Author: Uri Blass
Date: 11:48:06 05/29/01
Go up one level in this thread
On May 29, 2001 at 11:17:19, Robert Hyatt wrote: >On May 28, 2001 at 02:19:31, Uri Blass wrote: > >>On May 27, 2001 at 22:13:55, Bruce Moreland wrote: >> >>>On May 27, 2001 at 01:59:58, Uri Blass wrote: >>> >>>>On May 26, 2001 at 20:59:24, Bruce Moreland wrote: >>>> >>>>>On May 26, 2001 at 14:02:26, Uri Blass wrote: >>>>> >>>>>>On May 26, 2001 at 12:34:19, Bruce Moreland wrote: >>>>>> >>>>>>>I got a fail-high on Bxg7 after 80 hours, on a quad Xeon 450, but I didn't have >>>>>>>time to resolve it. >>>>>>> >>>>>>>bruce >>>>>> >>>>>>It is interesting to know the reason that it failed high. >>>>>> >>>>>>Unfortunately programs only give main line. >>>>>> >>>>>>If you can make your program remember all the lines that it considered for more >>>>>>than 1 minute than it will be interesting to know them. >>>>>> >>>>>>The number of these lines cannot be more than 80*60 in 80 hours and I am >>>>>>interested only in the lines that begin with Bxg7. >>>>>> >>>>>>Uri >>>>> >>>>>The only information I could have returned is that Bg7 failed high, and for each >>>>>possible response I could have remembered the move that refuted the response, >>>>>and for ever possible response to the move that refuted each response, I could >>>>>have remembered the move that refuted that move. >>>>> >>>>>You don't end up with a line when you have a fail high. You end up with >>>>>something with a little less direction to it. >>>> >>>>I understand it. >>>>I meant to say that even a main line is not the interesting information that I >>>>want to see but it is the information that all programs give after solving the >>>>fail high. >>>> >>>>The problem is that even after you solve the fail high and have the main line I >>>>believe that the interesting information is the lines that your program >>>>considered for a long time. >>>> >>>>If programs can give a tree with the idea that I suggest that it may help people >>>>to understand the reason for the move. >>>> >>>>I suggested the lines that were considered for more than a minute but a better >>>>idea is to post first lines that your program considered for more time. >>>> >>>>I mean something like >>>>1.Bxg7 20 hours >>>>1.Bxg7 Kxg7 2.Ne5 15 hours >>>>1.Bxg7 Kxg7 2.Ne5 Rfd8 3.Qg4+ 6 hours >>>>1.Bxg7 Rfd8 Rxd8+ 3 hours >>> >>>That is not how these work in single-variation mode. The searching process is, >>>for lack of a better word, more patchy than that. >>> >>>bruce >> >>Maybe I did not explain myself well >> >>The computer always has a line that it analyzes and I am interested in the >>relevant lines that it analyzes more often. >> >>Chessmaster6000 has an option to show every 1/2 second the line that it >>considers. >>I assume that other programs including ferret works similiraly. > >But the line it shows has _nothing_ to do with what is good and what is bad. IE >it is not the best line at this instant... it is the exact path of moves being >searched at this instant, without regard to whether the score is good or >bad. Even the second move displayed several times does not mean that is the >_best_ move in this position... it might just mean it is the hardest one to >search. > >Alpha/beta hides all the useful info besides "this is the best move"... > > > >> >>You can remember all these lines and post the begining of the lines that were >>considered more often. > >More often --> searched more times or more time spent? If (a) that would be >impossible since every ply=2 move (for example) is searched the same number >of times. If (b) this only measures the complexity of searching that move, >not whether it is good, bad, or equal. 1)I mean b and this is important information if we want to understand the reason for the move of the computer without asking it to do more long searches. I assume that the program remember the killer moves and I am interested only in lines with killer moves for white. If the program remember that the killer move after 1.Bxg7 Kxg7 is the move Ne5 then lines like 1.Bxg7 Kxg7 2.Rd7 are not important for me. I know that the lines may include moves that are not the best for white but the moves are good enough to explain the reason that the program played 1.Bxg7. The problem that I may have after a long search is that the program gives a main line that does not include Kxg7 after 1.Bxg7 and I have no information from the main line to know what the program planned to play after 1.Bxg7 Kxg7 inspite of the fact that it is something that was calculated for a long time and not something simple. I am interested to know the moves that are good enough to justify the score of the program. I remember a previous case in one of the nolot positions when the main line of Ferret included something that was obviously losing for black when the interesting information is what Ferret planned against normal lines. I am sure that in this case Ferret searched for a long time the relevant lines but the main line gave no information about it. 2)My idea to remember the information about the main line every 1/2 second was only an idea how to achieve b. There are 2*3600*80 main lines in 80 hours that are not close to be all of the tree. If you see that the move 1.Bxg7 Kxg7 2.Ne5 appears in 7845 main lines then you can say that the line 1.Bxg7 Kxg7 2.Ne5 was probably calculated for 7845/2 seconds(it is not exactly correct but it is good approximation) There may be better ideas to achieve the same target. It is possible to remember the exact time in every case that you call search(alpha,beta ,depth) when depth is big enough. 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.