9 ms·
Operations research has known this probably since the 60s and certainly since my dissertation in 2000[1] so I don’t really understand the article? [1] amusingl
by andy_ppp 3y ago
Operations research has known this probably since the 60s and certainly since my dissertation in 2000[1] so I don’t really understand the article?
[1] amusingly my dissertation was about routing taxis efficiently around a set of pickup points, I distinctly remember thinking in a cab from the station that it was a shame there weren’t any devices that had GPS, maps, Internet and a programming interface to build something that looked a lot like Uber… oh well, execution and timing is everything I guess.
- atq2119 3y ago> oh well, execution and timing is everything I guess. Don't forget a willingness to break the law.
- andy_ppp 3y agoHaha, probably not good to say this for my YC application but I’d probably have made Uber absolutely fantastic for the drivers and not made much money and gone out of business once other businesses were competing with me… or made V1 too much like my desired product that it takes me a year to build.
- cbsmith 3y agoThe article, as usual, is terrible. I think this paper is more about figuring out the optimal step sizes.
- wenc 3y agoTypically optimal step sizes (α) along a descent direction in gradient descent are determined using a line search, which is itself a (cheaper) univariate optimization problem. https://en.wikipedia.org/wiki/Line_search https://en.wikipedia.org/wiki/Line_search Most nonlinear optimizers already do this.
- hackerlight 3y agoMy guess is that they proved this for a class of target functions, and it's the proof that is novel (novel at least insofar as the chosen distribution of target functions is concerned). Practitioners always knew, but lacked the proof.