Nay Lin Tun wrote:It is simple, and without doubt, machine can learn better than human. First step, machines can calculate better than human ( people already accept it). This is the next step of technological revolution ( machines can learn better than human).It would be a bitter experience for many programmers who spent significant years of their life to teach a machine(program a machine) rather than let the machine learn by his own.
That's EXACTLY what's sticking in their craw.
Of course, Google not giving more information gives them an opportunity to downplay the achievement of Alpha-Zero.
Nay Lin Tun wrote:It is simple, and without doubt, machine can learn better than human. First step, machines can calculate better than human ( people already accept it). This is the next step of technological revolution ( machines can learn better than human).It would be a bitter experience for many programmers who spent signifiant years of their life to teach a machine(program a machine) rather than let the machine learn by his own.
machine can learn only if humans teach it to learn.
The interesting question is how to learn and if humans can use the same methods of alpha zero to learn and improve in chess without reading chess books or looking at games.
Of course machines are faster so it is obvious that machines will do better but can humans learn not to get to the level of alphago but to get to a level when they beat big majority of humans by using Alpha Zero methods?
CheckersGuy wrote:
I still think that calling the weights of the NN an "opening boook" is kinda misleading. It is not an opening book as you would think of one
Why, when NN evaluation function is exactly what an opening book gives you - for a given position a probability of selecting each move and wining probability of that position? And the closer position to the root, the better tuned weights are for it.
Aren't SF's eval weights essentially the same, tuned by playing hundreds of thousands of games? (Remember how engines used to be bad in openings when playing without a book? That was the pre-tuning era...) Eval weights are inseparable from the game-playing code, just as NN weights are the game-playing code. NN plays every part od the game - opening, middlegame, endgame - using the same wiring, just as SF plays the entire game by evaluating the board using essentially the same code. On the contrary, rather: it's SF that "cheats" by having explicitly hardcoded knowledge of certain endings, for example.
Opening books and EGTBs bypass the evaluation by providing an instant lookup on a fixed number of precalculated positions. Not the case with NN - it can evaluate all positions, and since one cannot say it's ever "out of the book", one cannot says its "in the book" either.
Branko Radovanovic wrote:Aren't SF's eval weights essentially the same, tuned by playing hundreds of thousands of games? (Remember how engines used to be bad in openings when playing without a book? That was the pre-tuning era...) Eval weights are inseparable from the game-playing code, just as NN weights are the game-playing code. NN plays every part od the game - opening, middlegame, endgame - using the same wiring, just as SF plays the entire game by evaluating the board using essentially the same code. On the contrary, rather: it's SF that "cheats" by having explicitly hardcoded knowledge of certain endings, for example.
Opening books and EGTBs bypass the evaluation by providing an instant lookup on a fixed number of precalculated positions. Not the case with NN - it can evaluate all positions, and since one cannot say it's ever "out of the book", one cannot says its "in the book" either.
Well SF eval "weights" are few kB in size max (PSTs taking most of it and even they are precomputed) while A0 NN weights are few hundred MBs if not even above 1GB is size.
Cerebellum, entirely self-generated book (so in principle not much different from the way NN weights in A0 are obtained) takes around 100MB and it contains best move in positions up to incredibly high depths (so very similar to general purpose evaluation of A0). Cerebellum is nothing but a self-learned eval file for SF.
Sorry, but I see no fair argument why it is ok for A0 to use its self-learned weights while it is not ok for SF to use Cerebellum, just because A0 uses it in a smarter way.
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
Price would, of course, be a good criterion for comparison, but from an engineering standpoint it would be interesting to compare power consumption of these two systems. If Google's TPU system does not draw (much) more power than a 64-core PC, then it might be argued that it is not (much) more powerful, both in a literal and in a metaphorical sense. Does anyone know the figures?
