Recursive Least Squares Parameter

Estimation for Linear Steady State and

Dynamic Models

Thomas F. Edgar
Department of Chemical Engineering
University of Texas
Austin, TX 78712

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 1
 Outline
• Static model, sequential estimation
• Multivariate sequential estimation
• Example
• Dynamic discrete-time model
• Closed-loop estimation

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 2
 Least Squares Parameter Estimation

Linear Time Series Models

ref: PC Young, Control Engr., p. 119, Oct, 1969

scalar example (no dynamics)

model y = ax
data y * = ax + ∈ ∈: error

least squares estimate of a: (aˆ )

∑ ( ax
ˆ −y )
k
* 2
min i i (1)
aˆ i =1

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 3
 Simple Example

The analytical solution for the minimum (least squares) estimate is

−1
 k
  k *
ˆak =  ∑ xi 2   ∑ xi yi  (2)

i
−1
  
i
−1

pk bk

pk, bk are functions of the number of samples

This is the non-sequential form or non-recursive

form

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 4
 Sequential or Recursive Form

To update aˆk based on a new data pt. (yi , xi), in Eq. (2) let
k
pk −1
= ∑ xi 2 = pk −1−1 + xk 2 (3)
i =1

and k
bk = ∑ xi y i * = bk −1 + xk y k * (4)
i =1

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 5
 Recursive Form for Parameter Estimation


(
aˆk = aˆk −1 − K k xk aˆk −1 − y k *
 
) (5)
estimation error
where

( )
−1
K k = pk −1xk 1 + pk −1xk 2 (6)

To start the algorithm, need initial estimates aˆ0

and p0. To update p,

( )
−1
pk = pk −1 − p k −1
2
xk 1 + pk −1xk
2 2
(7)

(Set p0 = large positive number)

Eqn. (7) shows pk is decreasing with k

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 6
 Estimating Multiple Parameters
(Steady State Model)

y = x a = a1x1 + a2 x2 + ⋯ + an xn (8)
T

( )
2
min k
∑ i − (non-sequential solution requires
T *
x ˆ
a y i
aˆ i =1 n x n inverse)

To obtain a recursive form for aˆ,

−1 −1
= P k −1 + x k x k
T
Pk (9)
B k = B k −1 + x k y k * (10)

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 7
 Recursive Solution

( ) ( )
−1
a k = a k −1 − P k −1 x k 1 + x k P k −1 x k x k aˆ k −1 − y k * (11)
T T
ˆ ˆ



Kk estimation
error

−1
 
P k = P k −1 − P k −1 x k 1 + x k P k −1 x k  x k P k −1
T T
(12)

need to assume aˆ 0 (vector) P11 0 

P0 =  P22 

P0 (diagonal matrix)
 0 Pnn 
Pii large

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 8
 Simple Example (Estimate Slope & Intercept)
Linear parametric model
input, u 0 1 2 3 4 10 12 18
output, y 5.71 9 15 19 20 45 55 78
y = a1 + a2 u

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 9
 Sequential vs. Non-sequential Covariance Matrix vs. Number
Estimation of a2 Only (a1 = 0) of Samples Using Eq. (12)

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 10
 Sequential Estimation of aˆ1 and aˆ2 Using Eq. (11)

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 11
Application to Digital Model and Feedback Control

Linear Discrete Model with Time Delay:

y (t ) = a1y (t − 1) + a2 y (t − 2) + ⋯ + an y (t − n )

+ b1u(t − 1 − N ) + b2u(t − 2 − N ) + ⋯ + b r u ( t − r − N ) + d (13)

y: output u: input d: disturbance N: time delay

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 12
 Recursive Least Squares Solution

y (t ) = Ψ T (t − 1)θ (t − 1) + ε (t ) (14)

where
Ψ T (t − 1) = [ y (t − 1), y (t − 2),. . . y(t − n ),
u(t − 1 − N ),. . . u(t − r − N ), 1]
θ T (t − 1) = [a1, a2 ,. . . an , b1, b2 , . . . br , d ].
min t 2
∑ 

Ψ (i − 1)θ (i ) − y (i )
T
(15)
θˆ i =1  
"least squares"
(predicted value of y)

θˆ(t ) = θˆ(t − 1) + P (t )Ψ (t − 1)[ y (t ) −Ψ T (t − 1)θˆ(t − 1)] (16)

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 14
 Recursive Least Squares Solution

P (t ) = P (t − 1) − P (t − 1)ψ (t − 1)[ψ T (t − 1)P (t − 1)Ψ (t − 1) + 1]−1

Ψ T (t − 1)P (t − 1) (17)

P (t − 1)Ψ (t − 1)
K (t ) = (18)
1 +Ψ T (t − 1)P (t − 1)Ψ (t − 1)

P (t ) = [I − K (t )Ψ T (t − 1)]P (t − 1) (19)

θˆ(t ) = θˆ(t − 1) + K (t )[ y (t ) − yˆ (t )] (20)

K: Kalman filter gain

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 15
 Closed-Loop RLS Estimation

• There are three practical considerations in

implementation of parameter estimation algorithms

- covariance resetting

- variable forgetting factor

- use of perturbation signal

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 16
 Enhance sensitivity of least squares estimation algorithms
with forgetting factor λ
t 2

J (θ (t )) = ∑ λ t −i ψ T (i − 1)θ (i ) − y (i ) (21)

i =1
1
P (t ) = [P (t − 1) − P (t − 1)ψ (t − 1)[ψ T (t − 1)P (t − 1)ψ (t − 1) + λ ]−1.
λ
ψ T (t − 1)P (t − 1)] (22)

θˆ(t ) = θˆ(t − 1) + P (t )ψ (t − 1)[ y (t ) − ψ T (t − 1)θˆ(t − 1)] (23)

λ prevents elements of P from becoming too small

(improves sensitivity) but noise may lead to incorrect
parameter estimates
0 < λ < 1.0 λ →1.0 all data weighted equally
λ ~ 0.98 typical

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 17
 Closed-Loop Estimation (RLS)
Perturbation signal is added to process input (via set point)
to excite process dynamics
large signal: good parameter estimates
but large errors in process output

small signal: better control but more sensitivity to noise

Guidelines: Vogel and Edgar, Comp. Chem. Engr., Vol. 12,

pp. 15-26 (1988)
1. set forgetting factor λ = 1.0

2. use covariance resetting (add diagonal matrix D to P

when tr (P) becomes small)

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 18
3. use PRBS perturbation signal only when
estimation error is large and P is not small.
Vary PRBS amplitude with size of elements
proportional to tr (P).
4. P(0) = 104 I

5. filter new parameter estimates

θc (k ) = ρθc (k − 1) + (1 − ρ )θ (k )
ρ: tuning parameter
(θc used by controller)

6. Use other diagnostic checks such as the sign of the

computed process gain

Thomas F. Edgar (UT-Austin) RLS – Linear Models Virtual Control Book 12/06 19