Should Stockfish continually investigate underpromotion in lines? (non-queen promotions) #4164
-
First, I'm curious if there is a simple classical/logical method to determine if promoting to a piece other than a queen is necessary. In my games, I would always promote a pawn to a queen unless it causes an immediate stalemate or when promoting to another piece leads to a checkmate in one. Are there positions in which more logic is needed beyond checking for a stalemate in one move or checkmate in one or two moves? I have most 7man Tablebase WDL (.rtbw) files and use them during deep 60+ depth searches. I've discovered some Tablebase WDL misses on files like KBBBvKRN.rtbw and began to wonder what the TT resource consumption and unnecessary time was being spent on underpromotion lines. Would it be more efficient to prune underpomotion lines if they don't lead to checkmate within two or three moves? While you could reach a KBBBvKRN position with the last move being an underpromotion to a Bishop, you could also reach it with a capture implying that the underpromotion to a Bishop happened much earlier. If Stockfish started with a position of 8 or more pieces, and reached a line with KBBBvKRN, that implies to me that it also has lines with KQBBvKRN in it's table lookups. How much resources are being used by those KBBBvKRN lines? |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 3 replies
-
I've found a few cases where underpromotion is required; mainly to avoid stalemate. You may be able to find some examples from puzzle collections. Those happen quite rarely, though. |
Beta Was this translation helpful? Give feedback.
-
I believe we have tried to search under promotion lines less deep, but that didn't lead to an Elo gain. I believe the search already deals with this efficiently enough. However, this is not an issue, feel free to test a patch on fishtest that explores these ideas. |
Beta Was this translation helpful? Give feedback.
I believe we have tried to search under promotion lines less deep, but that didn't lead to an Elo gain. I believe the search already deals with this efficiently enough. However, this is not an issue, feel free to test a patch on fishtest that explores these ideas.