Historic Milestone: AlphaZero

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
corres
Posts: 456
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: MCTS-NN vs alpha-beta

Post by corres » Thu Dec 07, 2017 7:27 pm

[quote="Albert Silver"]

They did more than that. They published a scientific paper anyone can read.
https://en.chessbase.com/post/the-futur ... arns-chess

[/quote]

Sorry, but this is not a scientific article, this is a report about DeepMind and they works and results.
Moreover it is inaccurate because AlphaZero do not start its learning from zero but it was feeded a lot of high quality human chess games before the self training. In this manner AlphaZero got a good quality specialized opening book but Stockfish did not use any opening book during tests and its hash table was only 1GB - instead of 64GB as normally used.
Another very important difference between Stockfish and AlphaZero is Stockfish does not preserve its games but AlphaZero has an incredible huge memory to preserve its "experience".
I think the difference between the power of AlphaZero's hardware and whichever hardware of chess engines is so huge that it is very incorrect any comparison between them. A comparison based on the position per seconds is very misleading because working of Stockfish and working of AplhaZero is different fundamentally.

supersharp77
Posts: 475
Joined: Sat Jul 05, 2014 5:54 am
Location: Southwest USA

Re:Alpha Zero's "Phantom Win over Stockfish 8"

Post by supersharp77 » Thu Dec 07, 2017 7:33 pm

corres wrote:
Albert Silver wrote:
Sorry, but this is not a scientific article, this is a report about DeepMind and they works and results.
Moreover it is inaccurate because AlphaZero do not start its learning from zero but it was feeded a lot of high quality human chess games before the self training. In this manner AlphaZero got a good quality specialized opening book but Stockfish did not use any opening book during tests and its hash table was only 1GB - instead of 64GB as normally used.
Another very important difference between Stockfish and AlphaZero is Stockfish does not preserve its games but AlphaZero has an incredible huge memory to preserve its "experience".
I think the difference between the power of AlphaZero's hardware and whichever hardware of chess engines is so huge that it is very incorrect any comparison between them. A comparison based on the position per seconds is very misleading because working of Stockfish and working of AplhaZero is different fundamentally.

=====================================

Note:To whom it may concern...We refuse to accept the conclusions of this "so called" defeat of Stockfish 8 by 'Alpha Zero' (a program no one seems to have heard of before this so called "result was published") If Google wants the chess engine community to accept these "conclusions" follow the 'RULES'of CHESS ENGINE TESTING..and Enter 'ALPHA ZERO" in TCEC or some other Recognized chess engine tournament...and let the Chips fall where they mayl....until then....The "Google Teams" results are 100% meaningless!! :D :wink:

CheckersGuy
Posts: 272
Joined: Wed Aug 24, 2016 7:49 pm

Re: MCTS-NN vs alpha-beta

Post by CheckersGuy » Thu Dec 07, 2017 7:41 pm

corres wrote:
Albert Silver wrote:
They did more than that. They published a scientific paper anyone can read.
https://en.chessbase.com/post/the-futur ... arns-chess
Sorry, but this is not a scientific article, this is a report about DeepMind and they works and results.
Moreover it is inaccurate because AlphaZero do not start its learning from zero but it was feeded a lot of high quality human chess games before the self training. In this manner AlphaZero got a good quality specialized opening book but Stockfish did not use any opening book during tests and its hash table was only 1GB - instead of 64GB as normally used.
Another very important difference between Stockfish and AlphaZero is Stockfish does not preserve its games but AlphaZero has an incredible huge memory to preserve its "experience".
I think the difference between the power of AlphaZero's hardware and whichever hardware of chess engines is so huge that it is very incorrect any comparison between them. A comparison based on the position per seconds is very misleading because working of Stockfish and working of AplhaZero is different fundamentally.
You said "A0 has an incredibly huge memory" that's still something we don't know, or do you know how much GB they used to store the weights of the neural-network ? I would be suprised if it was more then 8gb (which I wouldnt call incredibly huge) but we don't know

