Rodolfo Leoni wrote: I think SF could have a better performance with better settings tough.
This is trivial to test, just play SF at 1min/move against SF at 1min/move with all the paraphernalia that it was missing, if the new one beats the old one by 30 elo, A0 would only have won by 70 elo intead of 100.
I think you're right. But I'm using my limited free time and old PC for correspondence chess. I could try something like that on next spring, IMO. It could be interesting.
Latest SF Dev + Cerebellum + Syzygy vs naked SF8, 1 min/move.
I think that is easily more than 100 Elo stronger than SF 8 by now, but maybe not when SF 8 has 64 cores at the hardware they were using. The Cerebellum and Syzygy are not really fair IMO, although a form of learning should be allowed because that is what they did too.
I used "easily" too easily, I just want to correct that. Stefan Pohl's data suggest that the distance under his conditions, with his hardware, SF 8 to asmBrainfish is actually very close to 100 points,
Thanks to Stefan for the .jpg, I hope he does not mind that I reproduce it here!
Debugging is twice as hard as writing the code in the first
place. Therefore, if you write the code as cleverly as possible, you
are, by definition, not smart enough to debug it.
-- Brian W. Kernighan
Ovyron wrote:And they claim that A0 gets stronger with more time, so while Stockfish's strength gets stuck and doesn't improve with more time, allowing weaker engines to reach it, A0 takes off, and just beats Stockfish badlier and badlier.
Except that Stockfish does improve with more time and does not allow weaker engines to reach it. Or is there any solid evidence to the contrary?
Note that I am not contesting that AlphaZero improves more than Stockfish with more time. At least the AlphaZero paper suggests that it does.
I read what Ovyron wrote as that it scales better, improves more than Stockfish with more time. That can only mean that it really searches less 'junk' than normal chessprograms, is more selective. But that would be something related to their version of Monte Carlo -somehow, but the details seem to be vague?-, which is not Monte Carlo, but I have not read much about. And an eval without holes which I accept is true. That MCTS is less susceptible to the holes left -if any- I accept from others here mentioning that, without really understanding.
(My explanation for AlphaZero profiting more from having more time is that it needs that time to correct for the inevitable tactical deficiencies of its NN-based evaluation.)
I think I would have great difficulty accepting this. From what I have seen its tactics is formidable and much beyond Stockfish, this may again be only in places where we have holes and 'Aleph-Null' plays like Morphy as Ed put it, but it certainly is true in the games that "we" lost. That Aleph-Null does worse in Blitz is I think irrelevant, there could be 100's of reasons for that..
Last edited by Eelco de Groot on Wed Dec 27, 2017 1:57 am, edited 2 times in total.
Debugging is twice as hard as writing the code in the first
place. Therefore, if you write the code as cleverly as possible, you
are, by definition, not smart enough to debug it.
-- Brian W. Kernighan
syzygy wrote:......................................................
Note that I am not contesting that AlphaZero improves more than Stockfish with more time. At least the AlphaZero paper suggests that it does. (My explanation for AlphaZero profiting more from having more time is that it needs that time to correct for the inevitable tactical deficiencies of its NN-based evaluation.)
If you're right, we could then presume AlphaZero would heavily lose with blitz TCs.
Just take a look at the paper. Figure 2.
And there's another doubt: what if AlphaZero played without that 4 hours self-training session?
We'll never know, IMO...
Since AlphaZero was initialised with random weights, it played even worse than I before it was trained. Sorry, but did you not read the paper at all? Or am I misunderstanding your doubt?
Btw, the full training session was 9 hours. The 9-hour version played the 100-game match at 1 move/min. The 4-hour point was where AlphaZero surpassed SF in strength at 1 move/sec (see Figure 1 and the various explanations in the paper).
Rodolfo Leoni wrote: I think SF could have a better performance with better settings tough.
This is trivial to test, just play SF at 1min/move against SF at 1min/move with all the paraphernalia that it was missing, if the new one beats the old one by 30 elo, A0 would only have won by 70 elo intead of 100.
I think you're right. But I'm using my limited free time and old PC for correspondence chess. I could try something like that on next spring, IMO. It could be interesting.
Latest SF Dev + Cerebellum + Syzygy vs naked SF8, 1 min/move.
I think that is easily more than 100 Elo stronger than SF 8 by now, but maybe not when SF 8 has 64 cores at the hardware they were using. The Cerebellum and Syzygy are not really fair IMO, although a form of learning should be allowed because that is what they did too.
I used "easily" too easily, I just want to correct that. Stefan Pohl's data suggest that the distance under his conditions, with his hardware, SF 8 to asmBrainfish is actually very close to 100 points,
Thanks to Stefan for the .jpg, I hope he does not mind that I reproduce it here!
Very interesting. We could conclude asmBrainfish would have been, at least, competitive even without any learning. BTW, is asmBrainfish the same as asmFish?
Rodolfo Leoni wrote: I think SF could have a better performance with better settings tough.
This is trivial to test, just play SF at 1min/move against SF at 1min/move with all the paraphernalia that it was missing, if the new one beats the old one by 30 elo, A0 would only have won by 70 elo intead of 100.
I think you're right. But I'm using my limited free time and old PC for correspondence chess. I could try something like that on next spring, IMO. It could be interesting.
Latest SF Dev + Cerebellum + Syzygy vs naked SF8, 1 min/move.
I think that is easily more than 100 Elo stronger than SF 8 by now, but maybe not when SF 8 has 64 cores at the hardware they were using. The Cerebellum and Syzygy are not really fair IMO, although a form of learning should be allowed because that is what they did too.
