Lc0 question

[mclane]

Oh my god. Another „is LC0 stronger then Stockfish“ debate.

It’s so boring,

And how to choose the hardware. To make it possible.

Hardware. How boring.
I prefer decreasing the speed of hardware.
I don’t understand that you want faster hardware. Why ?

[Modern Times]

1. CCRL net was using 20x256 which is suitable for stronger cards only. ( I saw in discord chat that they sent 10x128 net for CCRL hardware. Once they test that net , the result may change)

No-one has ever sent us anything to test. They are more interested in the likes of TCEC and CCC and monster hardware.

However at 40/40 (40/15) the 20x256 net should be fine on the GTX1050 I'd have thought.

[kranium]

For what it's worth, CEGT also has SF on top
(and these are 1 CPU results)

http://www.cegt.net/40_4_Ratinglist/40_ ... liste.html
I suppose there's some confusing explanation...
I tend to just look at the list and use Elo ranking to make a determination.

[jp]

Stockfish plays machine chess, although developed by humans.
LC0 plays human chess, although developed by a machine.

The philosophical impact of this cannot be underestimated.

Dispute that LCZero plays human chess. It plays chess in a way that humans would like to imagine they play chess.

Yes, that's correct. It plays chess that many humans can easily make explanatory stories about, and the humans then attach human qualities to those moves, e.g. "positional", which only makes any sense for how humans think about chess, not engines, which just calculate.

But if we made a test of carefully chosen middlegame sections, probably no one would be able to distinguish which were AB and which were NN moves. Humans' best chance of distinguishing the engines is from their bad opening or endgame play, not their good middlegame play.

[kranium]

Humans' best chance of distinguishing the engines is from their bad opening or endgame play, not their good middlegame play.

Engines play the openings and endgames badly?

But, playing against the strongest humans, SF and/or LC0 would not lose a single game, and probably give up very few draws, if any...
even against the most 'opening' savvy/prepared GM.

How have some humans determined these engines are playing badly? I don't get it.
Seems a bit pretentious...or some desperate need to remain relevant.

[chrisw]

Stockfish plays machine chess, although developed by humans.
LC0 plays human chess, although developed by a machine.

The philosophical impact of this cannot be underestimated.

Dispute that LCZero plays human chess. It plays chess in a way that humans would like to imagine they play chess.

Yes, that's correct. It plays chess that many humans can easily make explanatory stories about, and the humans then attach human qualities to those moves, e.g. "positional", which only makes any sense for how humans think about chess, not engines, which just calculate.

But if we made a test of carefully chosen middlegame sections, probably no one would be able to distinguish which were AB and which were NN moves. Humans' best chance of distinguishing the engines is from their bad opening or endgame play, not their good middlegame play.

Well, that looks like a fun classification problem for a net. I don't suppose anyone ever tried it.
Data requirement - a lot of (human, AB, NNet) PGNs
process moves 15-40 (say) and save for each position
(fen_string, move, player type)
gather up, let's guess, 100 million positions for each type of player. split into a training and a test set.
train on training_set(fen_string, move). Quite possibly using a LC0 type resnet. predict type of player on a test set.
check result against actuality.

I guess, it would partly "learn" on both the type of position, and the "move played". Might work ....

[Modern Times]

For what it's worth, CEGT also has SF on top
(and these are 1 CPU results)

http://www.cegt.net/40_4_Ratinglist/40_ ... liste.html
I suppose there's some confusing explanation...
I tend to just look at the list and use Elo ranking to make a determination.

Well... 32930 is an old net, they have improved maybe +50 Elo or so since then

[kranium]

For what it's worth, CEGT also has SF on top
(and these are 1 CPU results)

http://www.cegt.net/40_4_Ratinglist/40_ ... liste.html
I suppose there's some confusing explanation...
I tend to just look at the list and use Elo ranking to make a determination.

Well... 32930 is an old net, they have improved maybe +50 Elo or so since then

for what it's worth, neither version is recent...

[jp]

But if we made a test of carefully chosen middlegame sections, probably no one would be able to distinguish which were AB and which were NN moves. Humans' best chance of distinguishing the engines is from their bad opening or endgame play, not their good middlegame play.

Well, that looks like a fun classification problem for a net. I don't suppose anyone ever tried it.
Data requirement - a lot of (human, AB, NNet) PGNs
process moves 15-40 (say) and save for each position
(fen_string, move, player type)
gather up, let's guess, 100 million positions for each type of player. split into a training and a test set.
train on training_set(fen_string, move). Quite possibly using a LC0 type resnet. predict type of player on a test set.
check result against actuality.

I guess, it would partly "learn" on both the type of position, and the "move played". Might work ....

Yeah, I was thinking of something simpler, just whether humans given a bunch of game fragments can guess whether the players are AB or NN engines. I think the humans would fail.

[ThatsIt]

For what it's worth, CEGT also has SF on top
(and these are 1 CPU results)

http://www.cegt.net/40_4_Ratinglist/40_ ... liste.html
I suppose there's some confusing explanation...
I tend to just look at the list and use Elo ranking to make a determination.

Well... 32930 is an old net, they have improved maybe +50 Elo or so since then

for what it's worth, neither version is recent...

Hi !

http://cegt.forumieren.com/t1152-testin ... da-nw42668

The next update for our 40/4 ratinglist will be online next week, perhaps on Thursday.

Best wishes,
G.S.
(CEGT team)