Which tablebases actually help engine performance the most?

[nine castles]

Hello everyone. I'm installing tablebases all the way up to 6 pieces, but I see that a great percentage of tablebases will not in fact help the engine enough to be worth the disk space. For example no engine will ever need a KRRRK tablebase. This is an extreme example, but there are also borderline cases that turn out to be less useful to engines than say KQPKQ, or KRPKR, etc. Is there any consensus on which tablebases from each set are most useful to increase engine performance in engine v engine matches? What are everyone's thoughts on this?

[Kirill Kryukov]

All 3 to 5 men tablebases should be installed, even if some look useless.

About 6-men tablebases, here is from the "Endgame Tablebases Online" site:

4. You may like to download tablebases by "importance" order, which is based on statistics of occurrance of each ending in real games. Several such lists exist: by Dieter Bürßner, Nelson Hernandez, and Peter Kasinski.

I hope it helps.

[nine castles]

This is some interesting information, thank you.

My only question would be whether it would actually make sense to download these tablebases in order of frequency, as some of them are very easy to win and unlikely to be incorrectly evaluated by the engine like KRPKPP.

[Dhanish]

According to the FAQ at the Rybka website: http://www.rybkachess.com/index.php?auswahl=Rybka+3+FAQ

>>Is it useful to load 3-4-5 Nalimov tablebases in Rybka 3?
>> Yes, slightly. I believe that tablebases improve the quality of endgame play by an Elo point or so. This is only an >>opinion, as we don't have a good way to measure improvements of this type of magnitude. "

For engine vs engine matches, tablebases may not be of much use compared to hardware or opening book. That is because the hard disk access time is higher, slowing down the engine. For blitz games, 5 piece bases should be sufficient. For longer time controls, common 6 piece bases are desirable. It also depends on the strength of the engines. When they are evenly matched with good opening books, more games are likely to reach the endgame stage! But for deep analysis of positions, say in correspondence chess, they are essential.

Regards,
Dhanish

[nine castles]

Yes, I agree with this. Currently I am playing with Rybka on FICS, which is not the strongest place for comp vs comp matches, and the game are mostly blitz. Therefore I have the tablebase usage set to "rarely" and may consider it set to "never" to simply allow the GUI to handle the tablebase, IE rybka gets into the endgame and then tablebases simply take over.

But your concerns are another reason besides disk space why I am hesitant to install a lot of TBs, and I only want to install the TBs that the engine is likely to incorrectly evaluate. KBPK is a classic example of a known bug where Rybka will think she is winning despite the fact that she has a wrong-color rook pawn. TBs are useful here to clue her into the fact that she should not seek this kind of endgame. Luckily this TB is tiny so I have included it. KRPKRP is huge, but all engines are notorious for being unable to play this kind of ending, so I am going to install that tablebase too. Same goes for KQPKQ.

But what about 5-1 or 4-2 endings or like KRPKPP? It is intesely unlikely that Rybka will fail to recognize that this is a win (when it is a win) or that she will play it inacurrately. Despite the fact that this ending is apparently common I don't think I want to include this TB since using it will only slow down the search where her own evaluation function would have been sufficient to evaluate the position. I can think of other similar cases, and it make me think that maybe I could achieve best performance by removing some of the 4 and 5 men TBs too. Am I off in my thinking here?

[Dhanish]

But what about 5-1 or 4-2 endings or like KRPKPP? It is intesely unlikely that Rybka will fail to recognize that this is a win (when it is a win) or that she will play it inacurrately. Despite the fact that this ending is apparently common I don't think I want to include this TB since using it will only slow down the search where her own evaluation function would have been sufficient to evaluate the position. I can think of other similar cases, and it make me think that maybe I could achieve best performance by removing some of the 4 and 5 men TBs too. Am I off in my thinking here?

Your point is valid, and I have read several people mentioning the same. But I don't think anybody has revealed the specific tablebases used by them.

It would be interesting, if you could compare your results with selected tablebases and without tablebases. Of course, you need thousands of games to reach conclusions. And whether you use tablebases should be randomized, to offset the effect of position learning / book improvements by yourself or your opponents.

Regards,
Dhanish

[Ray]

Chessbase clearly have some method for choosing the selection of 6-men that they include in their product DVD Endgame Turbo 3:

http://www.chessbase.com/shop/product.asp?pid=263

But it is a pretty small selection