Just run AutoGTP to help building NN for this Go-AI.

zero.sjeng.org

100 contributors and counting

Many thanks to Gian-Carlo Pascutto

Is this version ready to go live formally? Or is it still a test version that improves an network for just testing?
I am ready to notify to a club, but I am not sure how I should describe in a notification.

Just downloaded. So far it plays like crazily random.

The main author, Gian-Carlo Pascutto, has not made an official "it's released now" announcement. I characterize the situation as a soft release where tech minded people can start now. I have downloaded the source, compiled it, and it's running now generating and uploading games.

But I will wait for Gian-Carlo to make an official announcement before telling my friends about it.

After that in just silently crashes.

The code (in main() of autogtp) unfortunately checks for a bunch of conditions and silently exits if they are not all met.

For me, this happened when I forgot to copy leelaz into the current directory. I'm sure dozens of other things could cause it as well, though the log indicates that the problem happened in run_one_game(), which makes it possible that starting Leela is your problem as well.

Probably you're missing OpenCL drivers, or your GPU doesn't support OpenCL:

https://www.reddit.com/r/baduk/comments ... without_a/

Now 1000 contributors and more then 50.000 games played

At first it was frozen middle in the second game after one game played. Now I start command prompt first, then go to the leela zero folder and run autogtp. And it works.

Required client version: 2 (OK)
curl.exe -s -O -J -w %{filename_effective} http://zero.sjeng.org/best-network
Curl filename: fe3f6afd23aab715be6743fd42c378ddfcd9b049ebe305478c335b8a9401ac13.gz
Net filename: fe3f6afd23aab715be6743fd42c378ddfcd9b049ebe305478c335b8a9401ac13
Engine has started.
Infinite thinking time set.
1 (K10) 2 (R9) 3 (pass) 4 (pass) Game has ended.
Score: W+7.5
Winner: white
Writing 37f229673da44cdcb452f128f64f0624.sgf
Dumping 37f229673da44cdcb452f128f64f0624.txt
Stopping engine.
curl.exe -F networkhash=fe3f6afd23aab715be6743fd42c378ddfcd9b049ebe305478c335b8a9401ac13 -F clientversion=2 -F sgf=@37f229673da44cdcb452f128f64f0624.sgf.gz -F trainingdata=@37f229673da44cdcb452f128f64f0624.txt.0.gz http://zero.sjeng.org/submit
Game data 3a30f3b92d5f9327b8f56127ccf3e5338c99dcc4bd00beac73ef5ea97028cbed stored in database
1 games played.
Best network hash: fe3f6afd23aab715be6743fd42c378ddfcd9b049ebe305478c335b8a9401ac13
Required client version: 2 (OK)
Already downloaded network.
Engine has started.
Infinite thinking time set.

However, middle in second game it crashed again.

1 (N1) 2 (C6) 3 (K14) 4 (M5) 5 (C7) 6 (D2) 7 (A14) 8 (Q2) 9 (A9) 10 (F18) 11 (A4) 12 (L4) 13 (P18) 14 (Q13) 15 (F16) 16 (T9) 17 (H12) 18 (G5) 19 (O1) 20 (K13) 21 (R18) 22 (P3) 23 (M9) 24 (T6) 2

Are you on Windows 10? Did you click and/or highlight text in the output window? Try hitting enter:

https://github.com/gcp/leela-zero/issues/71

OK you are right. Now it seems to be normal. Well some games are actually not "normal", they are either too short because both sides played "PASS", or too long (almost 500 moves).

Yes, it definitely was the reason. On a more newer machine it works fine.

Where can I fing the game records? It looks like SGF file are being deleted after submitting to the database.

I saw in the Git repository that an option was added to keep SGF files, but currently it is not released as a binary.

Because I had connection problems, I could copy a SGF file before it was deleted. The game is some hours old. Don't expect too much:

I wonder about the result. B+22.5 does not seem to be correct. Leela 0.11.0 shows B+217.5, but does not mark dead stones correctly, but maybe this is a GUI issue. Zenith Go has similar problems.

And you can find a file all.sgf.xz here:
https://sjeng.org/zero/
I didn't check it, but the name suggest that … well, a big variant tree of random moves?

Edit: It is not a GUI issue. Leela always shows different scores.

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?

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.

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.

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.

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!