Author: Robert Hyatt
Date: 18:08:56 05/14/00
Go up one level in this thread
On May 14, 2000 at 14:31:13, blass uri wrote: >On May 14, 2000 at 13:04:18, Heiner Marxen wrote: > >>On May 14, 2000 at 01:46:54, blass uri wrote: >> >>>On May 13, 2000 at 23:15:33, Robert Hyatt wrote: >>> >>>>On May 13, 2000 at 21:12:21, blass uri wrote: >>>> >>>>>On May 13, 2000 at 18:02:13, Robert Hyatt wrote: >>>>> >>>>>>On May 13, 2000 at 16:49:59, blass uri wrote: >>>>>> >>>>>>>On May 13, 2000 at 16:30:27, Robert Hyatt wrote: >>>>>>> >>>>>>>>>I understood that the idea behind chest is that it has not to try all the ply=1 >>>>>>>>>moves to find that there is no mate in 2(it is obvious for mate in 1 when you >>>>>>>>>need to generate only threat king moves but I understood that it is also >>>>>>>>>possible to do it for mate in 2). >>>>>>>>> >>>>>>>>>Chest knows for every piece the squares that it controls so it knows the squares >>>>>>>>>need to be controled in order to do mate. >>>>>>>>> >>>>>>>>>If it is obvious from the starting position of the pieces that they cannot >>>>>>>>>control the relevant squares in 2 moves then you can discover that there is no >>>>>>>>>mate in 2 without generating moves. >>>>>>>>> >>>>>>>>>Uri >>>>>>>> >>>>>>>> >>>>>>>>what about zugzwang??? >>>>>>> >>>>>>>chest proves that there is a no mate without really executing the moves and it >>>>>>>does not do errors in zugzwang positions. >>>>>>> >>>>>>>I understand that it does not assume no move for the defender but assumes a >>>>>>>simple strategy(king move if possible) and try to prove that the squares cannot >>>>>>>be controled in 2 moves by the attacker. >>>>>>> >>>>>>>Uri >>>>>> >>>>><snipped> >>>>>>In my case, I need to know that there is a forced mate at the current position, >>>>>>so that the null-move won't be tried. There are lots of zugzwang mates where >>>>>>you have to confirm that no matter what you do you get mated, even if you are >>>>>>not in check. And if you do nothing, you don't get mated instantly. >>>>>> >>>>>>I am trying to point out that (a) the test is very expensive because it has to >>>>>>be done so many times >>>>> >>>>>I agree that the test has to be done many times but many times a small number is >>>>>not very expensive if the number is small enough. >>>>> >>>> >>>>Confucious say "if you multiply a big number times anything, you get a big >>>>number." :) >>>> >>>>I can add a single if statement and see the NPS change. If I add it in the >>>>right place... >>>> >>>> >>>> >>>> >>>>>The number of times you need to do the test is exactly the number of times that >>>>>you find that there is no threat and if the time that you need to discover that >>>>>there is no mate in 2 is 10 times smaller than the time that you need to >>>>>discover that there is no 3 ply threat then you can earn important information >>>>>because you sometimes can miss mate in 2 in searching for 3 ply threats because >>>>>of null move pruning or by the fact that the evaluation does not know that a >>>>>position is mate. >>>> >>>>I don't do any "threat" detection of any kind. When I enter a new node, I try >>>>a hash probe, followed by a null-move search to see if I can exit quickly. >>> >>>a null move search is a threat detection. >>> >>>What I suggest is that if the null-move search tell you that you can exit >>>quickly to do a mate threat search for the same depth that is clearly faster >>>than the null move search and only if the mate threat search tell you there is >>>no mate threat then you can exit. >>> >>>Sometimes null move search to depth 3 cannot see mate in 2 so this idea can be >>>productive against mate attacks. >>> >>>The reason that I suggest it is that I saw that Crafty has problem in detecting >>>some ideas about mate attacks. >>> >>>Crafty lost against tal in the millenium tournament after it fell victim to mate >>>attack and I saw that crafty could not see a sacrifice of chessmaster against >>>Insomniac because of mate threat prunning. >>> >>>I posted the position when crafty needed 10 plies to see Nf5 but the problem was >>>some plies early when crafty could not see the winning sacrifice of chessmaster. >>> >>>You can also decide to use the idea of mate search