Jouni
Posts: 1787
Joined: Wed Mar 08, 2006 7:15 pm

Re:Alpha Zero's "Phantom Win over Stockfish 8"

Post by Jouni » Thu Dec 07, 2017 7:44 pm

One question: how do they tell the rules to Alpha Zero? Give they PDF from FIDE page and start?
Jouni

Albert Silver
Posts: 2747
Joined: Wed Mar 08, 2006 8:57 pm
Location: Rio de Janeiro, Brazil

Re: MCTS-NN vs alpha-beta

Post by Albert Silver » Thu Dec 07, 2017 8:18 pm

corres wrote:
Albert Silver wrote:
They did more than that. They published a scientific paper anyone can read.
https://en.chessbase.com/post/the-futur ... arns-chess
Sorry, but this is not a scientific article, this is a report about DeepMind and they works and results.
Moreover it is inaccurate because AlphaZero do not start its learning from zero but it was feeded a lot of high quality human chess games before the self training.
No, it didn't.
In this manner AlphaZero got a good quality specialized opening book
No, it didn't.
Another very important difference between Stockfish and AlphaZero is Stockfish does not preserve its games but AlphaZero has an incredible huge memory to preserve its "experience".
No, it doesn't.
"Tactics are the bricks and sticks that make up a game, but positional play is the architectural blueprint."

Leo
Posts: 459
Joined: Fri Sep 16, 2016 4:55 pm
Location: Saint Paul, Minnesota USA

Re:Alpha Zero's "Phantom Win over Stockfish 8"

Post by Leo » Thu Dec 07, 2017 8:23 pm

supersharp77 wrote:
corres wrote:
Albert Silver wrote:
Sorry, but this is not a scientific article, this is a report about DeepMind and they works and results.
Moreover it is inaccurate because AlphaZero do not start its learning from zero but it was feeded a lot of high quality human chess games before the self training. In this manner AlphaZero got a good quality specialized opening book but Stockfish did not use any opening book during tests and its hash table was only 1GB - instead of 64GB as normally used.
Another very important difference between Stockfish and AlphaZero is Stockfish does not preserve its games but AlphaZero has an incredible huge memory to preserve its "experience".
I think the difference between the power of AlphaZero's hardware and whichever hardware of chess engines is so huge that it is very incorrect any comparison between them. A comparison based on the position per seconds is very misleading because working of Stockfish and working of AplhaZero is different fundamentally.

=====================================

Note:To whom it may concern...We refuse to accept the conclusions of this "so called" defeat of Stockfish 8 by 'Alpha Zero' (a program no one seems to have heard of before this so called "result was published") If Google wants the chess engine community to accept these "conclusions" follow the 'RULES'of CHESS ENGINE TESTING..and Enter 'ALPHA ZERO" in TCEC or some other Recognized chess engine tournament...and let the Chips fall where they mayl....until then....The "Google Teams" results are 100% meaningless!! :D :wink:
Great statement. If the Cerebellum Brainfish people would release their program publically we could train it around the clock on a server and build our own chess engine book. I have played with cerebellum and it goes very deep. Its potential is a lot bigger because they use tiny hardware to improve it. Give SF Dev 128 GB or more of DDR4 4000 Mhz Ram and 30 overclocked cores plus 7 man EGTB. Maybe use a liquid nitrogen setup if that would help however that works. The mere 3 losses with black would be eliminated and maybe turned to wins. The 25 losses would be chipped away at. Much more drawn and some would be wins. Deep Mind or whatever it is called has accomplished something amazing and shaken up the chess computer world but they have not conquered the chess engine world until further notice. Why in the world didn't they let Deep Zero run for a week and learn. I think Milos might be right. After 4 hours it maxed out. It was saturated.
Brainfish fan. AMD fan.

