Sparsity in system identification and data-driven control

Transcription

1 1 / 40 Sparsity in system identification and data-driven control Ivan Markovsky

2 This signal is not sparse in the "time domain" 2 / 40

3 But it is sparse in the "frequency domain" (it is weighted sum of six damped sines) 3 / 40

4 Problem: find sparse representation (small number of basis signals) 4 / 40 existence representation approximation

5 System theory offers alternative methods based on low-rank approximation 5 / 40 y(1) y(2) y(3) y(2) y(3) y(4) rank of y(3) y(4) y(5) y(l) y(l + 1) y(l + 3)

6 6 / 40 Plan Sparse signals and linear-time invariant systems System identification as sparse approximation Solution methods and generalizations

7 Sum-of-damped-exponentials signals are solutions of linear constant coefficient ODE 7 / 40 y = α 1 exp z1 + + α n exp zn exp z (t) := z t p 0 y + p 1 σy + + p n σ n y = 0 (σy)(t) := y(t + 1) y = Cx, σx = Ax x(t) R n state

8 The solution set of linear constant coefficient ODE is linear time-invariant (LTI) system 8 / 40 n-th order autonomous LTI system B := {y = Cx σx = Ax, x(0) R n } dim(b) = n complexity of B L n LTI systems with order n

9 9 / 40 y B L n is constrained/structured/sparse belongs to n-dimensional subspace is linear combination of n signals described by 2n parameters

10 We assume that sparse representation exists, but we do not know the basis 10 / 40 classical definition of sparse signal y y has a few nonzero values (we don t know which ones) basis: unit vectors y B L n with n # of samples y is sum of a few damped sines (their frequencies and dampings are unknown) basis: damped complex exponentials

11 The assumption y B L n makes ill-posed problems well-posed 11 / 40 noise filtering given y = ȳ + ỹ, ỹ noise find ȳ true value forecasting given "past" samples ( y( t + 1),...,y(0) ) find "future" samples ( y(1),...,y(t) ) missing data estimation given samples y(t), t Tgiven find missing samples y(t), t Tgiven

12 Noise filtering: given y = ȳ + ỹ, find ȳ with prior knowledge ȳ B L n, ỹ N(0,νI) 12 / 40

13 Heuristic: smooth the data by low-pass filter 13 / 40

14 Optimal/Kalman filtering requires a model The best (but unrealistic) option is to use B 14 / 40

15 Kalman filtering using identified model B, (i.e., prior knowledge B L n ) 15 / 40

16 16 / 40 Summary the assumption y B L n imposes sparsity the basis is sum-of-damped-exponentials with unknown dampings and frequencies y B L n "regularizes" ill-posed problems

17 17 / 40 Plan Sparse signals and linear-time invariant systems System identification as sparse approximation Solution methods and generalizations

18 18 / 40 System identification is an inverse problem simulation B y given model B Ln and initial conditions find the response y B identification y B given response y and model class Ln find model B Ln that "fits well" y

19 19 / 40 "fits well" is often defined in stochastic setting assumption y = ȳ + ỹ where ȳ B L n is the true signal ỹ N(0,νI) is noise (zero mean white Gaussian) maximum likelihood estimator minimize over ŷ and B y ŷ subject to ŷ B L n "The noise model is just an alibi for determining the cost function." L. Ljung

20 Example: monthly airline passenger data fit by 6th order LTI model 20 / 40

21 21 / 40 How well a given model B fits the data y? error(y,b) := min ŷ B y ŷ likelihood of y, given B projection of y on B validation error identification problem: minimize over B Ln error(y,b)

22 The link between system identification and sparse approximation is low rank 22 / 40 y B L n y(1) y(2) y(t n) y(2) y(3) y(t n + 1) rank... n y(n + 1) y(n + 2) y(t ) Hankel structured matrix H n+1 (y)

