Phillip Kerger

UC Berkeley IEOR faculty, ex-NASA, PhD in applied math from Hopkins. Optimization, math, tech, soccer, cats.

[Phillip Kerger @p_kerger]

I was curious: Do FIFA Hydration Breaks Change the Flow of the Game? I put together an article with some statistical evidence of how they do! https://t.co/McLK9tv1g1

[Phillip Kerger @p_kerger]

@octonion this reminds me of "the model doesn't care by how much it wins, only *that* it wins" in AlphaGo, where the AI would make silly looking moves because it was sure it already won. It doesn't care about making the result pretty, just about getting the result right

[Phillip Kerger @p_kerger]

@memecrashes Many of these are difficult-to-find counterexamples to conjectures. Without knowing those conjectures are false, people would have spent countless more hours trying to prove them with zero hope.

[Phillip Kerger @p_kerger]

@jasondeanlee @SebastienBubeck yes correct, this bound is a strengthening specifically assuming accuracy regime eps = O(R/sqrt(d))), where the optimal solution is in the Euclidean ball of radius R. Though nothing rules out a constant-error LB of d^2 existing, but would need a different construction definitely

[Phillip Kerger @p_kerger]

@SebastienBubeck I was stunned! Lots of credit to those who worked on the CDC conjecture result, since it was their approach that I adapted for this.

[Phillip Kerger @p_kerger]

@benellsworth01 @SebastienBubeck yes, the lean verification is of a full lower bound statement! I've fully verified for accuracy of order d^-3, but still capturing the main result of a d^2 lower bound, and the verification for the d^-1/2 is coming, since this is getting attention! Github is linked in the post

[Sebastien Bubeck @SebastienBubeck]

This is actually crazy, many of us thought about this lower bound for many years ...
https://t.co/20T6UMJ5lZ