Author: Ulrich Tuerke
Date: 03:51:13 03/08/01
Go up one level in this thread
On March 08, 2001 at 06:26:05, Ed Schröder wrote: >On March 08, 2001 at 06:17:40, Ulrich Tuerke wrote: > >>On March 08, 2001 at 05:08:03, Matt McKnight wrote: >> >>>oops, messed up that last post... >>> >>> I've been toying with some move ordering ideas, and I would like to hear some >>>feedback on them. >>> >>> First idea: >>> Inspired by the killers list, I tried keeping a list of two moves per ply >>>that seemed to be bad ideas, (score < alpha) and using that to sort those moves >>>in the future to the end of the list. This seems to give me a small gain. Is >>>this idea good, or will it cost me in some positions? >>> >>> Second idea: >>> Along with the history heuristic, how about sorting moves by the >>>positional gain they make? For instance: >>> score += piece_eval[to] - piece_eval[from]; >>> >>>this also seems to help a little. Again, is it helping or not? >>> >>>Also, are these ideas dumb and old? >> >>At least, your 2nd idea is not new. It has been tried in an old gnuchess >>version. I don't know if it's still used by the more recent versions of gnu. >> >>IMHO, the drawback is that you have to evaluate in addition to the leaves also >>the inner nodes in order to get these piece values. I experimented a bit with >>this, but I didn't make a thorough analysis. Since I seemed to observe that the >>gain from improved move order (which is there IMO) is not really worth the >>performance loss by evaluating all nodes, I decided not to use this. >> >>However I guess that your suggestion may deserve more attention. >> >>Uli > > >The second idea to sort "the remaining moves" by a piece_square table >has been in Rebel from the very early start. Last time (about 2-3 years >ago) I removed the algorithm (just to satisfy my curiosity) Rebel's >performance in ply-depth decreased with a factor of 2. Surprised by the >outcome I quickly activated the algorithm again. > >Ed Thanks for the sharing your precious experience. In the old gnuchess (IIRC), it wasn't just a piece-square table, but something like the "net result" of the positional evaluation (weak field, king safety, ...) had been attributed to the square. Also creation of a doubled pawn for instance resulted in a penalty for this move regarding move ordering. Using solely a piece-square table instead may be a good alternative without serious performance penalty. Uli
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.