Sie sind auf Seite 1von 40

Dept.

of Energy and Systems Engineering

University of Pisa (Italy)

System Identication by Nuclear Norm Minimization


eng. Sergio Grammatico grammatico.sergio@gmail.com

Class of Identification of Uncertain Systems 2011/12 held by Prof. Andrea Caiti


S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 1 / 31

Outline
1

Introduction Identication problem Subspace Identication Nuclear Norm Minimization Rank Minimization Nuclear Norm Minimization Identication by Nuclear Norm Minimization Basics and Variations Simulations Benchmark examples for System Identication Conclusion
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 2 / 31

Outline
1

Introduction Identication problem Subspace Identication Nuclear Norm Minimization Rank Minimization Nuclear Norm Minimization Identication by Nuclear Norm Minimization Basics and Variations Simulations Benchmark examples for System Identication Conclusion
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 3 / 31

Semi-Denite Programming for System Identication


Thanks to Zhang Liu and Lieven Vandenberghe

This presentation is based on Z. Liu and L. Vandenberghe, Interior-point method for nuclear norm approximation with application to system identication, SIAM Journal on Matrix Analysis and Applications, vol. 31, n. 3, pp. 12351256, 2009. Z. Liu and L. Vandenberghe, Notes on rst-order method for nuclear norm approximation problems and system identication, e-mail communications between Zhang Liu and Sergio Grammatico, November 2010.

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

4 / 31

Problem statement
State-space identication of a discrete-time LTI model

Discrete-time LTI system x(t + 1) = Ax(t) + Bu(t) y (t) = C x(t) + Du(t), x(t) Rn , u(t) Rm , y (t) Rp , t = 0, 1, ..., N, A Rnn , B Rnm , C Rpn , D Rpm . (1)

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

5 / 31

Problem statement
State-space identication of a discrete-time LTI model

Discrete-time LTI system x(t + 1) = Ax(t) + Bu(t) y (t) = C x(t) + Du(t), x(t) Rn , u(t) Rm , y (t) Rp , t = 0, 1, ..., N, A Rnn , B Rnm , C Rpn , D Rpm . (1)

Identication problem Given sequences u(0), u(1), ..., u(N) and y (0), y (1), ..., y (N), nd n, A, B, C , D and the initial state x(0).
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 5 / 31

Subspace Identication
Main ideas (1)

Dene the block Hankel matrices y (0) y (1) Y = . . . y (1) y (2) . . . y (2) y (3) . . . y (N r ) y (N r + 1) . . . y (N)

(2)

y (r ) y (r + 1) y (r + 2)

u(0) u(1) U = . . .

u(1) u(2) . . .

u(2) u(3) . . .

u(N r ) u(N r + 1) . . . u(N)


Identication of Uncertain Systems

(3)

u(r ) u(r + 1) u(r + 2)


S. Grammatico (DESE UNIPI) Nuclear Norm Minimization

6 / 31

Subspace Identication
Main ideas (2)

(1) Y = G X + HU, where G = C CA CA2 . . . CAr D CB CAB . . . 0 D CB . . . 0 0 D . . . .. . 0 0 0 . . . D

(4)

, H =

, (5)

CAr 1 B CAr 2 B CAr 3 B

X =

x(0) x(1) x(2)

x(N r ) .

(6)

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

7 / 31

Subspace Identication
Main ideas (3)

Multiply (4) on the right by U (UU = 0).

Main equality Y U = G X U smaller rank YU smaller identied n... (7)

B. De Moor, M. Moonen, L. Vandenberghe and J. Vandewalle, A geometrical approach for the identication of state space models with the singular value decomposition, Proc. of the IEEE Int. Conf. on Acoustics, Speech and Signal Processing, Washington (USA), 1988.
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 8 / 31

Subspace Identication
(Approximate) Solution (1)

Change of coordinates: x(t) = T x (t), T non singular. G = basis for the range of YU G = GT = G0 G1 . . . r G C CA = . T . . CAr

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

9 / 31

Subspace Identication
(Approximate) Solution (2)

C = G0

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

10 / 31

Subspace Identication
(Approximate) Solution (2)

C = G0 G1 G2 . . . r G G0 G1 . . . r 1 G A

A Least Squares solution of

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

10 / 31

Subspace Identication
(Approximate) Solution (2)

C = G0 G1 G2 . . . r G G0 G1 . . . r 1 G A

A Least Squares solution of

0 , B, D Least Squares solutions of x


t1

x C At 0 +
k=0

C Atk Bu(k) + Du(t) = y (t),

t = 0, 1, ..., N. (8)

