ZIB-IOL / FrankWolfe.jl

Julia implementation for various Frank-Wolfe and Conditional Gradient variants
MIT License
93 stars 18 forks source link

Vanilla Frank-Wolfe with convex combination representation #96

Closed matbesancon closed 3 years ago

matbesancon commented 3 years ago

As discussed, we need an implementation of vanilla FW which maintains the current iterate as a set of atoms and convex combinations.

This could make the nuclear norm problems more stable by explicitly representing a low-rank matrix as a weighted sum of rank 1, while the current results has a lot of near-zero singular values

pokutta commented 3 years ago

As discussed, we need an implementation of vanilla FW which maintains the current iterate as a set of atoms and convex combinations.

This could make the nuclear norm problems more stable by explicitly representing a low-rank matrix as a weighted sum of rank 1, while the current results has a lot of near-zero singular values

already implemented: run afw with "awaySteps = false" - then it is effectively vanillaFW + activeSet

matbesancon commented 3 years ago

ok yes I'll close this