Author: Uri Blass
Date: 21:30:04 12/21/05
Go up one level in this thread
On December 21, 2005 at 22:43:42, William Penn wrote: >On December 21, 2005 at 21:25:41, Uri Blass wrote: > >>On December 21, 2005 at 21:18:13, Tansel Turgut wrote: >> >>>The engine matches between programs seem to be done under fast time controls. >>>Does anyone know how good Rybka is in tournament mode (or in infinite mode) >>>compared to other programs? >> >>I do not know but it can change it's mind after a long time and I already saw it >>changing it's mind after 2 hours of analysis to a better move. >>I guess that it is better than other engines at all time controls. >> > >See >http://www.talkchess.com/forums/1/message.html?471773 > >I have also run a lot of infinite analyses with Rybka. The main thing I notice >is a reduction of the analysis outputs to 2 ply at longer time controls. This is >position-dependent to some extent, but it appears that eventually the analysis >will always truncate to 2 ply with any position. This can happen after only 20 >minutes, or may not have happened yet after 1000 minutes, but eventually it will >happen. The trend seems clear. So if you are looking for detailed analysis at >long run times (many hours), Rybka is not giving it in the current version. No program is giving details analysis that I want after long analysis and not only rybka. details analysis is not only about one line but about a tree of important lines that the programs considered. If the program suggested line that start with 1.e4 e5 2.Nf3 after 24 hours of search then it does not tell me what is the move that it has in the hash table after 1.e4 c5 and it must have some move in the hash when it starts to search 1.e4 c5 It is possible that the program considered 2.Nf3(after 1.e4 c5) for some hours and in previous iteration it decided that 2.Nc3 is good enough to refute 1...c5 and if it is the case then it means that I may need later hours to find what was the plan of the program after 1.e4 c5 I do not care about having no interface to support it and I would like to see some more information even in text file. The >tests in the link above show the same trend. The Rybka author has promised to >look at this. > >A separate question is how good the chosen move is at long run times, and >whether it actually improves if you let it run for a longer time. So is the move >selected better after 5 hours than after 5 minutes of run time? I don't have any >data on this question, although it's important to me. I have clearly data on this question and I know that the move that rybka changed it's mind to winning move after 2 hours in one analysis that I did. It was tactics but not tactics that is easy to find and I think that other programs will also need a very long time to find it. I'm not sure how to devise >a test(s) to answer this question. Because long run times take a long time, it >is difficult to evaluate them, because it takes such a long time to obtain any >such test data (a circular argument, to be sure) It is not difficult to evaluate that the program changed it's mind to a winning move. 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.