Did anyone notice the 100 Elo explosion in LCZero?

[Dann Corbit]

http://lczero.org/

I am keen for an explanation.

[mar]

bigger networks

[Guenther]

bigger networks

that was easy ;-)

[fierz]

I am playing it on a more or less regular basis in the easy mode (= as far as I can see just playing the move the policy network recommends without any search) - I am very impressed by what I see there. I have trouble beating it despite being a strong chess player, and I definitely do some calculations

I'm very curious to see how far this goes!

[Daniel Shawul]

wrong setup

Bigger network but tested with fixed nodes per moves

[mar]

wrong setup

Bigger network but tested with fixed nodes per moves

Yes, you're right, +100 elo but 2x slower. Also self-play inflates about 2x.

But I'm running a test now and so far it seems much stronger than the last one I tested (217) even if it runs at half speed.
Still too early to draw conclusions but ID 217 scored 35% against 2CPU Cheng in 100 games,
now it's 47% but only 34 games, still it seems to work very well.

I'm really excited about this project, it will soon pass 3k CCRL on consumer hardware, looking forward to see next TCEC.

[Werner]

Yes,
the nps are about 50% lower under the same conditions.
So maybe not much difference in strenght.

[Albert Silver]

Yes,
the nps are about 50% lower under the same conditions.
So maybe not much difference in strenght.

Yes, I think the self-play is not done in timed games, but a fixed nodes or depth or whatever, so 100 Elo in self-play would mean it is not taking into account the slower speed that will reduce strength in timed games.

[jkiliani]

Yes,
the nps are about 50% lower under the same conditions.
So maybe not much difference in strenght.

Yes, I think the self-play is not done in timed games, but a fixed nodes or depth or whatever, so 100 Elo in self-play would mean it is not taking into account the slower speed that will reduce strength in timed games.

Correct, all self-play matches are done at fixed 800 nodes. The strength at fixed time controls is roughly similar right now between the new 192x15 net and the previous 128x10 network. Still, this makes sense for self-play Elo, since the gain will mean much higher quality of the training data from now on. Also, the new network size has a lot of growth potential and should be able to gain a couple hundred more Elo over the coming weeks.

[Albert Silver]

Yes,
the nps are about 50% lower under the same conditions.
So maybe not much difference in strenght.

Yes, I think the self-play is not done in timed games, but a fixed nodes or depth or whatever, so 100 Elo in self-play would mean it is not taking into account the slower speed that will reduce strength in timed games.

Correct, all self-play matches are done at fixed 800 nodes. The strength at fixed time controls is roughly similar right now between the new 192x15 net and the previous 128x10 network. Still, this makes sense for self-play Elo, since the gain will mean much higher quality of the training data from now on. Also, the new network size has a lot of growth potential and should be able to gain a couple hundred more Elo over the coming weeks.

I wasn't complaining, just noting. The new 227 is about similar strength to the last editions of older NN, though a bit more boring to watch. Still, I expect that will change soon as it begins to take shape. Looking forward to it.