This preview shows pages 1–3. Sign up to view the full content.
This preview has intentionally blurred sections. Sign up to view the full version.
View Full Document
Unformatted text preview: 2.160 Identification, Estimation, and Learning Lecture Notes No. 3 February 15, 2006 2.3 Physical Meaning of Matrix P The Recursive Least Squares (RLS) algorithm updates the parameter vector ˆ ( θ ( t − 1) based on new data ϕ T ( t ), t y ) in such a way that the overall squared error may ˆ ( gain matrix which contains matrix P t1 . To better understand the RLS algorithm, let us examine the physical meaning of matrix P t1 . be minimal. This is done by multiplying the prediction error ϕ T ( t ) θ ( t − 1) − t y ) with the Recall the definition of the matrix: P t t − 1 T ( i ) T = ∑ ϕ ϕ ( i ) ΦΦ = (17) i = 1 m × t t × m m × m = Φ [ ϕ ( ).. 1 ϕ ( t ) ] ∈ R , Φ T ∈ R , ΦΦ T ∈ R Note that matrix ΦΦ T varies depending on how the set of vectors { ϕ ( i )} span the m  dimensional space. See the figure below. ) 1 ( ϕ ) 2 ( ϕ ) ( i ϕ ϕ –vector are ( ) T ΦΦ max λ ( ) T ΦΦ min λ ) ( min P λ ) ( P λ New data: ) ( t ϕ 1 ) ( − ΦΦ = T P m –dim space mdim space many in this direction Well traveled max less traveled direction Geometric Interpretation of matrix P1 . 1 mxm Since ΦΦ T ∈ R is a symmetric matrix of real numbers, it has all real eigenvalues. The eigen vectors associated with the individual eigenvalues are also real. Therefore, the matrix ΦΦ T can be reduced to a diagonal matrix using a coordinate transformation, i.e. using the eigen vectors as the bases. λ 0 L 0 1 ΦΦ T ⇒ D = 0 M λ 2 M ∈ R mxm ( 1 9 ) O 0 L λ m λ = λ ≥ λ ≥ L ≥ λ = λ min max 1 2 m 1/ λ 0 L 0 1 − 1 ⇒ D − 1 1/ λ 2 M mxm P = ( ΦΦ T ) = 0 M O ∈ R (20) 0 L 1/ λ m The direction of λ ( ΦΦ T ) = The direction of λ min ( P ) . max If λ min = 0 , then det( ΦΦ T ) = 0 , and the ellipsoid collapses. This implies that there is no input data ϕ ( i ) in the direction of λ min , i.e. the input data set does not contain any information in that direction. In consequence, the mdimensional parameter vector θ cannot be fully determined by the data set. In the direction of λ , there are plenty of input data: ϕ ( i ) L . This direction has been max well explored, well excited. Although new data are obtained, the correction to the ˆ parameter vector θ ( t − 1) is small, if the new input data ϕ ( t ) is in the same direction as that of λ . See the second figure above....
View
Full
Document
This note was uploaded on 02/27/2012 for the course MECHANICAL 2.160 taught by Professor Harryasada during the Spring '06 term at MIT.
 Spring '06
 HarryAsada

Click to edit the document details