Author: Vasik Rajlich
Date: 02:21:52 12/24/05
Go up one level in this thread
On December 24, 2005 at 03:19:46, Uri Blass wrote: >On December 24, 2005 at 03:05:23, William Penn wrote: > >>Vas, >> >>Sorry, "hanging" or "hung" was a poor choice of words. (I was a little bit >>despondent when I wrote that.) What that implies did not happen. The engine and >>GUI were still fully functional and active. There was no problem with the >>various operations. >> >>What I meant to say was that nothing further appeared in the analysis display >>window after about 9 minutes. The engine was still active. It just wasn't >>providing any more PV outputs at greater depths. Possibly it would have, if I >>had let it go for a longer time, but I doubt it. >> >>Actually in this position, there is no doubt that the best move was already >>selected after 9 minutes. The game is almost over, a quick win. In that case it >>looks like Rybka simply decided there is no purpose for further analysis. So it >>is not necessarily a bad behavior, it is just different compared to how other >>engines behave. Naturally we would like to see the analysis and PV outputs >>continue, but that's not essential if Rybka's primary objective is simply to win >>the game. >> >>--------- >> >>I just ran the following long analysis in infinite mode which shows something >>different. The nodes and nodes/second numbers are behaving strangely after level >>20.01. It looks like the nodes counter is recycling to zero after about >>500,000,000 nodes by my estimate. We used to see something similar in the older >>CB engines. Easy to fix, they just had to make their counter able to count to >>bigger numbers. I would guess this happened because it looks like you tripled >>those output numbers for cosmetic reasons... :) >> >> (game#243) >> >>[D]5qrk/pp1r3p/n1p2p2/P1p1pNp1/4P2P/1P1PQ1P1/2P2PK1/R6R w - - >> >>Engine: Rybka 1.01 Preview 2 32-bit (704 MB) by Vasik Rajlich >> (run 3-PV for a few minutes) >>15 2:36 +1.31 33.hxg5 Rxg5 34.Qf3 Nc7 35.Rh5 Ne6 >> 36.Rxg5 fxg5 37.Rh1 Qe8 38.Qd1 b6 >> 39.axb6 (16.943.250) 110 >>15 3:18 +1.06 33.Ra4 g4 34.Rc4 h5 35.Rd1 Rg6 >> 36.Ra1 Kg8 37.Kh2 Qe8 38.f3 gxf3 (21.431.195) 110 >>15 2:59 +1.03 33.Qd2 Nc7 34.hxg5 Rxg5 35.Qc3 Nb5 >> 36.Qc4 Nd4 37.Rh4 Nxf5 38.exf5 Qd6 >> 39.Qe6 Qd5+ (19.457.399) 110 >> (then reduce to 1 PV) >>11.01 0:01 +1.32 33.hxg5 Rxg5 34.Qf3 Nc7 35.Rh5 Ne6 >> 36.Rxg5 fxg5 37.Rh1 Ng7 38.Qg4 Ne6 >> 39.Rh6 Qe8 (195.954) 111 >>12.01 0:02 +1.33 33.hxg5 Rxg5 34.Qf3 Nc7 35.Rh5 Ne6 >> 36.Rxg5 fxg5 37.Rh1 Ng7 38.Qg4 Ne6 >> 39.Rh6 Qe8 (267.392) 110 >>13.01 0:04 +1.28 33.hxg5 Rxg5 34.Qf3 Nc7 35.Rh5 Ne6 >> 36.Rxg5 fxg5 37.Rh1 Ng7 38.Qg4 Ne6 >> 39.Rh6 Qe8 (513.071) 108 >>14.01 0:12 +1.30 33.hxg5 Rxg5 34.Qf3 Nc7 35.Rh5 Ne6 >> 36.Rxg5 fxg5 37.Rh1 Ng7 38.Qg4 Ne6 >> 39.Rh6 Qe8 (1.386.246) 110 >>15.01 0:24 +1.22 33.hxg5 Rxg5 34.Qf3 Nc7 35.Rh5 Ne6 >> 36.Rxg5 fxg5 37.Rh1 Ng7 38.Qg4 Ne6 >> 39.Rh6 Qe8 (2.608.983) 110 >>16.01 1:29 +1.30 33.hxg5 Rxg5 34.Qf3 Rg6 35.Rh5 Kg8 >> 36.Nh4 Rgg7 37.Rh6 Nb4 38.Rxf6 Qd8 >> 39.Rc1 (9.846.132) 113 >>17.01 5:49 +1.23 33.hxg5 Rxg5 34.Qf3 Nc7 35.Rh5 Rxh5 >> 36.Qxh5 Ne6 37.a6 b6 38.Rh1 Qg8 >> 39.Qh6 Qg5 (38.723.297) 113 >>18.01 12:09 +1.18 33.hxg5 Rxg5 34.Qf3 Nc7 35.Rh5 Qg8 >> 36.Rh6 Qf8 37.Rah1 Ne6 38.Kg1 Kg8 >> 39.R1h4 Qd8 (81.225.326) 114 >>19.01 26:31 +1.14 33.hxg5 Rxg5 34.Qf3 Nc7 35.Rh5 Qg8 >> 36.Rh6 Qf8 37.Rah1 Ne6 38.R1h4 Kg8 >> 39.Kg1 Qd8 (178.202.522) 114 >>20.01 46:03 +1.26 33.hxg5 Rxg5 34.Qf3 Nc7 35.Rh5 Qg8 >> 36.Rh6 Qf8 37.Rah1 Ne6 38.R1h4 Kg8 >> 39.Kg1 Qd8 (309.766.912) 114 >>21.01 84:55 +1.22 33.hxg5 Rxg5 34.Qf3 Nc7 35.Rh5 Qg8 >> 36.Rh6 Qf8 37.Rah1 Ne6 38.R1h4 b6 >> 39.Qd1 b5 (35.438.207) 7 >>22.01 184:10 +1.20 33.hxg5 Rxg5 34.Qf3 Nc7 35.Rh5 (188.955.970) 17 >>23.01 554:16 +1.19 33.hxg5 Rxg5 34.Rad1 (46.840.269) 1 >> >>WP > >This is strange because 500,000,000 nodes or 600,000,000 nodes that are supposed >to be searched after 554 minutes are not enough to be bigger than 2^31 so I see >no logical reason for the number of nodes to do down. > >If the real number of nodes is higher and in printing you divide them by 4 or 8 >then I can understand it but I see no reason to do it. > >Uri This was also fixed in Rybka 1.01. This version now gives credit for time-consuming "search like" work, as otherwise the knps can drop horribly in some positions. Vas
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.