ARMAX identification

ARMAX models and nonlinear LS

Let θ be the unknown vector of parameters. Let y(t) be the observed output and ŷ(t|θ) be the predicted
output (by our model). Then, we estimate θ by solving the optimization problem
1 X
θ̂N = argminθ (y(t) − y(t|θ))2 . (1)
N t

This is called the prediction error method. (Block diagram for prediction error, Figure 7.1 [3]) The identifi-
cation data is D = {(u(t), y(t))|t = 1, . . . , N }.
Recall that ARX model is given by

y(t) + a1 y(t − 1) + · · · + an y(t − n) = b1 u(t − 1) + · · · + bm u(t − m) + e(t)

and the parameters (a1 , . . . , an , b1 , . . . , bm ) can be found by solving an LS problem.

Consider the following ARMAX model

y(t)+a1 y(t−1)+· · ·+an y(t−n) = b1 u(t−1)+· · ·+bm u(t−m)+e(t)+c1 e(t−1)+· · ·+cl e(t−l). (2)

Let

A(q −1 ) = 1 + a1 q −1 + · · · + an q −n , B(q −1 ) = b1 q −1 + · · · + bm q −m , C(q −1 ) = 1 + c1 q −1 + · · · + cl q −l

θ = (a1 , . . . , an , b1 , . . . , bm , c1 , . . . , cl )T
B(q −1 ) C(q −1 )
G(q, θ) = , H(q, θ) = .
A(q −1 ) A(q −1 )
This model consists of the moving average (MA) part C(q −1 )e(t), the auto regressive (AR) part A(q −1 )y(t)
and the exogeneous input (X) part B(q −1 )u(t). The ARMAX model cannot be written as a linear regression.
Notice that
y(t) = G(q, θ)u(t) + H(q, θ)e(t). (3)
Let
v(t) = H(q)e(t) (4)
where H(q) is deterministic and e(t) is a zero mean uncorrelated white noise i.e., E[e(t)] = 0, E[e(t)e(s)] =
λδ(t − s). Note that
v(t) = H(q)e(t) = y(t) − G(q)u(t). (5)
P∞
Assume that H(q) is monic i.e., H(q) = 1 + k=1 h(k)q −k . Therefore,
∞
X
v(t) = e(t) + h(k)e(t − k) = e(t) + m(t − 1) (6)
k=1

where
∞
X
m(t − 1) := h(k)e(t − k) = (H(q) − 1)e(t). (7)
k=1

We refer to v(t) as the colored noise. It can be estimated as

v̂(t|t − 1) = m(t − 1) = (H(q) − 1)e(t)

= (H(q) − 1)H(q)−1 (y(t) − G(q)u(t)) = (1 − H(q)−1 )(y(t) − G(q)u(t)) (8)
= v̂(y(t − 1), y(t − 2), . . . , u(t − 1), . . .). (9)

1
Since H(q) is monic, H(q)−1 is also monic. Notice that v(t) − v̂(t|t − 1) = e(t) ⊥ v̂(t|t − 1) indicating
that v̂(t|t − 1) is the best MSE estimator of v(t).
The output can be estimated as

ŷ(t|t − 1) = G(q)u(t) + v̂(t|t − 1) = G(q)u(t) + (1 − H(q)−1 )(y(t) − G(q)u(t))

= H(q)−1 G(q)u(t) + (1 − H(q)−1 )y(t). (10)

The prediction error is

y(t) − ŷ(t|t − 1) = y(t) − H(q)−1 G(q)u(t) − (1 − H(q)−1 )y(t)

= H(q)−1 (y(t) − G(q)u(t)) = H(q)−1 H(q)e(t) = e(t). (11)

Notice that the one step prediction error is white noise which is unpredictable. Therefore, we have used all
predictable information in ŷ(t|t − 1). The one step predictor in (10) is the optimal. Observe that

A(q −1 ) B(q −1 ) A(q −1 ) B(q −1 ) A(q −1 )

ŷ(t|t − 1) = u(t) + (1 − )y(t) = u(t) + (1 − )y(t).
C(q −1 ) A(q −1 ) C(q −1 ) C(q −1 ) C(q −1 )
Therefore,

B(q −1 ) A(q −1 )
ŷ(t|θ) = u(t) + (1 − )y(t) (12)
C(q −1 ) C(q −1 )
⇒ C(q −1 )ŷ(t|θ) = B(q −1 )u(t) + (C(q −1 ) − A(q −1 ))y(t)
⇒ ŷ(t|θ) = (1 − C(q −1 ))ŷ(t|θ) + B(q −1 )u(t) + (C(q −1 ) − A(q −1 ))y(t). (13)

Since the parameters are not linear in ŷ(t|θ), ŷ(t|θ) 6= φT (t)θ. However, it can be written in a pseudo linear
regression form as

ŷ(t|θ) = (1 − C(q −1 ))ŷ(t|θ) + B(q −1 )u(t) + (C(q −1 ) − 1)y(t) + (1 − A(q −1 ))y(t)

= (1 − A(q −1 ))y(t) + B(q −1 )u(t) + (C(q −1 ) − 1)(y(t) − ŷ(t|θ)). (14)

Define
(t|θ) := y(t) − ŷ(t|θ). (15)
Then,

ŷ(t|θ) = −a1 y(t − 1) − · · · − an y(t − n) + b1 u(t − 1) + · · · + bm u(t − m) +

c1 (t − 1|θ) + · · · + cl (t − l|θ). (16)

Define the parameter and the regressor vectors as

θ := (a1 , . . . , an , b1 , . . . , bm , c1 , . . . , cl )T (17)
T
φ(t|θ) := (−y(t − 1), . . . , −y(t − n), u(t − 1), . . . , u(t − m), (t − 1|θ), . . . , (t − l|θ)) . (18)

Thus, we obtain the prediction in pseudo linear regression form ŷ(t|θ) = φT (t|θ)θ. This is a nonlinear
optimization problem
1 X
θ̂(N ) = argminθ (y(t) − ŷ(t|θ))2 (19)
N
where ŷ(t|θ) = φT (t|θ)θ.

1. Assume C(q −1 ) = 1 and solve the ARX problem with the same order of A(q −1 ) and B(q −1 ) to get
an initial estimate θ̄ = (a1 , . . . , an , b1 , . . . , bm ).

2. Set θ(0) := (θ̄, 0, . . . , 0)T and compute (t − 1|θ(0) ), . . . , (t − l|θ(0) ).

2
 3. For i = 1 to M ,
Construct φ(t|θ(i−1) ), t = 1, . . . , N
 −1  
PN PN
Compute LSE θ̂ =(i)
t=t∗ φ(t|θ̂
(i−1) T
)φ (t|θ̂ (i−1) ) t=t∗ φ(t|θ̂
(i−1) )y(t) where t∗ =max(m, n, l).
Compute (t − 1|θ(i) ), . . . , (t − l|θ(i) ).
End.

This is not a convex optimization problem and we may find a local solution.
Exercise: Take a first order model and implement the algorithm.

References
[1] H. Asada, Identification, Estimation and Learning, MIT, Lecture notes and video lectures, 2021.

[2] L. Ljung, System Identification, Theory for the user, PHI, 2nd Edition, 1999.

[3] T. Söderström, P. Stoica, System Identification, Prentice Hall, 1989.

[4] P. Van den Hof, Lecture notes on System Identification, 2020.

[5] M. Diehl, Lecture notes on Modeling and System Identification, Lecture notes and video lectures, 2020.