Ozymandias wrote:I didn't think they could output so many games per second, people argued that the HW difference between the PC used by SF and the machine in which the NN worked, wasn't that high. Unless they replaced the machine for the match, it clearly was they case.
They used lots of HW for training.
They used a big PC with 4 TPU expansion cards (each using just 28-40 Watt) for playing. SF likely played on the same big PC but obviously did not use the TPUs.
It's all documented quite well.
(Note that SF also uses lots of HW for tuning *and* a lot of human brains. The AlphaZero approach seems to be far more suitable for massive parallelisation.)
I haven't even read many of the other threads about A0, much less the paper. It's easier to ask any given question and trust the answer from select people, who have actually done that. Thnx.
Ozymandias wrote:I didn't think they could output so many games per second, people argued that the HW difference between the PC used by SF and the machine in which the NN worked, wasn't that high. Unless they replaced the machine for the match, it clearly was they case.
They used lots of HW for training.
They used a big PC with 4 TPU expansion cards (each using just 28-40 Watt) for playing. SF likely played on the same big PC but obviously did not use the TPUs.
It's all documented quite well.
(Note that SF also uses lots of HW for tuning *and* a lot of human brains. The AlphaZero approach seems to be far more suitable for massive parallelisation.)
I haven't even read many of the other threads about A0, much less the paper. It's easier to ask any given question and trust the answer from select people, who have actually done that. Thnx.
Rebel wrote:I don't care that SF lost, it's totally irrelevant in the light of the huge claim by the Deepmind company, the alleged 4 hours self-play, quoting the document again: without any additional domain knowledge except the rules of the game.
Have you already let it sink in what is stated here?
No mobility, no king safety, no passed pawn evaluation, no castling knowledge, not even piece values?
How would that first self-play game look like? Something 1.a3 a6 2.a4 a5 3. b3 b6 etc and how would that lead to anything for the second self-pay game?
And so I voted for option 3.
Hello Ed,
I must mention something else that bothers me about the AlphaZero - Stockfish match. Why haven't the AlphaZero team released the ENTIRE 100 games in pgn format (or did I miss them releasing it somewhere?) so we can scrutinize this AlphaZero further?
I suspect they just cherry picked the 10 best games that their program played without showing the ones where it might have made some buffoonish moves.
Why do you believe that a chess program or machines that is capable to beat stockfish in that specific way, is weak?or cheating ?
Didn’t you see what was going on ?
Stockfish (in the same way as Komodo and Houdini) rely on a very clever search,
Very clever and coming very deep. With lots of NPS.
With this search and search depth they were able to beat humans.
But search is only one method to gain chess strength,
There are other methods to build chess strength.
While Komodo and Houdini and stockfish are all very similar.
Alpha zero is NOT a beast like them. It’s a different idea.
A different idea to create strength.
If you do not build such a deep tree, and do. It compute 70.000.000 NPS, you need to find different ways to create chess strength,
And of course it is possible to do so, because humans do also not compute 70.000.000 NPS when they play very strong chess,
What seems like a fairy tale today may be reality tomorrow.
Here we have a fairy tale of the day after tomorrow....
mclane wrote:While Komodo and Houdini and stockfish are all very similar.
Alpha zero is NOT a beast like them. It’s a different idea.
A different idea to create strength.
If you do not build such a deep tree, and do. It compute 70.000.000 NPS, you need to find different ways to create chess strength,
You really have no clue about A0, do you?
You didn't read any paper, have no clue what MCTS is, you are exactly like a caveman that have seen thunderstruck for the first time and believing to have witnessed a divine act...
I am not a caveman. I wasn’t even in a cave.
I have done computerchess for 38 years.
In these 38 years (BTW. How old are you ?)
I have seen a lot of computerchess games of very many dedicated chess computers and chess software.
I watched how the engines played against each other, spoke with the chess programmers about their engines, visited computer chess events all over the world, had a computerchess department of my own, wrote in computerchess Magazins and tried to work with many computerchess people all over the world.
Btw there were also MCTS chess programs within these 38 years. I remember treebeard by Jeff rollason from AI factory.
As i said to you before, we build this forum to discuss computerchess.
