Just read this paper this morning:
https://arxiv.org/pdf/1712.01815.pdf
Quote:
"Starting from random play, and given no domain knowledge except the game rules, AlphaZero achieved within 24 hours a superhuman level of play in the games of chess and shogi (Japanese chess) as well as Go, and convincingly defeated a world-champion program in each case."
In a 100-game match against Stockfish 8 AlphaZero won 64-36 without losing a single game. Some games are at the end of the paper.
https://arxiv.org/pdf/1712.01815.pdf
Quote:
"Starting from random play, and given no domain knowledge except the game rules, AlphaZero achieved within 24 hours a superhuman level of play in the games of chess and shogi (Japanese chess) as well as Go, and convincingly defeated a world-champion program in each case."
In a 100-game match against Stockfish 8 AlphaZero won 64-36 without losing a single game. Some games are at the end of the paper.
Game 10 is just amazing. AlphaZero plays the Polugajevsky gambit vs the Queen's Indian, sacrifices a piece and SF has no chance (I didn't incude all the moves, stopped where black is clearly lost):
[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "AlphaZero"]
[Black "Stockfish"]
[Result "1-0"]
[PlyCount "83"]
1. d4 Nf6 2. c4 e6 3. Nf3 b6 4. g3 Bb7 5. Bg2 Be7 6. O-O O-O 7. d5 exd5 8. Nh4
c6 9. cxd5 Nxd5 10. Nf5 Nc7 11. e4 d5 12. exd5 Nxd5 13. Nc3 Nxc3 14. Qg4 g6 15.
Nh6+ Kg7 16. bxc3 Bc8 17. Qf4 Qd6 18. Qa4 g5 19. Re1 Kxh6 20. h4 f6 21. Be3 Bf5
22. Rad1 Qa3 23. Qc4 b5 24. hxg5+ fxg5 25. Qh4+ Kg6 26. Qh1 Kg7 27. Be4 Bg6 28.
Bxg6 hxg6 29. Qh3 Bf6 30. Kg2 Qxa2 31. Rh1 Qg8 32. c4 Re8 33. Bd4 Bxd4 34. Rxd4
Rd8 35. Rxd8 Qxd8 36. Qe6 Nd7 37. Rd1 Nc5 38. Rxd8 Nxe6 39. Rxa8 Kf6 40. cxb5
cxb5 41. Kf3 Nd4+ 42. Ke4 *
[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "AlphaZero"]
[Black "Stockfish"]
[Result "1-0"]
[PlyCount "83"]
1. d4 Nf6 2. c4 e6 3. Nf3 b6 4. g3 Bb7 5. Bg2 Be7 6. O-O O-O 7. d5 exd5 8. Nh4
c6 9. cxd5 Nxd5 10. Nf5 Nc7 11. e4 d5 12. exd5 Nxd5 13. Nc3 Nxc3 14. Qg4 g6 15.
Nh6+ Kg7 16. bxc3 Bc8 17. Qf4 Qd6 18. Qa4 g5 19. Re1 Kxh6 20. h4 f6 21. Be3 Bf5
22. Rad1 Qa3 23. Qc4 b5 24. hxg5+ fxg5 25. Qh4+ Kg6 26. Qh1 Kg7 27. Be4 Bg6 28.
Bxg6 hxg6 29. Qh3 Bf6 30. Kg2 Qxa2 31. Rh1 Qg8 32. c4 Re8 33. Bd4 Bxd4 34. Rxd4
Rd8 35. Rxd8 Qxd8 36. Qe6 Nd7 37. Rd1 Nc5 38. Rxd8 Nxe6 39. Rxa8 Kf6 40. cxb5
cxb5 41. Kf3 Nd4+ 42. Ke4 *
Wow! Amazing!
At the end of the article, they point out that AlphaZero is processing 80K nodes per second, while Stockfish 8 is processing 70,000K nodes per second...
It really shows how underappreciated a good evaluation function is...
At the end of the article, they point out that AlphaZero is processing 80K nodes per second, while Stockfish 8 is processing 70,000K nodes per second...
It really shows how underappreciated a good evaluation function is...
will the rating lists test it now ?!
Probably not, since AlphaZero requires specialized hardware...
I don't think Alphazero was programmed with a better evaluation function in 24 hours. It likely had superior search heuristics.
It uses an evaluation guided search, so of course it has vastly superior search heuristics which allow it to be better while evaluating almost 1000 times fewer nodes per second...
All top engines employ evaluation guided search.
Yes. And not nearly as effectively as AlphaZero. In fact, the difference in efficiency as indicated in the paper, is at least three orders of magnitude...
Here is game 9. First AlphaZero puts its king on e3 at move 17. Then 30.Bxg6! and 32.f5! clinches the game.
[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "AlphaZero"]
[Black "Stockfish"]
[Result "1-0"]
[PlyCount "103"]
1. d4 e6 2. e4 d5 3. Nc3 Nf6 4. e5 Nfd7 5. f4 c5 6. Nf3 cxd4 7. Nb5 Bb4+ 8. Bd2
Bc5 9. b4 Be7 10. Nbxd4 Nc6 11. c3 a5 12. b5 Nxd4 13. cxd4 Nb6 14. a4 Nc4 15.