For those who believe the matchup was not fair on the account of (supposedly or not) much stronger hardware used by AlphaZero, consider the following question: how many cores on a conventional PC running Stockfish 8 would it take to outplay a 64-core PC also running Stockfish 8 by as much as 100 Elo? 256 would certainly not suffice, and my guess is that even 512 would not be enough. It's not just that AlphaZero is stronger - it's by what margin, and I'm not sure this could be explained simply by having more powerful hardware.
To evaluate performance in chess, we used
Stockfish
version 8 (official Linux release) as a
baseline program, using 64 CPU threads and a hash size of 1GB"
so 64 threads! =32 cores! is ppl so dizzy from that hype to read that correctly?
and why 1 gb hash only ?
Ras wrote:The match was fair, and AlphaGo did NOT have massive hardware advantage, and it does NOT make sense to ask for it to run on x86. Just like it would not make sense to ask for Stockfish to run on TPUs.
Oh, the TPUs are more expensive? Of course they are because they are a brand new development, but the whole point of this match, from Google's point of view, was to prove that this hardware is suited to neural network applications. It was not specialised hardware like with Hydra or Deep Blue. It was general purpose NN hardware. The same hardware also plays Go and Shogi, something that Hydra or Deep Blue could not do - and neither could a Stockfish version boosted by FPGAs.
Oh, the hardware isn't for sale? Wrong point. Google doesn't sell them now because they can use their current monopoly better with cloud offers. But that creates a demand, and since Google has proven how capable these things are, others will catch up.
Just like others caught up when Intel started to put the 4004 into production, the first microprocessor actually available on the market. In fact, I think that Intel should finally wake up here after their failed adventures from Itanic to challenging ARM. Or maybe even AMD since they already have GPU competence. Or Nvidia.
Second, these TPUs with neural networks scale better than traditional cores which are subject to the law of diminishing returns, which is why Johnny with 2000 cores is still weaker than Stockfish. That's why these TPUs are not a one-time artifact like Hydra or Deep Blue - they are here to stay and change computing.
The energy consumption was comparable, that's the central point here, because that's what will determine the cost of the whole approach in the medium run.
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?
Zenmastur wrote:It's been a long time since I have seen such BS posted on this board. Did IQ's suddenly drop while I was away?
It's clear to me that few posts on this thread are objective. Many posts are filled with insults, misinformation, and in some cases, out right lies. What the hell is wrong with you people?
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.
+1 I believe you are 100% correct!
I still think that calling the weights of the NN an "opening boook" is kinda misleading. It is not an opening book as you would think of one
But it still won most of its 1.d4 games by fianchettoeing its kingside bishop, Bg2.
That is a huge boost in knowledge.
Many SF developers wondered once if kingside fianchetto is really a valid concept, or just a human peculiarity that works simply in human games.
Now you get the answer: Bg2 is winning games.
CheckersGuy wrote:
I still think that calling the weights of the NN an "opening boook" is kinda misleading. It is not an opening book as you would think of one
Why, when NN evaluation function is exactly what an opening book gives you - for a given position a probability of selecting each move and wining probability of that position? And the closer position to the root, the better tuned weights are for it.
Aren't SF's eval weights essentially the same, tuned by playing hundreds of thousands of games? (Remember how engines used to be bad in openings when playing without a book? That was the pre-tuning era...) Eval weights are inseparable from the game-playing code, just as NN weights are the game-playing code. NN plays every part od the game - opening, middlegame, endgame - using the same wiring, just as SF plays the entire game by evaluating the board using essentially the same code. On the contrary, rather: it's SF that "cheats" by having explicitly hardcoded knowledge of certain endings, for example.
Opening books and EGTBs bypass the evaluation by providing an instant lookup on a fixed number of precalculated positions. Not the case with NN - it can evaluate all positions, and since one cannot say it's ever "out of the book", one cannot says its "in the book" either.
SF start tuning their parameters at move 2 or even move 8, and with a horrible book at that, featuring completely random openings, and that is their main tuning deficiency.
You can not tune well opening terms with such a book.
This dies not show against wekaer engines like Komodo and Houdini, but there are stronger ones.
Would the conditions be right? A minimum of seriousness, please... 