Author: Robert Hyatt
Date: 11:46:29 10/19/01
Go up one level in this thread
On October 19, 2001 at 13:51:10, Rafael Andrist wrote: >On October 19, 2001 at 11:01:27, Robert Hyatt wrote: > >>I'm also interested. It has three major changes: >> >>1. Deep Blue search extension bound, two consecutive plies can't extend more >>than 2 plies, rather than any single ply being limited to extending 1 ply. >> >>2. New EvaluateWinner() endgame module that does a better job (certainly not >>perfect yet) of assessing whether a game is a draw, can be won by white or >>black, or by both. >> >>3. More stability in the endgame search. > >Does the last point mean that the score is more stable and less fluctuating? >What kind of improvements have led to this? > >Rafael B. Andrist The problem has been (for a long time) using the SEE scores to eliminate some captures in the q-search. But if this happened with little material left on the board, it could be wrong. Because removing the opponent's last piece could either leave you in a won position (which I caught) or leave you in a drawn position even though you are materially way behind (which I didn't catch). In that position we were looking at here, with the bishop + wrong rook pawn, one variation was very interesting and led me to fix this. It knew (by eval) that this was a draw. But it would start a search, and notice that after the king moved, and black played h8=R (=Q would not work as it was a check) that black would not take the rook, because black was so far down taking the rook still left the score below alpha. Which was wrong because taking the rook would leave the positional evaluation as "0.00". I fixed that particular case so that if the material is limited, and a capture will leave the opponent with one piece (or less, where a piece is a bishop or knight) then the delta pruning is disabled since with just a bishop left on the board can leave a dead drawn position and pull the score far farther along than expected just by looking at what was captured. Now there are far fewer bursts of fail highs, followed by no score because once the beta value hit +infinity, the delta pruning couldn't work until a good score dropped it...
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.