Time: January 12, 2026 17:00
online

Titel: Nonsmooth quasi‑Newton schemes for efficient large‑scale NLP

Abstract: 

Obtaining numerically efficient and robust solutions for the diverse (often nonsmooth, highly structured, large‑scale) nonlinear programming

(NLP) problems that arise in engineering and machine learning requires integrating problem-specific structure directly into the algorithmic steps. In this talk, I will present some quasi-Newton approaches from our recent work that allow a straightforward incorporation of such structures while preserving practical efficiency. For NLPs with general nonlinear equality and inequality constraints, I will describe an augmented Lagrangian method (with some convergence properties) that uses a simple rule for updating the penalty parameters, thereby alleviating the ill-conditioning caused by the classical linear update scheme.

I will show how these schemes have been implemented in open-source software by exploiting recent advances in automatic differentiation tools.

The talk will conclude with numerical experiments that illustrate the methods' efficiency and with a discussion of promising avenues for future research.