back


26341Mrert
2018-04-19 16:42:30
much progress in latest networks
26340queen13
2018-04-19 13:08:21
LC0 makes division 4 interesting for me
26339Mrert
2018-04-19 04:49:39
lol TCEC GUI broken
26338Mrert
2018-04-18 23:22:34
let's see how good LC0 will play this as white
26337queen13
2018-04-18 21:10:59
I wonder if using leela approach for the opening in stockfish could work.
Neural net plays the opening and stockfish takes it from there.
Or maybe if they could work together?
stockfish uses the cpus and leela the gpus so they could run simultaneously?
26336ahithophel ebejer
2018-04-18 20:42:10
after nearly 6, 000, 000 games

leela s elo rating is at its highest
26335Mrert
2018-04-18 20:11:26
not that bad from leela, but as usual, mistake in late middlegame kills her
26334Mrert
2018-04-18 20:06:47
usually it's code refactoring (simplifiy stuff, tidy up the code...)
26333Mrert
2018-04-18 20:05:51
@charles : you got it
26332charles
2018-04-18 20:01:22
all the lastest builds are non-functinal. what does that mean? no strength gained?
26331Mrert
2018-04-18 16:05:44
wait and see. Looks like leela gains a lot of strength through TB.
But for sure even in a couple of months it probably won't beat SF on equal hardware.
Remember that Deepmind gave powerful, tailor-made hardware to A0, a tiny hash for SF, and probably cherry-picked the match with best results.
26330queen13
2018-04-18 13:49:03
@ebejer
> when you indirectly say

that stockfish is invincible



I never said that. Neither directly nor indirectly.
At worst I said that leela chess can never surpass stockfish if they are trying only to reproduce alphazero because alphazero did not surpass stockfish.
However, they could improve on it or at the very least train it more.(There is a problem with diminishing returns as it gets stronger it gets harder to get stronger so maybe simply training it more won't work)
So let's see what will happen.
26329Omd
2018-04-18 07:59:02
The version of Leela Chess in TCEC

The version of Leela Chess Zero that will participate in TCEC is LC0 ID 125. This is the same version that played the test matches vs Scorpio and Fruit. For more about LC0, meet the experts and contributors in the TCEC chat

Leela Chess Zero will run entirely on CPU during TCEC Season 12. There are plans to make a match between a higher level engine and Leela Chess Zero on GPU/TPU, however, that entirely depends on the support of the community

More about Leela Chess Zero in TCEC

The “Zero” in LC0 signifies no human chess knowledge is added. This is the key difference between neural net engines and standard chess engines. Technically, this will still be true for the TCEC version of LC0. However, one modification will be made to LC0′s TCEC version. Tests have shown that LC0 performs as much as 300 Elo better when paired with 6-man Syzygy tablebases, which are not human knowledge but rather a table of endgame computations. Many other engines in the TCEC field refer to tablebases and we deem it most fair (and most fun) to have LC0 compete with tablebase access as well.

The LC0 project requires a huge amount of computation and the only feasible way for LC0 to advance is as a distributed project. If you want to help Leela achieve its full potential please visit https://groups.google.com/forum/#!forum/lczero or the Leela Chess wiki
26328Omd
2018-04-18 07:55:00
Leela Chess Zero (LC0), an open source adaptation of DeepMind’s recent Alpha Zero artificial intelligence demonstration project, will compete in TCEC’s Season 12. In so doing it will become the first chess-playing neural net in history to publicly challenge traditional human chess programming since the days of Turing and Shannon. This epochal event will commence April 19th at 19:00 CET. Follow the games on the official website

Leela Chess Zero is an open source project by large community of programmers, with the goal to build a strong UCT chess AI following the same type of techniques as AlphaZero, as described in Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm i.e. follow in the steps of Alpha Zero by using the contributions of the computer chess community. Leela Chess Zero uses Stockfish’s position representation and move generation although no heuristics or prior knowledge are carried over from Stockfish.

Leela Chess Zero is going to play in TCEC Season 12, The entry is possible due to Defenchess (Div 3) withdrawing in the last moment. This means ChessBrainVB will automatically promote to Division 3 based on it’s Season 11 performance, and this in turn opens a place in Division 4 for LC0 to enter the competition.

More about TCEC S12: Official website / Twitch video channel / Full list of participants

Test match of LC0 vs a TCEC engine

Leela Chess Zero is a rather new but rapidly developing project. With every version released the strength of the engine increases. But is it ready for TCEC and the division structure of the championship, running only on CPU and observing the TCEC rules set? TCEC is built for projects in active development of Elo strength estimated at 3000 or more. To prove that LC0 is ready for the event it was pitted against two of the Division 4 engines in the field and a classical engine.

Against Scorpio, LC0 achieved +4=3-13 (5.5 points) and a performance rating of ~2700 ELO. Among the wins of LC0 was a fantastic miniature on the black side of Caro-Kann Tartakower with 6.c3, a 36 moves win with white in Spanish: Closed, Smyslov, 12.Nf1 Bb7, and a win in King’s Indian: Fianchetto, Classical, 9.h3 Qb6 10.c5 See the full list of games of the Scorpio – LC0 match.

In a mini match against Fruit, Leela Chess Zero was more than convincing with +4 = 1 -1 and a performance of over 3000 ELO. Against the classical version of Stockfish 1.0, Leela Chess Zero achieved +7=3-10
26327Mrert
2018-04-18 05:55:02
Leela enters TCEC S12 !!
26326Mrert
2018-04-18 01:40:41
wow fruit got rekt
But fruit is vastly inferior to any division 4 engine tbh. Maybe leela will enter in season 13...
26325ahithophel ebejer
2018-04-17 23:47:19
after 5, 500, 000 games

leela's elo rating is 4900 (2428.3)
26324ahithophel ebejer
2018-04-17 22:21:24
queen13

don t you think that

when you indirectly say

that stockfish is invincible

you re being arrogant?

everybody knows what they used to say about the titanic

and what happened to it
26323queen13
2018-04-17 20:53:46
@ebejer
I don't think leela will be much stronger than what it is now in say 6 months.
I hope she will though. This new approach sounds interesting.
I would say max 200 elo gain in 6 months from now.

Then maybe they just want to recreate alphazero and as such can never reach stockfish 9.
Let's see.
26322ahithophel ebejer
2018-04-17 20:48:15
my contribution to the improvement of leela

http://lczero.org/user/AHITHOPHEL
26321ahithophel ebejer
2018-04-17 20:39:46
boys and girls

don t talk about leela's strength now

she's still in the very beginning

it would be wiser to talk about her in a few months time
26320Mrert
2018-04-17 19:57:58
Well looks like 1 ? Fruit 2.1 was written in like 2004-5. Not sure it can use several cores
26319queen13
2018-04-17 19:35:07
How many threads does fruit use?
26318Mrert
2018-04-17 16:52:04
@queen13: agree, and given hardware and TC it shows low quality of play compared to the top engines
26317queen13
2018-04-17 15:05:57
stockfish 1 leela draw rate is low

next page >
single page