Author: Robert Hyatt
Date: 15:27:07 10/12/98
Go up one level in this thread
On October 12, 1998 at 02:03:49, blass uri wrote: > >On October 11, 1998 at 10:35:41, Robert Hyatt wrote: > >>On October 11, 1998 at 03:27:59, blass uri wrote: >> >>> >>>On October 11, 1998 at 00:17:55, Robert Hyatt wrote: >>> >>>>On October 10, 1998 at 14:51:56, blass uri wrote: >>>> >>>>>I think it is possible to save space by storing only win,draw,loss in these >>>>>tablebases >>>>>when you go to the position with 4 pieces you can generate the tablebase >>>>> >>>>>Maybe it is not practical for blitz but it can be practical for tournament time >>>>>control. >>>>> >>>>>Uri >>>> >>>>I would not want to generate either KBNK or KQKR on the fly... because >>>>I probe *in* the search, and need the result then, not just at the root. >>> >>>You will have the result because I did not say not o have the result of >>>win/draw/loss >>>>And KQKR can't be built in a couple of minutes by anybody yet... the 3-4 >>>>piece files take 3-4 *hours* to build, in total... >>> >>>It is not important to build all the 3-4 pieces files but only the KQKR when you >>>get to the KQKR endgame and you can save it in the ram and not in the harddisk >>>so you are faster. >>> >>>I read that before some years(I think in 1989 but I am not sure) a 6 pieces file >>>was generated by a computer by 6 hours and it found mate in more than 200 in a >>>position of KRBvKNN or something similiar(it was done by a super computer andnot >>>by a personal computer but personal computers are today better) >>> >> >> >>this was done, but on a 50 million dollar computer. Which is not going to >>be available to any of us in a tournament... nor will the PC catch up to >>the speed of that machine for 20-30 years... So there's no use planning on >>being able to compute even a 5 piece file. IE KRP KR takes 80 hours plus >>on a machine faster than a PC today (big alpha machine). You also need to >>compute the promotion classes which take 12-24 hours each to build. > >You need to compute the distance to conversion for 24 KRPvKR >There are 24 possible places of the pawn(I do not count files e,f,g,h) >It should not be more than 24 time slower than KQvKR if you know all the sons. > >The number of positions is bigger because you cannot use symmetric consideration >when there is a pawn but the number of moves per position is smaller than KQvKR >and the maximal number of moves before conversion is lower. > >Do the 80 hours on big alpha machine including generating the sons? No. This requires an exhaustive search thru the 32*64^4 positions which has to be done N times where N is the max mate-in-N length. It is not the number of positions that is the problem, as you mistakenly put it above, it is the number of positions * the depth of the longest forced mate. In the 5 piece files, there are more positions, and the mates are deeper as well, which makes this a huge piece of work... > > >I think that building KRvKN or KRvKB is faster then KQvKR >The complexity of the problem is : >(number of positions *Number of moves per position*maximal number of plies to >force a conversion) > >How much time do the best pc need to build these tablebases >and How much time they need to generate a move(can I save time if I save in the >ram all the possible moves for every possible position)? > >I think it can be done if you have 256Mbytes ram. > >Uri easy to do, but it takes time. IE to build all the 3-4 piece files on a big memory pentium pro takes 4 hours roughly. Not usable in a tournament type game obviously. The 5 piece files are bigger and much harder to make. Ditto for sixes...
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.