back


28523Mrert
2018-12-14 20:07:41
@Phil: Yes exactly you cannot follow blindly a line output by the engine. You have to try a beginning of a variation then re-analyze to see where it's going.
Actually SF dev team made a lot of efforts to have a good PV output (it was much worse 2-3 years ago), but still you'll get more accurate results re-analysing often.
28522Phil
2018-12-14 16:38:43
@Mrert: thanks a lot for this information, this explains everything, this "improvement" feeling when analyzing.
I asked this question after having tried chessbase "deep analysis": after 4h on 1 position, it gave me short variations but when I chekced them in 30mn, I found it were plenty of imprecisions. Now I understand why this "deep analysis" cannnot be a good idea.
28521queen13
2018-12-14 11:52:38
or they can get alphazero to play on tcec and win it.

CCC as well, here: https://www.chess.com/computer-chess-championship

I think maybe they could still cheat if they wanted to but that would be more difficult.
28520Mrert
2018-12-14 01:06:31
Weights are just numbers, it's hardware independant. it's the implementation of the network/search that must be TPU-compatible as TPU support only a small amount of operations.
28519SilverFang
2018-12-14 00:18:30
@28518
I'm afraid that those weights are designed for TPU. Better if Lc0 team could cooperate with DeepMind.
28518Mrert
2018-12-14 00:09:27
If they want A0 to be acknowledged as the best chess playing entity, they just have grant open access to the weights (as we know the NN architecture and search algo that was used), so that people can try themselves.
Otherwise it's just playing with home advantage and tailwind

@Phil: The longer the line output the less accurate. Also, Stockfish stores variations in RAM. So if you launch an 2nd analysis after making a couple of moves it will use the stored variations containing thoses moves to get more accurate results
28517queen13
2018-12-13 20:03:56
never mind, "view crosstable" it is...
28516queen13
2018-12-13 20:02:21
does anyone know how I can see the results

from https://www.chess.com/computer-chess-championship

but only the games where lc0 played stockfish/houdini/komodo?

It seems to me that Lc0 is weak against weak engines
28515Kit
2018-12-13 17:49:41
On the same level as Stockfish, but not above and beyond. ;)

Still no one is over the level of Stockfish, which shows clearly that this program has the world leading chess developing community. A bow.
28514queen13
2018-12-13 15:52:41
then*
28513queen13
2018-12-13 15:52:29
@Phil

If you know that those 7-8 moves are the best for both sides than I think yes.
28512SilverFang
2018-12-13 14:37:07
From https://www.chess.com/computer-chess-championship we see CCC claims that Lc0 is at least on the same level with SKH, above Ethereal.
28511SilverFang
2018-12-13 14:34:58
Lc0's progress is supported by a growing network size. Hope it won't grow out of the average RAM before nosing out SF😄
28510Phil
2018-12-13 14:33:08
@queen13: thanks, yes it seems it "improves" while analyzing, but I cannot understand "mathematically" how it can do that.
Consequently, when analyzing with stockfish, it is clearly advisable to make moves in the variations, for example 7-8, and put last position to depth 35 BUT NOT let stockfish on the initial position to depth 42-43?
28509Kit
2018-12-13 11:58:06
In the case of chess AlphaZero is a commercial PR of Google, a well prepared advertisment campaign. One year later they are still on the "four hours" with so powerful hardware and no progress. Very obviously it is not stronger till proven the opposite. This is visible to a lot of people related close to the computer chess community. Good point is that DeepMind admit or recognize that their AI program lost games (with black) when the opponent is using a good opening book. The current true project now is Leela. It is interesting to see is it able to surpass Stockfish. Meanwhile in the center of the computer chess universe and chess itself still stands Stockfish (and opening theory). Thank you to all developers, involved in creating this conventional intelligent program.
28508queen13
2018-12-13 11:03:04
is lc0 stronger than latest stockfish?

is alphazero stronger than latest stockfish?

Is it going to be?
28507queen13
2018-12-13 10:58:00
@Phil

It might be necessary to restart stockfish for "variation 2"

They can still be different but if you don't restart I think stockfish remembers progress

from variation 1 and builds on what was searched..
28506SilverFang
2018-12-13 10:09:36
AlphaZero crushed Stockfish again, this time with fairer hardware and more convincing time control:
http://science.sciencemag.org/content/362/6419/1140
Unfortunately, the exact version of the SF used is not offered.
28505SilverFang
2018-12-13 09:54:01
@28504
Because they're estimated to have similar values, so SF will somewhat randomly choose one.
28504Phil
2018-12-13 05:13:35
I analyze with Stockfish, say it goes to depth 30. I call its first variation "variation 1".
Then I enter the two moves of variation 1, and I wait for its analysis in depth 28, then I name the first variation as "variation 2".

Why variation 1 and variation 2 are different?

Thanks
28503queen13
2018-12-11 22:54:20
I can't get lc0 to run with gpu. it refuses to work.

I wanted to see how it compared to cpu version.

I can now see nodes per second. there is a setting to change kn/sec to n/sec

I got 178 N/sec after 53 seconds of running lc0 with analyse...

in the initial position with syzygy and network 31962.

It started with low N/sec and was gradually increasing.

I don't know why it no longer works with gpu.

Cuda or opencl it just does not work.
28502SilverFang
2018-12-11 08:03:48
https://storage.googleapis.com/deepmind-live-cms/documents/AZ-Blog-Fig1-Generality-Performance-Across-Games.gif
It seems that AZ is not much stronger than SF, considering its unaffordable hardware requirement. For PCs, SF is the least handicapped god we can afford.
28501SilverFang
2018-12-11 08:02:42
<a href="https://storage.googleapis.com/deepmind-live-cms/documents/AZ-Blog-Fig1-Generality-Performance-Across-Games.gif">AZ Learning Statics</a>
It seems that AZ is not much stronger than SF, considering its unaffordable hardware requirement. For PCs, SF is the least handicapped god we can afford.
28500Kit
2018-12-10 02:53:59
Which Stockfish is the strongest?
28499WCN
2018-12-08 14:22:42
AlphaZero (Chess) Copied! The A.R.B Chess System A.R.B :)

https://www.youtube.com/watch?v=zSuCu9ZNv4w




WCN

next page >
single page