goal: track a reference and satisfy constraints
Cu(t) · o
L¢u(t) · q
Hy(t) · p
; 8t 2 Z Convex compact set
that contains origin
m
31. 2 4 6 8 10 12 14
-1.5
-1
-0.5
0
0.5
1
1.5
Problem definition
• Known bounds on disturbance and
noise signal magnitude
• True system unknown, but its impulse
response coefficients lie in a known set
• Discrete, open-loop stable, LTI,
SISO system, FIR model
Can generalize to MIMO systems, basis
transfer function parameterization
time step
'(t) = [u(t ¡ 1); : : : ; u(t ¡ m)]T
Length of the impulse response
polytope
y(t) =
mX
k=1
#(k)u(t ¡ k) + e(t)
= '(t)T µ + e(t)
µ = [#(1); : : : ; #(m)]T
je(t)j · ²; 8t 2 Z
µ 2 FPS(0)
FPS(0)
• Goal: track a reference and satisfy constraints
Cu(t) · o
L¢u(t) · q
Hy(t) · p
; 8t 2 Z Convex compact set
that contains origin
impulse response
coefficient
magnitude
m
32. • Bounded complexity polytopic update prevents the unlimited growth of face number
Recursive set membership
identification
• Recursive identification algorithm:
At each time step
add 2 new inequalities to FPS(t)
FPS(t) =FPS(t ¡ 1)
n
µ : jy(t) ¡ #(t)T µj · ²
o
33. • Bounded complexity polytopic update prevents the unlimited growth of face number
• Nominal model of the plant:
Calculate by solving an LP
Recursive set membership
identification
Center of maximum volume
norm ball inscribed inl2
• Recursive identification algorithm:
At each time step
add 2 new inequalities to FPS(t)
FPS(t) =FPS(t ¡ 1)
n
µ : jy(t) ¡ #(t)T µj · ²
o
FPS(t)
µc(t)
µc(t)
34. MPC algorithm
Estimate of the current
disturbance
• Finite horizon optimal control problem:
cost function based on the nominal model, prediction horizon:
subject to:
N ¸ m
min
U
t+N¡1X
k=t
(^y(kjt) ¡ ydes(k))T
Q (^y(kjt) ¡ ydes(k))T
+ u(kjt)T Su(kjt) + ¢u(kjt)T R¢u(kjt)
^y(k + 1jt) = '(kjt)tµc(t) + ^e(t)
^e(t) = y(t) ¡ '(t)T µc(t)
35. last m predicted control moves constant
MPC algorithm
Estimate of the current
disturbance
• Finite horizon optimal control problem:
cost function based on the nominal model, prediction horizon:
subject to:
N ¸ m
robust satisfaction of output constraints
• Can be formulated as a Quadratic Program (QP) of moderate size
• Recursive feasibility and integral action guaranteed
u(t + N ¡ mjt) = : : : = u(t + N ¡ 1jt)
min
U
t+N¡1X
k=t
(^y(kjt) ¡ ydes(k))T
Q (^y(kjt) ¡ ydes(k))T
+ u(kjt)T Su(kjt) + ¢u(kjt)T R¢u(kjt)
^y(k + 1jt) = '(kjt)tµc(t) + ^e(t)
^e(t) = y(t) ¡ '(t)T µc(t)
Cu(kjt) · o L¢u(t) · q
H'(kjt)T µ · p; 8µ 2 FPS(t)
36. Application case studies
• Quad-tank experimental results
• Non-minimum phase MIMO system
• Goal: track reference and enforce
constraints
• Building climate control simulation results
• Typical office room model
• Goal: minimize energy consumption and
enforce occupant comfort constraints
37. Quad-tank experimental example
Non-adaptive MPC
Adaptive MPC with set membership
Adaptive MPC with least squares
• Adaptive MPC with least squares:
Constraint violation during adaptation
• Adaptive MPC with set membership
No constraint violation
38. Conclusion and outlook
• Set membership identification can bring a lot of benefits to controller design
Deterministic and accurate uncertainty description
Can deal with hard constraints
Can make derivation of stability and performance guarantees easier
• Has to be used in a smart way prevent excessive conservativeness and
computational complexity
• Further research required to exploit all its benefits in controller design
39. Thank you!
• Supervisor
Prof. Manfred Morari
• Committee
Prof. Mario Milanese
• Collaborators
Dr. Lorenzo Fagiano
Prof. Sébastien Mariéthoz
Prof. Carlo Novara
Prof. Roy Smith
Dr. David Sturzenegger
Dr. Paul Goulart
Damian Frick
Oliver Schultes
• Students
Lawrence Minnetian
Isik Ilber Sirmatel
(and many others)
• Colleagues
George Xiaojing
Dr. Claudia Fisher
Dr. Manfred Quack
Alexander Liniger
Dr. Robert Nguyen
Tony Wood
(and many others)
• My family and friends