23 LTI system identification is equivalent to Hankel structured low-rank approximation minimize over ŷ and B y ŷ subject to ŷ B L n minimize over ŷ y ŷ subject to rank ( H n+1 (ŷ) ) n 23 / 40

24 24 / 40 Summary system identification aims at a map y B the map is defined through optimization problem equivalent problem: Hankel low-rank approx. (impose sparsity on the singular values)

25 25 / 40 Plan Sparse signals and linear-time invariant systems System identification as sparse approximation Solution methods and generalizations

26 26 / 40 Three solution approaches: nuclear norm heuristic subspace methods local optimization

27 The nuclear norm heuristic induces sparsity on the singular values 27 / 40 rank: number of nonzero singular values : l 1 -norm of the singular values vector minimization of the nuclear norm tends to increase sparsity = reduce rank leads to a convex optimization problem

28 Nuclear norm minimization methods involve a hyper-parameter minimize over ŷ y ŷ subject to H n+1 (ŷ) γ minimize over ŷ α y ŷ + H n+1 (ŷ) γ/α determines the rank of H n+1 (ŷ) we want α opt = max{α rank ( H n+1 (ŷ) ) n} α opt can be found by bijection 28 / 40

29 Originally the subspace identification methods were developed for exact data 29 / 40 L n class of LTI systems of order n state space representation B := {y = Cx σx = Ax, x(0) R n } exact identification problem y (A, C) given y B Ln exact data find (A,C) model parameters

30 Two steps solution method 1. rank revealing factorization C CA [ H L (y) =. CA L+1 } {{ } O ] x(0) Ax(0) A 2 x(0) A T L x(0) }{{} C 2. shift equation C CA. CA L 1 CA A = CA 2. CA L O(1:L 1,:)A = O(2:L,:) T = 2n + 1 samples suffice, L [n + 1,T n] 30 / 40

31 For noisy data, subspace methods involve unstructured low-rank approximation 31 / 40 do steps 1 and 2 approximately: 1. singular value decomposition of H L (y) 2. least squares solution of the shift equation L is hyper-parameter that affects the solution B

32 32 / 40 Local optimization using variable projections "double" optimization min B L n ( min ŷ B ) y ŷ "inner" minimization "outer" minimization error(y, B) = Π By min B L n error(y, B)

33 Representation of an LTI system as kernel of polynomial operator 33 / 40 p 0 y + p 1 σy + + p n σ n y = 0 (σy)(t) := y(t + 1) p(σ)y = 0, where p(z) = p 0 + p 1 z + + p n z n ] model parameter p = [p 0 p 1 p n

34 34 / 40 Parameter optimization problem optimization over a manifold min error(y, B) B L n min error(y,p) p = optimization over Euclidean spaces [ ] p = x 1 Π p 0 Π permutation Π fixed total least-squares Π can be changed during the optimization

35 35 / 40 Three generalizations systems with inputs missing data estimation nonlinear system identification

36 36 / 40 Dealing with missing data minimize over ŷ y ŷ v subject to rank ( H n+1 (ŷ) ) n weighted 2-norm approximation y ŷ v := k,t v k (t) ( y k (t) ŷ k (t) ) 2 with element-wise weights v k (t) (0, ) if y k (t) is noisy approximate y k (t) v k (t) = 0 if y k (t) is missing interpolate y k (t) v k (t) = if y k (t) is exact ŷ k (t) = y k (t)

37 Example: piecewise cubic interpolation vs LTI identification on the "airline passenger data" 37 / 40

38 38 / 40 Conclusion y is response of LTI system y sparse LTI identification low-rank approx. solution methods convex relaxation (nuclear norm) subspace (SVD + least squares) local optimization

39 39 / 40 DFT analysis suffers from the "leakage" 0 0

40 Gridding the frequency axis and using l 1 -norm minimization has limited resulution 40 / 40 given signal y select "dictionary" Φ(t) = [ ] sin(ω 1 t) sin(ω N t) minimize over a a 1 subject to y = Φa