Page 2 of 2
Re: AlphaZero vs. the human brain
Posted: Sun Oct 20, 2019 11:57 am
by Laskos
shrapnel wrote: ↑Sun Oct 20, 2019 10:23 am
corres wrote: ↑Sat Oct 19, 2019 7:47 pmAlone the bigger size is not a guarantee for better play.
Coupled with a strong GPU like 2080 or 2080 Ti or even two of them together, it IS.
Master Om gave you the link, try them, the last one is best I think.
Even the 32 Cores Players will be unable to beat you, if you use this Network, at least not easily, given at least a reasonably good Book and a 2080 or 2080 Ti.
No, there is a sweet spot where larger is not better. On 2080 Ti, 320x24 is better than 256x20 only in blitz and longer time control, not in bullet or faster. Larger nets do scale better with hardware and time control. I guess this 384x30 net can be the strongest only at some VERY long time controls, like 1 hour/move on an RTX GPU. Also, the strength is not well defined. I seem to have observed comparing best 256x20 and 320x24 nets, that in a direct match, larger net beats the smaller starting with some 5s/move, but against Stockfish it can perform worse, and performs better only at LTC. I am checking right now this. The problem with larger 320x24 nets with RTX GPU for Lc0 is that they tactically catch up only with very much time allotted, and at blitz TC are vulnerable against Stockfish tactical shots. Larger nets are positionally indeed better, but they need much more time for tactics.
Here are results from test suites, the best nets of their size, RTX 2070 GPU:
10s/position
256x20 T40B.4-200 net:
positional: score=157/200 [averages on correct positions: depth=4.6 time=0.87 nodes=18890]
tactical: score=121/199 [averages on correct positions: depth=7.1 time=0.95 nodes=13717]
320x24 J13B.4-150 net:
positional: score=160/200 [averages on correct positions: depth=4.9 time=1.11 nodes=9742]
tactical: score=108/199 [averages on correct positions: depth=6.3 time=1.35 nodes=7927]
Now a blitz gauntlet against Sockfish compares them.
Re: AlphaZero vs. the human brain
Posted: Sun Oct 20, 2019 7:00 pm
by Master Om
Laskos wrote: ↑Sun Oct 20, 2019 11:57 am
shrapnel wrote: ↑Sun Oct 20, 2019 10:23 am
corres wrote: ↑Sat Oct 19, 2019 7:47 pmAlone the bigger size is not a guarantee for better play.
Coupled with a strong GPU like 2080 or 2080 Ti or even two of them together, it IS.
Master Om gave you the link, try them, the last one is best I think.
Even the 32 Cores Players will be unable to beat you, if you use this Network, at least not easily, given at least a reasonably good Book and a 2080 or 2080 Ti.
No, there is a sweet spot where larger is not better. On 2080 Ti, 320x24 is better than 256x20 only in blitz and longer time control, not in bullet or faster. Larger nets do scale better with hardware and time control. I guess this 384x30 net can be the strongest only at some VERY long time controls, like 1 hour/move on an RTX GPU. Also, the strength is not well defined. I seem to have observed comparing best 256x20 and 320x24 nets, that in a direct match, larger net beats the smaller starting with some 5s/move, but against Stockfish it can perform worse, and performs better only at LTC. I am checking right now this. The problem with larger 320x24 nets with RTX GPU for Lc0 is that they tactically catch up only with very much time allotted, and at blitz TC are vulnerable against Stockfish tactical shots. Larger nets are positionally indeed better, but they need much more time for tactics.
Here are results from test suites, the best nets of their size, RTX 2070 GPU:
10s/position
256x20 T40B.4-200 net:
positional: score=157/200 [averages on correct positions: depth=4.6 time=0.87 nodes=18890]
tactical: score=121/199 [averages on correct positions: depth=7.1 time=0.95 nodes=13717]
320x24 J13B.4-150 net:
positional: score=160/200 [averages on correct positions: depth=4.9 time=1.11 nodes=9742]
tactical: score=108/199 [averages on correct positions: depth=6.3 time=1.35 nodes=7927]
Now a blitz gauntlet against Sockfish compares them.
On a slow GPU like GTX 1050 Ti which net will be good where amount of time it runs is ample?
Re: AlphaZero vs. the human brain
Posted: Sun Oct 20, 2019 8:54 pm
by Laskos
Master Om wrote: ↑Sun Oct 20, 2019 7:00 pm
Laskos wrote: ↑Sun Oct 20, 2019 11:57 am
shrapnel wrote: ↑Sun Oct 20, 2019 10:23 am
corres wrote: ↑Sat Oct 19, 2019 7:47 pmAlone the bigger size is not a guarantee for better play.
Coupled with a strong GPU like 2080 or 2080 Ti or even two of them together, it IS.
Master Om gave you the link, try them, the last one is best I think.
Even the 32 Cores Players will be unable to beat you, if you use this Network, at least not easily, given at least a reasonably good Book and a 2080 or 2080 Ti.
