Historic Milestone: AlphaZero

[clumma]

AlphaZero's total score against SF 8 seems to be 797 out of 1300 (318/958/24), which is +80 Elo.

[JJJ]

AlphaZero's total score against SF 8 seems to be 797 out of 1300 (318/958/24), which is +80 Elo.

Maybe it was Houdini 7 after all.

[Jhoravi]

I wonder if neural chess for PC is competitive against the best engines using this latest hardware

https://www.anandtech.com/show/12135/nv ... 00-dollars

[shrapnel]

I wonder if neural chess for PC is competitive against the best engines using this latest hardware

https://www.anandtech.com/show/12135/nv ... 00-dollars

A computer chess program which can harness the Power of the GPU !
My dream come true !
Hmm... a bit on the expensive side, but I might just be able to afford it.
Now all I need is Google to start selling the Alpha-Zero Program.

[Nay Lin Tun]

I wonder if neural chess for PC is competitive against the best engines using this latest hardware

https://www.anandtech.com/show/12135/nv ... 00-dollars

It would be a luxury feature and total cost may exceed $ 10000 ( including the price of Alpha Zero program). However,with these publicily available hardwares, a handful of people will be willing to spend that amont of money for their hobby.

[shrapnel]

would be a luxury feature and total cost may exceed $ 10000 ( including the price of Alpha Zero program). However,with these publicily available hardwares, a handful of people will be willing to spend that amont of money for their hobby.

I remember when Mobile Phones first appeared many years ago, they were a luxury item and were very expensive and out of the reach of most people.
Now, even a poor labourer walks around with one in his pocket !
The same thing will happen here...prices will come down....just a matter of time.

[clumma]

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...

According to Andreas here and here, SF 8 gains 128 Elo going from 1 to 4 cores. That is more than reported by CCRL 40/40, presumably because of time control or other differences. However if I apply the doubling behavior from Andreas' tests to the CCRL scores, I get 3475 for SF 8 on 64 cores. Probably still wrong but maybe a bit better.

[tpoppins]

According to Andreas here and here, SF 8 gains 128 Elo going from 1 to 4 cores. That is more than reported by CCRL 40/40, presumably because of time control or other differences.

Most likely because self-play inflates the differences.

[Lyudmil Tsvetkov]

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

It learned them by self-training.

[Lyudmil Tsvetkov]

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.

Evaluation easily suggests to me black is lost, but of course, it used just brute force.