H. G. Muller wrote on Sat, Feb 15, 2020 04:59 PM UTC:
This is very strange. Why do you need custom evaluation for KRK in the first place? Fairy-Max had absolutely no trouble to win KRK on its normal evaluation. (Which has Rooks completely neutral, and use the parabolic centralization table for Kings.) In the latest Fairy-Max version I use what you could call custom evaluation for checkmating a bare King (it just multiplies the centralization bonus for that King by a large factor), but that was only needed for end-games with many centralizing pieces (such as KFFFK in Makruk), because these did not bother to leave the center just for driving a single piece into a corner. Checkmating with Rook is very easy on any size board; you almost need no search depth for it at all. You just shephard the King towards the corner with your Rook, keeping the latter protected from behind by your King.
I suppose you do switch off null move?
The only thing I can think of that causes what you describe is contamination of the hash score with draw scores resulting from repetitions. What happens when you disable repetition detection (accepting PV hash cutoffs)? I don't really understand how a line that makes progress could get draw scores from this, though.
Can you post a game of how it typically plays KRK? (Best would of course be one that it bungles.)
Here is thinking output from a version of Fairy-Max that doesn't use the special bare-king evaluation, for position 8/8/8/5k2/3R4/8/5K2/8 w - 0 1 :
This is very strange. Why do you need custom evaluation for KRK in the first place? Fairy-Max had absolutely no trouble to win KRK on its normal evaluation. (Which has Rooks completely neutral, and use the parabolic centralization table for Kings.) In the latest Fairy-Max version I use what you could call custom evaluation for checkmating a bare King (it just multiplies the centralization bonus for that King by a large factor), but that was only needed for end-games with many centralizing pieces (such as KFFFK in Makruk), because these did not bother to leave the center just for driving a single piece into a corner. Checkmating with Rook is very easy on any size board; you almost need no search depth for it at all. You just shephard the King towards the corner with your Rook, keeping the latter protected from behind by your King.
I suppose you do switch off null move?
The only thing I can think of that causes what you describe is contamination of the hash score with draw scores resulting from repetitions. What happens when you disable repetition detection (accepting PV hash cutoffs)? I don't really understand how a line that makes progress could get draw scores from this, though.
Can you post a game of how it typically plays KRK? (Best would of course be one that it bungles.)
Here is thinking output from a version of Fairy-Max that doesn't use the special bare-king evaluation, for position 8/8/8/5k2/3R4/8/5K2/8 w - 0 1 :