Author: Robert Hyatt
Date: 08:21:18 08/21/02
Go up one level in this thread
On August 21, 2002 at 10:35:13, Vincent Diepeveen wrote: >On August 21, 2002 at 10:26:57, Terry McCracken wrote: > >hello, > >Another thing. > >In 1998/1999 Hyatt claimed that deep blue searched 11 to 12 ply, >but that their *extensions* were better than everyone else. > >They used singular extensions. > >In 2000/2001, i also was using singular extensions, as well as >several other programs. And as I said, your implementation is pitiful compared to _real_ SE as implemented in Deep Thought/Deep Blue, Cray Blitz and HiTech. You totally fail to handle the FH-singular case which is complex and expensive. And you _also_ fail to remember that you have said _many_ times "singular extensions suck and what I do is much better". Only later you discover that your _implementation_ sucked and than when you finally got it right, then it did work pretty well. Which is typical for you, of course... > Then suddenly when we searched above 11 to 12 ply >depths, it was said that the 12(6) of the machine which means 12 ply >nominal search depth, was excluding 6 ply hardware search depth. Directly from the team, remember. I posted the email _right here_ to make it public... > >Which is bloody idiocy in itself, because the thing had no hashtable. > >The big theoretician Knuth has written a lemma for game tree search. > >The minimum tree to search at 18 ply search depth using alfabeta >(without nullmove which wasn't used by deep blue): > >2 * (squareroot(number of legal moves) ^ depth) > >or: > >2 * sqrt(40)^18 = 524288000000000 nodes needed to search it *minimum*. > >As you can imagine, getting 11 to 12 ply fullwidth search was already a >very good achievement in 1997. Again, your theoretical explanation is wrong. How do you explain that Knuth's "lemma" (as you wrongly call it) predicts a branching factor of > 6, when we can _prove_ that DB's branching factor was under 4? Again, the logs reveal everything, but you _do_ have to open your eyes first. > >Also the (6) can't be hardware search depth as the paper clearly says >they searched nominal 11 to 12 ply including hardware search depth and >the hardware could do at most 4 to 5 ply in hardware. 4 in middlegame, >5 in endgame. *not* 6. Vincent, you are simply trying to make what the group wrote "bend" into saying what you want it to say. Belle searched _six_ plies in hardware. Two in software. So 6 was definitely possible. Hsu claimed that DT searched 4-5 plies in hardware. The quote you are citing pretty vague, taken in context. > >They optimistically estimated their number of nodes at 126 million >nodes a second. From that about 20% were *effective* the other 80% >was lost by parallellism, as you can see in the document. That is also wrong. He said the search was 7% effective, not 20%. And as I said, when I talked to him he used 7% as a number to be used against the peak performance of the machine, which was 1B nodes per second based on simple math. That turns into maybe 70M nodes per second in terms of single-processor equivalent. > >That is scientific not a very nice thing to do, as they cheated how they >got the 20% number. Even then it is 25MLN nodes a second, which >is not enough to search 524288000000000 nodes in 3 minutes. Of course, if they don't need to search that many nodes, then it isn't an issue. You haven't yet answered the branching factor question. I don't think you ever will either. Just wave hands and go back to the ranting... > >Best regards, >Vincent > > >>On August 21, 2002 at 09:28:03, Uri Blass wrote: >> >>>On August 21, 2002 at 09:00:19, Terry McCracken wrote: >>> >>>>On August 21, 2002 at 08:22:06, Vincent Diepeveen wrote: >>>> >>>>>On August 21, 2002 at 08:13:47, emerson tan wrote: >>>>> >>>>>>wHATA IS DEEP BLUES AVERAGE PLY LEVEL ON TOURNAMENT TIME CONTROL, I UNDERSTAND >>>>>>THAT IT CALCULATES 200 MILLION POSITIONS PER SECOND BUT NEVER HEARD OF ITS >>>>>>AVERAGE PLY. THANKS >>>>> >>>>>they estimate it at 126 million nodes a seconda gainst deep blue. >>>>> >>>>>It had a nominal search depth (depth limit) of 12 ply. With a lot of >>>>>tactical extensions a lot of lines were searched at 17 ply though. >>>>> >>>>>that's however true for all chess programs. I averaged in DIEP way >>>>>deeper. It depends basically whether you count hashtable cutoffs with >>>>>the depth or not. I tend to not count them. In hardware there was not >>>>>a hashtable. >>>>> >>>>>Average search depth says nothing. The nominal search depth is more important. >>>>>This was 12 ply. So the weakest link were lines of 12 ply simply. >>>>> >>>>>Best regards, >>>>>Vincent >>>> >>>> >>>>Deep Blue II did searches with extensions of 17 ply + at 3 min. per move and >>>>wasn't uncommon to actually hit 22 plys. >>>> >>>>Dr. Robert Hyatt watched this with his own eyes, so I'm sure he'll have >>>>something relevant to say on this matter. >>> >>>Nobody watched the tree of deep blue. >>> >>>I remember from previous post that people in the deep blue team >>>said that the brute force depth was only 12 plies. >>> >>>The output of deeper blue also does not say 17 plies. >>>There are people who claim that 12(6) means 18 plies >>>brute force but the deep blue team never claimed it. >>> >>>I do not know what it means. >>> >>>It may mean 12 plies brute force with some selective search >>>of the next 6 plies(of course there were extensions but >>>the number 12 6 are not about extensions because >>>it is clear that the longest lines that they searched >>>were clearly more than 18 plies). >>> >>>Uri >> >> >>My mistake here is Dr. Hyatt read the logs, and I should have stated he read >>the logs...as no, he didn't see the readouts live at the Kasparov Deep Blue II >>Match. >> >> >>Terry >> >>P.S. >>Deep Blue II >> >>Was the fastest, most powerful Chess Computer ever built with the exception of >>Deep Blue III which never played Kasparov and was scrapped, and 5 years later >>it's still true and you can take that to the bank! >> >>Terry
This page took 0.01 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.