Leela Zero

[Gomoto]

Leela Zero is so strong now. I am in awe of the network and its moves. We have not started to appreciate it yet accordingly.

Thanks to Google and Deepmind. Thanks to Gian-Carlo Pascutto. Thanks to the community supporting the developing of Leela Zero.

It is such a great learning tool. It feels like having Go Seigen for a teacher. In fact it feels like having an individual Go Seigen available to every go player. I feel so blessed to be able to witness this stepping stone for the development of the game of go.

I reviewed my own games and pro games with Leela Zero since a few days. Today I started reviewing AlphaGo games with Leela Zero. This was an eye opener to me how strong the Leela network has gotten already. You can feel its strength watching its moves. It is sharp like a knife, calm like the ocean and strong like the strongest of the four fundamental interactions

[sorin]

Leela Zero is so strong now. I am in awe of the network and its moves. We have not started to appreciate it yet accordingly.

Thanks to Google and Deepmind. Thanks to Gian-Carlo Pascutto. Thanks to the community supporting the developing of Leela Zero.

It is such a great learning tool. It feels like having Go Seigen for a teacher. In fact it feels like having an individual Go Seigen available to every go player. I feel so blessed to be able to witness this stepping stone for the development of the game of go.

I reviewed my own games and pro games with Leela Zero since a few days. Today I started reviewing AlphaGo games with Leela Zero. This was an eye opener to me how strong the Leela network has gotten already. You can feel its strength watching its moves. It is sharp like a knife, calm like the ocean and strong like the strongest of the four fundamental interactions

Indeed it is amazing - in my own tests, it is 2 stones stronger than Leela 11 and 1 stone stronger than AQ...

Unfortunately it is also dumb as a doorknob in some peculiar situations, such as this one (see last 2 moves): http://zero.sjeng.org/viewmatch/0b385a7 ... viewer=wgo

It just doesn't see it is in atari, I guess because of the weak white stones that are in contact with it?
This happens also with the latest version, I just checked (the version that played in the linked game is a little older I think).

[hydrogenpi7]

Leela Zero is so strong now. I am in awe of the network and its moves. We have not started to appreciate it yet accordingly.

Thanks to Google and Deepmind. Thanks to Gian-Carlo Pascutto. Thanks to the community supporting the developing of Leela Zero.

It is such a great learning tool. It feels like having Go Seigen for a teacher. In fact it feels like having an individual Go Seigen available to every go player. I feel so blessed to be able to witness this stepping stone for the development of the game of go.

I reviewed my own games and pro games with Leela Zero since a few days. Today I started reviewing AlphaGo games with Leela Zero. This was an eye opener to me how strong the Leela network has gotten already. You can feel its strength watching its moves. It is sharp like a knife, calm like the ocean and strong like the strongest of the four fundamental interactions

Indeed it is amazing - in my own tests, it is 2 stones stronger than Leela 11 and 1 stone stronger than AQ...

Unfortunately it is also dumb as a doorknob in some peculiar situations, such as this one (see last 2 moves): http://zero.sjeng.org/viewmatch/0b385a7 ... viewer=wgo

It just doesn't see it is in atari, I guess because of the weak white stones that are in contact with it?
This happens also with the latest version, I just checked (the version that played in the linked game is a little older I think).

Your link is to a self playing game right? is that not very low playouts? at high playouts LZ is already top level pro:

see

https://www.reddit.com/r/cbaduk/comment ... _per_move/

[Sneegurd]

Yeah, we're now at or close to a level where consumer hardware/software will win against any pro. Had to happen and the awesome part is that it happened quicker than anticipated.

[sorin]

Leela Zero is so strong now. I am in awe of the network and its moves. We have not started to appreciate it yet accordingly.

Thanks to Google and Deepmind. Thanks to Gian-Carlo Pascutto. Thanks to the community supporting the developing of Leela Zero.

It is such a great learning tool. It feels like having Go Seigen for a teacher. In fact it feels like having an individual Go Seigen available to every go player. I feel so blessed to be able to witness this stepping stone for the development of the game of go.

I reviewed my own games and pro games with Leela Zero since a few days. Today I started reviewing AlphaGo games with Leela Zero. This was an eye opener to me how strong the Leela network has gotten already. You can feel its strength watching its moves. It is sharp like a knife, calm like the ocean and strong like the strongest of the four fundamental interactions

Indeed it is amazing - in my own tests, it is 2 stones stronger than Leela 11 and 1 stone stronger than AQ...

