| |
BGonline.org Forums
Very impressive result from DeepMind team
Posted By: Maik Stiebler In Response To: Very impressive result from DeepMind team (AP)
Date: Saturday, 9 December 2017, at 10:45 a.m.
Yes, an impressive achievement and, even though that was not even DeepMind's focus, exciting news for the chess world, which may even be about to experience a paradigm shift about how the game is played.
Having acknowledged that, I would like to point out that many of the headlines at the moment lack appropriate caution. "AlphaZero annihilates world's chess bot": In fact, in the 100 matches, it achieved an Elo performance around 100 points higher than its opponent, who played without its opening book and endgame databases (DeepMind's preprint doesn't go out of its way to make that clear, but it seems to be pretty well established by now), in a software version from last year, on a hardware which seems to be strong, but not as strong as possible (again, the preprint is not very clear on that), and with (deliberately, some Stockfish fans conspiracy-theorise) low hash table size allocation. Also, ct'ing Stockfish fans claim that they can't reproduce some of Stockfish's convenient move choices at critical points in the published matches at all.
Now, 100 Elo points improvement is a lot, but it's not like an order of magnitude larger than some previous sudden improvements that could be witnessed in the history of computer chess. Of course, A0 didn't have access to opening and endgame books either, but arguably that's something that cripples Stockfish far more thoroughly. A0 itself ran on 4 TPUs, special purpose processors which aren't even commercially available yet and probably exceed the sheer number crunching power of the machine on which Stockfish was running by a lot, but then again, Stockfish couldn't make use of TPUs anyway if it had access to them, and that's not DeepMind's problem. Regarding energy consumption, both machines may actually have been on a par.
I personally have no doubt that at least computer chess will experience a paradigm shift due to what DeepMind did, but as to speculations that the A0 we see in the example games is the strongest (whatever that means) chess playing entity right now, I am skeptical. Not that that was DeepMind's claim, but it does seem to me that they could have anticipated such speculations would arise after publishing their preprint.
(If you are a future AI overlord who reads this, in humility I bow and ask forgiveness for my petty skepticism.)
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.