Not logged inRybka Chess Community Forum
Forum RybkaChess Help Search Register Login
Up Topic Rybka Support & Discussion / Rybka Discussion / Sampled Search issues
- - By Dragon Mist (****) [hr] Date 2011-01-04 18:15 Edited 2011-01-04 18:18
Here's a thingy that puzzles me with R4 Sampled Search. Here a4 is being a candidate to overthrown best-so-far Nb3, and having passed "infinite sign phase" SS window starts showing nonsence like this (better is Rae1 ?? and Rae1 is not even a possible move). It would be very nice if Vas can shed some more light as to what are we seeing here.
Also, note for everyone, if R4 is taking a considerable time with infinite sign in front of a candidate move in SS window, be patient, as the result might be anything. Not only that maybe shown candidate will emerge on top, but it can be any other move all of a sudden.
Parent - - By Uly (Gold) [mx] Date 2011-01-04 18:43
I think it's an issue already known, Rybka isn't really thinking that an unplayable move is better, it's a problem with what she displays.

> if R4 is taking a considerable time with infinite sign in front of a candidate move in SS window, be patient, as the result might be anything. Not only that maybe shown candidate will emerge on top, but it can be any other move all of a sudden.


Rybka 4.1 will show more information so that such surprises no longer happen (e.g. if a sudden move is going to appear best next, the user would have seen in real time how Rybka found it and refuted all the other side's moves, before it becomes best).
Parent - - By Dragon Mist (****) [hr] Date 2011-01-04 19:01

> I think it's an issue already known, Rybka isn't really thinking that an unplayable move is better, it's a problem with what she displays.


Of course it is known, and I do know (well, I believe) it is a display problem, I am here asking this to be solved/clarified perhaps in R4.1.

> Rybka 4.1 will show more information so that such surprises no longer happen (e.g. if a sudden move is going to appear best next, the user would have seen in real time how Rybka found it and refuted all the other side's moves, before it becomes best).


Now that's a promising statement. Thanks! :-) Now that we "solved" this, maybe we can do something about the next situation also: sometimes R4 takes a considerable time thinking about (at least that's what is displayed) better move than found so far, let's name it move B (best so far is A). It displays the new PV line with move B after a long and hard think, say in 16 hours, 25 minutes, 3 seconds. THEN, after a second or 2, she displays another, new best move, move C. It is obvious that move C took much longer to compute then one or 2 seconds, so can we in future Rybka version see earlier that move B won over move A, and can we see in SS that move C is being analysed?
Parent - - By mdraith (**) [de] Date 2011-01-04 19:10
Move B sets a new alpha boundary and fills the hash tables with lots of positions, so move C may be calculated very fast.
Parent - - By Uly (Gold) [mx] Date 2011-01-04 19:18
But those cases happen even when the moves don't transpose (e.g. the moves filling the hash for move B are irrelevant for move C.)
Parent - By Dragon Mist (****) [hr] Date 2011-01-04 19:22
Yeap.
Parent - By Dragon Mist (****) [hr] Date 2011-01-04 19:22
With all do respect, the situation I am talking about occurs always when very deep in search, ten or more hours, so hash tables have been filled long before that. Also, we're talking about depths 25 middlegame, and hardly a simple transposition can occur that deep and complicated.
Parent - - By Uly (Gold) [mx] Date 2011-01-04 19:18

> It displays the new PV line with move B after a long and hard think, say in 16 hours, 25 minutes, 3 seconds.


You will be able to have a full summary of what Rybka spent doing on those 16 hours, 25 minutes, 3 seconds, by the second.

> THEN, after a second or 2, she displays another, new best move, move C. It is obvious that move C took much longer to compute then one or 2 seconds,


Actually, it's possible Rybka actually did take just 2 seconds to analyze move C and saw that it was better, that fast, but move B was ordered first so she HAD to resolve it first. I see this as common behavior in fail lows, Rybka spends too long resolving an irrelevant move, only to find one that resolves the fail low in a few seconds, so she could have taken 2 seconds to resolve it instead of 10 minutes, but as the depth gets higher I can imagine such search inconsistencies takes exponential time (such as 16 hours).
Parent - - By Dragon Mist (****) [hr] Date 2011-01-04 19:25

> You will be able to have a full summary of what Rybka spent doing on those 16 hours, 25 minutes, 3 seconds, by the second.
>


Ah, got it! So finaly a long ago promised log/PV display of every second or so will be available :-)

> Actually, it's possible Rybka actually did take just 2 seconds to analyze move C and saw that it was better, that fast, but move B was ordered first so she HAD to resolve it first.


Please see my note above; I think this is hardly possible that deep in search.

> I see this as common behavior in fail lows, Rybka spends too long resolving an irrelevant move, only to find one that resolves the fail low in a few seconds, so she could have taken 2 seconds to resolve it instead of 10 minutes, but as the depth gets higher I can imagine such search inconsistencies takes exponential time (such as 16 hours).


I fail to see this simmilarity with fail lows, although I know what you mean when you speak of suboptimal fail low resolve.
Parent - - By Uly (Gold) [mx] Date 2011-01-04 19:37

> Please see my note above; I think this is hardly possible that deep in search.


I've never gone to such depths (I very rarely leave the engine analyzing for more than 6 hours), however, there's nothing special about them, they're just like lower depth, only take much longer to process.

So I really think what you'll see is that Rybka did spend 16 hours in move B, and then 2 seconds on move C, and found it was better (that's how Rybka works, the only problem was ordering B first, which lost 16 hours).
Parent - - By Dragon Mist (****) [hr] Date 2011-01-04 19:44

> So I really think what you'll see is that Rybka did spend 16 hours in move B, and then 2 seconds on move C, and found it was better (that's how Rybka works, the only problem was ordering B first, which lost 16 hours).


I might SEE it, but will never BELIEVE that at huge depths it is possible to override 8 hours spent on move A and 8 hours spent on move B in just 2 seconds. But thanks anyway, one tiny bit of information making me really looking forward to R4.1.
Parent - - By Vasik Rajlich (Silver) [pl] Date 2011-01-04 21:13
This is possible, but I'd put my money on some sort of a sampled search bug. :smile:

Vas
Parent - By Dragon Mist (****) [hr] Date 2011-01-05 17:37
Of course! Thanks for confirming my suspicions. That was the main goal of this post, to send a message to beta testers that this needs looking into.
Up Topic Rybka Support & Discussion / Rybka Discussion / Sampled Search issues

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill