Author: Robert Hyatt
Date: 11:36:00 08/05/99
Go up one level in this thread
On August 05, 1999 at 12:41:38, KarinsDad wrote: >On August 04, 1999 at 20:20:54, Robert Hyatt wrote: > >>On August 04, 1999 at 10:44:06, KarinsDad wrote: >> >>>On August 03, 1999 at 19:27:18, Ricardo Gibert wrote: >>> >>>[snip] >>>>> >>>>>If you are in a mate in 110 situation (probably extremely rare), it would >>>>>probably behoove you to do your own searching for capture/push moves that lead >>>>>to positions that maintain the win since you will probably find them relatively >>>>>close to the main line anyway. Compared to searching even something as simple as >>>>>an 8 ply alpha-beta, searching for those moves would take VERY little time. The >>>>>reason this works out so well is that you do not have to search outside the >>>>>tablebase once you are in it (or once you are close enough to it to force your >>>>>way into it). >>>> >>>>In many pawnless EGs a "good" capture can be hard to come by. 8 ply would not do >>>>it in that case. >>>> >>> >>>One other note on this. When you have moves that lead to mate, but they lead to >>>mate 110 moves (i.e. 220 ply, I believe this is what you were trying to discuss) >>>later, you have the option of searching for win preserving moves at ply 1, ply >>>3, ply 5, etc. Everytime it is your turn, you can attempt to find a win >>>preserving move which has a pawn push or a capture. >>> >>>Granted, you will not be able to search REAL far each time, but even if you only >>>had a second per move, it would take at most two minutes total to search down 6 >>>ply each and every time (i.e. 50 times at ply 1, 3, 5, 7, etc. for a maximum of >>>about 120 moves or 240 ply for your 220 ply best case) before a capture or push >>>was needed. This assumes that you can load up a given 5 piece tablebase into >>>memory (or most of it) and use it (since you will be in a position of not >>>capturing, you will continue to be in the same tablebase). It also assumes that >>>you can search just as far into a tablebase as you could search normally in your >>>search/eval which on the surface, appears to be a slam dunk assumption. Most >>>likely, you could search even faster and farther with the tablebase. >>> >> >> >>stop just a minute and think about what you just said. You _really_ think you >>can do a 220 ply search? Because that is _exactly_ what you have to do to >>follow that tablebase search to the end to handle 50 move and rep correctly. >> >>The tablebase contains _all_ possible positions. So if it says mate in 220, >>which some 6 piece files will definitely do, you are going to have to search >>far farther than you think to find that playable capture or pawn push. So >>far that it will be beyond our capability for at least a few zillion years. :) > >I believe you either misinterpreted what I wrote or I wrote it really lousy. > >If you have mate in 220, you could search side positions (which contain a pawn >push or a capture within the tablebase) as you continue making moves. > >Move 1: Don't bother to search for side position. >Move 2: Search for side position that have mate in 219 (you do not really need >to do this here, but read on), if successful, you have defered your mate in 220 >to mate in 219, but you have reset the 50 move counter. >. the problem here might be that if you move a pawn at move X, the next possible pawn move might well be more than 50 moves beyond X. But at X you "committed" and now you draw. You have to search all the way to the end and find a PV that doesn't have a 50-move draw or three-fold repetition in it, before you enter into that line. And that is just too far to search, not to mention the fact that all this stuff is on disk and is very slow to access.. >. >. >Move 87: Have to mate, push, or capture within 12 moves or you draw. Say you >find a side position with mate in 150. You are currently at mate in 133. From >where you were 86 moves ago, you can reset your 50 move counter and drop the >mate in 220 with 99 moves to make until a draw, to mate in 150 with 99 moves to >make until a draw. If you can get one more major drop like this (due to a >capture or a push), you can mate your opponent. > that will work in some cases, and break in others. Because when you make an 'irreversible' move, you can't take it back, and it might commit you to a path that has an unforseen 50-move draw that now you can't avoid. Because you didn't follow the complete path... >I think it would be EXTREMELY easy to find a capture or a push nearby which >maintains a win in a LOT of positions. This does not mean that it will be easy >to do this at any given ply. But sooner or later, it should be easy to do after >doing an 8 ply search 50 times over (for some positions). the question is "lot" = "most" or just "some"? This has to be tested, and at present, we don't have many endings where this is an issue (IE Eugene's KBNKN was one example that just barely has the problem). We need a position with a pawn to really see how this works out... > >Of course you cannot do a 220 ply search. But the point is that you do not have >to. You only have to find a push or capture that leads to a position that >maintains the win. But you can't verify this without the 220 ply search. Because what happens after you make that push/capture must allow a path without a 50-move draw in it, but you won't know if you don't search far enough to see it before you make the push now. > >Granted, there could be weird fortress positions or somesuch where this would >not work (as per your information on Lewis Stiller's work below). But as a >general rule, it will probably work at least occasionally. And, it is more >likely to work in a position where the side to win has one or more pawns. And, >it does not matter if it doesn't work for a given position. That position is >drawn anyway. The fact is that it probably will work for SOME positions and that >is the reason to do it (i.e. if it preserves 10% of wins in these rare cases >where the win is beyond 99 ply, then that is a good enough reason to do it). >Note: you do have to make sure the program has no timing bugs so that it never >loses on time attempting this and this includes the time it takes to read other >tablebases in from the hard drive. > Against humans this doesn't matter. I've never seen anyone play any sort of tablebase position against crafty in anywhere near an optimal manner. IE the KNN vs KP ending I saw was a mate in 103, but it went from 103 to maybe 80, to 55, to 30 over 5 moves, because the human didn't play it very well. If you are playing a computer with tablebases, it will be different, of course, as since it always goes for the deepest mate (if losing) that will tend to make it follow lines with potential 50 move draws in them... >> >>>Anyway, my basic point is that you would not HAVE to search far. Sooner or >>>later, you would most likely find a win preserving move within 6 or 8 ply >>>(depending on how much time you have and how far you can search the tablebase) >> >>Lewis Stiller disproved this. He found lots of positions where a playable pawn >>move (or capture) happened more than 70 full moves from the original position. >> >>It would be absolutely impossible to search 140 plies in a 6 piece ending. It >>would be impossible to search even 60 plies in most of them unless they are all >>pawns and they are locked up totally... >> > >I am not familiar with Lewis Stiller's work. I have not yet read his Berkeley >talk or his thesis (but I plan to now). > >However, the question comes down to whether or not positions in the graph close >to the PV of the tablebase (within 8 ply for example) can reset the counter. >From what you have stated here, I am not convinced that this cannot happen. Sure it can. But it also may commit you to a course of action that is bad, because if you push _here_ you don't have that pawn push to save you later. The only way to resolve this is to search to the mate position making sure you don't cross over the 50 move rule problem anywhere along the way... Your >example of 70 full moves does force one to conclude that there are 6 piece >positions where this cannot be done and a draw is forced. However, if this is >truly the case, then it does not matter. There is NO win preserving move within >50 full moves that can be used for that position, so it does not matter there. > >Nothing you have said convinces me that this technique cannot maintain a win in >at least some percentage of those rare cases where one side has mate in 100+. >How often this will do this is debatable without more data. > >KarinsDad :)
This page took 0.01 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.