User avatar
MikeGL
Posts: 694
Joined: Thu Sep 01, 2011 12:49 pm

Re: Much weaker than Stockfish

Post by MikeGL » Fri Dec 08, 2017 1:22 am

Lyudmil Tsvetkov wrote:That is what you have written: 100 game match versus SF8 was played on all 12 common ECO openings.
Don't put that word into my mouth.
My second reply to you on this thread already mentioned a total of 1,200 games which made you confused.
But I understand your confusion because Table 1 of the paper is conflicting with Table 2. That's one of the bug/error in the paper.


But if you analyze Table 2 it is 1,200 games total for SF8 match.
Image
Just add the highlighted numbers on the above pic and you will get 1,200.


You wanted to see a 5000 Elo engine, but if you read page 5 footnote you will notice that these researchers already claimed 3500 Elo seems to be the peak because of prevalent draws.

From the Paper:
3 The prevalence of draws in high-level chess tends to compress the Elo scale, compared to shogi or Go.
So perhaps these highly paid researchers/scientists and specialists from Google already
discovered something novel about 3500 being peak due to draw rates and possible draw
escapes by a very strong engine (ie Stockfish 8).

clumma
Posts: 151
Joined: Fri Oct 10, 2014 8:05 pm
Location: Berkeley, CA

Re: Historic Milestone: AlphaZero

Post by clumma » Fri Dec 08, 2017 2:03 am

Kai, can you estimate the FIDE-equivalent Elo for Stockfish 8 with 64 cores and 1 min per move?

My crude estimate goes...

* Shredder 12 won WCSC 2015 on 4 cores with 45 min and a 15s increment. Regan estimates its FIDE Elo in this tournament at 3075.

* On CCRL 40/40 list, Shredder 12 4CPU Elo is 3025. So this list is maybe 50 Elo below FIDE scale. But this is within Regan's margin of error and the time control is slightly less than for WCSC, so call it a wash.

* Stockfish 8 gains 89 Elo on this list going from 1 to 4 cores. Assuming the same improvement for each 4x increase in cores, we have

(SF8 4CPU Elo) + (89 * 2) = 3567

But probably this core scaling is wrong...

-Carl

User avatar
MikeB
Posts: 2526
Joined: Thu Mar 09, 2006 5:34 am
Location: Pen Argyl, Pennsylvania

Re: MCTS-NN vs alpha-beta

Post by MikeB » Fri Dec 08, 2017 2:59 am

MikeGL wrote:
kranium wrote:
Lyudmil Tsvetkov wrote:It is not at all clear to me where were books used and where not.
I'm sure opening books were not used...
In the early self-play games things like 1.a3, 1.a4, etc. were probably tried by AlphaZero...
eventually it learned that 1. e4 or 1. d4 had the highest success rates.
Books or no books, I think AlphaZero would still demolish SF8.
Just look at this game 9, it was a decent French Defence by SF8, but it was dismantled with
amazing tactical and strategic shots by AlphaZero which seems to be beyond the reach of alpha-beta engines.

[pgn]
[Event "?"]
[Site "?"]
[Date "2017.12.06"]
[Round "9"]
[White "AlphaZero"]
[Black "Stockfish"]
[Result "1-0"]
[TimeControl "40/1260:300"]
[Termination "normal"]
[PlyCount "103"]
[WhiteType "human"]
[BlackType "human"]

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
[/pgn]

not sure if 18.g4!, 30.Bxg6! and other would be found by current engines.

[d]r2qk2r/3bbppp/1p2p3/pP1pP3/P2P1P2/3BKN2/6PP/R2Q3R w kq - 0 18
After 17...b6 of black, can some engine consider 18.g4! in this position?



xxxxxxxxxxxxxxxxxxxxxxxxxxxxx
[d]4q2r/1b1kbp2/1p2p1p1/pP1pP1N1/P2P1PQP/3BK3/2R5/8 w - - 6 30
After 29...Be7, can current engines consider 30.Bxg6! here?


