Author: blass uri
Date: 09:09:16 07/29/00
Go up one level in this thread
On July 29, 2000 at 09:42:01, Jeremiah Penery wrote: >On July 29, 2000 at 03:29:35, blass uri wrote: > >>On July 28, 2000 at 20:22:01, Jeremiah Penery wrote: >> >>>On July 28, 2000 at 19:12:55, blass uri wrote: >>> >>>>On July 28, 2000 at 18:25:17, blass uri wrote: >>>><snipped> >>>>>Fritz5.32 that is not the latest Fritz gives the following evaluation: >>>>>Qxc6 -1.62 depth 14/30 >>>>>fail high Qe3 -1.59 depth 14/41 >>>>>solved fail high with -1.50 score. >>>> >>>>I can add that the valuation at depth 16 also failed high and the failed high >>>>was solved with -0.81 score at depth 15/45 >>> >>>I suggest an experiment: Let Fritz continue to search even deeper, say to 20 >>>plies or more, and then report about what Fritz thinks. If it ever finds the >>>drawing line, I will (in the style of Dann Corbit :) write a poem in Fritz's >>>honor, extolling its greatness of search and evaluation. >>> >>>I did this experiment with some versions of Crafty and some strange things >>>happened: It kept failing high on Qe3 until it found a draw-score (at ply 21 or >>>so), but it was not the real drawing line, so it started failing low again >>>because it could never find the right line. It eventually failed back to -1.5, >>>but I was forcing it to search only Qe3, so I couldn't know if it would change >>>its mind or not. There is no way to know that DB didn't search the Qe3 line >>>very deeply or not, but I contend that no computer will ever find the draw >>>because there are quiet moves near the end of the line. >> >>There is a way to know because you can search the alternative and see if the >>score of it is clearly lower than -1.5. >> >>I do not think that no computer will find the draw line because it is possible >>to find it by the right extensions. > >I don't think this is true because of the quiet moves near the end of the line. >How are you going to tell your search to extend this line, and if you do how do >you avoid extending a ton of useless moves and your search tree will explode? I can tell the computer to extend the right lines in a similiar way that I can find them with the help of the computer and some hours. I give the computer to play against itself and try to improve the line for white because white lost an advantage in the line. It may be not productive for commercial programs but if you have the same number of nodes that deep blue had then it is not a problem to tell the program to play against itself and learn from the experience in 10% of the time. I expect part of the commercial programs in the future to find the draw at tournament time control with only 20,000,000 nodes per second and some software improvement based on the knowledge that you can search 20,000,000 nodes per second. I believe that it may be done with the right software with the hardware of today at tournament time control(the fact that I know nobody who did it does not prove that it is impossible). 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.