New engine: a0lite

[dkappe]

For those wanting to tinker with their own NN engine, I present a0lite. It's written in python. The core of the algorithm is just 95 lines.

https://github.com/dkappe/a0lite

It's meant for tinkering. As such, more complex developments will be in branches rather than in the master. Also check out the forks and see what others are doing.

You will need to install badgyal for the neural nets.

By default it uses MeanGirl-8 (32x4) net on CPU. Plays about ~2050 CCRL in this configuration. The shell script a0lite.sh works as a uci engine. You can find a log file, a0lite.log in the same dir as the script.

[dkappe]

I've created a branch called 'fancy' that has tree reuse and 'smart' pruning. On a 2070 with Bad Gyal it plays about 2500 CCRL.

https://github.com/dkappe/a0lite/tree/fancy

[Jamal Bubker]

thank you !

[jhellis3]

This seems like it would be a good test bed for ideas; I wonder what happens if one only increments visits for own moves...

[jhellis3]

I don't really know python syntax but I am guessing

Code: Select all

if turnfactor > 0:
	current.number_visits += 1

in the while loop would get the job done.

[cucumber]

Your code is beautifully succinct.

[dkappe]

I've created a branch called 'batching' that uses virtual loss to gather batches of positions for the gpu. https://github.com/dkappe/a0lite/tree/batching

The speed now seems dominated by the python-chess code, and my inefficient implementation.

On a 2070, the difference with batching is

Code: Select all

info depth 1 seldepth 1 score cp 22 nodes 10000 nps 1170 pv e2e4
bestmove e2e4

and without

Code: Select all

info depth 1 seldepth 1 score cp 28 nodes 10000 nps 446 pv e2e4
bestmove e2e4

Will try to fold in tree reuse in another branch and also get a 256x20 net working.

[dkappe]

Ah, you'll need to reinstall badgyal, as it has some bulk eval support added.

[dkappe]

Added in a branch that combines batching with tree reuse.

https://github.com/dkappe/a0lite/tree/b ... tree-reuse

I will add in smart pruning in a day or two. Then I will need to think about simplifying the code.

[dkappe]

Tried out a0lite with batching and tree reuse at a tc of 30 m + 5 sec. Donna 4.1 lost on time, but it was a foregone conclusion.

[pgn]
[Event "?"]
[Site "?"]
[Date "2020.03.30"]
[Round "2"]
[White "Donna"]
[Black "A0Lite"]
[Result "0-1"]
[ECO "B00"]
[GameDuration "01:03:06"]
[GameEndTime "2020-03-30T21:52:18.464 CDT"]
[GameStartTime "2020-03-30T20:49:12.074 CDT"]
[Opening "KP"]
[PlyCount "106"]
[Termination "time forfeit"]
[TimeControl "1800+5"]
[Variation "Nimzovich defense"]

1. e4 {book} Nc6 {book} 2. d4 {book} e5 {book} 3. d5 {book} Nce7 {book}
4. Nc3 {+0.99/15 149s} Ng6 {-0.35/1 73s} 5. Bc4 {+0.94/14 139s}
Bc5 {-0.12/1 70s} 6. Nf3 {+0.84/14 72s} Nf6 {-0.06/1 68s} 7. h4 {+0.52/14 124s}
d6 {-0.12/1 65s} 8. h5 {+0.66/15 23s} Nf4 {-0.15/1 63s} 9. Bxf4 {+0.42/14 113s}
exf4 {+0.02/1 60s} 10. h6 {+0.43/14 105s} O-O {+0.05/1 58s}
11. hxg7 {+0.29/14 36s} Re8 {-0.03/1 56s} 12. Bd3 {+0.33/15 95s}
Ng4 {+0.11/1 54s} 13. O-O {+0.45/15 18s} Ne5 {+1.00/1 52s}
14. Be2 {+0.20/14 88s} a6 {+1.11/1 50s} 15. Qd2 {+0.12/15 41s} Qf6 {+1.10/1 49s}
16. Nxe5 {+0.12/15 16s} dxe5 {+1.14/1 47s} 17. Na4 {+0.11/15 78s}
Bd4 {+1.19/1 45s} 18. c3 {-0.07/15 72s} Ba7 {+1.23/1 43s} 19. b4 {-0.34/16 67s}
Kxg7 {+1.33/1 42s} 20. Nc5 {+0.11/15 13s} Kh8 {+1.25/1 40s}
21. Rfb1 {-0.25/14 16s} Rg8 {+1.28/1 39s} 22. Bf3 {-0.48/15 13s}
Bg4 {+1.07/1 38s} 23. Qd1 {-1.04/16 61s} Qh4 {+1.53/1 36s}
24. Qd3 {-1.04/16 16s} b6 {+2.13/1 35s} 25. Nxa6 {-0.12/17 18s} b5 {+1.46/1 34s}
26. Nc5 {-0.35/18 11s} c6 {+0.84/1 33s} 27. dxc6 {+0.19/17 54s}
Rad8 {+1.82/1 32s} 28. Qxd8 {-6.44/18 51s} Rxd8 {+2.95/1 30s}
29. Bxg4 {-6.45/19 47s} Qxg4 {+3.48/1 30s} 30. f3 {-6.28/19 44s}
Qg3 {+3.34/1 29s} 31. Rb2 {-6.39/18 8.5s} Bxc5+ {+2.71/1 28s}
32. bxc5 {-5.14/19 41s} Rg8 {+2.26/1 27s} 33. Rf1 {-5.56/17 38s}
Rc8 {+2.40/1 26s} 34. c7 {-4.85/16 20s} Rxc7 {+2.38/1 25s}
35. Rxb5 {-5.07/15 20s} Kg7 {+2.87/1 24s} 36. a4 {-5.17/18 33s}
Qg5 {+2.69/1 23s} 37. Rd1 {-3.87/18 31s} Qg3 {+2.58/1 23s} 38. a5 {-4.19/17 29s}
h5 {+2.20/1 22s} 39. a6 {-3.67/17 27s} h4 {+2.05/1 21s} 40. Rb2 {-4.88/17 10s}
Ra7 {+2.36/1 21s} 41. Ra1 {-5.67/18 25s} h3 {+2.25/1 20s} 42. c6 {-5.72/18 24s}
Qg6 {+1.60/1 19s} 43. c7 {-6.10/16 7.0s} Qc6 {+1.88/1 19s}
44. gxh3 {-5.33/15 12s} Qxc7 {+2.70/1 18s} 45. Rba2 {-5.90/15 22s}
Qc4 {+3.29/1 18s} 46. h4 {-5.90/15 12s} Qd3 {+3.66/1 17s} 47. Kf2 {-5.82/15 20s}
Qe3+ {+3.44/1 17s} 48. Kg2 {-5.89/15 19s} Qxc3 {+3.42/1 16s}
49. Rf1 {-6.46/15 18s} Kg6 {+3.55/1 16s} 50. Rff2 {-6.42/14 17s}
Kh7 {+3.47/1 16s} 51. Kh2 {-6.68/14 8.6s} Kh6 {+3.70/1 15s}
52. Rac2 {-6.73/15 3.5s} Qe3 {+3.82/1 15s} 53. Rce2 {-6.99/15 16s}
Qa3 {+4.05/1 14s, White loses on time} 0-1
[/pgn]