Least-Squares Methods For System Identification

9/18/2001
Soft Computing : Least-Squares Estimators
System Identification
The problem of determining a mathematical
model for an unknown system by observing its
input-output data pairs is generally referred to
as system identification
Least-Squares Methods for

System Identification
(Chapter 5)
The purposes of system identification are

to predict a systems behavior,
to explain the interactions and relationships
between inputs and outputs, and
to design a controller or simulation of the
system
Bill Cheetham, Kai Goebel

GE Corporate Research & Development
cheetham@cs
cheetham@cs..rpi.
rpi.edu
goebel@
goebel@cs.
cs.rpi.
rpi.edu
1
Why cover System Identification
Linear Regression
It is a well established and easy to use

technique for modeling a real life system.
Statistical method of fitting data to a linear

model (also called Least Squares Estimator)
+3
+2
+1
y 0
-1
-2
-3
It will be needed for the section on fuzzy-neural

networks.
There will be a homework assignment that
covers it.
x
y=x-3
3
Spring Data
Components of System Identification

Step1: Structure Identification
Experiment
1
2
Force(newtons)
1.1
1.9
Length(inches)
1.5
2.1
3
4
5
3.2
4.4
5.9
2.5
3.3
4.1
6
7
7.4
9.2
4.6
5.0
Determine the class of models within which the search for the
most suitable model is conducted. y = f(u; ) where u is the input
vector and is the parameter vector.
Example:
Linear
y= 0 + 1u1,
Second-order polynomial
1u 1
2u1
Step 2: Parameter Identification

Apply optimization technique to determine parameters
Example:
Linear model where
0
1
y=
What will the length be when the force is 5.0

newtons?
= -3
=1
Page 1
9/18/2001
Parameter Identification
training
data
Target System
to be Identified
System Identification Process
yi
yi - yi
Mathematical Model yi
f(u; )
Specify class of models representing the system
Identification
Techniques
Optimize parameters
Select another model
feedback
Conduct validation tests (Are they satisfactory?)
Training data is used for both system and model.
no
yes
Done
Structure and parameter identification may

need to be done repeatedly
Difference between Target Systems output, yi,

and Mathematical Model output, yi, is used to
update parameter vector, .
7
Least-squares parameter optimization
Least-squares: Error
We will restrict the least-squares discussion

to:
Least squared heavily weights error for data

points that are far from the expected value
+3
+2
+1
0
-1
-2
-3
linear models
models that have linear parameters
y=
1u1
2u 1
static (memory-less) systems
Expected
output depends on current inputs only.
Error 0
+1
-2
+3
output does not depend on history.
Error2 0
Sum of Squared Error = (0 + 0 + 1 + 0 + 4 + 9 + 0) = 14
10
Least-squares: Matrix method

To uniquely identify the unknown vector
necessary that
Least Squares: Matrix method (2)
it is
Example: Find a line from two points (1.1, 1.5),

(1.9, 2.1)
m (# of training items) >= n (# of parameters)

training data is {(ui;yi), i = 1, , m}
If m = n , then we can solve for using
= A-1 y
A=
1 1.1
1 1.9
y=
1
2
1.5
2.1
A-1 =
2.375
-1.25
-1.375
1.25
-1.375
1.25
1.5
2.1
If m = n = 2 where f1(u) = x0 and f2(u) = x1 then

f1(u1)
f2(u1)
A=
u1
=
f1(u2)
f2(u2)
= A-1 y =
y1
y=
u2
2.375
-1.25
.675
.75
y2
11
12
y = .675 + .75x
Page 2
9/18/2001
Least-squares: Matlab
Least-squared: m > n
>>A = [1 1.1;1 1.9]

ans = 1
1
When m > n there are more data pairs than

fitting parameters.
1.1
1.9
An exact solution, satisfying all m equations, is

not always possible.
>>y = [1.5 ; 2.1]
In order to handle this we need to incorporate

an error vector.
ans = 1.5
2.1
13
+e=y
>>inv(A) * y
f1(u1)
fn(u1)
ans = 0.675
0.750
f1(um)
fn(um)
e1
+
n
y1
=
em
ym
14
Least-squared: m > n (2)
Spring Example
The best set of parameters is the one that

minimizes the sum of the squared values of e.
Structure Identification can be done using

domain knowledge.
Theorem 5.1 states that the error is minimized

when
(AT
A)-1
The change in length of a spring is

proportional to the force applied.
ATy
- Hookes law
Proof on page 106 of Jang
length = k0 + k1 force
15
16
Spring Example
Example: Spring data plot
+e=y
1
1
1
1
1
1
1
k
k1
1.1
1.9
3.2
4.4
5.9
7.4
9.2
k0
k1
e1
e2
e3
e4
e5
e6
e7
= (AT A)-1 ATy =
1.5
2.1
2.5
= 3.3
4.1
4.6
5.0
1.20
0.44
When force = 5, length = 1.2 + 5 * .44 = 3.4
17
18
Page 3
9/18/2001
Regression is easy...
or maybe it isnt?
Exhaust gas temperature sampled every flight

(a typical day has about 5 flights).
Same regression performed for the last 10

points
Regression over 10 points:
6 0
5 5
5 0
5 0
4 5
4 0
4 0
3 5
3 0
3 0
2 5
2 0
2 0
1 5
1 0
1 0
5
0
0
20 0
4 00
60 0
8 00
1 00 0
12 00
1 40 0
16 0 0
2 0 0
4 0 0
6 0 0
8 0 0
1 0 0 0
1 2 0 0
1 4 0 0
1 6 0 0
1 8 0 0
1 80 0
19
20
Memory models
Memory models
output depends on history.

f(t) = Df(t - 1) + (1 - D)x(t)
N day running average (GE stock)
+3
+2
+1
0
-1
-2
-3
Expected
Yellow line is smoothed with
= .5
21
22
Page 4

Least-Squares Methods For System Identification

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Least-Squares Methods For System Identification

Uploaded by

Copyright:

Available Formats

9/18/2001

Soft Computing : Least-Squares Estimators

Soft Computing : Least-Squares Estimators