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

10 / 31

Subspace Identication
Some references

P. Van Overschee and B. De Moor, N4SID: subspace algorithms for the identication of combined deterministic-stochastic systems, Automatica, vol. 30, pp. 7593, 1994. M. Verhaegen, Identication of the deterministic part of MIMO state space models given in innovations form from input-output data, Automatica, vol. 30, pp. 6174, 1994. M. Viberg, Subspace-based methods for the identication of linear time-invariant systems, Automatica, vol. 31, pp. 18351851, 1995. P. Van Overschee and B. De Moor, Subspace identication for linear systems. Theory, implementation, applications, Kluwer Academic Publishers, 1996. L. Ljung, System Identication Toolbox, MathWorks, 2011.
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 11 / 31

Outline
1

Introduction Identication problem Subspace Identication Nuclear Norm Minimization Rank Minimization Nuclear Norm Minimization Identication by Nuclear Norm Minimization Basics and Variations Simulations Benchmark examples for System Identication Conclusion
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 12 / 31

Rank minimization
Problem formulation

Rank minimization problem Given B Rpq , A1 , A2 , ..., An Rpq , solve


xRn

min rank (A(x) B) ,

(9)

where A(x) = x1 A1 + x2 A2 + ... + xn An .

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

13 / 31

Rank minimization
Problem formulation

Rank minimization problem Given B Rpq , A1 , A2 , ..., An Rpq , solve


xRn

min rank (A(x) B) ,

(9)

where A(x) = x1 A1 + x2 A2 + ... + xn An . NP-hard M. Fazel, H. Hindi and S. Boyd, A rank minimization heuristic with application to minimum order system approximation; Rank minimization and applications in system theory, Proc. of the IEEE American Control Conference, Arlington (USA) 2001 and Boston (USA) 2004.
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 13 / 31

Nuclear norm minimization


Problem approximation

Nuclear norm (trace norm) approximation (9) min A(x) B n


xR

(10)

Nuclear norm Z

= sum of the singular values of Z .

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

14 / 31

Nuclear norm minimization


Problem approximation

Nuclear norm (trace norm) approximation (9) min A(x) B n


xR

(10)

Nuclear norm Z

= sum of the singular values of Z .

Nuclear norm minimization as SDP


x,U,V

min trace (U) + trace (V ) U (A(x) B) A(x) B V Rqq


Identication of Uncertain Systems 14 / 31

sub. to

0,

(11)

U = U Rpp , V = V
S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Nuclear norm minimization


l 1 -norm minimization as a particular case

Diagonal case Z = diag (zi ) Z

=
i

|zi | = z

(12)

Vector case of (9)


xRn

min Px q

(13)

P Rpn , q Rp , solved by the Linear Program min 1 y


x,y

(14)

sub. to y Px q y
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 15 / 31

Outline
1

Introduction Identication problem Subspace Identication Nuclear Norm Minimization Rank Minimization Nuclear Norm Minimization Identication by Nuclear Norm Minimization Basics and Variations Simulations Benchmark examples for System Identication Conclusion
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 16 / 31

Identication by Nuclear Norm Minimization


Back to the main equality of Subspace Identication

Main equality (7) Y U = G X U smaller rank YU smaller identied n...

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

17 / 31

Identication by Nuclear Norm Minimization


Back to the main equality of Subspace Identication

Main equality (7) Y U = G X U smaller rank YU smaller identied n...

Key idea min rank YU


(t) y

min
(t) y

YU

(15)

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

17 / 31

Identication by Nuclear Norm Minimization


Resolving Optimization

Trade-o: identied dimension rank(YU ) vs output deviation ((t) y (t)) y Nuclear Norm Minimization with Quadratic Regularization
N

min
(k) y

YU

+
t=0

(t) y (t) y

2 Wt

(16)

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

18 / 31

Identication by Nuclear Norm Minimization


Resolving Optimization

Trade-o: identied dimension rank(YU ) vs output deviation ((t) y (t)) y Nuclear Norm Minimization with Quadratic Regularization
N

min
(k) y

YU

+
t=0

(t) y (t) y

2 Wt

(16)

Constrained Fashion min


(k) y

YU

(17) t = 0, 1, ..., N
Identication of Uncertain Systems 18 / 31

sub. to
S. Grammatico (DESE UNIPI)

(t) y (t) y
Nuclear Norm Minimization

Identication by Nuclear Norm Minimization


Solving the Subspace Identication problem

= optimal solution of (16) y (0) y (1) y Y= . . . (1) y (2) y . . . (2) y (3) y . . . (N r) y (N r + 1) y . . . (N) y

(r) (r + 1) (r + 2) y y y

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

19 / 31

Identication by Nuclear Norm Minimization


Solving the Subspace Identication problem

= optimal solution of (16) y (0) y (1) y Y= . . . (1) y (2) y . . . (2) y (3) y . . . (N r) y (N r + 1) y . . . (N) y