Would be nice if we can try to feed some difficult epd positions into AlphaZero,
to estimate its ELO strength.
Hi Mike - nice positions for sure - near Final Release dev SF-McB - running on 12 cores with 18 threads. This version is more correlated to Corchess than into past - the very excellent SF fork created by Ivan Ivec designed for correspondence chess , the first position is way too hard

Code: Select all

dep	score	nodes	time	(not shown:  tbhits	knps	seldep)
 42	+0.91!	14.8G	11:12.14	Bxg6! 
 41	+0.80 	14.5G	10:55.80	Bxg6 Bxg5 Qxg5 fxg6 f5 Rg8 f6 Qf7 Qh6 Kd8 Kd2 Qd7 Kd1 Qf7 Rc3 Re8 Ke2 Rg8 Qc1 Qd7 Qa3 Ke8 Kd1 g5 hxg5 Rxg5 Rh3 Rg8 Qd3 Qf7 Ke2 Kd8 Qa3 Qf8 Qxf8+ Rxf8 Rh7 Bc8 Kf3 Bd7 Rg7 Be8 Rb7 Rh8 Rxb6 Rh4 Rd6+ Kc8 Rxe6 Bh5+ Ke3 Re4+ Kd3 
 41	+0.65!	13.8G	10:23.22	Bxg6! 
 41	+0.39?	13.2G	9:57.51	Bxg6 Bxg5? 
 41	+0.74!	12.3G	9:14.22	Bxg6! 
 41	+0.62!	11.9G	8:58.22	Bxg6! 
 41	+0.55!	11.7G	8:48.89	Bxg6! 
 40	+0.47 	10.8G	8:06.15	Bxg6 Bxg5 Qxg5 fxg6 f5 Rg8 f6 Qf7 Qh6 Kd8 Kd2 Qd7 Kd3 Qf7 Qc1 Rh8 Qa3 Qf8 Qxf8+ Rxf8 Rg2 Rh8 Rxg6 Ke8 Ke3 Bc8 h5 Bd7 h6 Rh7 Rg8+ Kf7 Rb8 Rxh6 Rb7 Ke8 Rxb6 Rh3+ Kf2 Ra3 Rb8+ Kf7 Rb7 Ke8 f7+ Kxf7 Rxd7+ Ke8 Ra7 Rxa4 Ke3 Kd8 Ra6 Ke7 
 40	+0.42!	9.26G	6:58.64	Bxg6! 
 40	+0.26!	8.80G	6:37.81	Bxg6! 
 40	+0.15!	8.77G	6:36.60	Bxg6! 
 40	+0.07!	8.53G	6:25.78	Bxg6! 
 39	  0.00 	5.80G	4:21.54	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 Qxd5+ Ke7 Qc6 Qd2+ Kf3 Qd1+ Ke4 Qg4+ Ke3 Qg1+ Kd3 Qb1+ Ke3 Qe1+ Kd3 Qd1+ Ke4 
 38	  0.00 	3.97G	2:59.62	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 Qxd5+ Ke7 Qc6 Qd2+ Kf3 Qd1+ Ke4 Qg4+ Ke3 Qg1+ Kd3 Qb1+ Ke3 Qe1+ Kd3 Qd1+ Ke4 
 37	  0.00 	3.02G	2:16.14	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 Qxd5+ Ke7 Qc6 Qd2+ Kf3 Qd1+ Ke4 Qg4+ Ke3 Qg1+ Kd3 Qb1+ Ke3 Qe1+ Kd3 Qd1+ Ke4 
 36	  0.00 	543.9M	0:24.94	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 Qxd5+ Ke7 Qc6 Qd2+ Kf3 Qd1+ Ke4 Qg4+ Ke3 Qg1+ Kd3 Qb1+ Ke3 Qe1+ Kd3 Qd1+ Ke4 
