Due Giorni di Algebra Lineare Numerica e Applicazioni

Randomized Gauss-Newton methods for large scale nonlinear least-squares

21 Jan 2025, 11:10

20m

Building E (Aula Magna "Fratelli Pontecorvo", Building E, Polo Fibonacci. Pisa)

Talk Morning Session IV

Speaker

Benedetta Morini (Università di Firenze)

Description

We address the solution of large-scale nonlinear least-squares problems by stochastic Gauss-Newton methods combined with a line-search strategy. The algorithms proposed have per-iteration computational complexity lower than classical deterministic methods, due to the employment of random models inspired by randomized linear algebra tools. Under suitable assumptions, the stochastic optimization procedures can achieve a desired level of accuracy in the first-order optimality condition. We discuss the construction of the random models and the iteration complexity results to drive the gradient below a prescribed accuracy.