Author: Robert Hyatt
Date: 21:01:01 05/13/98
Go up one level in this thread
On May 13, 1998 at 16:20:50, Don Dailey wrote: >On May 13, 1998 at 13:57:46, Robert Hyatt wrote: > >>On May 13, 1998 at 13:26:26, Don Dailey wrote: >> >>> >>>Before the 1993 ICCA tournament me and Larry put our last tweaks in >>>Socrates, played 2 test games with Genius on equal hardware and won >>>both games!!! Wow, that's 100% score. We can beat Genius EVERY >>>time! We went on to actually win this tournament, but subsequent >>>testing showed we were weaker than the Genius we tested against. >>> >>>- Don >> >>This is a "typical programmer analyis" by Vincent... but he forgets >>that >>*everyone* has bugs. In fact, in the last two weeks, against one of his >>three copies on ICC two games were lost by crafty in totally won >>positions, >>because I had a SMP bug that would get crafty/xboard out of sync, crafty >>would keep thinking and never see the move, letting its flag fall. So >>there's >>no way to say that "these 4 games" are more revealing than "those 100 >>games" >>because in those 100 games "I had a bug". In "those 100 games, probably >>*both* had bugs.... > >Hi Bob, > >I'm a firm believer that bugs can kill more rating points than anything >else. This program I mentioned in my post was the most debugged program >I ever brought to a tournament, lo and behold it manages to win! > >But just as important is the sample size. 4 games are ridiculous few to >base any strength claims on. In fact a single 5 round tournament is not >very telling either in my opinion, except that your program has to be >in good shape to win. In this same tournament, we beat Genius, Mchess >and >Rebel and got a draw against Cray Blitz remember? But I don't believe >my program was better that any of these programs. It was just "within >striking range" of them which means (by my definition) that's it's not >so much weaker that it's unlikely to win an individual game. > >I have a question for you which I'll bet you've answered before for >others: > >From a pure software standpoint, is Crafty a better program than Blitz? >Please don't consider the engineering factors like cray assembler or >special highly hardware dependent things. > >- Don It depends... because the pure software is really tied to the Cray. Cray Blitz could do some things that Crafty can't do yet... but the speed gap is now likely gone since the parallel search is working. I can't hit 10M nps on any known alpha/intel machine yet, as I could on a T90, but I can get close... So lets call it "speed: even" so long as we all understand that a T90 sells for 60 million dollars, while I currently run on a 4 processor P6/200 machine. search is interesting... CB had a traditional q-search with captures, checks, checkmate, and the like. Crafty's is much streamlined with only winning/even captures and nothing else (promotions to Q also, but that is roughly a capture). I like the simple approach as I am now only concerned with the "normal search" code... which simplifies debugging and modification. Here, the advantage is unknown. Extensions. The last version of CB used singular extensions, and it was really capable of some incredible long-range tactics... I don't do that in Crafty yet, but with the newly found horsepower in parallel searching, I am fixing to start this testing again to see whether it is worthwhile or not... So extensions favor CB. Null-move. Cray Blitz was *very* conservative at R=1, non-recursive, while Crafty is using R=2 recursively. As a result, for nominal search depths, Crafty would actually get deeper... but with singular extensions, it would be "no contest" in tactics... CB is *deadly*. So, which is better? No real idea. Practically, CB is, because of the hardware and the 20,000 lines of assembly that make everything vectorize and run like the blazes. But, Crafty does some things better, particularly in evaluation, because the bitmaps are faster and easier, although the vector hardware on the cray made the eval fly there too. I'd guess that CB and Crafty are "close" in positional skill, maybe... Crafty is probably much better "tuned" due to ICC, but CB's eval was along the same lines as far as things evaluated and number of terms in the eval... at least in the same ballpark. However, one huge advantage for crafty: platform. Getting a Cray is nearly impossible for a tournament. My 4 processor machine sits in my office.. it is *mine*... that's a *huge* difference. Crafty's parallel search will be better than CB's because of the amount of test time I have, where I had almost none for CB...
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.