23 March 2020

Tablebases and Fortresses

There's an interesting thread on the Fishcooking forum: Misplayed wins, fortresses, or wrongly evaluated positions (google.com/fishcooking). It starts,

I've done an analysis of ~1M recent Fishtest LTC [Long Time Control] games, looking for those games that had a draw result, but where the evaluation was >2.0 for at least four plies. The result is ~17000 FENs that have been misplayed wins, fortresses, or wrongly evaluated positions. I think this might be interesting to have a look at, and might yield some [Stockfish evaluation] improvements.

The 17K FEN positions are in a Dropbox.com file available to everyone. I downloaded the file and pasted the first dozen positions into the Analysis board (lichess.org). I first discussed this tool in Seven-piece Tablebase on Lichess (August 2018).

Of the positions I looked at, at least half were tablebase positions, thereby subject to exact solution. The rest were fortresses like the one shown below.

6k1/2p4r/1p1q2p1/rPpNp1Pn/P1P1Pp2/1KQP4/5P1R/R7 w

Because of the extra Pawn, this looks to me like a possible win for White -- play for Nb3 -or- try the exchange sacrifice on h5 -- but the engines have a remarkable talent for finding hidden resources in tough positions. More work is necessary, but first I have to develop a better tool than copy/paste into a web page.

No comments: