[ExI] alpha zero

Alejandro Dubrovsky alito at organicrobot.com
Sat Dec 9 05:06:06 UTC 2017


On 09/12/17 08:38, spike wrote:
> OK I worked on this from my own board, and now I change my assessment.  
> After f5, I now think black has a good chance of at least a draw in this 
> awful position, 50% or better.  I am having a hard time seeing how 
> StockFish could have played Kh8 in any of its versions regardless of the 
> platform and settings.
> 
> f5 picks up a knight for the pawn, then once one of the black knights is 
> sacked back, the position no longer puts the ug in ugly.
> 
> DeepMind has some esplainin to do?



I set the position up on StockFish 8. My computer doesn't have 64 CPU 
threads, so I only gave it one but I let it mull on it for a few hours 
to compensate. It flips between the two but it does seem to prefer Kh8 
after all.

It likes f5 for the first 3 seconds (14 ply), then switches to Kh8 till 
it reaches 35 ply at about 5 minutes (35 ply in 5 minutes!! I forget how 
insanely fast these things are), then back to f5 till depth 40, then Kh8 
only briefly, then back to f5 until depth 44 when it switches back to 
Kh8. This last switch is at around the 43 minutes mark on my computer 
and it sticks with it at least until depth 48 which is 3 hours on. The 
time control was 1 minute per move, and they were using around 100 more 
computation than I'm giving it, so that sounds about right.

It sees the f5 line going as follows:

20 ... f5 21. Qf4 fxe4 22. Bxe4 Rxf4 23. Bxh7+ Kxh7 24. Bxf4 c5 25. Re8 
Na5 26. Rae1 Bc4 27. Rxb8 Rxb8 28. Bxb8 Nc6 29. Bd6 Bxa2 30. f3 Be6 31. 
Kf2 Na5 32. Bf4 Kg8 33. Ra1 Nc6 34. Bd6 Kf7 35. g4 Bb3 36. Re1 Bc2 37. 
f4 Bd3 38. Ke3 c4 39. b3 b5 40. Ra1 Ke6 41. Bc5 a5 42. Bd4 Kf7 43. Bb6 
d5 44. Kd2 b4 45. Bxa5 which does look pretty tricky. If I play that 
through, its evaluation at the end is a solid 0.00, but it must have 
seen some improvement for white at higher depths from the original 
position, since it switches from this line that had an evaluation of 
-0.27 to the Kh8 line to which it assigns a -0.52.

The improvement I think is 27.Rf8 instead of Rxb8. When playing it 
through, if I give it some time on that move, it switches to Rf8 after a 
few minutes and its evaluation goes to +1.3 for white and rising. The 
line seems to be: 27. Rf8 Be6 28. Bd6 and then one of Nc4, Nac6 or g6 
but seems to lean towards g6, in which eventually Bxb8 is eventually 
played and then black loses another exchange by taking on that bishop 
(!!?) but white is left with only its a pawn. Here's the line:

28 ... g6 29. f4 Nc4 30. Bxb8 Nxb2 31. Re8 Nd3 32. Re3 c4 33. g4 Rxb8 
34. Rxb8 Nxf4 35. Kf2 Nd5 36. Re1 Bxg4 37. hxg6+ Kxg6 38. Re4 Be6 39. 
Rxc4 Ne7 40. Rc2 Nc6 41. Rh8 Nb4 42. Rc3 h5 43. Rg3+ Bg4 44. Ke3 Kf5 45. 
Ra8 a5 46. Rb8 Ke5 47. Kf2 Nd5 48. Rg1 Kd4 49. Rb1 Ke5 50. Kg3 Kd6 51. 
Rc1 Ke5 52. Kh4 Kf6 53. Rh8 Ke5 54. Kg3 Kd4, with an evaluation of 
around +2, which rises to +2.6 when you play it through but this 
position is so asymmetric and strange that I don't know what to think or 
how to go about winning it for white. Spike?

Here's the PV for Kh8 btw. I think that it just never finds Bg5, and 
it's not that it likes its position under Kh8, it just that it likes it 
less under f5 (I don't really understand Bg5. I'll leave that for 
another day)

20 ... Kh8 21. b4 d5 22. Nc3 Bc4 23. Bf4 Na6 24. Qd7 Nxb4 25. Qxb7 Qf5 
26. Re5 Qf6 27. Qe7 Nd3 28. Qxf6 gxf6 29. Re7 Nxf4 30. gxf4 Rg8 31. Rc1 
Rg4 32. Nd1 Rxf4 33. Ne3 b5 34. Rxf7 a5 35. Rc7 Rg8 36. Rxc6 Bxa2 37. 
Ra1 Bb3 38. Rc5 d4 39. Nf5 Rfg4 40. Ng3 d3 41. Bd5 Bxd5 42. Rxd5 R8g5 
43. Rxd3 a4 44. Kf1 Rf4 45. Re3 Kg7 46. Ne2 Rh4 47. Re7+.




More information about the extropy-chat mailing list