Leela has a new version

Garf · Post by **Garf** » Fri Feb 24, 2017 11:54 am

Pippen wrote:I rechecked and here is what Leela says in detail: D2(best move): Win% 42, MC win% 51.7, Net win% 31.8, Net prob% 29. What is the meaning of all these numbers anyway? I always thought that win% matters, the rest not so much.

It's explained on Leela's webpage.

FAQ: "What do the columns in the Analysis Window mean?"

goame · Post by **goame** » Fri Feb 24, 2017 1:18 pm

Garf wrote:
goame wrote:Is it possiple to use 4x nvidia geforce gtx 1080 ti with Leela?

No, because you cannot buy those yet

Is it possiple to use 4x nvidia geforce gtx 1080 with Leela?
And how big is the benefit from 1 to 2 to 4 gpus? +1, +2, +4 Dan

Pippen · Post by **Pippen** » Fri Feb 24, 2017 5:20 pm

Simulations: the number of Monte Carlo playouts used to investigate the move. More simulations means more confidence in the winrate as the move has been investigated deeper.
Win%: this represents Leela's best guess as to how likely it is for the player to move to win the game.
MC Win%: the likelyhood that the player to move wins the game, as determined by randomized Monte Carlo playouts from the current position. This is a factor in the Win% calculation.
Net Win%: the likelyhood that the player to move wins the game, as determined by analyzing the position with a neural network. This is a factor in the Win% calculation.
Net Prob%: the probability that a pro player would play this move, as estimated by the Neural Network.
PV: the principal variation. The sequence of suggested moves for both players that Leela believes is optimal.

How is Net win % and Net Prob win % calculated? What does Leela do (in simple words)? For example, in MC win%, Leela is just playing many random games and e.g. if Black wins 6000 out of 10.000 random games it will give Black 60% winrate.

Schachus · Post by **Schachus** » Sat Feb 25, 2017 1:49 am

Its a bit like human intuition. That net looked at millions of positions and who was winning in the end, and learned to develop an intuition saying "this looks like black is going to win" or "this looks like its still even". This intuition thinks black is likely to win from here(probably the net was trained with pro games or strong amateur games(KGS? I know Alphago started out with KGS games), maybe also with selfplay-games of Leela, I dont know.

yharzu · Post by **yharzu** » Sat Feb 25, 2017 3:51 am

haylee playing leela 0.9 :

https://www.youtube.com/watch?v=FFU7pEEMAjU

sadly not on optimal hardware and also recording while actually playing (instead of just reviewing her game or something like that).

hopefully she plays leela again under better conditions

still a cool video!

Pippen · Post by **Pippen** » Sat Feb 25, 2017 7:07 am

Schachus wrote:Its a bit like human intuition. That net looked at millions of positions and who was winning in the end, and learned to develop an intuition saying "this looks like black is going to win" or "this looks like its still even".

So basically its intuition depends on the quality of the database used, interesting. Obviously the AI has to look at my pattern and compare it with the same or similiar patterns of all the database games, and then look to how these games went. Is there any structure of how this pattern-searching and -matching goes or is this what they call "black box"?

Schachus · Post by **Schachus** » Sat Feb 25, 2017 10:51 am

I'm relly no expert there, so dont think that what I say is neccessarily correct, but it seems to be the general question how a neural network work, so you might want to read up on that(https://en.wikipedia.org/wiki/Artificial_neural_network). In general it tries so simulate the way a human brain works. Some go realted neural network usualy gets 19x19 "neurons" and has interconnections between them, that get more or less important as the net evolves and learns which connections between the intersections on the board are important and which arent.

If you do it a little more carefully, you probably have more than one layer with 19x19 neurons, so that not only the positions of the Stones(black/white/none) can be input, but also proprocessed information(can you capture there/is there a ladder there/ is a move there lagal), so that the net doesnt have to learn those things on its own(for example figuring out the rule "if that move captures a lot of stones, its probalbly good", is much easier to implement if you get an input point that tells you "capture here" and you dont have to combine the stones postion to find that out on yourself without someone telling you the capture rule).

At least AlphaGo has two of those nets, one to prdict, where the next move might be, and one to predict, who is likely to win, I suppose its the same with Leela, the first one being responsible for "NetProb"(how likely does it think a next move there is), the other one for "Net Win"

splee99 · Post by **splee99** » Sat Feb 25, 2017 5:04 pm

yharzu wrote:haylee playing leela 0.9 :

https://www.youtube.com/watch?v=FFU7pEEMAjU

sadly not on optimal hardware and also recording while actually playing (instead of just reviewing her game or something like that).

hopefully she plays leela again under better conditions
still a cool video!

However, Leela enjoyed a lot of pondering time because she talked a lot.

pookpooi · Post by **pookpooi** » Sun Feb 26, 2017 1:47 am

Wild card was offered to Leela from UEC organizers, does this mean it'll participate there as well.

goame · Post by **goame** » Sun Feb 26, 2017 9:16 am

yharzu wrote:haylee playing leela 0.9 :

https://www.youtube.com/watch?v=FFU7pEEMAjU

sadly not on optimal hardware and also recording while actually playing (instead of just reviewing her game or something like that).

hopefully she plays leela again under better conditions
still a cool video!

Why sadly not on optimal hardware?

For example
-can Leela use 2x xeon???
-can Leela use 2x gpus???

I actually have plans to buy 4x and 4x in some months.

pookpooi · Post by **pookpooi** » Mon Feb 27, 2017 11:14 am

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

zwim · Post by **zwim** » Sat Mar 04, 2017 2:39 pm

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 ?

pnprog · Post by **pnprog** » Sun Mar 05, 2017 7:23 am

Hi,

I just followed your link from my other thread.

Surprisingly the 'allmighty' Fuego plays M19 blindly and instantly.

In fact, even the venerable GnuGo would play M19

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

Same issue with Leela 6.2

splee99 · Post by **splee99** » Sun Mar 05, 2017 11:51 am

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

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.

pookpooi · Post by **pookpooi** » Thu May 18, 2017 7:00 pm

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

Life In 19x19

Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version

Re: Leela has a new version