Leela Zero + AutoGTP for Windows released

[splee99]

I don't know much of the neural network, but from the game posted I had to say, WOW! Zero knowledge is really scary. Don't we start with the basic rules of Go? Does the program know that a group with one eye is dead?

[yoyoma]

The games are scored using area rules and without removing dead stones. The bot must learn to actually capture all dead stones. So in your example, since the large white group on the left is still on the board, it's counted as points for white.

[Bas]

Thank you for the explanation, yoyoma. Than the result sounds reasonable.
It is interesting that the NN still has no sense of life and death.
But it tries to keep stones together – at least it places stones near to other friendly stones.

[yoyoma]

I made a spreadsheet tracking LeelaZero progress and a few other things:
https://docs.google.com/spreadsheets/d/ ... =806323235



This is a graph showing the Elo rating vs number of games. The tiny blue line in the lower left is where LeelaZero is, at around -3336 Elo. The short red line in the upper right is a target showing where the AlphaGoZero 20 block got after 5 million games. Google was able to do that in 72 hours. I'm measuring in games because time is dependent on how many machines LeelaZero gets.

Speaking of time and machines, I have another tab in the sheet with a rough calculation of the money required to use AWM p2.xlarge services. It would cost somewhere around $1.2 million!

Finally there is a tab showing historical game rate. I have a script scraping the status page once an hour, but I haven't updated it for awhile.

Let me know if you have feedback on the presentation or requests for more.

[sorin]

I don't know much of the neural network, but from the game posted I had to say, WOW! Zero knowledge is really scary. Don't we start with the basic rules of Go? Does the program know that a group with one eye is dead?

It can be confusing at first, but if you think of the difference between rules and tactics in detail: filling one's liberties in a silly way is not against the rules in any way, it is just very bad tactics.
So even if the program knows the rules, it fills its own liberties because it's lacking basic tactical knowledge. For now at least, hopefully it will learn that soon!

[Uberdude]

I don't know much of the neural network, but from the game posted I had to say, WOW! Zero knowledge is really scary. Don't we start with the basic rules of Go? Does the program know that a group with one eye is dead?

A group with one eye being dead is not a rule of Go (and stated so simply is false e.g seki counterexample). It is a consequence of the rule of capture and rule of alternating play. Most human beginners are capable of logical reasoning and abstract thought so can work that out fairly quickly (or are told it) but these go playing programs aren't AIs in that sense, they are learning patterns from repeated trial and error.

[Pio2001]

But why does the bot suddenly pass ?
Is passing at a given stage of the game something that the NN learns by itself, or is there external knowledge that tells it when it should pass ?
If not, then it has already learned that passing is not a good strategy in the beginning of the game.

[yoyoma]

But why does the bot suddenly pass ?
Is passing at a given stage of the game something that the NN learns by itself, or is there external knowledge that tells it when it should pass ?
If not, then it has already learned that passing is not a good strategy in the beginning of the game.

It's not told when to pass. But if there are two passes in a row, by rule the game is over, and the Network is told who won. Because of this the Network can discover when it's a good idea to pass (my opponent passed and I have more stones+territory).

[kityanhem]

Leela Zero 0.6 + AutoGTP v4

https://github.com/gcp/leela-zero/releases/tag/v0.6
The current status of L-Zero: http://zero.sjeng.org
Is much faster and with 288 clients we will have 2383 games in an hour (it mean 57k games/day).
Before we have about 500 clients.
If we reach 200k games/day (2000 clients) we can make L-Zero get strong as A-Zero (5185 elo) in 150 days.
A-Zero plays 750k games/day.
And the more clients, the more faster.
The author want to build a "Zero" reaching 6000 elo. And it is free.
We will have Zero for each level, too.
Need help!


~~
Zero for everyone
Thanks for your attention!
For any infomation you need: https://github.com/gcp/leela-zero
Somewhere discussion: https://www.reddit.com/r/cbaduk/

[Pippen]

Is it possible to write the author? I want to ask him for a self-play feature, so that Leela can play vs. Leela without using other clients and stuff. Does that mean a lot of work (question to guys who know a bit about programming)?

[kityanhem]

Is it possible to write the author? I want to ask him for a self-play feature, so that Leela can play vs. Leela without using other clients and stuff. Does that mean a lot of work (question to guys who know a bit about programming)?

you can ask him in here: https://github.com/gcp/leela-zero

[Vargo]

If you have a decent graphic card, you can run simultaneously 3-4 instances of autogtp, (I have 4 autogtp windows on my computer, without any problem) it seems to be the easiest way to have a snowball effect.

[kityanhem]

Leela Zero 0.6 + AutoGTP v4
https://github.com/gcp/leela-zero/releases/tag/v0.6

Print timing information.
Fixed a bug causing level in the endgame to be lower than intended.
Fixed a bug that could cause a hang in positions with very few legal moves.
Performance improvements.

~~
For any infomation you need: https://github.com/gcp/leela-zero
Somewhere discussion: https://www.reddit.com/r/cbaduk/

[Sneegurd]

I wonder if in chess, with the same concept, we could get an engine which destroys the currently best engines (Komodo, Houdini, Stockfish) easily. Or, we'll be close to perfect play (and only see draws?)...

[as0770]

I wonder if in chess, with the same concept, we could get an engine which destroys the currently best engines (Komodo, Houdini, Stockfish) easily. Or, we'll be close to perfect play (and only see draws?)...

In chess tactics are much more important than in go, so I guess it will be difficult to beat the best conventional chess engines.

One of the AlphaGo team created a chess engine based on neural networks: https://bitbucket.org/waterreaction/gir ... downloads/

Here is some information about it: http://talkchess.com/forum/viewtopic.ph ... ht=giraffe