Author: Uri Blass
Date: 10:30:40 03/03/02
Go up one level in this thread
On March 03, 2002 at 11:34:25, Sune Fischer wrote: >On March 03, 2002 at 08:46:11, Uri Blass wrote: > >>On March 03, 2002 at 08:16:14, Sune Fischer wrote: >> >>>On March 03, 2002 at 07:52:43, Uri Blass wrote: >>>>>I think is is more interesting to ask these question at the higher plies. >>>> >>>>I think that you prefer to ask questions that are practically impossible to >>>>answer. >>> >>>Well, yours is pretty far fetched also ;) >>> >>>>We cannot play games at 20,21,22 plies today. >>> >>>I suppose a million games to ply three is more realistic ;) >> >>Yes >> >>one game at depth 3 may take 1 second with the speed of the >>computers of today and wasting one million seconds or even >>few million seconds is something realistic. > >Hmm, maybe, but I think you will need a faster interface than winboard. > >>>>I think that before asking the question about the rating of the perfect player >>>>we need to ask if rating has a clear meaning. >>>> >>>>The rating is based on some assumption about the expected result of A against C >>>>based on the expected result of A against B and the expected result of B against >>>>C. >>>> >>>>This assumption is based on the normal distribution and not on games. >>>>It is possible that we can get better assumption by playing games in the way >>>>that I suggest. >>>> >>>>suppose A1 beats A0 600-400 >>>>A2 beats A1 600-400.... >>>>A1000 beat A999 600-400 >>>> >>>>We may assume that the rating of A0 is 0. >>>> >>>>The question is what is the rating of A1000 and if we can expect the same rating >>>>for A1000 if we are going to play only 500 matches(A2 against A0,A4 against >>>>A2...A1000 against A998) >>>> >>>>Uri >>> >>>I'm guessing there would be 0.1 elo difference between A1000 and A999, you need >>>more than 500 games to reveal that within the standard diviation. >> >> >>Your assumption is wrong because if there is probability of 0.1% for a random >>move then I expect a random move to happen one time for 1000 moves. >> >>If we asssume that the average game take 60 moves then it means that >>there will be a random move in more than 1/20 of the games. >> >>I do not expect every random move to be decisive but I expect to see it often >>enough to get at least a result of 50.5-49.5 forn A1000 >> >>50.5-49.5 means 4 elo difference so I agree that the difference is going to be >>small but not 0.1 elo. >> > >So, because An+1 will in 1 of 20 games make 1 less random move than An, then >that will be enough to get 4 point more elo? >I seriously doubt it, it probably does translate to less than an elo, one less >random move does not translate to a won game. Even if it was 4 elo, you still >would need more than 1000 games to observe this within the standard diviation. > >>>In principle the list should be ordered as A1,A2,....A1000, you will probably >>>see some strange things like A500 beating A523, so I'm not sure what you hope to >>>prove by this experiment. >> >>If the match is long enough I am not going to be strange things but I agree that >>with only 1000 game per match with these small difference I may see strange >>things so it is maybe better to use matches of 10000 games and probabilities of >>100%,99%,98%,....1%,0%(I believe that with these probabilities I am not going to >>see strange things even with match of 1000 games so we can even save time. > >I think you should do the experiment, I wonder if they wouldn't all end in a >draw due to insufficent material, seems unlikely one would get mated. I know that most of the games between players who play random moves are draws based on a previous experiment but wins happen and there is no practical chance that 1000 games are going to end in a draw(I remember that 80% or 90% of the games between players who played random moves ended in a draw). I do not think to do the experiment(at least not in the near future) I prefer to waste programming time in improving my chess program and not in this experiment. It is today probably slightly better than tscp and I do not feel sure that with this level it can go to the 4th division. Today a summery of it's knowlegde is 1)piece square table evaluation including special piece square table for the endgame 2)qsearch 3)history tables 4)fail high,fail low 5)check extension 6)futility pruning. 7)some knowledge about time management I hope to add killer moves and some more extensions and evaluation of passed pawns and iterative internal deepening before the beginning of the 5th division in april. I hope also to do some speed improvement and to improve the futility pruning. I hope that it is going to be enough to get one of the first 6 places in the 5th division but I am not sure because it is dependent on the oponents. I am optimistic because a slightly slower version than the version of today that does not use futility pruning and has worse time management could get 3th place out of 10 new programs. I guess that the version of today(movei005f is 50 elo better than the version that got 3th place movei0041. I did not do enough tests to prove it and some evidence is: 1)movei005f-movei0041 7-5(1,2,3,4,5,6 minutes per game) 2)movei005f-tscp1.73 5.5-4.5(1,2,3,4,5 minutes per game) when movei0041 could only draw 5-5 under the same conditions 3)movei005f solved 283 position in wac when the previous version solved only 280. 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.