after null move search only >>>in part of the positions when you can see by evaluation king safety problem so >>>you will do it after sacrifices like Bxh2+ >>> >>>I was careful now to suggest only mate in 2 after 3 ply search but it is >>>possible that mate in 3 after 3 ply search is also not a waste of a long time if >>>you do it only in positions when there is a king safety problem. >>> >>>It is possible to stop mate attacks also by better evaluation but I do not think >>>that you are going to use big evaluation in order to sacrifice a bishop for mate >>>attack because of positional reasons and if you do not do it then the only way >>>that I see to do something against mate attacks is by improving the search. >>> >>> I >>>>don't do any kind of 'searching' or 'analysis' to determine if a null-move is >>>>safe to try... >>>> >>>> >>>> >>>> >>>>> >>>>>I think that searching for mate in n by chest is usually clearly faster than >>>>>searching 2n-1 plies forward by crafty. >>>>> >>>>>Uri >>>> >>>> >>>>I wouldn't be surprised at all. Finding mates is far easier if you are only >>>>looking for them and not trying to play a real game... But the question is >>>>still how much does it cost. >>>> >>>>I like the relatively simple search I am using. as simple -> bug-free. Yes, >>>>I wish it could do some things quicker. It will when I take the time to get >>>>the singular extension stuff implemented... >>> >>>I do not know if singular extension will help much if you prune threat mate move >>>after null move search when the search is not deep enough. >>> >>>Even if they help I think that my idea can also help. >>> >>> >>>Uri >> >>I am very pleased that the ideas implemented in Chest stimulate such an >>interesting discussion :-) >> >>I tend to agree with you, Uri, that a special mate threat detection can >>help, if designed carefully. Some time ago I have also tried to suggest >>a verification procedure for null move prunings. A mate threat search >>could help some problems, which really are there, but also are not exactly >>frequent, they are somewhat rare. >> >>Also, the methods of Chest use a comparatively big board: for every field/ >>position there are maintained two attack sets, which are quite crucial for >>the operation of Chest. That makes the move execution inside chest an >>expensive operation, quite unlike most playing programs, including crafty. >> >>I tend to agree with Bob that throwing in an additional mate threat search >>can easily slow things down a lot. >> >>Embedding the ideas for fast mate search into a playing program would need >>to completely re-evaluate the ideas and to completely re-design the >>implementation, IMO. That is not a small/quick mini-project. Obviously >>Bob has not yet run out of ideas he considers easier to implement. >>So... who volunteers to work out these things until we all see it can be >>done and is quick enough? I do not volunteer :-( What about you, Uri ? >> >>Heiner > >I also do not volunteer to do it. >It is of course easier to think about ideas than to write them. > >The reason that I suggested this idea is the fact that >I did not like to see programs become victims of mate attack of tal because of >the behaviour of the programmer of tal and I found that in the millenium >tournament tal often won because of mate attacks. > >I thought that my idea could help to stop part of the mate attacks of tal(it is >better to do it by evaluation but I do not see a way to know by evaluation when >a sacrifice of a piece is right). > >I think that this idea may be important also against humans who try to win >programs by mate attacks. > >Some mate attacks are too deep and cannot be stopped by search but some mate >attacks can be stopped by search. > >In the Israeli league 2 programs were victims of mate attacks(Fritz and Rebel) >and it seems to me that in this case the mate attack could be stopped also by >search. >In the last game of Fritz that was posted today the mate attack could not be >stopped by search. > >Uri I am under the impression that Rebel doesn't use null-move. IE a null-move program typically goes 2 plies deeper than without it...
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.