Author: Robert Hyatt
Date: 21:49:38 09/11/01
Go up one level in this thread
On September 12, 2001 at 00:00:21, Uri Blass wrote: > >1)The size of the tree is not enormous. >If it was so big humans had no chance to prove the draw or equal material by a >tree. First, the size of the tree _is_ enormous. It took many humans, searching all night long, to find the draw in game 2. It was _not_ found in minutes. It was not even found in a few hours. It took _many_ hours and _many_ people. Second, people are _far_ better at selective searching than computers are. This is why some players like Mike Valvo have totally thrashed programs like deep thought in correspondence-type time controls. The selectivity of the human mind is still amazing. > >I proved it by a tree when programs could see draw evaluation at every leaf of >the tree after a short search. Your analysis didn't prove a thing. It depends on _who_ wants the draw, as to who is going to favor 0.0 in the tree. If the program mis-evaluates a key position, and thinks the wrong side is winning, then it may well say "draw is best here" when it is not. Selectively running a program around in a tree doesn't prove a single thing, other than that it can be done. > >If you count the number of nodes in all the relevant small trees of short search >you can see that the tree is not enormous. If there are only two possiblities at every other node in the tree, that leaves a tree of 2^30 positions. That is enormous. And it is far smaller than the real tree that has to be searched. > >programs of today do not use singular extensions and my point is that with >singular extensions the size of the tree is not enormous so saying that programs >of today cannot see the draw is not a convincing claim. Neither is saying that the tree is not enormous, or that by following just a few moves you can force a draw according to a program. This is an argument that is going nowhere. > > >2)It is also interesting that you did not claim that the size of the tree is >enormous about the c5 move. The win there was _not_ 60 plies deep. At least not by my counting. It was closer to 20. There is a huge difference between 60 plies with queen checks and 20 plies without. > >This is exactly the claim of people who believe that deep thought could not see >+2 score against Cray blitz. > >For some reason there is no limit for the number of plies that Deep thought >could see by singular extensions and even if the +2 evaluation happens only 60 >plies after c5 Gian-Carlo Pascutto is going to use it in the discussion if >singular extension could help deep thought to see +2.xx evaluation. Nope... I just keep things sane. First, I _saw_ deep thought pop out that eval. It is not a claim. It is not a belief. It is a simple statement of fact, as seen by my own eyes with no hear-say involved at all. Second, I _know_ how deep the repetition is. Ed analyzed this and posted it on his web site. And I know that I don't see any feasible way for a program to search 60 plies deep when queens are on the board and there are checks everywhere that don't all have just one legal way out. > >Gian-Carlo Pascutto used lines of more than 40 plies after c5 and I say 60 plies >because the evaluation in the leaf was not static evaluation but evaluation that >is based on search. I didn't pay a lot of attention to his analysis. There is a fine line between what a program sees tactically with search and positionally with evaluation. I am not sure the distinction in his analysis was very clear. Because when some- one says "this is won for black" then off they go into deeper analysis to prove or disprove it. My program makes lots of positional assumptions that are based on material considerations. Such as a bishop trapped at a2, for example. And it can spot that with a shallow search even though winning the bishop might take a far deeper search than it can ever hope to complete... All I know about the DT game is that it said +2.5 roughly. And that somewhere around 10 moves later (it may have been 10, it may have been 5, that is a pretty vague memory) our score dropped to -2 and we lost the game. I saw it happen. It happened to my program. And it wasn't pretty, nor was I proud of it when it happened. But it _definitely_ happened, whether it was due to great play by them, lucky play by them, buggy play by them, or poor play by my program is something else to discuss. But it definitely did happen. > >Deep thought was also clearly weaker than Deeper blue so we can assume that if >Deeper blue could not see 60 plies forward then Deep thought could not see even >50 plies forward even with singular extensions. I'm sure DT _and_ DB could see 50 plies deep. Just not in positions where they were looking for a perp check. 50 forcing plies is one thing, but with a perp, every other ply generally has multiple legal moves, which makes the tree grow at an alarming rate. And with nothing being captured, the tree is not tapering down. That's why I don't believe someone can find that 60 ply perp from the starting position. Now or in the forseeable future. However I _know_ DT burned me in the game I mentioned. Because unfortunately, I had to sit through it. > >Uri
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.