Bd3 Nxd2 16. Kxd2 Bd7 17. Ke3 b6 18. g4 h5 19. Qg1 hxg4 20. Qxg4 Bf8 21. h4 Qe7
22. Rhc1 g6 23. Rc2 Kd8 24. Rac1 Qe8 25. Rc7 Rc8 26. Rxc8+ Bxc8 27. Rc6 Bb7 28.
Rc2 Kd7 29. Ng5 Be7 30. Bxg6 Bxg5 31. Qxg5 fxg6 32. f5 Rg8 33. Qh6 Qf7 34. f6
Kd8 35. Kd2 Kd7 36. Rc1 Kd8 37. Qe3 Qf8 38. Qc3 Qb4 39. Qxb4 axb4 40. Rg1 b3
41. Kc3 Bc8 42. Kxb3 Bd7 43. Kb4 Be8 44. Ra1 Kc7 45. a5 Bd7 46. axb6+ Kxb6 47.
Ra6+ Kb7 48. Kc5 Rd8 49. Ra2 Rc8+ 50. Kd6 Be8 51. Ke7 g5 52. hxg5 1-0
[Event "?"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "AlphaZero"]
[Black "Stockfish"]
[Result "1-0"]
[PlyCount "103"]
1. d4 e6 2. e4 d5 3. Nc3 Nf6 4. e5 Nfd7 5. f4 c5 6. Nf3 cxd4 7. Nb5 Bb4+ 8. Bd2
Bc5 9. b4 Be7 10. Nbxd4 Nc6 11. c3 a5 12. b5 Nxd4 13. cxd4 Nb6 14. a4 Nc4 15.
Bd3 Nxd2 16. Kxd2 Bd7 17. Ke3 b6 18. g4 h5 19. Qg1 hxg4 20. Qxg4 Bf8 21. h4 Qe7
22. Rhc1 g6 23. Rc2 Kd8 24. Rac1 Qe8 25. Rc7 Rc8 26. Rxc8+ Bxc8 27. Rc6 Bb7 28.
Rc2 Kd7 29. Ng5 Be7 30. Bxg6 Bxg5 31. Qxg5 fxg6 32. f5 Rg8 33. Qh6 Qf7 34. f6
Kd8 35. Kd2 Kd7 36. Rc1 Kd8 37. Qe3 Qf8 38. Qc3 Qb4 39. Qxb4 axb4 40. Rg1 b3
41. Kc3 Bc8 42. Kxb3 Bd7 43. Kb4 Be8 44. Ra1 Kc7 45. a5 Bd7 46. axb6+ Kxb6 47.
Ra6+ Kb7 48. Kc5 Rd8 49. Ra2 Rc8+ 50. Kd6 Be8 51. Ke7 g5 52. hxg5 1-0
Post some games alphazero lost.
> In a 100-game match against Stockfish 8 AlphaZero won 64-36 without losing a single game. Some games are at the end of the paper.
>Here is game 9. First AlphaZero puts its king on e3 at move 17. Then 30.Bxg6! and 32.f5! clinches the game.
Those are quality games, especially this one, and it is playing more or less the moves I would have chosen in this position.
I will remain open to the value of AlphaZero and their approach until we get some more conclusive evidence.
In the meantime, I would like to share the following:
- 64/36 is not a colossal score against Stockfish 8(latest SF dev beats it 55/45), especially bearing in mind Alpha ran on a faster hardware, when we factor in the hardware advantage, it would beat current dev just slightly or maybe even not, who knows how big the hardware advantage is, and how Stockfish would perform with special algorithms for that hardware?
- this is not AI, some patterns have been introduced into the engine first, some evaluation parameters, which I suppose have not been the most basic ones, they have probably taken some advanced chess knowledge as base,
maybe even Stockfish code, as there are a couple of chess programmers in the team; why would they start with a weaker software? This is actually a large self-tuning exercise, with tremendous computer power, what is so new about it? Only the big hardware, probably.
- I guess this large tuning exercise will have some success, but will not discover the deeper aspects of the game, as it will need to formulate specific evaluation patterns for more complex positions, and it will be unable to do so; for example, it will not do well in closed positions
So I am not fully convinced this is some kind of a breakthrough in science/chess, though I remain open to new data and indications either way.
I would like to know more about the software/evaluation base Alpha started from.
I would also love to see latest SF and Alpha competing under fully equal conditions.
Supposedly they started from zero knowledge. At the start that means random neural network that is the dumbest possible one.
> why would they start with a weaker software?
That's research, without any idea to get the first place at TCEC. Why wouldn't they test if it's possible?
We won't get anything runnable anyway. :D :(
You either didn't read their paper, or you didn't understand it at any level...
From their paper:
"Domain Knowledge
1. The input features describing the position, and the output features describing the move,
are structured as a set of planes; i.e. the neural network architecture is matched to the
grid-structure of the board.
2. AlphaZero is provided with perfect knowledge of the game rules. These are used during
MCTS, to simulate the positions resulting from a sequence of moves, to determine game
termination, and to score any simulations that reach a terminal state.
3. Knowledge of the rules is also used to encode the input planes (i.e. castling, repetition,
no-progress) and output planes (how pieces move, promotions, and piece drops in shogi).
4. The typical number of legal moves is used to scale the exploration noise (see below).
5. Chess and shogi games exceeding a maximum number of steps (determined by typical
game length) were terminated and assigned a drawn outcome; Go games were terminated
and scored with Tromp-Taylor rules, similarly to previous work (29).
AlphaZero did not use any form of domain knowledge beyond the points listed above."
As far as running on faster hardware, neural networks are well suited to running on massively parallel processors (e.g. Google's TPU) while alpha-beta search is not. That isn't going to change, so an approach based on MCTS will always run on much faster hardware if we are talking about comparable hardware budgets.
I would also love to see latest SF and Alpha competing under fully equal conditions.
Face palm...
From their paper:
"Domain Knowledge
1. The input features describing the position, and the output features describing the move,
are structured as a set of planes; i.e. the neural network architecture is matched to the
grid-structure of the board.