 35	  0.00 	445.5M	0:20.48	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 Qxd5+ Ke7 Qc6 Qd2+ Kf3 Qd1+ Ke4 Qg4+ Ke3 Qg1+ Kd3 Qb1+ Ke3 Qe1+ Kd3 Qd1+ Ke4 
 34	  0.00 	335.9M	0:15.50	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 Qxd5+ Ke7 Qc6 Qd2+ Kf3 Qd1+ Ke4 Qg4+ Ke3 Qg1+ Kd3 Qb1+ Ke3 Qe1+ Kd3 Qd1+ Ke4 
 33	  0.00 	283.5M	0:13.15	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 Qxd5+ Ke7 Qc6 Qd2+ Kf3 Qd1+ Ke4 Qg4+ Ke3 Qg1+ Kd3 Qb1+ Ke3 Qe1+ Kd3 Qd1+ Ke4 
 32	  0.00 	241.8M	0:11.24	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 Qxd5+ Ke7 Qc6 Qd2+ Kf3 Qd1+ Ke4 Qg4+ Ke3 Qg1+ Kd3 Qb1+ Ke3 Qe1+ Kd3 Qd1+ Ke4 
 31	  0.00 	219.0M	0:10.20	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 Qxd5+ Ke7 Qc6 Qd2+ Kf3 Qd1+ Ke4 Qg4+ Ke3 Qg1+ Kd3 Qb1+ Ke3 Qe1+ Kd3 Qd1+ Ke4 
 30	  0.00 	203.9M	0:09.51	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 Qxd5+ Ke7 Qc6 Qd2+ Kf3 Qd1+ Ke4 Qg4+ Ke3 Qg1+ Kd3 Qf1+ Ke3 Qe1+ Kd3 Qd1+ Ke4 
 29	  0.00 	134.7M	0:06.33	Nxe6 fxe6 Bxg6 Qf8 h5 Rg8 Ke2 Bb4 f5 exf5 Qxf5+ Qxf5 Bxf5+ Kd8 h6 Rg2+ Ke3 Rxc2 Bxc2 Bf8 h7 Bg7 Kf4 Bc8 Bg6 Ke7 Kg5 Be6 Bf5 Bxf5 Kxf5 Bh8 Kf4 Bg7 Kg5 Kf7 Kf5 Ke7 Kg6 Bh8 Kh5 Kf7 Kh6 Bg7+ Kg5 Bh8 Kf5 
 28	  0.00 	128.0M	0:06.02	Nxe6 fxe6 Bxg6 Qf8 h5 Rg8 Ke2 Bb4 Kf3 Be7 Ke2 
 27	  0.00 	93.2M  	0:04.41	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 Qxd5+ Ke7 Qc6 Qd2+ Kf3 Qd1+ Ke4 Qg4+ Ke3 Qg1+ Kd3 Qf1+ Ke3 Qe1+ Kd3 Qd1+ Ke4 
 26	  0.00 	86.7M  	0:04.11	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 Qxd5+ Ke7 Qc6 Qd2+ Kf3 Qd1+ Ke4 Qg4+ Ke3 Qg1+ Kd3 Qf1+ Ke3 Qe1+ Kd3 Qd1+ Ke4 
 25	  0.00 	73.4M  	0:03.50	Nxe6 fxe6 Bxg6 Qf8 h5 Ba3 Ke2 Qh6 f5 Rc8 fxe6+ Kd8 Rxc8+ Bxc8 Qf3 Bb4 e7+ Bxe7 Qxd5+ Kc7 Qf7 Qg5 Be4 Kd8 Kd3 Qg4 Qd5+ Bd7 Qa8+ Bc8 Qd5+ 
 25	+0.11?	63.9M  	0:03.07	Nxe6 fxe6? 
 24	+0.93 	39.6M  	0:01.96	Nxe6 fxe6 Bxg6 Qf8 h5 Rg8 Ke2 Qg7 f5 Qh6 Kf3 Rc8 fxe6+ Kd8 Qf4 Qg7 Qf7 Qxf7+ exf7 Rxc2 Bxc2 Bf8 Kg4 Ke7 Kg5 Kxf7 Bg6+ Kg8 Kf6 Bc8 Bf7+ Kh7 Be6 Bb7 Bf5+ Kh6 e6 Bc8 Kf7 Bb4 e7 Bxe7 Bxc8 Bb4 Ke6 Kxh5 Kxd5 Bc3 
