YILDIZ TECHNICAL UNIVERSITY
FACULTY OF ELECTRICAL AND ELECTRONICS ENGINEERING
DEPARTMENT OF CONTROL AND AUTOMATION ENGINEERING
KOM 4570 System Identi cation Spring 2025-2026
Homework 6
1. The parametric models we identify are typically input/output models. Consider a system
modeled by the dierence equation:
y(t) + ay(t − 1) = bu(t − 1) + e(t)
where a and b are known constants, and u(t) and e(t) are realizations of independent white
noise sequences of variances σu2 and σe2 , respectively.
v[n] − 0.5v[n − 1] = 2e[n] + e[n − 1].
(a) Determine the transfer functions G(q) and H(q) in
y(t) = G(q)u(t) + H(q)e(t)
when y(t), u(t) and e(t) obey the dierence equation above.
(b) Determine the appropriate model structure (ARX, ARMAX, OE, FIR, or BJ) that
should be used for identifying this system.
(c) What is the major advantage of using an ARX model over a Box-Jenkins (BJ) model
when tting this data?
2. Let an input signal u(t) be made up of the sum of k sinusoids at dierent pulsations ωi (for
i = 1...k ) with ωi < π .
k
X
u(t) = sin(ωi t)
i=1
What is the minimum value of k required in order to be able to identify appropriately a
full-order model of the true system using a model structure like OE(nb = 2, nf = 4, nk = 3)?
Explain your reasoning.
3. Prediction error consists entirely of the noise innovations that enter the system.
(a) Determine the 3-step ahead predictor for the system y(k) = e(k) + ce(k − 1).
(b) What is the variance of the associated prediction error for this 3-step ahead prediction?
4. In the process of picking the best model, we must balance goodness of t vs. complexity
(a) Why does Akaike's Information Criterion (AIC) error function (or the log loss function
on training data) continue to decrease with an increasing number of model parameters?
(b) Is this continuous decrease of error on the training set a reliable indication of increased
model accuracy? Explain the risk of "over tting".
5. To prove a mathematical model is a good representation of the real physical system, we
perform residual tests (auto-correlation and cross-correlation)
(a) Explain why the auto-correlation function Rε (τ ) of the residual signal might not have
the ideal shape of the auto-correlation function of a pure white noise process, even if
the model is deemed acceptable.
1
(b) What does a non-zero value at τ = 0 indicate in the cross-correlation test between the
input and the residual?
6. Consider a true system S :
y(t) = b0 u(t − 1) + e(t)
You have identi ed the parameter estimate b̂N = 0.5 with an estimated variance σb2 = 0.01
(a) Determine a con dence interval containing the true parameter b0 with a probability of
95%.
(b) How does this interval relate to the Cramér-Rao Lower Bound (CRLB) if the chosen
model structure perfectly encompasses the true system and the noise is Gaussian?