Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Nalimov access
- - By Vasik Rajlich (Silver) [hu] Date 2007-01-09 13:59
h1a8 wrote:

-----------------

[quote="Vasik Rajlich"]

1) don't need both white and black
2) don't need distances to mate
3) don't need "ridiculous" positions

Vas[/quote]

I don't know much about this topic. But can 2) effect the engine
in the 50 move rule? And will getting rid of either black or white in 1)
cut the total 1.5tb into half (like 750gb)? If so, then this is a huge
plus.

----------------------

Yes, both points are correct. There are 50-move issues with #2, and #3 will make the probes slighlty less efficient (the search may need to wait half a move), but the net effect should be positive.

Vas
Parent - - By Vasik Rajlich (Silver) [hu] Date 2007-01-09 14:03
Nelson Hernandez wrote:

----------------------

I'm coming late to this discussion but since my name was mentioned several times I probably ought to weigh in.

When I referred to "engine games" I was actually saying "games between two chess engines". I made this distinction because games between two human players are, on average, about 25 moves shorter and attain 6 pieces much less frequently than engine games, which often play on to checkmate. In fact the percentage of games that reach six pieces is about 22% for engines vs. 4% for humans. (However if you add all the individual six-piece combinations that occur together you get a considerably higher number than 22% and 4% because in some games you see multiple six-piece combinations through promotion.)

When I started collecting EGTBs my approach was to cram the highest percentage of occuring combinations onto a 400GB external drive that I could fit. To achieve this I divided the likelihood of the combination by the file-size of the combination, thereby creating an index. I then systematically downloaded the best bang-for-buck combinations and managed to fit the top 55% occurring combinations onto that drive (and I could have optimized it even better).

I subsequently went on to collect the whole set, and passed the final sixteen combinations to folks affiliated with the eMule project with critical help from the Man himself, Eugene Nalimov. That was surely the highlight of 2006 for me!

There's not much doubt that the full 1.12TB collection would benefit from compression, though I don't know what Vas means by "ridiculous" positions. OK, a KNNNN-K cavalry charge position is inherently ridiculous. But that kind of a combination takes up very little file space. I suspect there will be hardware solutions before software, though. It will just get easier to load and manipulate data over time with the advent of next-generation hard drives in a few years.

The more interesting question, in my view, is the future of 7-man EGTB development. Nalimov-style solutions are at a dead-end; the file-size of a complete 7-man Nalimov collection would make 1.12TB seem puny, and it would be utterly unplayable in any case. Different things have been proposed; I think a WLD table is the way to go--all you would know in such a table would be which moves were surely winning. You would not mate as efficiently but you couldn't miss a victory unless the 50-move rule applied. The size of such a table would be drastically smaller than Nalimov.

---------------------

Re. the "ridiculous positions", I haven't worked this out yet. It's a non-trivial task.

Some positions which could go in this category are:

1) Really weird pawn configurations
2) Weird places for the king
3) Easily scored (and won) by the engine anyway
4) Etc ..

The number of tablebase positions where the exact score (and solution) make a significant contribution is _really_ small.

Vas
Parent - By Vasik Rajlich (Silver) [hu] Date 2007-01-09 14:05
Caissafan wrote:

--------------------

Hi Vas,

that isn't understandable because too short explained again.

More than a year is gone and we do not have a correct Rybka description so far! Mad
A correct description is more much more than some information given on Rybka web site, much more than some "perls of wisdom" fallen via Rybka forum here and there.

In this regard:

      What a "depth is <=X (where X depends on this parameter)"?
      What a "right X" using?
      What happens exactly when using NalimovUsage = Rarely instead NalimovUsage = Normally regarding Rybka's search?

Thank you.

----------------------

What I meant is that you can decrease the rate of Nalimov probing via the "NalimovUsage" parameter. If your current setting decreases Rybka's branch factor significantly, then it would probably help to decrease the probe rate.

Just to put this in perspective, though - it's not really important for a casual user to worry about it. We're talking here about 1 rating point.

Vas
Parent - - By Nelson Hernandez (Gold) [us] Date 2007-01-10 20:54
Vas, your statement...

The number of tablebase positions where the exact score (and solution) make a significant contribution is _really_ small.

...is arguable.  Empirically speaking, there are an enormous number of engine vs. engine games that end "incorrectly".  If we accept my figure of 22% as the percentage of engine games that attain six pieces, then the real question is what percentage of those games, if EGTB-adjudicated the instant they reached six pieces, have the same 1-0, 1/2 or 0-1 end-result as they actually attained in the game.  I don't have the exact answer to this but I would guess up probably 10% of 6-piece games (or about 2% overall) are misplayed in the endgame resulting in an outcome other than would have been the case with 6-man EGTBs.

Meaning: maybe 2% of the game-results in a typical database/opening book are misleading from this single cause.  (There are other causes as well that drive this number to more like 10%, but there is no point in digressing.)  Now, I grant you, these errors are randomly spread throughout a book so this is just static noise.  But it is something to think about when you are at the far limit of your move-tree and aren't really sure about the veracity of your data!