We did not build it to allow trolls to insult people. It was especially
This reason to have CCC Beeing moderated to handle the abuse of this computerchess scene by trolls who like to insult instead of discuss a certain topic.
What seems like a fairy tale today may be reality tomorrow.
Here we have a fairy tale of the day after tomorrow....
Rebel wrote:I don't care that SF lost, it's totally irrelevant in the light of the huge claim by the Deepmind company, the alleged 4 hours self-play, quoting the document again: without any additional domain knowledge except the rules of the game.
Have you already let it sink in what is stated here?
No mobility, no king safety, no passed pawn evaluation, no castling knowledge, not even piece values?
How would that first self-play game look like? Something 1.a3 a6 2.a4 a5 3. b3 b6 etc and how would that lead to anything for the second self-pay game?
And so I voted for option 3.
I've always thought NNs had a potential for playing chess, especially in generalizing positional factors. I even had a plan to test the theory by teaching an NN the static evaluation from a free engine, then replacing the static eval of the free engine with the NN and see if it could play as well as the algo-version. I never got around to that project.
I think this much larger project with a more vast and comprehensive NN, with a lot more NN experience behind the design is a very plausible one. I'm very interested that someone has figured out how to effectively use NNs in chess.
Rebel wrote:I don't care that SF lost, it's totally irrelevant in the light of the huge claim by the Deepmind company, the alleged 4 hours self-play, quoting the document again: without any additional domain knowledge except the rules of the game.
Have you already let it sink in what is stated here?
No mobility, no king safety, no passed pawn evaluation, no castling knowledge, not even piece values?
How would that first self-play game look like? Something 1.a3 a6 2.a4 a5 3. b3 b6 etc and how would that lead to anything for the second self-pay game?
And so I voted for option 3.
Ok, I repeat it again, because you didn't seem to have read it.
4 hours training time is another PR BS from Google.
4h on 5000TPUs where each TPU is equivalent to roughly 2 new GV100 or 10 1080Ti which is currently the top of the range graphics card normal individuals can afford. So those 4h of training time is like over 30 years of training on 1080Ti. I really wonder why they only used 5000 TPUs, they could have used 50000 instead and claim 24minutes of training time only. Or Google actually only owns 5000TPUs in total and they used all available TPU resources of the 500 billion dollar company for training a chess engine?
My guess is they actually didn't use 5000TPUs but much less and used much more time, but just extrapolated training time to what would be in case they used all of their existing TPUs just to make an impressive statement, and they obviously succeeded judging by the wrong impression you got from it.
It's not about that, it's about the self-play from scratch:
No mobility, no king safety, no passed pawn evaluation, no castling knowledge, not even piece values?
How would that first self-play game look like? Something 1.a3 a6 2.a4 a5 3. b3 b6 etc and how would that lead to anything for the second self-pay game?
Not even a million years could do the job.
neural networks are like your brain. No one had to tell you piece square table values to play better chess. they learn from experience, finding and exploiting patterns in the data.
And why do polls on facts? Whether we believe or do not believe, a fact is a fact. If you feel google made it up, then ask for proof. Taking opinion polls on facts among people without the information to verify them or not is not getting us closer to truth. Maybe a poll about was the match fair?
Rebel wrote:I don't care that SF lost, it's totally irrelevant in the light of the huge claim by the Deepmind company, the alleged 4 hours self-play, quoting the document again: without any additional domain knowledge except the rules of the game.
Have you already let it sink in what is stated here?
No mobility, no king safety, no passed pawn evaluation, no castling knowledge, not even piece values?
How would that first self-play game look like? Something 1.a3 a6 2.a4 a5 3. b3 b6 etc and how would that lead to anything for the second self-pay game?
And so I voted for option 3.
I've always thought NNs had a potential for playing chess, especially in generalizing positional factors. I even had a plan to test the theory by teaching an NN the static evaluation from a free engine, then replacing the static eval of the free engine with the NN and see if it could play as well as the algo-version. I never got around to that project.
I think this much larger project with a more vast and comprehensive NN, with a lot more NN experience behind the design is a very plausible one. I'm very interested that someone has figured out how to effectively use NNs in chess.
I agree one can do a lot of nice things with NN's in various areas. As for chess Giraffe has proven it.
