[WIP] Nearest Extreme Point Oracle variant

[pokutta]

[pokutta]

basically implements https://arxiv.org/pdf/2102.02029.pdf

-> not done yet. need to change theory also to be more useful

[matbesancon]

looking quickly at the paper, doesn't this need new LMOs?

[pokutta]

no for e.g., 0-1 problems or feasible regions with fixed norms as it can then be simulated via the LMO. see my implementation. This is actually quite smart as usually we would not pay any attention to „where“ we are but this one does. I can explain tomorrow.

Best, Sebastian On 14. Feb 2021, 23:01 +0100, Mathieu Besançon [email protected], wrote:

looking quickly at the paper, doesn't this need new LMOs? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[matbesancon]

@pokutta I think you have merged this into another branch that got merged into master, it is already there - it is not quite ready because of the interfacing as the gradient change depends on the iteration. i will fix it for vanilla FW today.