Ke Kern rnel el man anifold ifold re regress gression ion fo for r th the co e coupled upled el elec ectric tric dri rives ves dat ataset aset Mirko ko Mazzolen leni Matteo Sca cande della lla Fabi bio Previd vidi mirko.mazzoleni@unibg.it 11 April 2018 matteo.scandella@unibg.it fabio.previdi@unibg.it
Ou Outline line 1. Introduction and motivation 2. A new framework for nonparametric system identification 3. Application to the coupled electric drive dataset 4. Conclusions and future developments 2/32
Ou Outline line 1. 1. Introd oduc uctio ion n and mo d motiva vation ion 2. A new framework for nonparametric system identification 3. Application to the coupled electric drive dataset 4. Conclusions and future developments 3/32
In Introduction oduction an and d mot otiv ivation ation Data Distance Model 4/32
In Introduction oduction an and d mot otiv ivation ation Data Distance Mod odel el 5/32
In Introduction oduction an and d mot otiv ivation ation Mo Model del definition inition Consider the NARX RX system tem: ๐ฏ: ๐ง ๐ข + 1 = ๐ ๐ฆ ๐ฃ ๐ข , ๐ฆ ๐ง (๐ข) + ๐ ๐ข , where: โข ๏ผ ๐ง ๐ข โ โ is the system output ๏ผ ๐(๐ข) is a nonlinear inear function ction ๐ โ โ ๐ร1 is a regressor vector of past ๐ inputs ๏ผ ๐ฆ ๐ฃ ๐ข = ๐ฃ ๐ข , โฆ , ๐ฃ ๐ข โ ๐ + 1 ๐ โ โ ๐ร1 is a regressor vector of past ๐ outputs ๏ผ ๐ฆ ๐ง ๐ข = ๐ง ๐ข , โฆ , ๐ง ๐ข โ ๐ + 1 ๐ โ โ ๐+๐ร1 ๏ผ ๐ฆ ๐ข = ๐ฆ ๐ง ๐ข , ๐ฆ ๐ง ๐ข ๏ผ ๐ ๐ข โ โ is an additive white noise 6/32
Le Lear arning ning from om dat ata Repr Re producing oducing Kernel rnel Hilbert bert Spaces aces (R (RKHS) KHS) An RKHS is a Hilber bert space ce โ such that: โข a. Its elements are functions ๐ฃ: ฮฉ โ โ , where ฮฉ is a generic set b. โ๐ฆ โ ฮฉ, ๐ ๐ฆ : โ โ โ is a continuous linear functional ๐ฃ โ ๐ฃ(๐ฆ) ๐ฆ โ โ s.t. ๐ ๐ฆ ๐ฃ = ๐ฃ ๐ฆ = ๐ฃ, ๐ Rieszโs repr pres esentation entation theorem em ๏ฎ โ ๐ โข ๐ฆ The function ๐ ๐ฆ (โ ) is called the repr pres esenter enter of evaluatio aluation in ๐ฆ โข 7/32
Lear Le arning ning from om dat ata Repr Re producing oducing Kernel rnel Hilbert bert Spaces aces (R (RKHS) KHS) The reproducing kernel is defined as: ๐ฟ ๐ฆ, ๐จ = โข ๐ ๐ฆ , ๐ ๐จ , ๐ฟ = ฮฉ ร ฮฉ โ โ a. Symmetric: ๐ฟ ๐ฆ, ๐จ = ๐ฟ ๐จ, ๐ฆ b. Semidefinite positive ฯ ๐,๐=1 ๐ ๐ ๐ ๐ ๐ ๐ฟ(๐ฆ ๐ , ๐ฆ ๐ ) โฅ 0 โ๐, ๐ ๐ โ โ, โ๐ฆ ๐ โ ฮฉ Mo Moor ore-Ar Aronsza onszajn jn theo theorem ๏ฎ A RKHS KHS defines a corresponding repr eproducing oducing โข kernel rnel. Conversely, a reproducing oducing kernel rnel defines a unique RKH KHS 8/32
Lear Le arning ning from om dat ata Exa xampl mples of kerne nels ls โข Constant tant kernel: rnel: ๐ฟ ๐ฆ, ๐จ = 1 โข Polyn ynomial mial kernel el: : ๐ฟ ๐ฆ, ๐จ = ๐ฆ โ ๐จ + 1 ๐ rnel: ๐ฟ ๐ฆ, ๐จ = ๐ โ ๐ฆโ๐จ 2 โข Linear ear kernel: rnel: ๐ฟ ๐ฆ, ๐จ = ๐ฆ โ ๐จ โข Gauss ssian an kernel: 2๐2 Kernel rnel comp mposi osition tion th theorem eorem: โข A linear combination of valid kernel functions is a valid kernel function โข The space induced by this kernel is the span of the spaces induced by the single ones ๐ผ =โ ๐ ๐ผ ๐ 9/32
A new ew fram amewo ework rk for or sy syst stem em iden entifi tification cation Kernel rnel me method thods in syste tem identi entific fication ation Stab able spline ine kernel el ๐ก Represe resenters ters Pillonetto , Gianluigi and Giuseppe De Nicolao. โA new kernel - based approach for linear system identification.โ Automatica 46 (2010): 81-93. โข Pillonetto, Gianluigi et al. โA New Kernel-Based Approach for NonlinearSystem Identification .โ IEEE Transactions on Automatic Control 56 (2011): 2825-2840. โข 10/32 10
In Introduction oduction an and d mot otiv ivation ation Data Di Dist stan ance ce Model 11 11/3 /32
Lear Le arning ning from om dat ata No Nonpa parametric rametric learning rning Consider the variational riational formulation: โข ๐ 2 + ๐ ๐ โ ๐ โ 2 ๐ = arg min เท เท ๐ง ๐ โ ๐ ๐ฆ ๐ ๐ง ๐ = ๐ง ๐ข ๐ ; ๐ฆ ๐ = ๐ฆ ๐ข ๐ ๐โโ ๐=1 Tikhonov regularization: ๐ 2 โ penalizes the norm of the fitted function โข The minimization problem is on the RKHS space โ ๏ฎ infinite number of parameters! โข 12/32 12