 23	+0.47 	28.9M  	0:01.49	Nxe6 fxe6 Bxg6 Qf8 h5 Bb4 Kf3 Qh6 f5 Rc8 fxe6+ Kd8 Qh4+ Be7 Qf4 Qg7 Qf7 Qxf7+ exf7 Rxc2 Bxc2 Bf8 Kg4 Ke7 Kg5 Kxf7 Bg6+ Ke7 h6 Bc8 Bh5 Bd7 h7 Bg7 Kg6 Kf8 Bd1 Bh8 Bb3 Be6 Bc2 Bf7+ Kf5 
 22	+0.70 	22.8M  	0:01.21	Nxe6 fxe6 Bxg6 Qg8 h5 Qf8 f5 Qh6+ Kf3 Rc8 fxe6+ Kd8 Qf4 Qxf4+ Kxf4 Rc4 Rxc4 dxc4 Bf5 Bf8 Kg5 Ke7 h6 c3 Kg6 Bxh6 Kxh6 Bd5 Bc2 Kxe6 Kg5 Bc4 Kf4 
 21	+0.57 	19.4M  	0:01.05	Nxe6 fxe6 Bxg6 Qf8 h5 Rg8 Ke2 Qg7 f5 Qh6 Kf3 Rf8 f6 Bb4 Bd3 Kd8 Kg2 Bd2 Qg7 Qg5+ Qxg5 Bxg5 Kf3 Bh6 Rg2 Ke8 Rg6 Rh8 Rg7 Bc8 Rc7 Bd7 Bg6+ Kd8 
 20	+0.34 	14.5M  	0:00.83	Nxe6 fxe6 Bxg6 Qf8 h5 Bb4 Kf3 Qh6 f5 Rf8 Qf4 Qxf4+ Kxf4 Rc8 Rg2 Rh8 fxe6+ Ke7 Bf5 Kf8 Bg4 Bc8 Rc2 Ke7 Rc7+ Kd8 Rc6 Bd2+ Kg3 Ke7 Rc7+ Kd8 Rg7 Bb4 
 19	+0.12 	12.2M  	0:00.72	Nxe6 fxe6 Bxg6 Qf8 h5 Bb4 Kf3 Qh6 f5 Rf8 Qf4 Qxf4+ Kxf4 Rc8 Rg2 Rh8 fxe6+ Ke7 Rc2 Rc8 Kg5 Bc3 Rf2 Rb8 Rf6 Bxd4 Kf5 
 18	+0.25 	8.80M  	0:00.56	Nxe6 fxe6 Bxg6 Qf8 h5 Bb4 Kf3 Qh6 f5 Rf8 Qg3 Be7 Ke2 Qg5 fxe6+ Kd8 Qxg5 Bxg5 Bf7 Rh8 Kf3 Rh7 Kg3 
 17	  0.00 	4.49M  	0:00.33	Rh2 Ba3 Rc2 Be7 Rh2 
 16	  0.00 	3.16M  	0:00.25	Rh2 Ba3 Rc2 Be7 Rh2 
 15	  0.00 	3.11M  	0:00.25	Rh2 Ba3 Rc2 Be7 Rh2 
 14	  0.00 	2.04M  	0:00.18	Rh2 Ba3 Rc2 Be7 Rh2 
 13	  0.00 	1.54M  	0:00.14	Rh2 Ba3 Rc2 Be7 Rh2 
 12	 -0.12 	996245	0:00.10	Kf2 Qf8 Kg2 Bd8 Bf1 Qe7 Bd3 Ke8 Rc3 Rh6 Rb3 Qd7 
 11	 -0.23 	569904	0:00.06	Kf2 Qf8 Qg3 Bd8 Kg2 Qe7 Ra2 Kc7 Rc2+ Kb8 Qf2 Ka7 Qg3 
 10	  0.00 	167480	0:00.02	Kf2 Qf8 Ke3 Bd8 Ke2 Be7 Ke3 
  9	+0.13 	107447	0:00.02	Kf2 Qf8 Kg3 Bxg5 fxg5 Qa3 Qf3 Rh7 Qe3 
  8	+0.05 	95511  	0:00.01	Be2 Qf8 Kf2 Bxg5 fxg5 Qa3 Qf4 Rh7 
  7	+0.17 	65531  	0:00.01	Kf2 Qf8 Be2 Bxg5 fxg5 Qa3 Qf4 
  6	+0.27 	31767  	0:00.01	Bf1 Qf8 Bh3 Bxg5 fxg5 Qa3+ Kf2 
  5	+1.45 	11738  	0:00.01	Kf2 Ba8 Rc1 Bb7 Be2 
  4	+2.07 	5243    	0:00.00	Bf1 f5 exf6 Bxf6 Bh3 
  3	+1.79 	4474    	0:00.00	Bf1 f5 exf6 Bxf6 
  2	+1.42 	3629    	0:00.00	Bf1 Ba8 
  1	+1.44 	1865    	0:00.00	Bxg6 
  0	# 

