Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Will Rybka 3 correct this?
- - By Zadig (**) Date 2007-12-10 22:39
6k1/3q4/7P/6P1/4Q2K/8/8/8 w - -


White to move. Rybka cannot see the draw here. It gives a fanciful +4 for ever. Hmm...

(You're right: my tablebases stop at 5 men!)
Parent - - By Banned for Life (Gold) Date 2007-12-10 23:01
This is one of those extremely very rare TB cases that don't contribute to Elo so don't worry about it! :-)

Alan
Parent - By Zadig (**) Date 2007-12-10 23:18
Forget TBs. What I'm after is a chess engine algorithm smart enough to detect that it goes into pointless, progressless loops with a totally dead flat evaluation. Rybka has become brilliant in tough and deep zugzwang positions that used to be a pain for engines. That's the next frontier: circular draws. Will Rybka 3 get smarter at detecting them? There has to be a way!
Parent - - By turbojuice1122 (Gold) Date 2007-12-11 00:09
Yes, as Alan says, it makes no difference that Rybka calls this +4--Rybka will eventually draw it, and it would be a draw with or without tablebases.  However, the problem is that Rybka will wish to go to this endgame seeing evaluation of +4 in the analysis, but with tablebases, Rybka would avoid this and try to win in a different way.
Parent - By Banned for Life (Gold) Date 2007-12-11 00:23
Turbo,

I think TBs are critical for this position and many others. I was taking a not so subtle dig at those that believe that TBs offer no advantage, even in fixed depth games.

Regards,
Alan
Parent - - By Zadig (**) Date 2007-12-11 00:44 Edited 2007-12-11 00:59
My point precisely, turbojuice: this is coming from a game where Rybka has given a pawn on the d column for all the black pawns on the kingside, thinking to win and possibly missing a superior variation, ie winning, with the d pawn... "+4 doesn't matter, it's gonna draw anyway" is beside the point. Ok, 6-men TBs might save the day in this very example, but then, this blindness will screw up 7-men evals elsewhere. So detecting "draw loops" is relevent. Their telltale sign is always: identical material for ever and absolutely flat eval at depth n, n+3, n+10, n+15, etc... Can't that be exploited heuristically somehow?
Parent - - By Permanent Brain (*****) Date 2007-12-11 02:38
This is an old problem which occurs in many types of positions engines don't understand (typical: fortresses). I always thought that a kind of 'progress rule' is required. Like: If the static evaluation of a position is quite high, +3 or bigger, then it should grow after a couple of moves. Engines calculate very deep nowadays, in short time. But if the current (board) position is +4 (statically) and after a depth of 11 or 13 plies it still has NOT grown to 4.5 at least, it isn't really such a big advantage because there is no progress. Maybe a 'no progress penalty' could be assigned.

That would be my logic behind it, but I am no programmer and actually, that logic is so simple that I must assume it would have been done already if it is possible. So, there must be complications or problems about it, or I miss something important.

But I know that engines like Patzer and Little Goliath have special algorithms to detect long check series, evaluating them 0.00 disregarding the material balance, without having to calculate them until absolute certainty. I think those failed very rarely, and I think the benefit - often finding deep draws which sometimes would involve surprising sacs in the beginning - is clearly bigger than the risk. That may be something comparable.
Parent - - By Zadig (**) Date 2007-12-11 09:47 Edited 2007-12-11 09:50
Why doesn't the following approach work:

a. if evaluation dead flat for each depth from n to n+10 (with say n>10 or whatever), with no material change, circular draw/fortress alert
b. pick main line position after n+10 plys (let's call it P)
c. evaluate P on its own to depth n+10
d. if evaluation of P static flatliner again
e. optionally do b. again on P+n+10 plys in the newly obtained main line
f. if another flatliner is obtained, fortress likely identified, evaluate the initial position as 0.

