Author: Alessandro Damiani
Date: 05:55:51 06/04/98
Go up one level in this thread
On June 03, 1998 at 17:04:27, Robert Hyatt wrote: >On June 03, 1998 at 06:28:41, Alessandro Damiani wrote: > >>On June 02, 1998 at 06:49:16, Inmann Werner wrote: >> >>>Hello! >>> >>>Making programs faster is making them playing better.(?!) >>>OK >>>But in a match, if there is a time limit, it is important, that the >>>program thinks long at the right time, and not, when everything is >>>clear. >>>But to implement that is more then difficult. When is a move "clear"? >>>Maybe deeper, there is a fine combination. When should the program >>>take time to look deeper? >>> >>>any suggestion, which are common? >> >>Hi Werner! >> >>One idea I had some days ago was to use the definition of singularity at >>the root to decide whether to start the next iteration or not: >> >>If the best move is at least S better than all other moves, then stop >>the search. Since we know the score of the best move, the cost is >>testing the siblings with a null-window. >> >>So, when one move is obviously better than the others, the program will >>stop and play that move. Seems ok to me. But: is the cost worth it or >>not? I think not, but perhaps I am wrong? >> >>Ciao >> >>Alessandro > >this will make you look like a genius at times, and a fool at other >times. > >Here is a sample position: > >5r1k/6p/1n2Q2p/4p//7P/PP4PK/R1B1q/ w > >This is from the game Cray Blitz vs Belle, at the 1981 ACM event. This >has >shown up in more than one "test suite" position. At this tournament, we >were running in a "batch processing backup system" with no thinking on >the >opponent's time or anything. I had to submit a batch job, and then keep >checking to see when it finished. We were using about 30-45 seconds per >move because the machine was not dedicated (this was a backup machine as >our primary machine had died). > >Here's the gist: at short searches, Qxb6 seems to win material. If you >notice this and use your idea to notice that this move is way better >than >any other move at the root, you can quit the search early and make a >really >silly mistake, because you can win a piece, but get killed in the >process. > >In this game, Cray Blitz actually played Qxb6 and lost. Crafty finds >that >Qb6 is bad and finds the forced draw (Bxh6) in just a few seconds. But >if >it quit because it thought that Qxb6 was far better than all the other >moves (root singular) it would be "surprised..." Actually I don`t use this idea, because of the cost. I forgot to say that of course there should be an additional condition to stop the search, for instance: stopping is allowed if half of the time has been used. Alessandro
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.