The point being...EGTBs are important because they contribute to opening book accuracy.  Paradoxical, isn't it, that I got 6-man EGTB to strengthen my opening book, not my endgame.  Alan, I know you're maniacally smiling out there in Vegas.
Parent - - By Banned for Life (Gold) Date 2007-01-11 01:12
You're right Nelson!

Like you, I'm only interested in the openings, but rather than relying on the games outcome, I pick a somewhat arbitrary point in the middlegame and adjudicate at that point based on the eval. This has some problems, but so does looking at the final score, which will be more heavily influenced by engine choice, engine speed, and EGTBs. Since my reference point is earlier than yours (and most others), I expend a lot of effort trying to decorrelate the game score from the results at my middlegame adjudication point. At first, I would resign games where I had picked up a large advantage in the opening. This was very effective in pissing off Eros (which is worth a lot to me all by itself), but its too much work. Losing games on time would be a good option, but the CB server turns off auto. So now I either play with WinFinder (which I would do exclusively if it were stable) or play with Rybka 2.2n2 with crazy settings. My goal is to maximize the number of games that have an incorrect eval without driving my Elo down to a point where the top engines won't play against my engine (~2700). WinFinder would actually be the best choice for me, because it not only loses points relative to R2.2n2, but also wins a small number of points that R2.2n2 would miss.

So I'm probably the only one running an engine in the CB engine room that is really happy when they are up a pawn at move 30 and then end up losing the game! I'm still relatively new at this, but I'm hoping you'll be able to see the results of my work in this year's games.

Regards,
Alan
Parent - - By Vasik Rajlich (Silver) [hu] Date 2007-01-13 10:07
By the way, a 600-game match between WinFinder 2.2 and Fritz 10 in 32-bits ended 303 to 297 in favor of Fritz 10.

I kind of expected WinFinder to be a bit stronger than that, hopefully the next version will have a bit more overall chess strength.

Vas
Parent - - By Banned for Life (Gold) Date 2007-01-13 23:05
Whereas Rybka is a well balanced engine that plays strongly from most openings, WinFinder has strong strengths and weak weaknesses making it much more important to play opening lines that lead into suitable positions. For this reason, I'm not sure that using standard openings developed for testing purposes really makes sense for this type of test.

My experience after playing several hundred WinFinder games on the CB server has been more positive with the biggest problem being that WinFinder crashes every four games on average. I'm wondering how you get around this? In contrast, I haven't had a single engine crash with 2.2n2 over thousands of games.

In any event, its a real thrill seeing Rybka 2 on an overclocked quad core Intel chip getting 4 times as many kn/s drop a piece when its playing against WinFinder, even if it doesn't happen too often.

Regards,
Alan
Parent - By Vasik Rajlich (Silver) [hu] Date 2007-01-15 13:23
Ok, this makes sense. Getting the WinFinder to stop crashing is certainly a high priority.

Actually, I kind of suspect that trying to make an opening book for WinFinder 2.2 would just lead to frustration. There are some limits to what can be done.

Vas
Parent - By Uri Blass (*****) [il] Date 2007-01-11 14:37 Edited 2007-01-11 14:40
My guess is that less than 2% of the games when you get tablebase 6 piece position end incorrectly that means that programs do important mistakes in 6 piece position in less than 2% of the cases that they get there(a mistake of mating in 31 when the shorter mate is mate in 30 is not an important mistake)

If I am correct then it means that only using 6 piece tablebases at the root may change the result of not more than 2%*20%=0.4% of the games.

I think that it may be possible to check it because 6 piece tablebases are available online so you can see if programs made practical mistakes in 6 piece tablebases based on comparison between the tablebases and the moves.

Uri
Parent - - By Vasik Rajlich (Silver) [hu] Date 2007-01-13 10:05
I agree with Uri - 2% of all games (and 10% of 6-piece games) seems far too high.

If it were true, then just probing 6-piece tablebases at  the root (ie. not even in the search) would give you 14 Elo.

I've seen countless computer chess experts (even very knowledgeable engine authors) overestimate tablebases. It's probably a psychological thing - chess is such an unclear and uncertain game, so when we finally have something we can be really sure about, it just "feels" really good.

Vas
Parent - - By Nelson Hernandez (Gold) [us] Date 2007-01-13 14:21
Honestly, I don't know.  But fairly soon I may have some concrete numbers.  When I get them we'll continue this discussion.

I never said anything about ELO to begin with.  To infer that fixing inaccurate games in a database would lead to +xx ELO would be a speculation on top of a speculation.  All I was saying was that as a general principle an opening book ought to be better if your database has correct outcomes.  Whether the error rate is 2% or 0.2%, the same thing holds.  Of course this remark assumes your book is largely driven by empirical results, which not all are.
Parent - By Vasik Rajlich (Silver) [hu] Date 2007-01-15 13:20
Ok, this is true. It's just that the 10% really jumped out :)

Vas
Up Topic Rybka Support & Discussion / Rybka Discussion / Nalimov access

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill