Author: Brian Kostick
Date: 08:18:25 04/11/01
Go up one level in this thread
On April 10, 2001 at 19:45:34, John Merlino wrote: >On April 10, 2001 at 17:59:11, Brian Kostick wrote: > >>On April 10, 2001 at 12:18:33, John Merlino wrote: >> >>>On April 10, 2001 at 01:50:29, Paul Doire wrote: >>> >>>>Hardware:541Mhz celeron 128 mb ram win98se >>>>G/5, All engines have 32 MB Hash except CM default. >>>>All engine settings found here at this site >>>> >>>>Results: W D L Games Score >>>>CM KKND 32H 10 8 6 24 14.0 >>>>Chessmaster(default) 7 12 5 24 13.0 >>>>CM 32H 7 9 8 24 11.5 >>>>CM DEEP 32H(Kostick bk)6 10 8 24 11.0 >>>>CM 8777 32H 5 11 8 24 10.5 >>>> >>>>CM KKND is clearly stronger in this sampling. >>>>I have also played many offhand games with CM KKND against >>>>Gambit Tiger 2...G/5 same pc ponder on, and CM KKND appears >>>>to win 60-70% of the time! >>> >>>This (in some way) goes to prove one VERY IMPORTANT point that Johan and I have >>>stated time and time again. A 32MB hash table is 31MB too many for a G/5 game! >>>Johan uses the general rule of thumb of 1 byte per average position searched per >>>move. Since, on the average, CM8000 searches between 50K and 100K moves per >>>second (depending on processor speed -- your Celeron is going to be closer to >>>50K), this means that, unless the engine is going to have more than 10 seconds >>>average PER MOVE, there is NO REASON for anything more than a 1MB hash table. >>>Johan believes that, unless you're on a very fast processor, a 1MB hash table is >>>fine for everything up to G/10. >>> >>>The reason for the point that I'm making is that, with each new move, CM clears >>>out its hash table. A 32MB hash table takes approximately 1/2 second to clear. >>>In a G/5 game that goes 80 moves, this means that the engine has lost 40 >>>seconds, compared to a personality with an appropriate 1MB hash table. This >>>explains why the default personality had a better score than the 32MB version of >>>the default personality. >>> >>>HOWEVER, the next version of the patch (due out very soon), is going to fix this >>>problem to some degree. Johan has made the clearing of the hash table almost >>>"instant", meaning that using a too large hash table in a blitz game will no >>>longer be a detriment. >>> >>>As for your stating that KKND is "clearly stronger", I see absolutely no >>>evidence of that -- 1 point in a 24 game tournament? This is, statistically, >>>meaningless. >>> >>>jm >> >>To follow up, here are some game results. A bit over 100 games in each >>tournament. Please note WAY different machines AND time controls, so no use to >>compare one to the other. Err... the way it looks, maybe no need to compare >>different TT sizes either. Regards, BK >> >> >>Player Wins Draws Losses Games Score >>ChessmasterTT02 12 11 5 28 17.5 >>ChessmasterTT16 7 16 5 28 15.0 >>ChessmasterTT08 9 11 8 28 14.5 >>ChessmasterTT64 9 11 8 28 14.5 >>ChessmasterTT32 10 8 10 28 14.0 >>ChessmasterTT04 10 7 11 28 13.5 >>ChessmasterTT01 6 13 9 28 12.5 >>ChessmasterTT00 7 7 14 28 10.5 >> >>Computer: PIII 733MHz, 256MB Ram total >>Time Control: 5/3 Fischer Time >> >> >>Player Wins Draws Losses Games Score >>ChessmasterTT08 11 15 4 30 18.5 >>ChessmasterTT02 7 21 2 30 17.5 >>ChessmasterTT16 6 20 4 30 16.0 >>ChessmasterTT512kb 7 16 7 30 15.0 >>ChessmasterTT00 4 19 7 30 13.5 >>ChessmasterTT01 4 17 9 30 12.5 >>ChessmasterTT04 5 14 11 30 12.0 >> >>Computer: Pentium 60MHz, 64MB RAM total >>Time Control: 40moves/40minutes > >Quite honestly, I would suspect that the 1MB or 2MB versions would win a Fischer >5/3 tournament, and the 4MB or 8MB would win the 40/40 tournament (over the LONG >term, of course), but for different reasons. In the short time control, larger >hash tables are at a disadvantage because of the time to clear them for each >move (apart from just being wasted space). At longer time controls larger time >controls are more useful because they can store more data, and the time to clear >them is not as much of a percentage of the overall average move time. > >The longer the time control, the less this problem is detrimental to the engine. >It really only comes into play when we're talking about 15 seconds or less to >make a move, on average. > >Try this: 32 (or more) games at G/5 (or faster) with ChessmasterTT32 against >ChessmasterTT01. This may (or may not ;-) illustrate my point.... > >jm Well I don't think we need to really do this, especially if the engine code is going to be changed. For example, memset( ) for 64MB should not be too expensive. You might not recall but I've never been one of the bigger is always better people, sometimes quite the opposite. I think the tournament results I posted help support this. I think there are other settings that hold more interest. In the meanwhile I will delay any major time consuming test until after the next patch is released. Brian K.
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.