2. AlphaZero is provided with perfect knowledge of the game rules. These are used during
MCTS, to simulate the positions resulting from a sequence of moves, to determine game
termination, and to score any simulations that reach a terminal state.
3. Knowledge of the rules is also used to encode the input planes (i.e. castling, repetition,
no-progress) and output planes (how pieces move, promotions, and piece drops in shogi).
4. The typical number of legal moves is used to scale the exploration noise (see below).
5. Chess and shogi games exceeding a maximum number of steps (determined by typical
game length) were terminated and assigned a drawn outcome; Go games were terminated
and scored with Tromp-Taylor rules, similarly to previous work (29).
AlphaZero did not use any form of domain knowledge beyond the points listed above."
As far as running on faster hardware, neural networks are well suited to running on massively parallel processors (e.g. Google's TPU) while alpha-beta search is not. That isn't going to change, so an approach based on MCTS will always run on much faster hardware if we are talking about comparable hardware budgets.
I would also love to see latest SF and Alpha competing under fully equal conditions.
Face palm...
Lyudmil's book is basically on polynomial scoring via additions of features each of which gets a score. He is about as deeply invested in the old paradigm as any current chess programmer, or even owner of lots of chess programs. Unfortunately or not, depending, the old work is headed for history's dustbin. Doesn't matter how "accurately" an individual feature gets scored any more, when the addition polynomial concept is busted against integrated evaluation.
The rule based scoring approach based on a small number of weighted features will probably still remain as a human method of training of course.
As an alternative, one could imagine having a document with a million positions in it, each with a NN generated score. A person learning to be a better player would look at each position and try to guess the NN generated score, and then look at the NN score to see how well they did. In this manner, a person could attempt to improve their pattern recognition skills for chess (by learning to better approximate the NN outputs) without any definitive rules...
As an alternative, one could imagine having a document with a million positions in it, each with a NN generated score. A person learning to be a better player would look at each position and try to guess the NN generated score, and then look at the NN score to see how well they did. In this manner, a person could attempt to improve their pattern recognition skills for chess (by learning to better approximate the NN outputs) without any definitive rules...
how it works in human groups, is that a few smarter ones, prepared to focus and equipped with the necessary neurology, do the pattern recognising and then explain them to the rest via media or whatever. it may be that there are new heuristics waiting to be found,
I'm sure new heuristics will be found. But as you imply, the smarter ones will try to train themselves to evaluate using pattern recognition rather than rule based evaluation...
You gotta love that fighting King 


How does it count nodes, though? 
Saw this article and posted it but saw that Jeroen had already. If confirmed, it's the biggest news I've seen for a while in computer chess and I wonder, could this tabula rasa method be applicable to many other fields?
Happy xmas!

Saw this article and posted it but saw that Jeroen had already. If confirmed, it's the biggest news I've seen for a while in computer chess and I wonder, could this tabula rasa method be applicable to many other fields?
Happy xmas!
Presumably by positions visited, just as in an alpha-beta search. The difference is that it is much better than the best alpha-beta searchers in shutting down bad lines, but still less efficient than good human players (but it is looking at 1000 times more positions per second).
with more training, there is every reason it may become better than humans at candidate move selection. Maybe NN chess that looks N positions total per move, where N is similar to humans is possible?
That's an interesting proposition. Of course the man-made NN has far, far fewer nodes, which might put it at a significant disadvantage, but it might make up for that by being more consistent, having better control over the network and firing mechanism, and being able to train for longer time periods...
no reason why it shouldn't be doable for games with humans, even very strong ones, but I suspect quick positional assessments with limited nodes N, would have problems with deeper tactics. When do GMs start getting into trouble and miss things? 15 ply comp searches? Perhaps a NN program could be allowed N=200 nodes say, and allocate them wide or deep as it sees fit.
Deeper tactics can be handled (inefficiently) by increasing the depth of the NN, rather than through a tree search. This might cause a drop in playing strength though, since seeing the deeper tactics might result in less effective evaluation in the much more common conditions where deep tactics don't exist...
200 nodes would be too few to even express the game state. I don't recall seeing how many nodes were used, but I'm guessing it was many orders of magnitude greater.
200 nodes would be too few to even express the game state. I don't recall seeing how many nodes were used, but I'm guessing it was many orders of magnitude greater.
How do you know it's less efficent than human players at selectivity? It played 99.9% strong moves. I have examined the games.
Selectivity is another name for how extreme the tree pruning can be. I suspect that if you looked at the search, you would see some real trash being considered for several plies...
You ought to tweak Vas and see if you can get him interested in implementing something related to this. It could be his ticket back to the top...
AlphaZero is being discussed everywhere, so I am sure the top programmers will pick up this idea!
This could really be a rebirth in computer chess! An entirely new approach, superior to the current approach, where the best methods of implementing many aspects of the algorithms are still up in the air...
Of course it also means we will all have to get new hardware!
Of course it also means we will all have to get new hardware!

Have a look at this game: All Rybka versions struggle to see the combination after g6??
How the search functions is just as relevant as the evaluation heuristics guiding it. Deep Fritz 10.1 searches more extensively, but less thorough.
