“Optimizing smooth objectives on convex sets without projections”
Wednesday 19 January 2022 h. 14:30 - Room 2AB45 and Zoom - Damiano Zeffiro (Padova, Dip. Mat.)
Abstract
The well known gradient descent method for smooth unconstrained optimization can be extended in a straightforward way to problems with convex constraints by using projections. However, in many cases there are more effective ways to generate feasible descent directions. One of the most popular alternatives to the projected gradient method is the Frank-Wolfe method, characterized by a linear minimization subproblem replacing the projection subproblem.
In this seminar, after a brief review of the above mentioned methods, some examples of sets commonly used in optimization where linear minimization is cheaper than projection will be discussed. Then, variants to improve the convergence rate of the Frank-Wolfe method will be presented, together with a general framework to study such variants. Finally, an algorithm for fast cluster detection in networks based on a Frank-Wolfe variant will be described.
The video of the seminar will appear shortly afterwards in this Mediaspace channel.