Milos
Posts: 2993
Joined: Wed Nov 25, 2009 12:47 am

Re: MCTS-NN vs alpha-beta

Post by Milos » Fri Dec 08, 2017 3:54 am

MikeB wrote:
MikeGL wrote:
kranium wrote:
Lyudmil Tsvetkov wrote:It is not at all clear to me where were books used and where not.
I'm sure opening books were not used...
In the early self-play games things like 1.a3, 1.a4, etc. were probably tried by AlphaZero...
eventually it learned that 1. e4 or 1. d4 had the highest success rates.
Books or no books, I think AlphaZero would still demolish SF8.
Just look at this game 9, it was a decent French Defence by SF8, but it was dismantled with
amazing tactical and strategic shots by AlphaZero which seems to be beyond the reach of alpha-beta engines.

[pgn]
[Event "?"]
[Site "?"]
[Date "2017.12.06"]
[Round "9"]
[White "AlphaZero"]
[Black "Stockfish"]
[Result "1-0"]
[TimeControl "40/1260:300"]
[Termination "normal"]
[PlyCount "103"]
[WhiteType "human"]
[BlackType "human"]

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
[/pgn]

not sure if 18.g4!, 30.Bxg6! and other would be found by current engines.

[d]r2qk2r/3bbppp/1p2p3/pP1pP3/P2P1P2/3BKN2/6PP/R2Q3R w kq - 0 18
After 17...b6 of black, can some engine consider 18.g4! in this position?



xxxxxxxxxxxxxxxxxxxxxxxxxxxxx
[d]4q2r/1b1kbp2/1p2p1p1/pP1pP1N1/P2P1PQP/3BK3/2R5/8 w - - 6 30
After 29...Be7, can current engines consider 30.Bxg6! here?


Would be nice if we can try to feed some difficult epd positions into AlphaZero,
to estimate its ELO strength.
Hi Mike - nice positions for sure - near Final Release dev SF-McB - running on 12 cores with 18 threads. This version is more correlated to Corchess than into past - the very excellent SF fork created by Ivan Ivec designed for correspondence chess , the first position is way too hard

Code: Select all

dep	score	nodes	time	(not shown:  tbhits	knps	seldep)
 42	+0.91!	14.8G	11:12.14	Bxg6! 
Problem here is that even at this point
[d]3k2r1/1b3q2/1p2pPpQ/pP1pP3/P2P3P/4K3/2R5/8 w - - 1 35
SF needs depth of 32 to realize it is lost. Since due to move ordering 30.Bxd6 comes quite low on the list of moves (even though all of them are scored as 0.00), due to LMR SF needs an extreme depth (like 42 in your case) to see that move is bad.
I don't believe A0 actually sees it due to its evaluation, but due to MCTS. While SF sees only 3-fold repetitions MCTS explores all the root moves from the given position till the end and finds an actual win, back-propagating it and boosting probability of Bxg6 move.

Post Reply