(435) Deep Fritz 10 - Rybka 1.2n [D28]
DESKTOP-M7G326U, Blitz:5'+3" DESKTOP-M7G326U (1), 06.12.2017
[0.66;0.23]
Intel(R) Celeron(R) CPU N3050 @ 1.60GHzW=12.5 plies; 1 595kN/s; 50 089 TBAsB=13.3 plies; 193kN/s; 99 032 TBAs 1.d4 White last book move 1...d5 2.c4 c6 3.Nf3 Nf6 4.Nc3 a6 5.e3 e6 Black last book move 6.Bd3 0.66/13 11 dxc4 0.23/13 12 7.Bxc4 0.45/15 27 c5 0.19/13 12 8.0–0 0.36/14 28 Nc6 (b5) 0.20/12 30 9.Qe2 0.52/13 13 b5 0.17/14 17 10.Bb3 (Bd3) 0.44/14 16 10...c4 (cxd4) 0.09/12 8 11.Bc2 0.39/15 19 Nb4 (Bb7) 0.09/13 11 12.Bb1 0.50/15 21 Bb7 (Be7) 0.24/12 17 13.e4 0.50/14 18 Be7 (Nd3) 0.29/12 8 14.Rd1 (a3) 0.59/14 14 14...Qa5 (0–0) 0.27/11 9 15.Bg5 (a3) 1.10/13 16 15...Nd3 0.30/11 13 16.Bxd3 0.70/13 10 cxd3 0.27/12 2 17.Qxd3 0.76/15 13 b4 0.23/14 8 18.Bxf6 0.68/15 9 Bxf6 0.19/15 14 19.Ne2 (e5) 0.68/14 14 19...0–0 0.16/13 14 20.a3 (a4) 0.71/14 6 20...Rac8 (Rfc8) 0.14/12 13 21.Nf4 0.77/13 8 Rfd8 0.14/12 12 22.axb4 (Rac1) 0.74/14 9 22...Qxb4 0.12/10 4 23.Nh5 (d5) 0.81/14 5 23...Be7 -0.03/10 2 24.Qe3 (Rdc1) 0.76/13 5 24...Rc2 0.03/11 11 25.Ne5 (Qf4) 0.69/13 11 25...g6 (Qb5) -0.09/11 6 [25...f6] 26.Qf4 (Nf4) 1.51/12 5 26...gxh5 -0.03/12 7 27.Ra3 (Ra4) 1.78/11 1 27...h4 0.94/13 11 28.Qg4+ (Rb3) 2.88/11 2 28...Kf8 0.39/14 6 29.Qh5 3.20/11 1 Kg7 0.58/13 10 30.Rb3 3.20/12 2 Qd6 (Qxd4) 1.07/13 4 31.Qg4+ 2.80/11 3 Kf8 1.13/14 4 32.Qf4 2.86/12 2 f6 1.14/14 6 33.Rxb7 2.95/13 3 Ke8 (Kg8) 1.17/14 5 34.Qg4 2.80/13 3 Kf8 1.20/13 2 35.Nf3 (Qf3) 3.36/12 2 35...Qc6 1.05/12 6 36.Rxe7 2.71/11 2 Kxe7 1.25/13 9 37.Qg7+ (e5) 2.71/11 3 37...Ke8 1.13/11 1 38.Qg8+ 2.43/11 2 Ke7 1.13/12 1 39.Qxh7+ 2.27/12 2 Ke8 1.16/13 5 40.Qg8+ (Qh8+) 2.27/12 3 40...Ke7 1.16/14 3 41.Qg7+ 2.46/13 2 Ke8 1.17/14 4 42.Qxf6 2.46/13 6 Qxe4 1.21/14 5 43.b4 (Ne5) 2.46/12 3 43...h3 0.60/12 4 44.Re1 (Qh8+) 2.64/13 4 44...Re2 1.26/13 3 45.Rc1 2.64/13 3 Rc2 1.31/14 7 46.Rxc2 2.70/13 2 Qxc2 1.34/12 1 47.Qxe6+ 2.90/13 5 Kf8 1.25/5 0 48.gxh3 3.00/13 5 Qb1+ 1.21/13 5 49.Kg2 2.97/13 11 Qxb4 1.22/11 1 50.Qf6+ 2.83/11 2 Ke8 1.17/12 3 51.Qxa6 (Ne5) 2.83/11 2 51...Ke7 (Rd6) 1.01/11 9 52.Qe2+ (h4) 3.14/11 6 52...Kd6 0.95/12 6 53.Qe5+ 3.14/11 9 Kc6 (Kd7) 0.97/12 2 54.h4 (Qe6+) 3.18/12 9 54...Rg8+ (Qb3) 1.05/12 10 55.Ng5 3.59/12 4 Rd8 (Qb3) 0.95/12 3 56.Ne6 (Qe4+) 3.60/10 2 56...Rg8+ 0.81/11 1 57.Kf3 3.32/12 3 Qc3+ 0.84/13 6 58.Kf4 3.51/12 6 Qd2+ (Qc1+) 0.84/12 6 59.Qe3 3.70/3 5 Qxe3+ (Qa2) 1.16/13 8 60.fxe3 3.59/13 0 Rh8 1.14/17 3 61.Ng5 (Kg4) 3.78/14 2 61...Rxh4+ (Kd5) 0.58/14 1 62.Kg3 3.64/16 2 Rh6 0.75/18 4 63.e4 3.66/16 3 Kb5 0.83/16 4 64.Nf3 3.66/16 3 Kc4 0.87/17 3 65.d5 3.65/16 2 Rg6+ (Kd3) 0.98/18 6 66.Kf4 4.36/14 2 Rf6+ 0.98/16 3 67.Kg4 4.36/15 1 Kc5 (Rh6) 0.98/16 5 68.h4 4.48/14 0 Rh6 (Rg6+) 1.45/19 7 69.Kg5 4.48/14 1 Rh7 (Rh8) 1.41/18 4 70.Kg6 (h5) 4.43/15 1 70...Rh8 1.16/20 2 71.h5 4.44/16 3 Rg8+ 1.40/21 5 72.Kf5 4.22/16 1 Rh8 (Rf8+) 1.45/17 4 73.Kg4 4.47/15 2 Rh7 1.57/16 2 74.Kg5 4.47/15 1 Re7 1.94/19 12 75.Nd2 4.45/14 1 Rg7+ 2.13/19 5 76.Kf6 4.45/15 1 Rg3 (Rh7) 3.44/19 9 77.h6 (e5) 5.00/15 2 77...Rh3 1.28/19 3 78.Kg5 3.91/16 5 Rg3+ 1.78/20 3 79.Kh4 3.63/18 6 Rg6 1.93/21 4 80.Nb3+ (Kh5) 4.65/15 3 80...Kc4 2.93/17 7 81.Kh5 5.18/14 1 Rb6 3.62/17 3 82.Na5+ 6.05/14 2 Kc5 (Kd3) 4.29/18 2 83.h7 (Nc6) 6.25/14 1 83...Rb1 7.13/18 6 84.Nb7+ 5.97/15 2 Kb6 7.13/15 4 85.Nd6 (Nd8) 6.20/15 1 85...Rh1+ 7.13/18 2 86.Kg6 5.89/17 2 Rg1+ 7.92/19 5 87.Kf6 6.73/16 2 Rh1 7.92/19 3 88.Nf7 7.50/14 2 Rxh7 7.92/18 1 89.d6 7.96/12 1 Kc6 7.92/17 2 90.Ke7 7.68/12 1 Rh4 7.92/13 4 91.Ne5+ 9.09/12 1 Kc5 7.92/16 4 92.d7 9.27/11 0 Rh8 (Kb4) 11.68/13 4 93.Nf7 9.45/10 0 Rh2 (Ra8) 12.68/15 4 94.d8Q 9.66/10 0 Rh1 (Rg2) #21/15 2 95.Qd2 (Qd6+) 13.42/10 2 95...Kc4 #22/11 0 96.Qe2+ 14.08/9 1 Kb4 (Kc5) #21/10 0 97.e5 (Qb2+) 20.40/10 1 97...Ka5 #19/11 0 98.Qd2+ 22.51/9 1 Ka4 #19/10 0 99.Qd4+ 24.52/8 1 Ka3 #18/9 0 100.Qa7+ 21.80/9 1 Kb4 #17/8 0 101.Qb7+ 24.09/9 1 Kc5 #16/7 0 102.Qxh1 23.81/9 3 Kd4 #14/5 0 103.Qe4+ (Qf3) #11/7 0 103...Kc3 (Kxe4) #14/4 0 104.Nd6 (Qd3+) #8/6 0 104...Kd2 #13/4 0 105.Qf3 (Qe3+) #5/7 0 105...Kc1 #4/3 0 106.Qd3 #4/5 0 Kb2 #3/3 0 107.Nc4+ #3/5 0 Ka2 (Kc1) #2/3 0 108.Qc2+ (Qc3) #2/5 0 108...Ka1 #1/3 0 109.Qb2# #1/5 0 1–0
How the search functions is just as relevant as the evaluation heuristics guiding it. Deep Fritz 10.1 searches more extensively, but less thorough.
(435) Deep Fritz 10 - Rybka 1.2n [D28]
DESKTOP-M7G326U, Blitz:5'+3" DESKTOP-M7G326U (1), 06.12.2017
[0.66;0.23]
Intel(R) Celeron(R) CPU N3050 @ 1.60GHzW=12.5 plies; 1 595kN/s; 50 089 TBAsB=13.3 plies; 193kN/s; 99 032 TBAs 1.d4 White last book move 1...d5 2.c4 c6 3.Nf3 Nf6 4.Nc3 a6 5.e3 e6 Black last book move 6.Bd3 0.66/13 11 dxc4 0.23/13 12 7.Bxc4 0.45/15 27 c5 0.19/13 12 8.0–0 0.36/14 28 Nc6 (b5) 0.20/12 30 9.Qe2 0.52/13 13 b5 0.17/14 17 10.Bb3 (Bd3) 0.44/14 16 10...c4 (cxd4) 0.09/12 8 11.Bc2 0.39/15 19 Nb4 (Bb7) 0.09/13 11 12.Bb1 0.50/15 21 Bb7 (Be7) 0.24/12 17 13.e4 0.50/14 18 Be7 (Nd3) 0.29/12 8 14.Rd1 (a3) 0.59/14 14 14...Qa5 (0–0) 0.27/11 9 15.Bg5 (a3) 1.10/13 16 15...Nd3 0.30/11 13 16.Bxd3 0.70/13 10 cxd3 0.27/12 2 17.Qxd3 0.76/15 13 b4 0.23/14 8 18.Bxf6 0.68/15 9 Bxf6 0.19/15 14 19.Ne2 (e5) 0.68/14 14 19...0–0 0.16/13 14 20.a3 (a4) 0.71/14 6 20...Rac8 (Rfc8) 0.14/12 13 21.Nf4 0.77/13 8 Rfd8 0.14/12 12 22.axb4 (Rac1) 0.74/14 9 22...Qxb4 0.12/10 4 23.Nh5 (d5) 0.81/14 5 23...Be7 -0.03/10 2 24.Qe3 (Rdc1) 0.76/13 5 24...Rc2 0.03/11 11 25.Ne5 (Qf4) 0.69/13 11 25...g6 (Qb5) -0.09/11 6 [25...f6] 26.Qf4 (Nf4) 1.51/12 5 26...gxh5 -0.03/12 7 27.Ra3 (Ra4) 1.78/11 1 27...h4 0.94/13 11 28.Qg4+ (Rb3) 2.88/11 2 28...Kf8 0.39/14 6 29.Qh5 3.20/11 1 Kg7 0.58/13 10 30.Rb3 3.20/12 2 Qd6 (Qxd4) 1.07/13 4 31.Qg4+ 2.80/11 3 Kf8 1.13/14 4 32.Qf4 2.86/12 2 f6 1.14/14 6 33.Rxb7 