Benefit of tablebases (probably old news)

[Jouni]

I run 100 positions EET suite and also private 100 positions endgame suite
with 5 piece tablebases and without (some top engines). Result: on average NO benefit whatsoever from tablebases! Some programs are
worse, some better with Nalimov bases. So I quess no measurable effect
in playing strength. But I have feeling these 200 positions haven't ANY
position where You need tablebases for best play Are they so rare
in practise may be?


But I still like Tbs to save time and energy, when engines don't try to win
draw positions.

Jouni

[ilari]

I have mixed feelings about tablebases/bitbases as well. Recently I added support for Scorpio's bitbases to Sloppy, and after about a thousand games (I run 100-game contests) I'm still not sure how much they help because of the great variation in results.

Against Slowchess Blitz the average score went up from about 48% to 52%, which would suggest a noticeable increase in playing strength. Against Wildcat 7 the score went down from 56% to 53%, and against Sloppy (same version but no egbbs) it gets about 54% of all points. That's all with 4-men egbbs. I tried 5-men as well, but actually got worse results, probably because probing them externally is slow. I tried loading the 5-men egbbs to RAM as well, but that didn't seem to help much.

I still need to run a lot more test games. Sloppy should be an excellent candidate for egbbs because it doesn't have any separate evals for special endgames like KBNK.

But I have feeling these 200 positions haven't ANY
position where You need tablebases for best play Confused Are they so rare
in practise may be?

By looking at Sloppy's logfile I can tell that positions where tablebases/bitbases are used are pretty common if the players are close in strength. How often do they change the outcome, that's another question.

[Kirill Kryukov]

Here are results of Little Goliath with and without tablebases (3-4-5-men Nalimov compressed, on a RAID 5 array, 8 GB total RAM should help with disk caching), in CCRL 40/4 blitz time control (128 MB for each engine, general book).

Results

4 points difference on current data.

I think different engines may benefit differently from Nalimov tables, depending on how well is engine understanding endgame by itself, and on how it is using tablebases.

I believe that bitbases should give noticeably more ELO increase than Nalimov tables.

Best,
Kirill

[Matthias Gemuh]

I use 4-men EGTB and 4-men EGBB.
There is no reason to waste more disk space and memory.

Matthias.

[Mike S.]

I use 4-men EGTB and 4-men EGBB.

As a compromise, in addition to 3/4-piece tables I use the Nalimovs for the most important 5-piece endings rook vs. rook (all KR?KR and KQRKR), but on my new computer from an USB memory stick, only. These are 30+252 MB. Also, I use a tablebase cache size of 128 MB because I have enough RAM, although it is still not clear to me if it has any performance impact. But at least it shouldn't harm.

On my older computer, I have the 3/4-piece Nalimovs on USB memory and only the r-r 5-piece endings on HD, and at least 32 MB tbs. cache.

I didn't make methodical performance tests or comparisons with/without tbs. Anyway, I think having rook vs. rook 5-piece tables supports analysis for difficult positions when it is about transposing into that ending. That should be a good investment of storage and memory because it is a frequent ending, while I agree that the rest is not necessary (or not much). As for those, I enjoy to avoid a suspected slowdown effect more than to get a doubtful gain from them.

[Matthias Gemuh]

As a compromise, in addition to 3/4-piece tables I use the Nalimovs for the most important 5-piece endings rook vs. rook (all KR?KR and KQRKR), but on my new computer from an USB memory stick, only. These are 30+252 MB. Also, I use a tablebase cache size of 128 MB because I have enough RAM, although it is still not clear to me if it has any performance impact. But at least it shouldn't harm.

On my older computer, I have the 3/4-piece Nalimovs on USB memory and only the r-r 5-piece endings on HD, and at least 32 MB tbs. cache.

I didn't make methodical performance tests or comparisons with/without tbs. Anyway, I think having rook vs. rook 5-piece tables supports analysis for difficult positions when it is about transposing into that ending. That should be a good investment of storage and memory because it is a frequent ending, while I agree that the rest is not necessary (or not much). As for those, I enjoy to avoid a suspected slowdown effect more than to get a doubtful gain from them.

Yes, some selected few 5-men tablebases are indeed helpful.


Matthias.