Author: Robert Hyatt
Date: 08:52:15 01/26/04
Go up one level in this thread
On January 26, 2004 at 10:40:29, Tony Werten wrote: >On January 26, 2004 at 06:13:10, Tord Romstad wrote: > >>On January 25, 2004 at 17:55:13, Robert Hyatt wrote: >> >>>On January 25, 2004 at 13:58:24, Uri Blass wrote: >>> >>>>I see in Crafty's code that Crafty extends mate threats but how can it know >>>>usually about mate threats when it is using fail hard and not fail soft? >>>> >>>>I do not see in crafty's code a special search after null move(even to reduced >>>>depth) in order to find if there is a mate threat. >>> >>>This is done _in_ the null-move search. It simply says "If I do nothing (play a >>>null-move) and my opponent mates me instantly, then this position has a threat >>>and should be searched deeper... >>> >>>I detect the threat when doing the null-move score and getting a mate score >>>backed >>>up. Crafty is not really "fail hard". It just is not fail-soft... >> >>I think you didn't quite understand Uri's question. I also don't understand >>how you can use the null-move search to detect mate threats. Let me try >>to explain why: >> >>After making the null move and calling search(), I don't necessarily search >>all moves. I stop searching as soon as I find a move which fails high >>(just like in all other nodes of the tree, of course). It is obviously >>possible that there exists a mate-in-one in the position, but this mate >>might not be found because the mating move is not searched first and one >>of the earlier moves fails high. What am I missing? >> >>Instead of using the null-move search to find mate threats, I try to >>detect them in the evaluation function. This is very cheap and works >>reasonably well, but it's not perfect. It happens that a mate threat >>is not found, and it happens (not often) that the evaluation function >>detects a mate threat which isn't there. >> >>While writing this post, I got an idea which perhaps could work. >>Before the ordinary null-move search, do a null-move search with the >>depth reduced to zero (or one, if you don't search checks in the qsearch), >>and beta lowered to -MATE_VALUE+Ply+2. This search should terminate >>very quickly, and should also reliably detect all threats of mate in >>one. If a mate is found, you extend and disable the ordinary null move >>search. > >From my experience ( wich is the same with singular extensions ) it's not the >detecting that will burn nodes, it's the extending that will. > >Tony Depends (singular extensions). The "cheapo" method really burns a lot of nodes in detecting singularity. It is way expensive even if you get zero extensions triggered. That's the thing I don't like about it. Hsu's approach is significantly better in terms of overall overhead, but it is a _lot_ more complicated than the "cheapo" approach. > >> >>Tord
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.