2.95/13 3 Ke8 (Kg8) 1.17/14 5 34.Qg4 2.80/13 3 Kf8 1.20/13 2 35.Nf3 (Qf3) 3.36/12 2 35...Qc6 1.05/12 6 36.Rxe7 2.71/11 2 Kxe7 1.25/13 9 37.Qg7+ (e5) 2.71/11 3 37...Ke8 1.13/11 1 38.Qg8+ 2.43/11 2 Ke7 1.13/12 1 39.Qxh7+ 2.27/12 2 Ke8 1.16/13 5 40.Qg8+ (Qh8+) 2.27/12 3 40...Ke7 1.16/14 3 41.Qg7+ 2.46/13 2 Ke8 1.17/14 4 42.Qxf6 2.46/13 6 Qxe4 1.21/14 5 43.b4 (Ne5) 2.46/12 3 43...h3 0.60/12 4 44.Re1 (Qh8+) 2.64/13 4 44...Re2 1.26/13 3 45.Rc1 2.64/13 3 Rc2 1.31/14 7 46.Rxc2 2.70/13 2 Qxc2 1.34/12 1 47.Qxe6+ 2.90/13 5 Kf8 1.25/5 0 48.gxh3 3.00/13 5 Qb1+ 1.21/13 5 49.Kg2 2.97/13 11 Qxb4 1.22/11 1 50.Qf6+ 2.83/11 2 Ke8 1.17/12 3 51.Qxa6 (Ne5) 2.83/11 2 51...Ke7 (Rd6) 1.01/11 9 52.Qe2+ (h4) 3.14/11 6 52...Kd6 0.95/12 6 53.Qe5+ 3.14/11 9 Kc6 (Kd7) 0.97/12 2 54.h4 (Qe6+) 3.18/12 9 54...Rg8+ (Qb3) 1.05/12 10 55.Ng5 3.59/12 4 Rd8 (Qb3) 0.95/12 3 56.Ne6 (Qe4+) 3.60/10 2 56...Rg8+ 0.81/11 1 57.Kf3 3.32/12 3 Qc3+ 0.84/13 6 58.Kf4 3.51/12 6 Qd2+ (Qc1+) 0.84/12 6 59.Qe3 3.70/3 5 Qxe3+ (Qa2) 1.16/13 8 60.fxe3 3.59/13 0 Rh8 1.14/17 3 61.Ng5 (Kg4) 3.78/14 2 61...Rxh4+ (Kd5) 0.58/14 1 62.Kg3 3.64/16 2 Rh6 0.75/18 4 63.e4 3.66/16 3 Kb5 0.83/16 4 64.Nf3 3.66/16 3 Kc4 0.87/17 3 65.d5 3.65/16 2 Rg6+ (Kd3) 0.98/18 6 66.Kf4 4.36/14 2 Rf6+ 0.98/16 3 67.Kg4 4.36/15 1 Kc5 (Rh6) 0.98/16 5 68.h4 4.48/14 0 Rh6 (Rg6+) 1.45/19 7 69.Kg5 4.48/14 1 Rh7 (Rh8) 1.41/18 4 70.Kg6 (h5) 4.43/15 1 70...Rh8 1.16/20 2 71.h5 4.44/16 3 Rg8+ 1.40/21 5 72.Kf5 4.22/16 1 Rh8 (Rf8+) 1.45/17 4 73.Kg4 4.47/15 2 Rh7 1.57/16 2 74.Kg5 4.47/15 1 Re7 1.94/19 12 75.Nd2 4.45/14 1 Rg7+ 2.13/19 5 76.Kf6 4.45/15 1 Rg3 (Rh7) 3.44/19 9 77.h6 (e5) 5.00/15 2 77...Rh3 1.28/19 3 78.Kg5 3.91/16 5 Rg3+ 1.78/20 3 79.Kh4 3.63/18 6 Rg6 1.93/21 4 80.Nb3+ (Kh5) 4.65/15 3 80...Kc4 2.93/17 7 81.Kh5 5.18/14 1 Rb6 3.62/17 3 82.Na5+ 6.05/14 2 Kc5 (Kd3) 4.29/18 2 83.h7 (Nc6) 6.25/14 1 83...Rb1 7.13/18 6 84.Nb7+ 5.97/15 2 Kb6 7.13/15 4 85.Nd6 (Nd8) 6.20/15 1 85...Rh1+ 7.13/18 2 86.Kg6 5.89/17 2 Rg1+ 7.92/19 5 87.Kf6 6.73/16 2 Rh1 7.92/19 3 88.Nf7 7.50/14 2 Rxh7 7.92/18 1 89.d6 7.96/12 1 Kc6 7.92/17 2 90.Ke7 7.68/12 1 Rh4 7.92/13 4 91.Ne5+ 9.09/12 1 Kc5 7.92/16 4 92.d7 9.27/11 0 Rh8 (Kb4) 11.68/13 4 93.Nf7 9.45/10 0 Rh2 (Ra8) 12.68/15 4 94.d8Q 9.66/10 0 Rh1 (Rg2) #21/15 2 95.Qd2 (Qd6+) 13.42/10 2 95...Kc4 #22/11 0 96.Qe2+ 14.08/9 1 Kb4 (Kc5) #21/10 0 97.e5 (Qb2+) 20.40/10 1 97...Ka5 #19/11 0 98.Qd2+ 22.51/9 1 Ka4 #19/10 0 99.Qd4+ 24.52/8 1 Ka3 #18/9 0 100.Qa7+ 21.80/9 1 Kb4 #17/8 0 101.Qb7+ 24.09/9 1 Kc5 #16/7 0 102.Qxh1 23.81/9 3 Kd4 #14/5 0 103.Qe4+ (Qf3) #11/7 0 103...Kc3 (Kxe4) #14/4 0 104.Nd6 (Qd3+) #8/6 0 104...Kd2 #13/4 0 105.Qf3 (Qe3+) #5/7 0 105...Kc1 #4/3 0 106.Qd3 #4/5 0 Kb2 #3/3 0 107.Nc4+ #3/5 0 Ka2 (Kc1) #2/3 0 108.Qc2+ (Qc3) #2/5 0 108...Ka1 #1/3 0 109.Qb2# #1/5 0 1–0
How the search functions is just as relevant as the evaluation heuristics guiding it.
