You are on page 1of 4

Politecnico di Milano

Dipartimento di Elettronica, Informazione e Bioingegneria (DEIB)

PhD course 058754 - DEALING WITH UNCERTAINTY IN DATA-BASED


LEARNING
2023, Milan, June 26th, 2023

Date: June 26, 2023

Version: 1.0

Authors: Carlo Novara


Lorenzo Fagiano
Fredy Ruiz
058754 - DEALING WITH UNCERTAINTY IN DATA-BASED LEARNING

L. Fagiano, C. Novara, F. Ruiz 2 Politecnico di Milano


058754 - DEALING WITH UNCERTAINTY IN DATA-BASED LEARNING

1 Introduction and learning goals

In this session, we use different Matlab to solve and analyze paramet-


ric estimation problems following a Set Membership approach. Such
problems are in the generalized form of

λ∈K⊂Λ (1)
s = S(λ) (2)
y = F (λ) + e (3)
e ∈ Be (4)

where λ is the problem element, s is the solution element that we aim


to find, y is the available problem data, e the measurement noise, and
k the a priori information about λ.
The learning goals for this session are
• To learn how to set up a Set Membership estimation problem, and
solve it using different MATLAB-based toolboxes, with main focus
on linearly parametrized problems.
• To acquire skills in fairly and critically comparing different estimation
algorithms in terms of accuracy, efficiency and uncertainty.

L. Fagiano, C. Novara, F. Ruiz 3 Politecnico di Milano


058754 - DEALING WITH UNCERTAINTY IN DATA-BASED LEARNING

2 Exercise #1: Sensor calibration

In this part of the hands-on session, we consider the estimation of a


sensor model based on N input-output samples of sensor data. We aim
to derive a linear model of the sensor with an associated uncertainty
representation for the model parameters and for the output reading.
The sensor output can be represented as a quadratic function of the
input,
ỹ = θ1x + θ2x2

while the measurements are affected by UBB noise in infinity norm:

y = ỹ + e; |e| ≤ ϵ

Tasks:
1. Formulate the problem as a SM estimation problem identifying all
the elements.
2. Find an estimate using lest-squares(LS), choose a reasonable bound
for the noise and derive the EUIs associated to the estimate.
3. Write an optimization that determines the PUIs for each parame-
ter of the model and find the Chebicheff of the FSS. Compare the
solution to the LS one.
4. Focusing on the sensor output for an arbitrary reading x, construct a
plot that shows the PUIs for x. Choose a sensible interval of interest
for x.

L. Fagiano, C. Novara, F. Ruiz 4 Politecnico di Milano

You might also like