Author: Robert Hyatt
Date: 12:20:01 10/09/98
Go up one level in this thread
On October 09, 1998 at 11:58:37, blass uri wrote: > >On October 09, 1998 at 10:37:29, blass uri wrote: > >> >>On October 09, 1998 at 10:25:28, blass uri wrote: >> >>> >>>On October 09, 1998 at 09:56:20, blass uri wrote: >>> >>>> >>>>On October 09, 1998 at 09:35:55, Robert Hyatt wrote: >>>> >>>>>On October 09, 1998 at 08:54:02, blass uri wrote: >>>>> >>>>>> >>>>>>On October 09, 1998 at 08:10:51, blass uri wrote: >>>>>> >>>>>>>I think it is better to store the number of moves to win in the tablebases only >>>>>>>if this number is divisible by 2 because otherwise you can compute the number of >>>>>>>moves by search. >>>>>>> >>>>>>>Maybe it is possible practically to store it only if it is divisble by 3 or 4 (I >>>>>>>do not know what is the speed of reading the tablebases from harddisk) >>>>>>> >>>>>>>Uri >>>>>>I think it is easy to save at least 1 bit per position by this idea if in >>>>>>positions when there is win in N when N is divisible by 2 you store win in N/2 >>>>>>and in positions when there is a win in N when N is not divisible by 2 you store >>>>>>a win in 1(you can discover by search that it is not truth). >>>>>> >>>>> >>>>> >>>>>this is imposible to do. KNN vs KP is a mate in over 100 moves. How do >>>>>I separate mate in 50 from mate in 100? I certainly can't search to find >>>>>this out... >>>> >>>>I mean that when you store the number N/2 you mean win in N >>>> >>>>you have only to search 2 plies because if you have mate in 99 >>>mate in 99 is not a good example because of the 50 move rule but >>>I understand that there is a problem here because the opponent may do a move >>>that let you only to do mates in an odd number of moves so 2 plies cannot help >>> >>>Uri >> >> >>you can store mate in N by the number N/2 or N-1/2 > > >of course I mean (N-1)/2 iand not N-1/2 >I choose N/2 if N is even and (N-1)/2 if N is odd >if 2 plies search are not enough than 4 plies are enough >the number of times you should call the tablebases if you have a good order of >moves is something like 1000 >I do not know the time that you need for it > >Uri > > Near the root of the tree, you are right. But imagine a position from the root where the best you can do is mate in 40. And there are some mate in 41 positions out there too. I can *never* search deeply enough to actually see those mates without tablebase hits, and all I get from the tablebase for either type of move is "mate in 40". If I follow toward the "mate in 41" because the scores are identical, the next search I will again have the same problem, because it will still be a mate in 40, but I might choose the mate in 41 since the scores are the same... I don't see a way around this... > and find by search the right >>move >>you need all the time to decrease the number >>I think that you have to compute sometimes 4 plies forward but if you fast >>enough in reading the tablebases then it can help >> >>Uri >> >> >> >>> >>>>the tablebases will tell you the number 1 >>>>You know by search that there is no mate in 2 but you have in the tablebase >>>>mates in 98 so by searching 2 plies you can read from the tablebases the mate >>>>in 98 after 2 plies. >>>>Searching 2 plies and reading the tablebases after every 2 possible plies(you do >>>>not need every 2 possible because of the alpha beta is not impossible >>>>> >>>>> >>>>> >>>>>>By this idea you can use the same number of bits to every position. >>>>>> >>>>>>If you want to use different number of bits to every popsition you must add 1 >>>>>>bit >>>>>>to tell if the is a position with win in number of moves divisible by 2 or not. >>>>>> >>>>>>of course if you lose it is not important to know the number of moves. >>>>>> >>>>>>I use the word win and not mate because it is a mistake to use mate >>>>>>because of the 50 move rule >>>>>> >>>>>>Uri >>>>> >>>>> >>>>>distance to conversion has the same 50-move problem... >>>> >>>>why? >>>>If the distance is more than 50 then you know it is a draw. >>>> >>>>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.