I used "easily" too easily, I just want to correct that. Stefan Pohl's data suggest that the distance under his conditions, with his hardware, SF 8 to asmBrainfish is actually very close to 100 points,
Thanks to Stefan for the .jpg, I hope he does not mind that I reproduce it here!
Very interesting. We could conclude asmBrainfish would have been, at least, competitive even without any learning. BTW, is asmBrainfish the same as asmFish?
Rodolfo Leoni wrote: I think SF could have a better performance with better settings tough.
This is trivial to test, just play SF at 1min/move against SF at 1min/move with all the paraphernalia that it was missing, if the new one beats the old one by 30 elo, A0 would only have won by 70 elo intead of 100.
I think you're right. But I'm using my limited free time and old PC for correspondence chess. I could try something like that on next spring, IMO. It could be interesting.
Latest SF Dev + Cerebellum + Syzygy vs naked SF8, 1 min/move.
I think that is easily more than 100 Elo stronger than SF 8 by now, but maybe not when SF 8 has 64 cores at the hardware they were using. The Cerebellum and Syzygy are not really fair IMO, although a form of learning should be allowed because that is what they did too.
I used "easily" too easily, I just want to correct that. Stefan Pohl's data suggest that the distance under his conditions, with his hardware, SF 8 to asmBrainfish is actually very close to 100 points,
Thanks to Stefan for the .jpg, I hope he does not mind that I reproduce it here!
Very interesting. We could conclude asmBrainfish would have been, at least, competitive even without any learning. BTW, is asmBrainfish the same as asmFish?
Hi Rodolfo
Yes, normal asmFish with Cerebellum as a (read-only, you need their software, it's private, meant for sale at some time in the future, to modify any moves) .bin book, see Stefan Pohl's explanation
Debugging is twice as hard as writing the code in the first
place. Therefore, if you write the code as cleverly as possible, you
are, by definition, not smart enough to debug it.
-- Brian W. Kernighan
Eelco de Groot wrote:Yes, normal asmFish with Cerebellum as a (read-only, you need their software, it's private, meant for sale at some time in the future, to modify any moves) .bin book, see Stefan Pohl's explanation
Which means Aleph-Null with Cerebellum will be 200 elo stronger than Stockfish
People get so involved in discussions about ELO, TBs and TimeControls and other technicalities, they forget the basics of chess.
If you just go through the Games as a chess player, its completely obvious that Stockfish was completely out-classed and simply didn't have a clue about what was going on.
It is obvious to any average chess player that AlphaZero was playing at a completely different level and was seeing the Chessboard as a whole in a completely different way than Stockfish.
You can grumble about the version of SF used, TC used and anything else, but in my humble opinion, AlphaZero was so completely superior to Stockfish, that it wouldn't have made any significant difference to the outcome.
Its the Dawn of a New Age in computer chess, whether the nay sayers agree or not.
Of course, unfortunately since DeepMind isn't showing much interest in chess, the voices of these negative people will only grow stronger.
Soon they will claim that there was nothing like AlphaZero and it was all a big hoax.
But the people who really understand what Chess is about, have seen what actually happened on the ChessBoard and know the Truth.
Rodolfo Leoni wrote: I think SF could have a better performance with better settings tough.
This is trivial to test, just play SF at 1min/move against SF at 1min/move with all the paraphernalia that it was missing, if the new one beats the old one by 30 elo, A0 would only have won by 70 elo intead of 100.
I think you're right. But I'm using my limited free time and old PC for correspondence chess. I could try something like that on next spring, IMO. It could be interesting.
Latest SF Dev + Cerebellum + Syzygy vs naked SF8, 1 min/move.
I think that is easily more than 100 Elo stronger than SF 8 by now, but maybe not when SF 8 has 64 cores at the hardware they were using. The Cerebellum and Syzygy are not really fair IMO, although a form of learning should be allowed because that is what they did too.
I'm not a mathematician you'd have to ask Tord and Joona but I prefer to think of Alpha Zero as Aleph-Null but that is just the German form of the mathematical term. Alpha Zero is more natural to English speakers. A form of infinity like ∞, where Google gets its name from. Aleph is used in the mathematical symbol, I suppose because Georg Cantor used that? Not sure.
Great find! Aleph-null would be a wonderful logo for A0. Thanks Eelco!
Ideas=science. Simplification=engineering.
Without ideas there is nothing to simplify.
Rodolfo Leoni wrote: I think SF could have a better performance with better settings tough.
This is trivial to test, just play SF at 1min/move against SF at 1min/move with all the paraphernalia that it was missing, if the new one beats the old one by 30 elo, A0 would only have won by 70 elo intead of 100.
I think you're right. But I'm using my limited free time and old PC for correspondence chess. I could try something like that on next spring, IMO. It could be interesting.
Latest SF Dev + Cerebellum + Syzygy vs naked SF8, 1 min/move.
I think that is easily more than 100 Elo stronger than SF 8 by now, but maybe not when SF 8 has 64 cores at the hardware they were using. The Cerebellum and Syzygy are not really fair IMO, although a form of learning should be allowed because that is what they did too.
I'm not a mathematician you'd have to ask Tord and Joona but I prefer to think of Alpha Zero as Aleph-Null but that is just the German form of the mathematical term. Alpha Zero is more natural to English speakers. A form of infinity like ∞, where Google gets its name from. Aleph is used in the mathematical symbol, I suppose because Georg Cantor used that? Not sure.
Great find! Aleph-null would be a wonderful logo for A0. Thanks Eelco!
I was thinking more like Alpha FortyFourMillion and Stockfish Zero, but that is just me.
If you are on a sidewalk and the covid goes beep beep
Just step aside or you might have a bit of heat
Covid covid runs through the town all day
Can the people ever change their ways
Sherwin the covid's after you
Sherwin if it catches you you're through