Author: blass uri
Date: 10:23:13 04/03/00
Go up one level in this thread
On April 03, 2000 at 11:52:08, Dave Gomboc wrote: >On April 03, 2000 at 02:59:43, Ed Schröder wrote: > >>On April 02, 2000 at 18:41:22, Dave Gomboc wrote: >> >>>On April 02, 2000 at 18:24:06, Ed Schröder wrote: >>> >>>>On April 02, 2000 at 16:33:24, blass uri wrote: >>>> >>>>>On April 02, 2000 at 15:46:54, Dave Gomboc wrote: >>>>> >>>>>>It sounds like your tunable policy paid off, in terms of people discovering good >>>>>>adjusted parameters? >>>>>> >>>>>>Dave >>>>> >>>>>Chess knowledge=25 is not an idea of me. >>>>> >>>>>I suggested to use it only after I read in Ed's site that Ed found that it is an >>>>>improvement against computers and maybe also an improvement against humans. >>>>> >>>>>I found also (after I read that Ed found that knowledge=25 is an improvement) >>>>>that Rebel(knowledge=0) does not lose in a result of almost 60:0 against >>>>>Rebel(chess knowledge=500) at 7 plied depth and it convinced me that Rebel knows >>>>>some important things even with knowledge=0. >>>>> >>>>>I think that the numbers of the chess knowledge parameter were misleading and >>>>>the minimal number should be clearly bigger than 0. >>>>> >>>>>I believe that this is the reason that people did not try to reduce the chess >>>>>knowledge parameter in the Rebel century personality contest. >>>>> >>>>>Uri >>>> >>>>I also did not expect it either that lowering the Chess Knowledge parameter >>>>would make Rebel stronger. In my tests I always increased the value of >>>>the Chess Knowledge parameter. When I did a test with [Chess Knowledge=25] >>>>(it was just curiosity) I was surprised to see the enormous speed gain of >>>>the search. Then [Chess Knowledge=25] suddenly had my full attention. >>>> >>>>So the improvement was discovered by accident. No real surprise as most >>>>of the time it goes that way. Chess remains a mystery, it is like a maze >>>>of 2^64 entries and no exit. >>>> >>>>Ed >>> >>>That's kind of funny, but I suppose it shouldn't be completely unexpected. If >>>you try increasing the amount of knowledge, and it performs worse, it makes >>>sense to try reducing it instead... and there have been other stories of people >>>removing knowledge from their programs before. >>> >>>Dave >> >>It is more complicated than that. The higher one set the [Chess Knowledge] >>parameter the better quality of moves the thing will produce. Just play >>a 100 games engine-engine match based on a fixed depth time control. The >>higher you set the [Chess Knowledge] parameter the bigger the victory. >> >>So the [Chess Knowledge] parameter works and does the job it is supposed >>to do. On the other hand a high value of the [Chess Knowledge] parameter >>is responsible for loss in speed during the search which may lead to a >>loss of a complete (iteration) ply. And losing an iteration ply seems to >>be more valuable than the gain of the extra chess knowledge. No doubt this >>is true for comp-comp but what if the topic is human-comp? Until now I >>really can't tell. So far I have chosen that [Chess Knowledge=25] is also >>better against humans. Am I completely sure? No... >> >>Ed > >I wouldn't be sure of it either. I do think it's a good idea to trade off some >search prowess to get better pattern recognition vs. humans, though. Even if >it's not stronger (and that is a possibility, as you suggest), it's probably >more pleasing to play against. The question is if increasing the chess knowledge from 25 to a bigger value gives important knowledge that you can practically never see by search because you need to search 20-50 plies forward to see the same thing by search with chess knowledge=25. I guess that it does not happen or almost does not happen otherwise using chess knowledge=25 could not be a good idea. 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.