Lear Le arning ning from om dat ata Re Representer presenter th theorem orem The minimizer imizer of the variational problem is given by: โข ๐ ๐ For some ๐ -tuple ๐(๐ฆ) = เท เท ๐ ๐ ๐ฟ ๐ฆ, ๐ฆ ๐ = เท ๐ ๐ ๐ ๐ฆ ๐ (๐ฆ) ๐ = ๐ 1 , ๐ 2 , โฆ , ๐ ๐ ๐ โ โ ๐ร1 ๐=1 ๐=1 Linear comb mbination ination of the representer esenters of the training points ๐ฆ ๐ , evaluated in the point ๐ฆ โข The solution is expressed as combination of ยซbasis functionsยป which properties are โข determined by โ 13/32 13
ฦธ Le Lear arning ning from om dat ata No Nonpa parametric rametric learning rning โ Solution ution Using the representer theorem it possible to express the variational problem as: โข โข ๐ โ โ ๐ร1 : vector of observations 2 + ๐ ๐ โ ๐ ๐ ๐ง๐ ๐ = arg min ๐ โ ๐ง๐ 2 โข ๐ง โ โ ๐๐ฆ๐ : semidefinite positive and ๐โโ ๐ symmetric matrix, s.t. ๐ง ๐๐ = ๐ฟ(๐ฆ ๐ , ๐ฆ ๐ ) Since the expression is quadratic in ๐ we have the closed osed-fo form solution ution: โข ๐ง + ๐ ๐ โ ๐ฝ ๐ โ ฦธ ๐ = ๐ 14 14/3 /32
Outline Ou line 1. Introduction and motivation 2. A new fr 2. frame mework work fo for nonpa parame rametric ric sy syst stem m ide dentif ific icati ation on 3. Application to the coupled electric drive dataset 4. Conclusions and future developments 15 15/32
Introduction In oduction an and d mot otiv ivation ation Manifol Ma ifold d learning rning (s (sta tatic tic system tems) s) Suppose that the regressorsโ belong to a โข manif ifold old in the regressorsโ space The position ition of the regressors adds prio ior r โข information mation How to incorporate this information in a โข learning ning framework? amework? 16/32 16
Introduction In oduction an and d mot otiv ivation ation Manifol Ma ifold d learning rning (s (sta tatic tic system tems) s) Suppose that the regressorsโ belong to a โข manif ifold old in the regressorsโ space The position ition of the regressors adds prio ior r โข information mation How to incorporate this information in a โข learning ning framework? amework? 17/32 17
In Introduction oduction an and d mot otiv ivation ation Incorpor orporating ating th the ma manifol fold d infor formation mation Semi Semi-su super pervi vise sed d smoothne thness ss assumption ption If two regressors ๐ฆ ๐ and ๐ฆ ๐ in a high-density region are close, then so should be their corresponding outputs ๐ง ๐ and ๐ง ๐ 18/32 18
In Introduction oduction an and d mot otiv ivation ation Link Li k to to dynamic namical al syste tems ms In dynamical systems, regressors can โข be stron ongly gly corr rrelated elated It is meaningful to think that they lie lie on โข a a manifold fold of the regressorsโ space PCA reveals how 91% of variance riance โข explained by one component 19 19/32
Introduction In oduction an and d mot otiv ivation ation Ma Manifol ifold regular ularizati ization on One way to enforce the smoothness ness assumption ption is to minimize the quantity: โข ๐ผ โ ๐ 2 ๐๐ ๐ฆ = เถฑ ๐ ๐ = เถฑ ๐ โ ฮ โ ๐ ๐๐ ๐ฆ ๐ฃ ๐ฃ ๐ ๐ฆ : probability distribution of the regressors, ๐ผ : Gradient, ฮ : Laplacian-Beltrami operators โข Minimizing ๐ ๐ means mi minim imizing izing th the gradient dient of the function โข The term can rarely be computed since ๐ ๐ฆ and ๐ฃ are unkno nown wn โข 20/32 20
Introduction In oduction an and d mot otiv ivation ation Ma Manifol ifold regular ularizati ization on The term ๐ ๐ can be modeled led using the reg egres esso sor r graph ph, encoding connections and โข distasnces between points: ๐ ๐ is a tuning parameter โข Higher value ๏จ similar โข ๏ผ with the regressors as its vertices regressors 2 โ ๐ฆ ๐ โ๐ฆ ๐ 2 ๏ผ the weights on the edges are defined as: 2๐ ๐ ๐ฅ ๐,๐ = ๐ Considering the Laplacian matrix ๐ = ๐ธ โ ๐ โ โ ๐ร๐ , where: โข ๏ผ ๐ธ โ โ ๐ร๐ is a diagonal matrix ๐ธ ๐๐ = ฯ ๐=1 ๐ฅ ๐,๐ ; ๐ โ โ ๐ร๐ contains the ๐ฅ ๐,๐ ๐ 21/32 21
Recommend
More recommend