Leela Zero progress thread


#123

Okay ELF v2 has about 90% winrrate against LZ current best

https://www.lifein19x19.com/viewtopic.php?f=18&t=15718&start=60

20-game match LZ15#160 v LZ15_ELF2 at time parity.

5 minutes per game (1xGTX1080), noponder, komi 7.5

ELF2 wins 18:2 (90% , 10 wins as W, 8 wins as B)


#124

I think our testing so far has shown often 20 games is not enough to see true difference


#125

For anyone curious, LZ progress continues unabated. While the increase to 20-block nets didn’t yield an initial strength improvement, predictions that the increased network size would allow for further improvement seem to have been correct. There have now been 2 natural promotions of stronger net-to-net trained Leelas with a self-play Elo approaching 12000.

Interestingly, there’s also an incredibly strong 40-block network that’s been trained from a combo of the Elf and Leela games, which is on par with the new version of Elf for strength. While the project isn’t quite ready to switch from 20-block networks to 40-block networks just yet, in part because 40-block networks are twice as slow, game data generated by the 40-block network will be used to speed up progress in the 20-block series, in a similar manner to how Elf data is being used.


#126

Very fascinating. Always love to read updates in this thread, so please feel free to keep them coming :slight_smile:


#127

And somewhere out there is a Deepmind researcher who lets AZ play the currently strongest open source bots, just for fun…


#128

So. are there more games played by 9p players against LZ on decent hw? Are we now securely in superhuman zone with LZ?


#129

I think we’re mostly there, but sometimes ladders still ruin the day even for the 40b networks.

If it was me, I’d just rename the project Leela One and teach it to deal with ladders. ;D


#130

I am not sure if everybody realises, but Google removed ladder knowledge in Zero, just to show it can self-learn all concepts including ladders. Not because technically it is better, but because it is more beautiful / elegant in terms of self-learning.

The Master version, had the concept of what moves where ladder breakers and ladder captures. Here is the original feature information, the information what goes into Master’s network to determine what are good moves, and who is ahead:

MASTER FEATURES - NOT ZERO _ SHOWN TO COMPARE

The Leela-Zero creator, has copied google’s method as close as it could. But there is nothing (technically) restricting to create a Leela-Almost-Zero version, who is ladder aware.

So, not being able to read ladders is a design choice, not a restriction because of some technique or method.


#131

I’m not sure, I would call it “removed ladder knowledge”. There are 2 input layers telling the neural network where ladders are and how to capture/escape them.
What is the remaining knowledge, the network has to learn about ladders?


#132

that table is from Master - not zero. Zero doesn;t have that. I was trying to show the difference not confuse :stuck_out_tongue:

This is from the Zero paper, indeed no ladder knowledge


#133

Yes, we are aware.