Page 2 of 2
Re: Different Quality of Play between Current AI and AlphaGo
Posted: Thu Oct 10, 2019 3:18 am
by iopq
jlt wrote:Bill Spight wrote:
How many playouts are necessary to avoid problems with ladders, life and death, semeais, etc.?
Well, what do you mean by a problem?
Let me ask a more precise question. Are there examples of human games (pro or amateur) in which a recent version of LeelaZero with 100000 playouts chooses a wrong move because of a misread ladder, or life-and-death problem, or semeai?
If the answer is yes, then by which number should I replace 100000 so that the answer becomes "no"?
Of course. There are huge problems with complicated fights.
The number depends on how FAR the complicated fight is. If it's right now, maybe 1 million can resolve it. If the fight is in ONE MOVE it might take ten million to see and move elsewhere. If it's in three moves, then one hundred million may be necessary.
In other words, there's a horizon effect where it takes much more power to see the surprising way you can lose the game.
Re: Different Quality of Play between Current AI and AlphaGo
Posted: Thu Oct 10, 2019 3:25 am
by jlt
If you or someone has a concrete example of such a complicated fight, I would be interested (not an artificial example, but an example from a real game).
Re: Different Quality of Play between Current AI and AlphaGo
Posted: Thu Oct 10, 2019 6:00 am
by Bill Spight
jlt wrote:Let me ask a more precise question. Are there examples of human games (pro or amateur) in which a recent version of LeelaZero with 100000 playouts chooses a wrong move because of a misread ladder, or life-and-death problem, or semeai?
I don't know about LeelaZero, but Elf has misread a semeai in its commentaries, which use more than 100k playouts. See
viewtopic.php?f=15&t=16641
Edit: zermelo reported that a recent version of Leela Zero did not find the correct play in 1.5 million playouts.
viewtopic.php?p=244857#p244857
Re: Different Quality of Play between Current AI and AlphaGo
Posted: Thu Oct 10, 2019 1:29 pm
by gennan
Uberdude wrote:To be fair we don't know how strong AG0 is on typical computers, just on high end hardware. The AlphaGo teaching tool (better Master not Zero) is on 10 million playouts, maybe it too made embarrassing ladder mistakes on 3000 playouts like LZ does and DeepMind wanted to hide that bad publicity. Does anyone know how many playouts Master was getting in the online series? I seem to recall that was on a single machine with a TPU or two in sub 30 seconds.
I recall that the teaching tool took 10 minutes for 10M playouts, so with 30 seconds on the same hardware it would reach 500k playouts.