Author: Robert Hyatt
Date: 10:50:14 06/29/98
Go up one level in this thread
On June 29, 1998 at 12:42:20, Bruce Moreland wrote: > >On June 29, 1998 at 00:35:09, Robert Hyatt wrote: > >>Good question. But I just got some of the latest results from my chess partner >>at Livermore Labs... CB (running on a T932 (32 processor T90, 2 gigawords (16 >>gigabytes) of memory) gets 297 in under 1 second (it reports the time as 0 >>seconds, but only displays whole seconds. two take 1 second, and 1 takes 7 >>seconds. An older version didn't have that 7 second spike, not sure what caused >>it. This was on the last version we did of Cray blitz, called Cray Blitz >>version 43i. It is *very* fast, the output says it averaged just over 11M nodes >>per sec on the 300 positions. > >You can get Crafty going 200K nps on that 4x200. So CB is 55x faster than >Crafty, at least. > >55x7 = 385 > >You should get that tough one in 5-10 minutes with Crafty. > >You don't get it in a lot longer than that, and neither do I. Why do you think >this is? > >bruce It seems to be a combination of several things: 1. evaluation. CB was more "speculative" than most are aware of. My "two connected passed pawns on the 6th" that everyone comments about was not new to Crafty, for example. Came right out of CB. It has lots of other such things that I have not "tried" yet. 2. singular extensions. An example is that CB (and Crafty) can solve wac2 instantly, due to evaluation. Crafty then solves it tactically at somewhere between 10 and 12 plies depending on how it is tuned, extension-wise. I believe the current version actually sees material gain at depth = 10 or 11. CB sees material gain at depth=5!. Crafty is nowhere close to that, because of all those rook checks. But they are all "singular" because any non-check lets black promote easily. 3. other extensions. We did/do a lot of things that are maybe not safe if you don't have large amounts of compute power. IE with a king at g8, most everyone extends on Bxh7+ of Bxf7+. CB also extends on Bxg7 even though it is not a check. The score for wac230 is not a huge win, even by CB's estimation, as the score is only +3.3, while the score for Rf7 (the best move before it changes) is +2.5. So it's "a close call" as to whether it is really getting to the bottom of this or not. scores go +2.1 at depth=10, +2.3 at depth=11, change to rb4, +3.1 at depth=12, +3.2 at depth=13, and +3.2 (fail high but possibly a fail low following) at depth=14, (but it completes 14 without changing from Rb4). Note that it is *much* faster than crafty to reach these depths, even though it is extending like mad. I don't have the log file showing the search statistics or I could tell you how many of each extension was applied. I should know in a month or two whether this is a result of singular extensions alone, since that is next on my list for Crafty. But it is more likely simply a synergistic effect between lots of things we do there, perhaps even a bug or two for all I know. I do remember that when Harry and Hans Berliner were working on WAC several years ago (mid 80's or so) wac230 was one of the ones we both considered "hard" but which we solved. In fact, we've solved this one for a long time, but managed to miss others in various versions, due to excessive extensions driving the search wild. Those are my "guess" as to how we get this one. Crafty doesn't solve it, but it is *close*. IE at depth=12, Rf7=+2.3, while Rb4 is +1.6, and Rb4 is getting better every iteration (13=1.86 for example). So it *might* just be a positional thing that I do better in CB (or maybe not better in CB for all I know.) I *hate* tactical positions that can be solved positionally as it is never convincing to me that the program is really "right." And Harry has my old WAC book, so I'm not even sure what Rb4 is supposed to accomplish, since black is already winning. It looks like a variation on WAC2 in fact, where cxb4 leads to two strong connected black passers. Looked more closely at the "summary" output Harry sent. We play Rb4 but don't think white will recapture. the pv had Rb4 Ra4 and Ra2 with the black pawn advancing... so this could well be a simple passed pawn score that is higher in CB than in Crafty. Think I'll look more carefully, however, as I'd like to be the first micro to get 300/300. :) (if you don't beat me to it.. :) Bob
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.