Dissertation defense (December 27, 2021): Thiago Rangel Pesset Gonzaga

Student: Thiago Rangel Pesset Gonzaga

Title: STOCHASTIC MODELING OF ONLINE LEARNING OF AN ADAPTIVE LINEAR STRUCTURE IMPLEMENTED IN BLOCKS

Advisors: Diego Barreto Haddad (advisor) and Felipe da Rocha Henriques (CEFET/RJ) (co-advisor)

Committee: Diego Barreto Haddad (president), Felipe da Rocha Henriques (CEFET/RJ), Eduardo Bezerra da Silva (CEFET/RJ), Tadeu Nagashima Ferreira (UFF).

Day/Time: December 27, 2021 / 14h

Room: https://teams.microsoft.com/l/meetup-join/19%3AAWU0sWZQqcAbmYFqYLCqv91CSvd0jaTXILPSCFcQ9hc1%40thread.tacv2/1639414430868?context=%7B%22Tid%22%3A%228eeca404-a47d-4555-a2d4-0f3619041c9c%22%2C%22Oid%22%3A%2202b6b3d1-811f-4650-909a-d2a79310ba31%22%2C%22MessageId%22%3A%221639414430868%22%7D

Abstract:

Adaptive filtering algorithms are a family of techniques with wide application in highly relevant problems, such as channel equalization, acoustic echo cancellation, noise cancellation, systems identification and time series. This work proposes a stochastic model capable of predicting the learning characteristics of the block LMS algorithm. The analysis is simplified by a model that divorces the radial distribution of the input vectors from the angular distribution, which is discretized. Despite this simplification, the model used for the input signal is consistent with the original autocorrelation matrix of the input data. From this analysis it was possible to model the divergence behavior of the studied algorithms, relating the excess MSE to the amount of blocks used and proportionally to the filter size. Theoretical predictions will be compared with performance curves derived from simulation, in order to assess the accuracy of the resulting estimates.

Comments are closed.