Author: Robert Hyatt
Date: 21:30:49 06/28/98
Go up one level in this thread
On June 28, 1998 at 23:48:43, Bruce Moreland wrote: > >On June 28, 1998 at 20:32:58, jonathan Baxter wrote: > >>So all this recent discussion about DB says it doesn't null move, and I saw >>another post which said in some positions it was getting no more than 11-ply. >>Since PC based programs now reach that easily (without heavy forward pruning >>apart from Null move), but PC programs are not close to challenging Kasparov, >>can't we only conclude one of two things: >> >>1) DB has a much better evaluation function than the PC programs. >> >>and/or >> >>2) Null moving makes *lots* of mistakes. > >There's a lot more than this. Their search isn't just a normal full-width >search without null moves. In the past they have relied upon the singular >extension heuristic to produce trees that are extended at points they consider >crucial. > >So, the tree they search isn't just the normal full-width tree with the >null-move pruned branches filled in. > >Next, the strength of DB hasn't been accurately evaluated. What do we know >about the strength of something like Rebel or Hiarcs, who have played hundreds >or even thousands of games that we can examine? How do you really evaluate >these programs? If there is any question about how strong they are, then how >could there not be many more questions about how strong DB is, when we only have >six games from its newest incarnation? > >Also, even if it is very strong, the project was a hardware project, so it is >hard to compare it with software projects. The issue is computer power and its >effect upon rating against various opponents at various time controls, and while >this has been studied and conjectured about, I don't think that there has been >any conclusive result, instead all we have is an incomplete set of lemmas that >are repeated by a lot of people but aren't really proven true. > >So even if we had a better idea what they were doing, it would be hard to >evaluate the effectiveness of their algorithms. It is difficult to know how >their project would compare to a similar hardware project conducted by someone >who came from microcomputers, and chose to use techniques that are prevalent on >microcomputers. > >I prefer to look at the DT/DB projects as sources of ideas, rather than as a set >of blueprints. There are interesting ideas scattered throughout the project, I >think. > >But I don't think that the DB project proved anything at all about search or >evaluation techniques. I don't think that you can look at the project from our >point of view as outsiders, and say that any technique is shown to be good or >bad because of what they did. Anyone who tries to say the project proved >anything ends up making arguments based upon painfully little real evidence, >most of the evidence is anecdotal *at best*. > >bruce perhaps to you. Unfortunately, I have had the opportunity to sit across the board from them, with a program of known capabilities (Cray Blitz) and have seen first-hand what they can do. I don't know of *any* program that has only lost 2-3 games in 10 years of ACM and WCCC tournaments. That is a *huge* record. I don't know how good they are, to be sure, but I have a good idea of what Cray Blitz can do, and they are certainly better than it.
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.