Graham Banks wrote:The only thing I question is that AlphaZero had an opening book based on learning over millions of games, whereas it looks as though Stockfish may not have used an opening book at all.
Stockfish was fine out of the opening (in the 10 games at 1 min/move). AlphaZero also played 1200 games against SF from known opening positions and won with a score of 733 (see page 6 of the paper).
-Carl
SF was DEAD LOST in 40% straight out of the opening, and very bad in another 40%.
The myth Alpha is very strong in the endgame and won its games there, is just a myth.
Lyudmil Tsvetkov wrote:Most of your statements make very much sense, except the first one. So why do they weigh boxers and weight-lifters before each and every competition?
But they did - the energy consumption was about the same. You can't really object NN software using NN optimised hardware. Stockfish, being classic branch-oriented, sequential software uses an x86 that is optimised for this kind of SW, too, so both have the hardware that is appropriate for their way of computing.
The only exception would be, of course, if one had a much more advanced production process. You wouldn't compare the energy consumption plus performance of, say, a chip produced with 130nm process to a chip produced with 22nm. But I don't think this is the case because Intel's production facilities are well-known for being state of the art.
Graham Banks wrote:The only thing I question is that AlphaZero had an opening book based on learning over millions of games, whereas it looks as though Stockfish may not have used an opening book at all.
Learning over millions of (self-)games is the only way for AlphaZero to raking up its ELO. You don't say about the billions of games Stockfish uses to improve itself on Fishtest.
Graham Banks wrote:The only thing I question is that AlphaZero had an opening book based on learning over millions of games, whereas it looks as though Stockfish may not have used an opening book at all.
Learning over millions of (self-)games is the only way for AlphaZero to raking up its ELO. You don't say about the billions of games Stockfish uses to improve itself on Fishtest.
This was over the course of many years, and besides, SF tunes with a (very crappy) book, so it does not tune for the very first moves, while tuning a lot of irrelevant stuff for the rest.
SF has to get rid of that book, if it has to progress.
They think this is not their main problem currently, but it is.
stavros wrote:to make it more practical: i want to know the price of 64core st hardware and the price of alpha zero hardware pls anyone? to make a more fair comparison
+ 10, most probably larger than 20/1, just as in the case with hardware speed.
And it used significantly more memory.
And an opening book, which actually decided the whole match.
Lyudmil Tsvetkov wrote:Most of your statements make very much sense, except the first one. So why do they weigh boxers and weight-lifters before each and every competition?
But they did - the energy consumption was about the same. You can't really object NN software using NN optimised hardware. Stockfish, being classic branch-oriented, sequential software uses an x86 that is optimised for this kind of SW, too, so both have the hardware that is appropriate for their way of computing.
The only exception would be, of course, if one had a much more advanced production process. You wouldn't compare the energy consumption plus performance of, say, a chip produced with 130nm process to a chip produced with 22nm. But I don't think this is the case because Intel's production facilities are well-known for being state of the art.
By about the same you mean 5 times higher for Alpha?
As you know pretty well, TPUs are precisely designed to run at couple of times lower tension, but that only improves memory and computations.
Lyudmil Tsvetkov wrote:By about the same you mean 5 times higher for Alpha?
No, about equal means about equal. Both in the 300W ballpark or so. You are completely mistaken if you think that Google threw half a computing centre at the match. In the self-training, yes, that was a different story, but so was Stockfish's development (e.g. fishtest).
Ras wrote:You are completely mistaken if you think that Google threw half a computing centre at the match. In the self-training, yes, that was a different story, but so was Stockfish's development (e.g. fishtest).
In particular, they played 44 million games in 9 hours. Fishtest does that in about 3 weeks.