Looking forward to leela-0.9.2-4c performance in cgos.
Maybe we have to wait until v.1.0 for public release.

There was an interesting game today on KGS between Ilya Shikshin and Leela.
In fact Leela seemed to be ahead and in control during the whole game but it miserably failed a few moves before the end and resigned there.

The position that interests us really begins at move 188 but the climax is move 191. D17

When analysing the game with my home version 0.9.0, I noticed that in every variation displayed by Leela before this move, then D17 was systematically assumed to be answered by E17, so that an eventual ko at Q19 could be handled.

As we can see at move 191. from this screenshot, Leela only considers 3 potential answers and still estimates his chances of success very high :



But Ilya played next 192. M19 and Leela realized this was hopeless and winning % droped from 70% to less than 10% !

I ran the position in various engines, Zen 9.7 is also overseeing this move for 30s before it appears effectively in its candidates move list and almost immediately realizes that it is a dreadful threat. Surprisingly the 'allmighty' Fuego plays M19 blindly and instantly.

I was wondering how such powerful engines not only cannot see this immediately but do not even consider a move as important as an atari onto a big group ?! I let Leela run for more than 5 minutes, and it never considered candidates moves other than the 3 above.

For engines to oversee a threat because it is behinf the search horizon is a thing, but to not even assess a potential atari seems quite a big flaw in the design.

What do you think ? Are you able to reproduce this on other versions of Leela ?

Hi,

I just followed your link from my other thread.
In fact, even the venerable GnuGo would play M19

Same issue with Leela 6.2

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!

After viewed the games with Haylee, I observed similar behaviors of Leela. It has a strong tendency to follow wherever the opponent plays locally. Unlike zen, it doesn't play save (defensively) even when it is far ahead. So after battles around the board from here to there, it will forget what the most important move is globally.

Leela 0.10.0 is available for download. (I kinda hope they'd name this 1.0 instead)

Does anybody know how strong Leela is on older computers? I have Xeon X5460 3,16GHz, 16GB RAM, GF GTX760 and today I won against 1 dan twice. (Leela 0.9.0, rated game 6 handi, 19x19 I've tried GPU and CPU also). I'm 11 kyu EGF by the way. Engine settings are all ticked.

Thank you.

I think the 1.0 will be commercial. So I am fine with 0.10

Nice to see a Linux version of the UI!

Leela 0.10.0 is SUPERB

I dunno, it looks weak, just played a game, look for yourself, White (Leela) needs to be able to bring a game like that home.

I wanted to check if and when Leela resigns, so I played useless moves as well. Again: From this game I'd give Leela 1k or 2k strength. No dan player would lose a game like that after capturing a bulk of black stones. Such a player would just simplify the game and take it home.

What is your hardware, cpu and gpu and did you use gpu version.
Also did you put it on maximum simulations?

There is some interesting discussion of the tactical issues with Leela 0.10 on the computer-go mailing list, including the author, in the thread "mini-max with Policy and Value network" at http://computer-go.org/pipermail/comput ... hread.html (in particular, start with Marc Landgraf's posts).

If you are playing 'rated' games with leela, I think it adjusts it's playing strength to your assumed rating. Then it's no wonder it does funny things.

To make it play strongly, select 'new game', unimited playouts, and then adjust the clock in the menu so that it has enough time. The window will show the number of nodes (in the MCTS tree? I'm not sure). I don't know how many nodes it needs it needs to be e.g. 6d, but that would be interesting to know.

I gave it 30 min on my 4-core laptop and lost, as euro 1 dan, taking 4 handi. I think the number of nodes it showed were typically upwards of 100k.

I'm still waiting for Leela on iOS.

Is anyone else having trouble getting the openCL version to run properly on macOS? The CPU-only version is running five to ten times faster than the openCL version on my machine. I've verified that Leela's detected my discrete GPU (The "About" window says "OpenCL: Geforce GT 650M @ 900 MHz), and I've verified that my drivers are up to date.

This is 0.10.0 running on macOS Sierra 10.12.5.

Even the non-GPU version is incredible, hitting 100k nodes in about 25 seconds on my extremely dated machine. I'm just very keen to see what Leela can do once she's tapped into all my hardware.

EDIT: Just noticed that this is only the case when "use neural network" is checked. When neither Leela nor Leela-openCL are using neural networks, Leela-openCL is much faster. But when they're both using neural networks, non-openCL Leela is way faster.

Hi,
I've used Leela to make my first game analyze.

The software is extremely useful ! It showed me many vital points that I didn't see.

I'm not yet comfortable with the exploration of variations, since there is no tree view, like the one in CGoban.

For the time being, if I want to try a variation from a recorded game while the analysis is running, I click "Remember this position on the stack", then play some moves, using the "take back one move" button to change variations. Leela remembers all the MC calculations that it already made.
When I'm done, I click the "revert to stored position" button, and I can go on with the game.

There is a little bug. If I open a game and start the analysis, when I click "Remember this position on the stack" to remember to position and try some variations, when I click on the "revert to stored position" button, the analysis engine goes on analyzing the older position, that is no longer displayed. I have to go back one move and go forward one move to force the reload of the current position into the analysis tool.

If the variation that I explored is in the list, then its evaluation takes into account the analyzes done during the variation explorations. That's the way I force Leela to evaluate some moves that it doesn't.

For example if there are 3 favorite moves with 10000 iterations each, then another one with a higher winning probability, but only 10 iterations, and just one follow up, this 4th move can be ignored for minutes.
So I click "+", play the move, let the analyze explore 10000 iterations, then I "revert to stored position", "take back one move", "go forward one move" (to reload the current position).
Now, the 4th move has 10000 iterations done, and the subsequent variations are stored too.

Near the end of the game, I also noticed some oddities.



Here, move 219 is a blunder where a whole group changes from "alive" to "ko". The neural network (red curve) seems much better at finding vital points than the Monte-Carlo (blue curve).

On the opposite, it seems to me that the MC curve is more reliable at estimating the winner during the late endgame, when only small endgame move remains, than the neural network, but in this situation, Leela didn't see the possibility of seki, so I can't be sure yet.

Which brings me to the topic of scoring.
At the end of the game, how can I inform Leela that a position is seki ?
If I'm not mistaken, the bottom left group is the attached position is in dual life, but Leela scores it as dead, and I can't change that

Leela 0.15.0 started playing on CGOS, currently winning 15 out of 15 games!

http://www.yss-aya.com/cgos/19x19/cross ... .15.0.html

I hope for a self-play-mode in the next versions.

Leela's 20 game win streak has finally been ended by Zen, after it lost a ko. Leela seems to be much stronger at life and death compared to version 0.10.0.

What is the version currently playing on Tygem? It seems to be quite annoying sometimes that it doesn't resign even when it is far behind.