TECHNISCHE UNIVERSITÄT MÜNCHEN
LEHRSTUHL FÜR INFORMATIONSTECHNISCHE REGELUNG
ORDINARIA: UNIV.-PROF. DR.-ING. SANDRA HIRCHE
Adaptive and Learning-based Control SS2025
Author: Tzu-Yuan Huang 25.05.2025.
EXERCISE 1
Problem 1: Parameter Identification
Consider the stationary process described by y(k) = c1 + c2 u(k) + e(k) where c1 ∈ R and c2 ∈ R
are unknown parameters, u(k) ∈ R and y(k) ∈ R are the input and output states respectively, and
e(k) ∈ R is an external disturbance. The figure below illustrates the relation between u(k) and y(k)
for k = 1, . . . , 5 and the table provides their numerical values. The unknown parameters c1 and c2
have to be found.
10
6
y(k)
0
0 1 2 3 4 5 6 7
u(k)
u(k) 1 2 3 4 5
y(k) 1 3 3 6 7
a) The least-squares estimation (LS) technique uses a loss function V (θ, k) which is minimal for a
vector of unknown parameters θ in sense of least squares. Derive Y , Φ, and θ necessary for
constructing V (θ, k).
b) Derive a general solution to minimize V (θ, k). Using the obtained expression, solve for θ using the
numerical values given in the table above.
c) A new set of data u(6) = 6 and y(6) = 7 is added. Using the recursive least-square estimation
(RLS) formulas, calculate the new estimate of the parameter vector θ̂(k).
Note: Run simulations S1 P1 param ident.m and S1 P1 param ident recurs.m
Problem 2: Recursive least squares estimation
Consider a linear model
y(k) + a1 y(k − 1) + a2 y(k − 3) = b0 u(k − 2) + b1 u(k − 3) (1)
with the output equation
where a1 , a2 , b0 , and b1 are the parameters of the model.
a) Identify the parameter θ and regressor φ vectors for the recursive least squares estimation.
b) State the equations of the algorithm necessary to obtain the estimates of the parameters online.
Denote the order of the identity matrix needed for successful transformation. Which variables are
needed to initialize the algorithm?
c) State the requirement on the input signal u(k) for parameter convergence in a recursive least squares
estimation as an inequality constraint.
d) What is the influence of the forgetting factor on the quality of parameter identification and when
is it necessary? How would you modify the above algorithm in order to take the forgetting factor
into account and obtain better parameter estimates?
e) Assume that some input-output data can be obtained offline. How could this data be used in order
to improve the behavior of the online algorithm?
Note: Run simulations S1 P2 offline ident.m and S1 P2 online ident.slx
Problem 3: Regressor Models
Write the following systems in the form of a regressor model y(t) = φ(t)T θ.
a) Continuous-time transfer function G(s)
Y (s) c
= 2 (2)
U (s) s + as + b
where y(t), u(t) ∈ R are the system output and the control input respectively, and a, b, c ∈ R
are unknown constants. Assume that the derivatives of the system output up to the order 2 are
available.
b) Non-linear system
x(k + 1) = −ax(k)2 + 2bsin(u(k)) (3)
with unknown parameters a, b ∈ R.
c) State space model
d x1 0 1 x1 0
= + u (4)
x
dt 2 a b x 2 c
x1
y = 1 0 (5)
x2
with unknown parameters a, b, c ∈ R.