(r) (r + 1) (r + 2) y y y

As for standard Subspace Identication: x G = basis for the range of YU , A, B, C, D, 0 from G

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

19 / 31

Outline
1

Introduction Identication problem Subspace Identication Nuclear Norm Minimization Rank Minimization Nuclear Norm Minimization Identication by Nuclear Norm Minimization Basics and Variations Simulations Benchmark examples for System Identication Conclusion
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 20 / 31

Simulations
Benchmark examples for System Identication

Experimental data sets Experiment 1 2 3 . . . 10 CD player arm Continuous stirred tank reactor Hair dryer # Inputs 2 1 1 # Outputs 2 2 2 N 800 950 550

Industrial dryer

850

B. De Moor, P. De Gersem, B. De Schutter and W. Favoreel, DAISY: a database for the identication of systems, Journal A, vol. 38, no. 3, pp. 4-5, 1997.
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 21 / 31

Simulation results
MIMO CD Player Arm

2 Inputs (forces of the mechanical actuators) 2 Outputs (tracking accuracy of the arm) Identied dimension: n = 3 N4SID default: n = 10 Validation error denition: eV =
NV 1 t=0 NV 1 t=0

y (t) y (t) y (t)

2 2 yV 2 2

1 2

(18)

Comparison: 0.18 (Nuclear Norm method) vs 0.19 (N4SID with n = 3)


S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 22 / 31

Simulation results
MIMO CD Player Arm in Python (1)

Singular Values (log scale)


100

Original Optimized

10-2 Normalized singular values

10-4

10-6

10-8 0

10

15 Singular value index

20

25

30

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

23 / 31

Simulation results
MIMO CD Player Arm in Python (2)

(Estimated) Outputs
1.5 1.0 Output y0 0.5

S. Grammatico (DESE UNIPI)


0.0 0.5 1.0 1.50 1.5 2.0 1.0 0.5 0.0 0.5 1.0 1.50

100

200

300 Time t

400

500

600

Measured Estimated

Output y1

100

200

300 Time t

400

500

600

Nuclear Norm Minimization

Identication of Uncertain Systems

24 / 31

Simulation results
SISO Heating system

1 Input (input drive voltage) 1 Output (temperature) Identied dimension: n = 3 N4SID default: n = 2 Validation error Comparison: 0.08 (Nuclear Norm method) vs 0.09 (N4SID with n = 2, 3)

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

25 / 31

Simulation results
SISO Heating system in Matlab (1)

Singular Values (log scale)


10^(0) Normalized singular values

10^(2)

10^(4)

10^(6)

10^(8)

10^(10) 2 4 6 8 10 12 14 Singular value index 16 18 20

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

26 / 31

Simulation results
SISO Heating system in Matlab (2)

(Estimated) Outputs
200

180

160 Output y0

140

120

100

80 0

50

100

150

200 Time t

250

300

350

400

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

27 / 31

Simulations
Software implementation

Input/Output experimental data from http://homes.esat.kuleuven.be/smc/daisy/ CVXOPT (Python) sol = sysid.sysid(y[:,:N],u[:,:N],1/gamma,threshold) CVX (Matlab) cvx begin variable y(N,1) minimize( norm nuc(hankel(y(1:r),y(r:end))*Un) + gamma*(y-ym)*(y-ym) ) cvx end
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 28 / 31

Outline
1

Introduction Identication problem Subspace Identication Nuclear Norm Minimization Rank Minimization Nuclear Norm Minimization Identication by Nuclear Norm Minimization Basics and Variations Simulations Benchmark examples for System Identication Conclusion
S. Grammatico (DESE UNIPI) Nuclear Norm Minimization Identication of Uncertain Systems 29 / 31

Conclusion
Summarizing

(Linear) Systems Identication can be casted into SDP Main advantage of the Nuclear Norm technique: automatic selection of the identied-model order

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

30 / 31

Conclusion
Summarizing

(Linear) Systems Identication can be casted into SDP Main advantage of the Nuclear Norm technique: automatic selection of the identied-model order

Convex Optimization System Identication

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

30 / 31

Conclusion
System Identication by Nuclear Norm Optimization: still investigated

K. Mohan and M. Fazel, Reweighted nuclear norm minimization with application to system identication, Proc. of the IEEE American Control Conference, Baltimore (USA), 2010. P.M.O. Gebraad, J.W. van Wingerden, G.J. van der Veen and M. Verhaegen, LPV subspace identication using a novel nuclear norm regularization method, Proc. of the IEEE American Control Conference, San Francisco (USA), 2011.

S. Grammatico (DESE UNIPI)

Nuclear Norm Minimization

Identication of Uncertain Systems

31 / 31

Das könnte Ihnen auch gefallen