Author: enrico carrisco
Date: 01:15:33 02/08/06
Go up one level in this thread
On February 07, 2006 at 06:59:21, Uri Blass wrote: >On February 07, 2006 at 01:37:23, Dann Corbit wrote: > >>On February 06, 2006 at 22:25:35, Ryan B. wrote: >> >>>On February 06, 2006 at 21:58:39, Dann Corbit wrote: >>> >>>>On February 06, 2006 at 21:47:09, Ryan B. wrote: >>>> >>>>>On February 06, 2006 at 17:20:00, Alvaro Jose Povoa Cardoso wrote: >>>>> >>>>>>Some time ago I saw someone reporting (don't remember if it was here) Rybka >>>>>>doing 30Knps, while crafy 19.x doing 800Knps and Fritz8 doing over a million nps >>>>>>on his pc while beating both crafty and fritz. >>>>>> >>>>>>Over the years computer programmers said that nps was an importante factor in an >>>>>>engine performance, (and I agree, look at Deep Blue and Hydra). >>>>>>In face of Rybka nps, my question is: >>>>>>What the hell is going on? >>>>>>In this department Rybka seams to prove that nps isn't a factor for performance >>>>>>at all. >>>>>> >>>>>>Please state you opinion on this. >>>>>> >>>>>>best regards, >>>>>>Alvaro Cardoso >>>>> >>>>>Rather it is to hide extra hidden search or for marketing propaganda I do not >>>>>know but I do know Rybka manipulates its node count. >>>> >>>>By what means do you know this? >>>> >>>>>What is really important >>>>>is the tactical strength of Rybka not the NPS. It will be interesting to see >>>>>what other engines have similar success after going back to try “old” ideas. >>>>>Also not that Rybka is bad in endgames, I think it is rather good, but it will >>>>>be interesting to see how Vas handles the endgame issues in Rybka. It does not >>>>>seem that it will be as simple as just adding endgame knowledge. >>>>> >>>>> >>>>>Ryan >>> >>> >>>At low depths the node count can go down from one depth to the next. That >>>should be proof enough as node count should never go down. >> >>It could conceivably happen due to hash table and forced replies. > >Reduction in nodes per second can happen but not reduction in the total number >of nodes and I remember a post of enrique that showed reduction of total number >of nodes in rybka at small depth and you probably need slow hardware to >reproduce it. > >If you see Reduction in nodes per second after few seconds of analysis with >rybka then the conclusion is that Rybka probably manipulates its node count. > >Uri Hello Uri. I've realized that the technical side of this forum has been lost -- no one really cares about such things and the question of "why?" As long as they can set up engine tournaments and receive the end results -- who cares about the actual games, what makes the engines play the moves they do and search at the speed they do (or report.) It's no wonder Anthony decided to quit posting (as did a few others, seen by their inactivity...) As far as the post you were talking about above, here was the basic info: P233MHz Laptop with Rybka 00:00:00.3 0.31 3 24056 exf6 00:00:00.6 0.18 4 47432 exf6 00:00:02.8 0.26 5 200296 exf6 Bxf2+ Kxf2 Qxf6 00:00:06.1 0.33 6 6837 exf6 Bxf2+ Kxf2 Qxf6 Nb3 00:00:12.4 0.33 7 15424 exf6 Bxf2+ Kxf2 Qxf6 Nb3 Ne5 00:00:28.0 0.40 8 35355 exf6 Bxf2+ Kxf2 Qxf6 Nb3 Ne5 Nc5 00:01:03.7 0.47 9 80989 exf6 Bxf2+ Kxf2 Qxf6 Nb3 Ne5 Nc5 Rae8 00:02:12.3 0.45 10 174555 exf6 Bxf2+ Kxf2 Qxf6 Nb3 Ne5 Nc5 Rae8 Kg1 00:00:03.9 -0.34 3 95224 Bf5 00:00:04.2 -0.33 4 118472 Bf5 00:00:05.1 -0.33 5 178384 Bf5 Nb3 00:00:08.7 -0.40 6 8895 Bf5 Bxf5 Bxf2+ Qxf2 Qxf5 00:00:19.2 -0.44 7 22847 Bf5 Bxf5 Bxf2+ Qxf2 Qxf5 Qc5 00:00:36.4 -0.49 8 50814 Bf5 Bxf5 Bxf2+ Qxf2 Qxf5 Qc5 Qe6 00:01:12.5 -0.71 9 114651 Bf5 Bxf5 Bxf2+ Qxf2 