Author: Robert Hyatt
Date: 09:08:31 05/21/01
Go up one level in this thread
On May 21, 2001 at 10:13:58, Uri Blass wrote: >On May 21, 2001 at 08:58:41, José Carlos wrote: > >>On May 21, 2001 at 06:52:52, Uri Blass wrote: >> >>>This position could happen in kasparov-Deeper blue but kasparov did not play >>>40.gxh5 >>> >>>Based on Deeper blue logfiles we know that Deeper blue evaluated the position >>>as 0.5 pawn advantage for white at depth 9(6)=15 >>> >>>Crafty's evaluation at depth 15 is 2.93 pawns for white so we can see that the >>>difference in evaluation is 2.43 pawns. >>> >>>Part of the difference can be explained by different evaluation but most of the >>>difference is clearly because of the fact that Crafty can see more than Deeper >>>blue. >>> >>>Deeper blue found the line 40...Bxh5 41.Rh1 only at depth 10(6)=16 when Crafty >>>could see it at depth 14 and the line was not changed at the next 2 plies. >>> >>>Crafty can see the line 40...Bxh5 41.Rh1 Bf3 42.Rg1 Bh5 43.g7 Nf3 44.Rg3 c5 >>>45.bxc5 Rc8 46.f6 Rxc5 at depth 15 and 16 when Deeper blue can see it only at >>>depth 17,18(deeper blue can see a different line but the difference is only >>>41...Kg5 42.Rg1 Kh6 that leads to the same position as 41..Bf3 42.Rg1 Bh5 and >>>moves 43-46 are the same. >>> >>>It seems that Crafty is a very good program and it can outsearch deeper blue by >>>2 plies. >>> >>>I do not know the meaning of the plies of deeper blue but it seems to be not >>>brute force search. >>> >>>I also doubt if deeper blue could search 200M nodes per second otherwise I >>>cannot understand the reason that it is outsearched by other top programs again >>>and again based on the pv's >>> >>>It is possible that the claims about 200M nodes per second were only a >>>psychological war against kasparov >>> >>>Here is the analysis of Crafty >>[snip] >> >> Not sure what you mean. If, for example, I choose not to prune at all and >>extend a lot, I'll see more than anyone at the same depth. The problem is that >>I'll never reach the same depth as the rest with the same time. >> So, maybe Crafty sees more at the same depth, but that means nothing. The >>right comparison would be >> >>who sees more _with the same time_? >> >> José C. > > >The point is that the claim about deeper blue was that deeper blue did more >extensions than other programs and no pruning(at least no null move pruning). I'm not sure who claimed that. I only claimed that they did SE, and not many are doing it, and those that are doing it are not doing it the way DB did. Their search _is_ better than a null-move search, for a simple reason. All you need to do to see why is a position where you sacrifice a queen at ply=1, and mate your opponent at ply=23. At many points in the tree, you reach a position with your opponent to move, and he says "if I make a null-move here, and search a shallower-than-normal tree, what happens? I get a value of beta back, which means that even if I let my opponent make two moves in a row, I am still winning (in reality I am a queen up so it looks like I am winning). Therefore I am simply going to return beta here and quit after the shallower search." Of course, if a _real_ search was done, the program would realize it is a queen up but it is getting mated. But chopping off 2-3 plies is enough to hide the mate and convince it everything is ok. A non-null-move program will _never_ make that mistake. Null-move doesn't just fail with zugzwang positions, it fails for any position where one side is ahead in material signficantly, while the other side has a threat that can only be seen with a very deep search. The shallow null-move search misses the threat and returns a bogus score. DB doesn't suffer from that _ever_. Crafty does. As does Fritz, and anybody else doing null-move in _any_ form whatsoever. > >It is clear that Crafty can see deeper with the same number of nodes that deeper >blue searched. Based on my positional evaluation of a particular sequence of moves? Suppose that passed pawn doesn't win? > >If you talk about the same time than it is dependent on the hardware and deeper >blue could see deeper than Crafty(p800) but it seems that Crafty on good alpha >could see deeper than deeper blue. I can realistically see a 30M node per second alpha box today. Which will act like 20M nodes. I won't ever think that will be better than DB tactically. 10:1 is a horrendous time handicap. You can try this yourself with any two programs you want. > >Deeper blue seems to be only 2-7 times faster based on the relevant position. > Again, you are drawing conclusions with _exactly_ 1/2 of the relevant data. What did DB evaluate? What does its PV look like to the endpoint where it did the evaluation? You only have info from crafty, and you aren't paying attention to how much positional bonus there is in that PV/score. As I said before, Crafty is very speculative. It has to be to survive with GM players at the speed it currently runs. That doesn't make it better or worse than DB, just "different". >The time that deeper blue got depth 10(6)=16 is 27 seconds to fail low and 65 >seconds to solve the fail low when Crafty got depth 14 that is eqvivalent based >on the main line after 179 seconds. Again, the eval of crafty is well-known. DB's is not. Until you know what they evaluate and how big the scores are, comparing them is wasted time. Is this a tactical fail low or a positional fail low? No way to know without their PV and we don't have it. what about crafty? Did the score change materially or positionally? you can find that out. > >179/27 is the biggest estimate for the speed ratio based on the relevant >position I can find positions where Crafty sees something 10 plies before _any_ other program in the endgame. I can also find positions where it sees something 10 plies _after_ any other program too. I'm not sure what picking one of the former would prove about overall chess skill... > >The time that deeper blue got depth 11(6)=17 is 229 seconds when Crafty could >get similiar main line at depth 15 after 536 seconds. Find an older version of Crafty. It could solve WAC2 in 3 plies. It took DB 12 plies the last time I saw the output. Is crafty that much better? Or was that two-connected-passed-pawns-on-sixth-rank term too speculative? I don't see how you can compare them with only 1/2 of the data you need. > >536/229 is the smallest estimate for the speed difference based on the relevant >position. > >Uri
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.