You are looking at two alpha-beta search routines in DF10 and R1.2n and then trying to generalize about an entirely different search approach MCTS (Monte-Carlo tree search).
As you may infer from the name, MCTS isn't going to be effective without a much more powerful evaluation function than that wielded by today's best traditional chess engines...
You are looking at two alpha-beta search routines in DF10 and R1.2n and then trying to generalize about an entirely different search approach MCTS (Monte-Carlo tree search).
As you may infer from the name, MCTS isn't going to be effective without a much more powerful evaluation function than that wielded by today's best traditional chess engines...
> This could really be a rebirth in computer chess!
Indeed, you might even be able to play 1.b3 again!

As I replied elsewhere, it's a shame that the developers of this didn't restrict their chess engine to playing sound, but unpopular openings, to show how these could be used to beat the strongest conventional engines!

I'm sure it would have also made Nelson very happy to see 1.h3 in the mainstream again.

Wow!! That's just so insane. I have never seen Stockfish being outplayed like that.
Wow, is this the End of chess now
.?
Can someone put all this games in PGN format, please?
Game Result Moves Year Event/Locale Opening
1. Stockfish vs AlphaZero 0-1 67 2017 AlphaZero - Stockfish C65 Ruy Lopez, Berlin Defense
2. AlphaZero vs Stockfish 1-0 52 2017 AlphaZero - Stockfish C11 French
3. AlphaZero vs Stockfish 1-0 68 2017 AlphaZero - Stockfish E16 Queen's Indian
4. AlphaZero vs Stockfish 1-0 100 2017 AlphaZero - Stockfish E16 Queen's Indian
5. AlphaZero vs Stockfish 1-0 70 2017 AlphaZero - Stockfish E17 Queen's Indian
6. AlphaZero vs Stockfish 1-0 117 2017 AlphaZero - Stockfish E17 Queen's Indian
7. AlphaZero vs Stockfish 1-0 95 2017 AlphaZero - Stockfish C11 French
8. AlphaZero vs Stockfish 1-0 60 2017 AlphaZero - Stockfish E15 Queen's Indian
9. Stockfish vs AlphaZero 0-1 87 2017 AlphaZero - Stockfish C65 Ruy Lopez, Berlin Defense
10. AlphaZero vs Stockfish 1-0 56 2017 AlphaZero - Stockfish E17 Queen's Indian
best regards,

Can someone put all this games in PGN format, please?
Game Result Moves Year Event/Locale Opening
1. Stockfish vs AlphaZero 0-1 67 2017 AlphaZero - Stockfish C65 Ruy Lopez, Berlin Defense
2. AlphaZero vs Stockfish 1-0 52 2017 AlphaZero - Stockfish C11 French
3. AlphaZero vs Stockfish 1-0 68 2017 AlphaZero - Stockfish E16 Queen's Indian
4. AlphaZero vs Stockfish 1-0 100 2017 AlphaZero - Stockfish E16 Queen's Indian
5. AlphaZero vs Stockfish 1-0 70 2017 AlphaZero - Stockfish E17 Queen's Indian
6. AlphaZero vs Stockfish 1-0 117 2017 AlphaZero - Stockfish E17 Queen's Indian
7. AlphaZero vs Stockfish 1-0 95 2017 AlphaZero - Stockfish C11 French
8. AlphaZero vs Stockfish 1-0 60 2017 AlphaZero - Stockfish E15 Queen's Indian
9. Stockfish vs AlphaZero 0-1 87 2017 AlphaZero - Stockfish C65 Ruy Lopez, Berlin Defense
10. AlphaZero vs Stockfish 1-0 56 2017 AlphaZero - Stockfish E17 Queen's Indian
best regards,
> Wow, is this the End of chess now
.?
Not at all. I doubt if my 60+ yr. old partner & myself can remember all AZ's analysis,

I'm sure 5 years from now AZ will be like the last wunderkind Rybka. But right now it's fun to see. I wish the games will be published soon.
Of course, I easily could be way wrong!
Yeah, that's an amazing game, but at the end I'm less impressed by this victory than by AlphaGo / AlphaZero wins over Ke Jie at Go. In the match vs. Stockfish, I'm afraid that the computer's speed made a difference :/
*some commentary about it suggests that Alpha go have very large computational power advantages over stockfish, running on specialized TPU processors. In any case, the advance is showing the generality of the approach.
https://forums.spacebattles.com/threads/alphago-learns-chess-pwns-stockfish.593908/
https://forums.spacebattles.com/threads/alphago-learns-chess-pwns-stockfish.593908/
Your statement is both true and irrelevant...