Unfortunately it is also dumb as a doorknob in some peculiar situations, such as this one (see last 2 moves): http://zero.sjeng.org/viewmatch/0b385a7 ... viewer=wgo

It just doesn't see it is in atari, I guess because of the weak white stones that are in contact with it?
This happens also with the latest version, I just checked (the version that played in the linked game is a little older I think).

Your link is to a self playing game right? is that not very low playouts? at high playouts LZ is already top level pro:

see

https://www.reddit.com/r/cbaduk/comment ... _per_move/

I agree it has reached pro level, but it still has these weird blind spots.
This example indeed comes from an older version and with few playouts (1600 or so).
I tried the same position with the current version (1ccb7342) and 24000 playouts, and its preferred move is to play double-atari on the outside, it just doesn't seem to "realize" that it gets captured in one move if it does that.

[Tryss]

Or maybe it realise it, but the other options are as bad as this.

What winrate does it expect before these moves, and what sequence does it consider?

[sorin]

Or maybe it realise it, but the other options are as bad as this.

What winrate does it expect before these moves, and what sequence does it consider?

Its winrate for black is in low 40s at lower playouts - so it is definitely not in "desperate mode".
The rate drops a lot (high 20s) if ran with more playouts.

I expected it to at least consider the capturing for white in some variations, but it seems to only do so when ran with many playouts.
I re-ran with about 100K playouts, it finally defended against capturing as first choice now.

The way I interpret this is that the capturing move is not considered a "natural move" in this position, but rather seems to only be discovered during MCTS.

[moha]

Unfortunately it is also dumb as a doorknob in some peculiar situations, such as this one (see last 2 moves): http://zero.sjeng.org/viewmatch/0b385a7 ... viewer=wgo
It just doesn't see it is in atari, I guess because of the weak white stones that are in contact with it?

I think some problems related to atari are because of a slight flaw in it's policy training (the loss function), which makes it blind to certain contact moves. Unlike AGZ and (presumably) most other bots, it also expects the net to tell illegal moves from legal ones, which it gets wrong sometimes.

[drmwc]

Leela Zero is amazing. I have zero chance in a 2 stone games. In a three stone game, it's close to a fair fight.

It is also unclear to me that optimising on even games gives weights the are optimal for handicap games. However, it does seem to have resulted in a machine that is amazing at handicap as well as even.

I'll be interested in how it evolves from here - if further weight optimisation will help make it noticably stronger than now.

[Gomoto]

@drmwc ... and also an example: european 4 dan is a force to be reckoned with


What kind of hardware are you using drmwc by the way?

[drmwc]

A Ryzen 8-core processor, with an Nividia 1080ti.

[as0770]

Leela Zero is amazing. I have zero chance in a 2 stone games. In a three stone game, it's close to a fair fight.

It is also unclear to me that optimising on even games gives weights the are optimal for handicap games. However, it does seem to have resulted in a machine that is amazing at handicap as well as even.

I'll be interested in how it evolves from here - if further weight optimisation will help make it noticably stronger than now.

How are your experiences with other bots? Original Leela? AQ? Rn?

[sorin]

Other than being stronger than all other free bots, I find LZ better for two other reasons:

• 1. even with time-limits set, it plays fast when confident (as opposed to waiting the full time to play an obvious move which is very frustrating)
  2. the variations it gives are long, and even deeper moves make sense - while AQ has very short variations, and Leela 11's moves deeper in the variations are just noise, one can only trust the first few moves shown.

[hydrogenpi7]

So it had a 256*20 version of Leela Zero that just hit about 10000 ELO today.

[ez4u]

It seems the regular project moved from 128x10 to 192x15 today (March 9th). This resulted in an immediate jump of about 340 points in the rating chart. And as I write, I can see in the list of test matches what seem to be some irregular matches between a 192x15 net and both 256x20 and 128x20 nets. These two games are incomplete right now. Oh the suspense!!!

Unfortunately my old box, that has been chugging out not quite 4 games per hour (128x10) is now plodding through not quite 2 games per hour (192x15). I am not sure I want to see what happens with 256x20.