FENs to test engine endgame knowledge?

[xmas79]

Hi Thomas,
I had a very fast run at draws and wins epds. Draws allowed to me to add quickly some missing endgames (KNR-KR etc.., as I was doing only up to 4 pieces endgames).
But it seems to me that some positions are not well suited for static evaluation (like the following, taken from wins.epd)
[d]8/8/n7/2R5/8/8/k7/6K1 w - - 0 1
This is clearly a won position, but from a static point of view this will have a winning score, but not a huge win score. If we land here after q-search (and no check extensions) this position gives "only" +2.00, and not +5.00. Now I'm wondering if it is worth turning on all these exceptions (that actually can be easily discovered by search, and IMHO must be...). So in this types of position I would simply return 0.00 as score (0.01?) , as I don't mind where the pieces actually are. I only want to know that this endgame is very hard to win, but if search is smart enough will catch up the correct sequence.

What do you think?

Thanks,
Natale.

[xmas79]

...an internal "test"...

Yes, I think this will be my next step, as it will allow some sort of automatic tuning.

Thanks,
Natale.

[xmas79]

Hi Steven, Downloaded. thank you! Same question for you: time limit for each position? I couldn't find anything on the net (I admit I did only a very fast search...)

Thankyou,
Natale.

[sje]

Hi Steven, Downloaded. thank you! Same question for you: time limit for each position?

1 second, 2 seconds, 4 seconds, 8 seconds, etc.

[xmas79]

Hi Steven, Downloaded. thank you! Same question for you: time limit for each position?

1 second, 2 seconds, 4 seconds, 8 seconds, etc.

mhhh... Position 1797: 2^1796 seconds?? LOL!!!

[sje]

Hi Steven, Downloaded. thank you! Same question for you: time limit for each position?

1 second, 2 seconds, 4 seconds, 8 seconds, etc.

mhhh... Position 1797: 2^1796 seconds?? LOL!!!

Run the whole suite iteratively at 1 second, 2 seconds, 4 seconds, 8 seconds, etc. for each position.

[xmas79]

Hi Steven, Downloaded. thank you! Same question for you: time limit for each position?

1 second, 2 seconds, 4 seconds, 8 seconds, etc.

mhhh... Position 1797: 2^1796 seconds?? LOL!!!

Run the whole suite iteratively at 1 second, 2 seconds, 4 seconds, 8 seconds, etc. for each position.

The lower, the better. I got it.

Thanks,
Natale.

[tpetzke]

Hi Natale,

it really depends on the design of your engine. In iCE I like to know when a position is a draw so I can terminate search in that node.

So in your position iCE would say

iCE 2.0 v270 x32 [2013.9.9]
position fen 8/8/n7/2R5/8/8/k7/6K1 w - - 0 1
eval
KRKN score : 68
go depth 1
info depth 1 seldepth 0 time 0 nodes 26 pv c5a5 a2b3 a5a6 nps 25999 score mate 17 hashfull 0 tbhits 3
bestmove c5a5 ponder a2b3
info time 0 nodes 26 nps 25999

iCE recognizes that this position has the potential of not being drawn (because of the positions of the black knight and king). static eval says 68. If this position is handed to search it will see that most of the white moves end in a drawn position so the winning path and a winning score is found already at depth 1 (searching 26 nodes).

As I terminate the search in a known draw position it is very important that I don't announce draws in non drawn positions (false positives).

Therefore I assembled those sets which help me to check my special endgame modules for different piece sets.

Most people don't bother and just use table bases but I want my engine to also do without. (although not fully, iCE has internal access to KPK, KRK and KQK)

Thomas...