No, there is a sweet spot where larger is not better. On 2080 Ti, 320x24 is better than 256x20 only in blitz and longer time control, not in bullet or faster. Larger nets do scale better with hardware and time control. I guess this 384x30 net can be the strongest only at some VERY long time controls, like 1 hour/move on an RTX GPU. Also, the strength is not well defined. I seem to have observed comparing best 256x20 and 320x24 nets, that in a direct match, larger net beats the smaller starting with some 5s/move, but against Stockfish it can perform worse, and performs better only at LTC. I am checking right now this. The problem with larger 320x24 nets with RTX GPU for Lc0 is that they tactically catch up only with very much time allotted, and at blitz TC are vulnerable against Stockfish tactical shots. Larger nets are positionally indeed better, but they need much more time for tactics.
Here are results from test suites, the best nets of their size, RTX 2070 GPU:
10s/position
256x20 T40B.4-200 net:
positional: score=157/200 [averages on correct positions: depth=4.6 time=0.87 nodes=18890]
tactical: score=121/199 [averages on correct positions: depth=7.1 time=0.95 nodes=13717]
320x24 J13B.4-150 net:
positional: score=160/200 [averages on correct positions: depth=4.9 time=1.11 nodes=9742]
tactical: score=108/199 [averages on correct positions: depth=6.3 time=1.35 nodes=7927]
Now a blitz gauntlet against Sockfish compares them.
On a slow GPU like GTX 1050 Ti which net will be good where amount of time it runs is ample?
Probably 256x20 net (B4 run of JHorthos) up to 10 minutes runs. Longer runs than that, probably 320x24 net (again B4 of JHorthos).
Re: AlphaZero vs. the human brain
Posted: Mon Oct 21, 2019 2:35 pm
by Master Om
Laskos wrote: ↑Sun Oct 20, 2019 8:54 pm
Master Om wrote: ↑Sun Oct 20, 2019 7:00 pm
Laskos wrote: ↑Sun Oct 20, 2019 11:57 am
shrapnel wrote: ↑Sun Oct 20, 2019 10:23 am
corres wrote: ↑Sat Oct 19, 2019 7:47 pmAlone the bigger size is not a guarantee for better play.
Coupled with a strong GPU like 2080 or 2080 Ti or even two of them together, it IS.
Master Om gave you the link, try them, the last one is best I think.
Even the 32 Cores Players will be unable to beat you, if you use this Network, at least not easily, given at least a reasonably good Book and a 2080 or 2080 Ti.
No, there is a sweet spot where larger is not better. On 2080 Ti, 320x24 is better than 256x20 only in blitz and longer time control, not in bullet or faster. Larger nets do scale better with hardware and time control. I guess this 384x30 net can be the strongest only at some VERY long time controls, like 1 hour/move on an RTX GPU. Also, the strength is not well defined. I seem to have observed comparing best 256x20 and 320x24 nets, that in a direct match, larger net beats the smaller starting with some 5s/move, but against Stockfish it can perform worse, and performs better only at LTC. I am checking right now this. The problem with larger 320x24 nets with RTX GPU for Lc0 is that they tactically catch up only with very much time allotted, and at blitz TC are vulnerable against Stockfish tactical shots. Larger nets are positionally indeed better, but they need much more time for tactics.
Here are results from test suites, the best nets of their size, RTX 2070 GPU:
10s/position
256x20 T40B.4-200 net:
positional: score=157/200 [averages on correct positions: depth=4.6 time=0.87 nodes=18890]
tactical: score=121/199 [averages on correct positions: depth=7.1 time=0.95 nodes=13717]
320x24 J13B.4-150 net:
positional: score=160/200 [averages on correct positions: depth=4.9 time=1.11 nodes=9742]
tactical: score=108/199 [averages on correct positions: depth=6.3 time=1.35 nodes=7927]
Now a blitz gauntlet against Sockfish compares them.
On a slow GPU like GTX 1050 Ti which net will be good where amount of time it runs is ample?
Probably 256x20 net (B4 run of JHorthos) up to 10 minutes runs. Longer runs than that, probably 320x24 net (again B4 of JHorthos).
Okay thanks.
Re: AlphaZero vs. the human brain
Posted: Mon Oct 21, 2019 5:46 pm
by Jouni
Has human GMs so far learned something new from NN engines!?
Re: AlphaZero vs. the human brain
Posted: Mon Oct 21, 2019 10:59 pm
by Ovyron
Jouni wrote: ↑Mon Oct 21, 2019 5:46 pm
Has human GMs so far learned something new from NN engines!?
Yes, the concept of "thorn pawn", where you leave an uncapturable pawn on your sixth rank, often near the opponent's king, to block and cripple the opponent, was used by Leela early on to beat the strongest A/B engines, in the times where wins by Leela were rare and noteworthy (and it was playing in a style "out of this world", at least on those games it won.)
This concept was picked up by human GMs to beat their opponents much more frequently since then (if even the strongest chess engines had a problem with the concept, it could work against humans, and it did.) Not that the concept was new, but it became mainstream thanks to Youtuber Kingscrusher, who also helped make Leela herself mainstream (and the proof is, that many people think these kind of pawns are called "fawn pawns." Why? Because of Kingscrusher's thick accent where "thorn" sounds like "fawn"!)
Since then you can even buy
thorn pawn t-shirts.
Other than this and some opening novelties Leela discovered that were played by human GMs later on, I think this is it.