Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / Stockfish 10 Released
- - By ChiefPushesWood (**) [us] Date 2018-11-29 15:49
Stockfish 10 was officially released. Have fun.

http://abrok.eu/stockfish/

CPW
Parent - By Lazy Frank (****) [gb] Date 2018-11-29 16:23
Bravo, SF team.
Impressive and hard work!
Thanks to all ...
Parent - - By Labyrinth (*****) [us] Date 2018-11-29 22:21
Anyone know the Elo vs. Stockfish 9 offhand?
Parent - - By pokerpawn (***) [be] Date 2018-11-29 23:23
no idea what to make of that page
but making a guess i would say ±60 elo :)
Parent - By pokerpawn (***) [be] Date 2018-11-30 10:13
thx

was not far off ! ;)
Parent - By Fulcrum2000 (****) [fr] Date 2018-11-30 15:17
See top left:
ELO: 54.21 +-1.9
Parent - - By Sesse (****) [no] Date 2018-11-30 15:47
My MMAP_RANDOM patch made it in, just in time :-)

Too bad I couldn't figure out what to do with analysis contempt. Warning to everybody upgrading from SF 9; SF 10 has a default contempt of 24 centipawns even in analysis mode, and it works by adjusting the score of side-to-move-at-root (not adjusting the score of a draw).
Parent - - By gsgs (***) [de] Date 2018-11-30 16:57
so, can we just substract 24 from the eval or is there a better formula
Parent - By Sesse (****) [gb] Date 2018-11-30 22:30
Just turn off analysis contempt. That's the simplest fix.
Parent - - By Uly (Gold) [mx] Date 2018-12-02 06:18
No, no, no, you learn to implement Contempt into your analysis, and have independent scores for white and for black. So there can be positions that are 0.24 from white's perspective and -0.24 from black's perspective, that is, a position you'd want to reach from either side! (because both sides have nice winning chances)

This is a new paradigm for chess analysis ("new" since analysis Contempt was raised for Stockfish) and lazy people just setting C=0 and using the old paradigm are missing out.
Parent - - By Carl Bicknell (*****) [gb] Date 2018-12-02 12:24
:grin:

Interesting.

I suppose my pushback would be that adding +0.24 automatically (which is what SF basically does with contempt, right?) doesn't achieve your goals at all. The ending K + B vs K doesn't give any more winning chances by adding +0.24 to it.

But maybe the SF contempt is more complex than I have described it? Do they have some sort of statistical measure now?
Parent - By Lukas Cimiotti (Bronze) [de] Date 2018-12-03 16:29
Contempt in general is a bit more complex. Even with contempt on an inevitable draw gives a 0.0 eval. But to be honest I can't imagine any benefit of contempt in analysis. I think it should be off by default.
Parent - - By Dragon Mist (****) [hr] Date 2018-12-05 21:46
I fail to see the reason for the sarcasm. C24 is at least not worse than C0, as demonstrated in regression tests. It significantly changes the way SF moves through the search tree as it aims for more complex positions, where that particular engine excels. That means it also produces different analysis lines than C0, which I personaly prefer to dullness of C0. The annoyance of having non simetrical evaluation during analysis is more than offset by having more complex and interesting, and most likely better, lines. Against other engines it worked miracles.

I for one use default contempt for my analysis, and did exactly that i.e. changed my mindset accordingly.
Parent - - By ChiefPushesWood (**) [us] Date 2018-12-06 15:33
Dragon,
How would you think leaving C24 affect IDEA analysis?

CPW
Parent - - By Dragon Mist (****) [hr] Date 2018-12-06 16:58
Sorry, I don't use IDEA, so no idea.
Parent - By Carl Bicknell (*****) [gb] Date 2018-12-07 23:42
Wouldn’t it mean extending the tree one more ply would totally change the evaluation by half a pawn?
Parent - - By Fulcrum2000 (****) [nl] Date 2018-11-30 17:05

> My MMAP_RANDOM patch made it in, just in time :-)


Do you happen to know the windows implementation for the same behavior?
Parent - By Sesse (****) [gb] Date 2018-11-30 22:30
I believe there's none. Like I said in the pull request, you could try opening the file with FILE_FLAG_RANDOM_ACCESS, but I believe it doesn't actually affect mmap behavior. Someone with Windows would have to try.
Parent - - By Labyrinth (*****) [us] Date 2018-11-30 22:17

>My MMAP_RANDOM patch made it in, just in time :-)


Can you explain roughly what this does? Just curious.

>Too bad I couldn't figure out what to do with analysis contempt. Warning to everybody upgrading from SF 9; SF 10 has a default contempt of 24 centipawns even in analysis mode, and it works by adjusting the score of side-to-move-at-root (not adjusting the score of a draw).


I thought SF9 had something similar, it's one of the reasons why I switched to SugaR for analysis.
Parent - By Sesse (****) [gb] Date 2018-11-30 22:35 Upvotes 3

> Can you explain roughly what this does? Just curious.


When Stockfish needs to read a Syzygy block from disk (typically 32 bytes plus some seeking, occasionally 64), the default behavior is that the OS reads that block plus ~256 kB around it, since it thinks you're going to need nearby data soon anyway. That's obviously meaningless since tablebases have zero locality, so MADV_RANDOM reduces it to the minimum possible, which is 4 kB. For an SSD, this means a world of difference, since the primary cost is not the number of I/O requests, but the amount of blocks read. (For a spinning drive, it doesn't matter that much. Also, of course, if you have enough RAM to hold the entire tablebase, the readahead could be useful, but then you should probably just preload the entire tablebase sequentially into RAM anyway.)

There are heuristics in place that should have kept this 256 kB scenario from occurring, but evidently, the access patterns of the TBs would seem to fool those heuristics. Thus, the explicit hint.
Up Topic The Rybka Lounge / Computer Chess / Stockfish 10 Released

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill