Author: Vasik Rajlich
Date: 03:45:15 08/01/05
Go up one level in this thread
On July 31, 2005 at 19:17:47, Uri Blass wrote: >On July 31, 2005 at 04:52:44, Drexel,Michael wrote: > >>On July 30, 2005 at 17:48:19, George Tsavdaris wrote: >> >>>On July 30, 2005 at 06:01:06, Drexel,Michael wrote: >>> >>>> >>>>...at all with this time controls and hardware used? >>> >>>No we can't! >>> >>>The list says the follwing about S9-UCI and S8-CB: >>> >>>S9-UCI has a rating that with a 95% probability it is between 2821-27 and >>>2821+28 >>>that means: >>>S9-UCI has a rating that with a 95% probability it is between 2794 and 2849 >>> >>>S8-CB has a rating that with a 95% probability it is between 2805-22 and 2805+23 >>>that means: >>>S8-CB has a rating that with a 95% probability it is between 2783 and 2828 >>> >>> >>> So it is clear that S9-UCI CAN actually be even by 2849-2783 = 66 ELO points >>>stronger than S8-CB.......... >>> But not only this. There is a 5% probability that S9-UCI can be X points >>>stronger or weaker than S8-CB and this X can be any positive number........ >> >>What's all this good for? >>We all know about the meaning of SSDF ratings. >> >>>>http://chessprogramming.org/cccsearch/ccc.php?art_id=364117 >>>> >>>>Sandro claimed Shredder8.bkt and more effective learning should increase the >>>>rating by about 35 points over Shredder 8 CB. >>> >>> Well, at the aforementioned link i just see a claim that S8-UCI on Shredder >>>Classic GUI, will be stronger from S8 at Chessbase GUI due to better learning >>>and better book, with 35 ELO. I don't see the same claim about S9-UCI and S8-CB. >>>Why this ELO increase has to be valid with these new version also......?!?!? >> >>Ever heared about logic?? >>Shredder 8 CB has a rating of 2805 and Shredder 8 UCI is supposed to be about 35 >>ELO stronger. Shredder 9 UCI has a rating of 2821. >> >>Is it likely that Shredder 9 UCI is stronger than Shredder 8 UCI then? >> >>Michael > > >I think that there is one problem that prevent accurate rating for Shredder. >Programs need to play also against stronger opponents to get accurate >rating(otherwise players who are stronger against weaker opponents but weaker >against stronger opponents never meat the stronger opponents and get too high >rating). > >Shredder never played against stronger opponents so it cannot get accurate >rating. > >Shredder8(A1200) and Shredder9(A1200) need to play against opponents like >Deep Fritz8(Quad opteron of WCCC2004) and Deep Junior9 (same quad opteron) in >order to get more accurate rating. > >Note that the rating of Deep Fritz8 and Deep Junior9 on quad opteron is >relatively unimportant because almost no user has them but they should get >rating only to help shredder to get more realistic rating. > >Unfortunately I have no money to give the ssdf quad opteron to test so you can >expect the ssdf of the best programs to continue to be not accurate and >comparison between them can be misleading. > >Uri IMHO the most meaningful test would be: 1) Huge round-robin tournament between the top X engines, where X is 4 or so. 2) Fixed starting date (maybe annual basis). 3) Official title of World Champion (if the organizere is bold enough). In principle a new engine can be made to clobber old engines, especially exploiting things like learner problems or big eval problems. In the SSDF list you sometimes see some really huge scores. In fact, here is a nice algorithm for making your program look good: If opponent is future version of me, play weak. Then it looks like you are always progressing. 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.