The specialized TPU processors suitable for implementation of deep neural networks provide much more processing power (by using highly parallel architectures) than traditional computers, but in a very specialized manner, and wouldn't be useful for the algorithms used by Stockfish. The specialized processors optimized for implementing deep neural networks are progressing rapidly, and there is no reason to believe that systems available for home use today couldn't host AlphaZero like algorithms and outperform today's best traditional chess engines running on state of the art computer hardware.
The specialized TPU processors suitable for implementation of deep neural networks provide much more processing power (by using highly parallel architectures) than traditional computers, but in a very specialized manner, and wouldn't be useful for the algorithms used by Stockfish. The specialized processors optimized for implementing deep neural networks are progressing rapidly, and there is no reason to believe that systems available for home use today couldn't host AlphaZero like algorithms and outperform today's best traditional chess engines running on state of the art computer hardware.
so, what does its hardware cost ?
-------------edit---------------------
https://cloud.google.com/tpu/
12 May 2017 ... In short, we found that the TPU delivered 15–30X higher performance and 30–
80X higher performance-per-watt than contemporary CPUs and GPUs. These
advantages help many of Google's services run state-of-the-art neural networks
at scale and at an affordable cost. In this post, we'll take an in-depth
----------------------------------------
Nvidia compared its Tesla P40 GPU against Google's TPU and it came out on top.
However, the comparison may not be quite fair given that we don't know how
much a TPU costs and that it uses three times less power than the Tesla P40, costing less to run.
---
a single server with 8 Tesla P40s
delivers the performance of over 140 CPU servers
---
Max Power:250W
Used & new (2) from $6,500.00 + $9.49 shipping
---
PNY Tesla P40, 24GB GDDR5 (TCSP40M-24GB-PB) ab € 7098,90
-------------------------------------------------
using 5,000 first-generation TPUs (15) to generate self-play games and 64 second-generation
TPUs to train the neural networks.1
AlphaZero and the previous AlphaGo Zero used a single machine with 4 TPUs
Stockfish and Elmo played at their strongest skill level using 64 threads and a hash size of 1GB.
========================
Interesting and remarkable IMO that they wrote nothing about hardware
performance factor estimates in the paper.
I estimate a factor of ~10 in favour of alpha zero
-------------edit---------------------
https://cloud.google.com/tpu/
12 May 2017 ... In short, we found that the TPU delivered 15–30X higher performance and 30–
80X higher performance-per-watt than contemporary CPUs and GPUs. These
advantages help many of Google's services run state-of-the-art neural networks
at scale and at an affordable cost. In this post, we'll take an in-depth
----------------------------------------
Nvidia compared its Tesla P40 GPU against Google's TPU and it came out on top.
However, the comparison may not be quite fair given that we don't know how
much a TPU costs and that it uses three times less power than the Tesla P40, costing less to run.
---
a single server with 8 Tesla P40s
delivers the performance of over 140 CPU servers
---
Max Power:250W
Used & new (2) from $6,500.00 + $9.49 shipping
---
PNY Tesla P40, 24GB GDDR5 (TCSP40M-24GB-PB) ab € 7098,90
-------------------------------------------------
using 5,000 first-generation TPUs (15) to generate self-play games and 64 second-generation
TPUs to train the neural networks.1
AlphaZero and the previous AlphaGo Zero used a single machine with 4 TPUs
Stockfish and Elmo played at their strongest skill level using 64 threads and a hash size of 1GB.
========================
Interesting and remarkable IMO that they wrote nothing about hardware
performance factor estimates in the paper.
I estimate a factor of ~10 in favour of alpha zero
Comparing Alpha Zero 'performance' to a high end workstation:
If you are measuring performance by IPS, it will be much, much greater than 10X due to the huge number of simple processors.
If you are measuring performance by power consumption, it will be comparable or maybe less.
If you are measuring performance by hardware cost, it would probably be comparable.
Google's TPU will compete head to head with products from Nvidia and AMD, and the prices will no doubt fall quickly as these products get used more in server farms. There is little reason to believe that this won't be the death knell for computer chess based on alpha-beta techniques...
If you are measuring performance by IPS, it will be much, much greater than 10X due to the huge number of simple processors.
If you are measuring performance by power consumption, it will be comparable or maybe less.
If you are measuring performance by hardware cost, it would probably be comparable.
Google's TPU will compete head to head with products from Nvidia and AMD, and the prices will no doubt fall quickly as these products get used more in server farms. There is little reason to believe that this won't be the death knell for computer chess based on alpha-beta techniques...
64-36 isnt really a boast we need more games.
It is when there are zero losses and huge hardware like Stockfish had.
Wikipedia :
The hardware cost for a single AlphaGo Zero system, including custom components,
has been quoted as around $25 million.[12][13]
Silver :
> While Alpha Go runs on about $25 million in hardware ...
The hardware cost for a single AlphaGo Zero system, including custom components,
has been quoted as around $25 million.[12][13]
Silver :
> While Alpha Go runs on about $25 million in hardware ...
The references in the Wikipedia article are pretty dubious, and not backed up by any explanation of calculation.
The AlphaGo Zero system is based on custom ASICs, so there was certainly a lot of NRE involved in this development. $25 million still sounds high, but this is such a tiny drop in the bucket for Alphabet that they would almost certainly have spent the money just for the good PR...
Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill