| |
BGonline.org Forums
Blue's 22 play before the ho-hum 61
Posted By: Timothy Chow In Response To: Blue's 22 play before the ho-hum 61 (Nack Ballard)
Date: Sunday, 22 June 2014, at 6:53 p.m.
Nack wrote:
Is it possible that bots have a built-in bias towards slightly unsound shifts and multiple slots (that modern players try to mimic) in preparation for a shot that somehow don't correspond to real equity? (For example, it might sometimes make less bold plays than it should for the opponent that don't take full advantage of the player's temporarily compromised board.)
My opinion is that the bot is a donkey when it comes to these sorts of plays. Especially on 3-ply or less, it will make all kinds of nutty-looking moves, e.g., dumping blots deep in the board while leaving gaps on higher points. This sort of strategy can then create some strange decisions for the opponent, such as whether to make a spectacular bold play that creates a slightly smoother distribution, counting on the assumption that the opponent won't hit due to the blotty board. Sometimes 3-ply will get such decisions spectacularly wrong. While such spectacular mistakes are rare, I think they are common enough that there is essentially a certain amount random noise in a rollout. It's not uncommon that I will analyze one of my matches and find that a perfectly reasonable move that I made didn't even make it through XG's move filter, but when I forced it through, it was 4-ply's top choice. I suspect that there's just not enough equity at stake for the neural net to tune itself properly during training.
Anyway, bottom line is that I would ignore equity differences of 0.002 in such positions. Just for comparison I checked with GNU. On 2-ply, its top 15 candidates were the same as in your rollout, but in a totally different order. When I ran them through 4-ply, 13/11 10/8 6/2 came out on top, with 13/11 6/2/ 4/2 in 3rd place, 0.003 back. Doesn't prove anything, of course, but it reinforces my belief that it's random noise.
1. Cubeful 4-ply 13/11 10/8 6/2 Eq.: -0.533
0.335 0.030 0.000 - 0.665 0.026 0.000
2. Cubeful 4-ply 13/9 6/2 Eq.: -0.535 ( -0.002)
0.334 0.030 0.000 - 0.666 0.026 0.000
3. Cubeful 4-ply 13/11 6/2 4/2 Eq.: -0.536 ( -0.003)
0.334 0.030 0.000 - 0.666 0.028 0.000
4. Cubeful 4-ply 10/2 Eq.: -0.536 ( -0.003)
0.334 0.030 0.000 - 0.666 0.026 0.000
5. Cubeful 4-ply 13/7 10/8 Eq.: -0.537 ( -0.004)
0.333 0.030 0.000 - 0.667 0.025 0.000
6. Cubeful 4-ply 6/2(2) Eq.: -0.547 ( -0.014)
0.333 0.030 0.000 - 0.667 0.038 0.001
7. Cubeful 4-ply 10/8 6/2 4/2 Eq.: -0.548 ( -0.016)
0.330 0.029 0.000 - 0.670 0.031 0.000
8. Cubeful 4-ply 13/7 6/4 Eq.: -0.549 ( -0.016)
0.329 0.029 0.000 - 0.671 0.026 0.000
9. Cubeful 4-ply 13/5 Eq.: -0.549 ( -0.017)
0.328 0.030 0.000 - 0.672 0.026 0.000
10. Cubeful 4-ply 13/7 4/2 Eq.: -0.550 ( -0.017)
0.331 0.029 0.000 - 0.669 0.035 0.001
11. Cubeful 4-ply 13/11 10/4 Eq.: -0.551 ( -0.018)
0.327 0.029 0.000 - 0.673 0.026 0.000
12. Cubeful 4-ply 13/9 10/8 6/4 Eq.: -0.552 ( -0.020)
0.327 0.029 0.000 - 0.673 0.026 0.000
13. Cubeful 4-ply 13/9 10/8 4/2 Eq.: -0.554 ( -0.021)
0.330 0.029 0.000 - 0.670 0.038 0.001
14. Cubeful 4-ply 13/9 10/6 Eq.: -0.554 ( -0.022)
0.326 0.029 0.000 - 0.674 0.026 0.000
15. Cubeful 4-ply 13/9 4/2(2) Eq.: -0.562 ( -0.030)
0.323 0.029 0.000 - 0.677 0.028 0.000
| |
BGonline.org Forums is maintained by Stick with WebBBS 5.12.