What are the odds that the engine won't be smarter with this in place, ie get it more frequently right than wrong?
Parent - - By Uri Blass (*****) Date 2007-12-11 21:36
It does not work because the problem is when this is a leaf position and not a root position.
The important thing is not to evaluate it as 0.00 but not to get this position in the first place when you have a winning alternative.

Uri
Parent - - By Zadig (**) Date 2007-12-11 23:58
Imagine these millions of leaf positions fanning out, explored and all of them return the exact same valuation, don't they? So if the "sampling" that I suggest into that part of the tree still tells you static evaluation having seperately tripled the depth search on the best line, -and yes, sacrificing the exhaustivity of the conventional tree search - bring back the eval of initial position from where all eval go identical to zero and check if the said line remains the best after that. This initial position has not to be the root position. I just want the node from which all eval go identical. Could be further into the decision tree. Am I talking rubbish, Uri? Would not be so surprising, but I would really love to know where I get it simplistically wrong.
Parent - By Uri Blass (*****) Date 2007-12-13 11:53
I do not understand exactly what you mean but
programs basically know only evaluation of the best line and do not know exact evaluation of something that is not in the main line
but only bounds.
The program may expect the opponent not to go to the fortress and it does not have evaluations of the fortress position at different depth to say that the evaluation does not improve.

It may be possible to evaluate the full path and not to evaluate position only based on the leaf but the problem is that in this case you will not be able to prune based on hash tables when the path is different.

Uri
Parent - By Uly (Gold) Date 2007-12-12 09:15

> But if the current (board) position is +4 (statically) and after a depth of 11 or 13 plies it still has NOT grown to 4.5 at least, it isn't really such a big advantage because there is no progress. Maybe a 'no progress penalty' could be assigned.


I don't think it would work for these reasons:

1. The line could have a slight swing in where the first plies had a great rise (from about +4 to +5), and then plenty of "not as good" (to +3.5) and "but better than that" (+3.9) and this should not be penalized the same as one that jumped from +4 to +3.9 directly.

2. The statically position at the root means nothing, you are talking about 0 ply, and it could be totally wrong (Like ply 0 thinks you're winning by a rook, but ply 1 show you lose the queen and are suddenly losing by 4 pawns), so the score from there can't be used for anything. Also, if you go from hash, only the last moves have a score, so you get lost.

3. You are comparing PVs, that assume that the opponent will do best play, but in an actual game the opponent most probably won't follow these PV moves, so when you add a 'no progress penalty' and your engine avoids those 'no progress' moves, it's probably avoiding best play itself without a real reason.

4. When the engine changes best move, the evaluations of previous plies suddenly becomes important, so it should store them somewhere (store the evaluations of all the plies of all possible moves just in case that they become the new Principal Variation), and this would slow down the engine as to nullify the possible improvement.

5. What's this margin that should be achieved to not mark the position "no progress"? Maybe the position takes 14 plies to improve 0.49, but if you penalize it, the engine has to choose another move that doesn't improve as much.

So, I don't think that this is a good solution, though it sounds logical at first sight.
Parent - By Vasik Rajlich (Silver) Date 2007-12-13 12:44
This is not an easy problem.

Assigning evaluation scores based on what happened earlier in the search tree introduces all sorts of complications with the transposition table. This is called 'path-dependent evaluation'.

Rybka does this in only one case - when the 50-move rule comes into the picture. I've decided that it's better in that case to make the adjustment abruptly (ie. 50 moves is assigned a draw score) than gradually, because the manoevering does often have a purpose.

If I could determine when further manoevering is pointless, then something better would be possible. Eventually, this will be improved.

Vas
Parent - By Permanent Brain (*****) Date 2007-12-11 02:47
6-piece tbs. acknowledge the draw, but if the wK was not on h4 but on h5 or on f4, White can win...
Parent - By Zadig (**) Date 2007-12-11 21:00
Could an admin move this to the Rybka forum by the way? Sorry for the mistake.
Up Topic Rybka Support & Discussion / Rybka Discussion / Will Rybka 3 correct this?

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill