Author: Roberto Waldteufel
Date: 18:13:07 10/14/98
Go up one level in this thread
On October 14, 1998 at 12:45:08, Robert Hyatt wrote: >On October 14, 1998 at 12:18:37, John Coffey wrote: > >>On October 13, 1998 at 20:37:43, Robert Hyatt wrote: >> >>>On October 13, 1998 at 18:41:22, Mark Young wrote: >>> >>>>On October 13, 1998 at 17:35:34, Robert Hyatt wrote: >>>> >>>>> >>>>>After reading this thread, I decided to run a test myself. I started >>>>>off with 20 games of crafty vs crafty, white having KQ and no tablebases, >>>>>black having tablebases. At 20 seconds per move, this ended in 20 wins >>>>>for white which surprised me. I took several different positions (20 in >>>>>total) most of which were mate in 30 or greater and gave crafty 20 secs/move >>>>>using 1 cpu on my ALR. All wins. I didn't expect this. >>>>> >>>>>I then repeated this at 10 secs/move. All wins. >>>>> >>>>>I then repeated it at 5 secs/move. All wins. I was going to try 1 sec >>>>>but decided that to the computer, this is far easier than I thought. It >>>>>seems that the simple heuristic drive king to the edge, then to the >>>>>corner, is enough. It didn't play it perfectly, but it never slipped more >>>>>than 2 moves from optimal at any single move. And when it started off at >>>>>mate in 34, the "game" never went past 40 moves. >>>>> >>>>>So, I retract my original feeling, that KQ vs KR is hard with the KR >>>>>side having a database. It seems it is a "trivial" ending regardless of >>>>>having them or not. >>>>> >>>>>Most surprising... >>>> >>>>It happens to the best of us. I wonder if this simple idea of driving the king >>>>to the edge, then to the corner is what the Grandmaster's caught onto so quickly >>>>after only drawing the first few times, or did they find some other idea that >>>>also works. >>> >>> >>>No.. there's more to it than that... the main problem is the rook on the >>>opposite board, checking the king at the right time, also preventing the >>>king and queen from getting on the same rank or file for a skewer... and >>>the check lets the king "out" to the other side where it has to be done >>>again, and once it it out, the 50 move counter usually ends the game. >>> >>>I am utterly amazed that a 5 second search can solve this. Had you told >>>me this a year ago I'd have laughed. As it was I was skeptical. And when >>>I saw it I was *still* skeptical and was looking for a setup error in my >>>experiment. >>> >>>I still want to try at 1 sec / move however... >>> >>>but this is definitely easier than KBN K now... to the computer, because >>>it takes a special piece/square table to win, since you have to drive to >>>the right corner as well... >> >>Although i doubt that any program can see the solution all the way through in >>such short time, I am sure that they can find a position that >>brings them closer to a solution (such as getting the king to the side of the >>board or corner, as you have said.) >> >>John Coffey > > >this is the point... a few moves by crafty (non-tablebase) looked pointless and >I would peek at the other program's log file only to discover the mate distance >went down by 1. After studying a minute or two, the pointless move covered a >square so that the "great flying rook" couldn't check from the rear and let >the king out of the box. > >Another thing that is interesting is that it is really not difficult to play >this ending perfectly. One thing I noticed, in studying the games, is that at >any given position, there are always a couple of bummers that let the king back >out to the center and start things over. But there are also *lots* of moves >that are either optimal or within 1 of optimal. Which means that this is a >pain for people, to avoid missing those 1/2 moves that break the king back >out, while for a computer, only interested in squeezing the king to the corner >with no idea it can be mated, these odd moves are obvious and are avoided with >ease. > >I just would never have believed this is easier than KBN vs K for a computer. >KBN can be a problem in worst-case positions, because a very short search >doesn't shed much light on forcing the opponent to the edge... you have to first >get yourself off the edge and into the center, so it takes a weak centralization >score for the winning side, and a stronger de-centralization score for the >losing side, to make this ending happen without tablebases and using a very >shallow search... I use this technique for mating a lone king with major pieces or pair of bishops (except I now use self generated tablebase for straight KQK or KRK with no other pieces), but surely for KBNK you need two evaluations, depending on the bishop you have, since the king needs to be forced into the right corner. This could easily be handled by a preprocessor, and would be well suited to piece-square value tables. There are also a small number of positions in KBNK which are drawn even with the stronger side to move, due to the knight being trapped, which can never happen in KQK, KRK or KBBK mates, so the evaluation function becomes more complex if this is to be taken into account as well. Has anyone succesfully done KBNK with an evaluation like this and a shallow search? My attempts were not terribly encouraging. Of all the lone king mates, I think KBNK is the one best suited to solution by tablebase, since the search and evaluation based on centrality works so well for the others. Best wishes, Roberto
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.