Author: Vincent Diepeveen
Date: 05:36:28 08/08/00
Go up one level in this thread
On August 07, 2000 at 05:58:44, Andrew Williams wrote: >On August 06, 2000 at 20:10:49, Vincent Diepeveen wrote: > >>On August 06, 2000 at 19:17:18, Tom Kerrigan wrote: >> >>>On August 06, 2000 at 16:37:24, Vincent Diepeveen wrote: >>> >>>>On August 06, 2000 at 12:45:11, Dan Andersson wrote: >>>> >>>>>Vincent has had this idea of MTD and never managed/bothered to defend it. I >>>>>believe it to be an unsupported opinion. >>>> >>>>No commercial program uses MTD. End of proof man. >>> >>>I thought the MP version of Fritz does. >>> >>>-Tom >> >>I never saw any MP version of Fritz in the shops so far, >>perhaps someone is gonna state soon that DB used MTD too. > >Oddly enough, this seems to be what Hsu says in his IEEE Micro article. >Unfortunately, he doesn't say quite enough to be clear: > > "The search control does not really implement the regular > alpha-beta search algorithm [Ref: Knuth & Moore 1975]. Rather, > it implements a minimum-window alpha-beta search algorithm > [Ref: Pearl 1984]" He's talking about the hardware processors. Using RAM at hardware is tough or something (i'm not a hardware designer). It's easier to give it a single search value and a position. >This is a bit ambiguous, because of course PVS could be called a "minimum >window algorithm". But the rest of the paragraph (which is too long to type >here) does seem to suggest that DB was using something more like MTD than >PVS. I don't know if Bob knows for sure (maybe it's in Hsu's book?). Either >way, I'd recommend looking at the article, "IBM's Deep Blue Chess Grandmaster >Chips", Feng-hsiung Hsu, IEEE Micro March-April 1999. The relevant section >is "Search Control" on page 80. > >Having said all that, I think your argument about commercial programs and MTD >is flawed (whether DB used MTD or not). The problem is that MTD is a relatively >new technique, like bitboards. AFAIK, no commercial program uses bitboards Where is that WORST CASE ANALYSIS OF MTD? I can design in an easy way something that solves all testsets, if i know the testset in advance. Just enter the position, give it a huge bonus to reach a certain position, and there you go. You search a few plies deeper then with the same program. It's like in opening. You can search easily a ply more by giving a knights at f3 and c3 a HUGE bonus. >either. I know you don't like that technique, Vincent, but no sane person >would say that the fact that they're not widely used in commercial programs >"proves" that they're no good as an approach to creating chess programs. >Andrew What i mean is: the whole publication of all those articles is a bit let's call it this way: "amateurish". Testing at positions that always fail high is not exactly a good design point to start with. Not to mention proof. However commercial programmers, all investing bunches of time, and investigating all kind of new algorithms, they all have spent more likely more time in MTD as the inventors of it. And they all don't use it. Greetings, Vincent
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.