Qxf5 Qc5 Qd7 a4 00:01:58.2 -0.66 9 182030 b4 Ba4 Bg4 Bxc6 Bxf2+ Qxf2 Qxc6 cxb4 00:00:01.1 0.37 3 78768 a4 00:00:01.8 0.54 3 132152 Qd3 00:00:03.4 0.55 4 245176 Qd3 00:00:05.5 0.55 5 390048 Qd3 Qf5 00:00:10.5 0.61 6 10084 Qd3 Qf5 Qxf5 00:00:14.0 0.61 7 15315 Qd3 Qf5 Qxf5 Bxf5 Bb3 00:00:42.3 0.66 8 49569 Qd3 Qf5 Qxf5 Bxf5 Bb3 Rad8 00:01:35.7 0.69 9 128294 Qd3 Qf5 Qxf5 Bxf5 Bb3 Rad8 Nh4 00:02:41.9 0.72 10 225286 Qd3 Qf5 Qxf5 Bxf5 Bb3 Be6 a4 Rae8 [D]r4rk1/2p3pp/p1n2q2/1pbp4/6b1/2P2N2/PPBN1RPP/R1B2QK1 w - - 2 1 Raw UCI output: position r4rk1/2p3pp/p1n2q2/1pbp4/6b1/2P2N2/PPBN1RPP/R1B2QK1 w - - 2 1 go depth 8 info depth 3 info depth 3 score cp 7 time 660 nodes 47096 nps 0 tbhits 0 pv b1c3 info depth 3 time 665 nodes 47536 nps 0 tbhits 0 info depth 4 info depth 4 score cp 3 time 1216 nodes 86824 nps 0 tbhits 0 pv b1c3 info depth 4 time 1374 nodes 98376 nps 0 tbhits 0 info depth 5 info depth 5 score cp 8 time 1559 nodes 111808 nps 0 tbhits 0 pv b1c3 b8c6 info depth 5 time 1669 nodes 119848 nps 0 tbhits 0 info depth 6 info depth 6 score cp 3 time 1934 nodes 139232 nps 0 tbhits 0 pv b1c3 b8c6 g1f3 info depth 6 time 2489 nodes 180056 nps 0 tbhits 0 info depth 7 info depth 7 score cp 5 time 3719 nodes 6159 nps 1695 tbhits 0 pv b1c3 b8c6 g1f3 g8f6 info depth 7 time 6625 nodes 9833 nps 1519 tbhits 0 info depth 8 info depth 8 score cp 6 time 9459 nodes 17667 nps 1912 tbhits 0 pv b1c3 b8c6 g1f 3 g8f6 d2d4 info depth 8 time 13654 nodes 23246 nps 1743 tbhits 0 info time 13658 nodes 23242 nps 1742 tbhits 0 bestmove b1c3 ponder b8c6 isready readyok position r4rk1/2p3pp/p1n2q2/1pbp4/6b1/2P2N2/PPBN1RPP/R1B2QK1 w - - 2 1 go depth 9 info depth 3 info depth 3 score cp 7 time 2890 nodes 205424 nps 0 tbhits 0 pv b1c3 info depth 3 time 2900 nodes 206224 nps 0 tbhits 0 info depth 4 info depth 4 score cp 3 time 4224 nodes 300392 nps 0 tbhits 0 pv b1c3 info depth 4 time 5716 nodes 406656 nps 0 tbhits 0 info depth 5 info depth 5 score cp 8 time 7690 nodes 547104 nps 0 tbhits 0 pv b1c3 b8c6 info depth 5 time 8755 nodes 622952 nps 0 tbhits 0 info depth 6 info depth 6 score cp 3 time 11895 nodes 846464 nps 0 tbhits 0 pv b1c3 b8c6 g1f3 info depth 6 time 20727 nodes 1474960 nps 0 tbhits 0 info depth 7 info depth 7 score cp 5 time 34393 nodes 6153 nps 183 tbhits 0 pv b1c3 b8c6 g1f3 g8f6 info depth 7 time 45825 nodes 9833 nps 219 tbhits 0 info depth 8 At ply 5 it was doing almost 71kn/s (but didn't show it), ply 6 it drops the total nodes searched from 390kn to *10kn*, then it shows kn/s (which was 960 nodes/second? 70 times slower?) In the raw UCI output you can see it not counting NPS until it gets "modified", then reports low end kn/s... Nothing more to really say about this -- the proof of the nps being manipulated is quite clear. Even the "don't confuse me with the facts, my mind is made up" folks should see that. What I don't understand is why not just display the truth (nps, pv, etc.) and say look how _fast_ *and* _strong_ it plays? Why deliberately _spend_ _effort_ to hide it? It's clear Rybka relies much more on search than eval. Search can make it look like the eval is "intelligent" as it sees many things coming much earlier than a good eval can -- at least for non-strategical things anyway, e.g. push passed pawn. A good example of the "Rybka high" causing people to exaggerate it's eval: http://www.talkchess.com/forums/1/message.html?484491 My response [in vain, of course]: http://www.talkchess.com/forums/1/message.html?484675 Anyway, enough wasting time -- let the fools be fooled and wallow in their ignorance. Regards, -elc.
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.