Author: Uri Blass
Date: 10:20:08 02/27/04
Go up one level in this thread
On February 27, 2004 at 10:54:07, Robert Hyatt wrote: >On February 27, 2004 at 05:59:16, Sune Fischer wrote: > >>On February 27, 2004 at 05:43:48, Uri Blass wrote: >> >>>I can add that I also have discontinuity and I see no special problems with it. >>>I also have discontinuity from middlegame to endgame and I have different piece >>>square tables for the king in the middle game and the endgame. >> >>I think what could be a potential problem is if you create a zig-zag behavior of >>the scores. >> >>If you use a very un-smooth eval there is a chance it will create local optimums >>here and there and the engine can get stuck in one of those if the search can't >>see beyond to a better optimum. >> >>It might also destabilize the search, not sure about that :) >> >>Bob might be right that discontinuites can be a trouble maker in this respect. > >Bob is definitely right here. > >The piece/square idea from Uri is a bit off the mark, because I doubt if his >piece/square values change _significantly_. The biggest change is for king at the center of the board and it can be changed from middle game to end game by 0.8 pawns. All evals have discontinuities, ie >a rook on the 7th is either there or not. That is somewhat akin to "unit step >functions" in math. But the larger the "gap", the larger the difficulty it will >cause. Usually the transition from middlegame to endgame can produce large >score swings. King safety goes away. King in the center is good. IE if your >king is in the center, you will try to trade material to get to the endgame to >make the king location better. If your king safety score is (say) -1.0 for the >king in the center, but the endgame score is +.5 for the same king position, you >will toss a pawn to reach the endgame. You might totally wreck your position, >giving your opponent a won ending, just to turn off that king safety score. I understand but I think that the problem here is bad evaluation. I think that the fact that I have only 2 piece square tables for the king is not best evaluation because practically the danger of having king in the centre is simply smaller in cases of less material so the piece square table score for this problem should be smaller but inspite of this I do not see it as a practical problem in most games. > >A slow transition reduces (but doesn't eliminate) this. > >> >>>It is possible that I can do my program better by not doing it. >>>The problem is that changing things without bugs is not a simple task when the >>>piece square table score is part of the evaluation that is changed incrementally >>>after every move(I thought when I started to have incremental evaluation but I >>>gave up that idea not because I think that it is a bad idea but because I do not >>>consider myself good enough to implement things fast without bugs but part of >>>the incremental evaluation is still there and I did not get rid of it when only >>>the rest of the evaluation is calculated from scratch at every node). >>> >>>Movei does not use a lot of time for the evaluation so maybe it may be better to >>>get rid of all the incremental evaluation. >> >>I don't use incremental eval, not even piece squares :) > >I don't see what incremental eval has to do with this, anyway. :) The point is that with incremental evaluation it is harder to do the change of slow transition between endgame piece square table and middle game piece square table. Uri
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.