Author: Robert Hyatt
Date: 12:51:30 05/28/02
Go up one level in this thread
On May 28, 2002 at 09:30:48, Sune Fischer wrote: >On May 28, 2002 at 09:09:18, Daniel Clausen wrote: > >>On May 28, 2002 at 04:05:47, Sune Fischer wrote: >> >>>I believe it is then safe to stop the search at 150 plies, that's 50 plies left >>>for extensions(!), it will simply have to do, even if I get a crash in 1 of >>>10000000 games ;) >> >>If you'd be in my programming team, either you or me would be looking for a new >>job now. ;) A simple additional if-statement at the beginning of search makes >>this 100% correct in all cases and slows down your engine by a _very_ small >>factor. So basically you're just careless here - and therefore maybe also >>careless in other places. ;) >> >>Remember that even though these small speedups do add up over time (and maybe >>give you an additional +2 ELO boost in the next 10 years *grins*) the risk for >>crashes or misbehaviours also adds up. And Murphy will make sure it will happen >>during the most important game. =) [should it happen during a game against my >>engine (in the future.. maybe..) I'll remind you of this thread =)] >> >>HTH :) > >"careless" :) >Hmm, I don't know. I need to extend ~100 plies(!!) to get a crash for a total of >200 plies search! >Can you come up with any kind of position that would even slightly challenge >this limit? To get many extensions you need checks or so, but then the position >won't transpose as good (because it won't be blocked pawns), so really you need >maybe 170 extensions in a 30 ply search, in all it's extremely unlikely. I got zapped by some "straightjacket" problems Harry Nelson used to work on. The initial problem was to define a position where black and white had the most consecutive forcing moves (one possible reply). I think Harry had one that went 32 plies, which meant that in Cray Blitz, it extended 32+ plies as well. And beyond that most of the remaining moves were forced for a while. It blew our 64 move limit quite nicely (we had 64 because early crays had 64 as a max vector limit due to hardware vector register size). 200 might be safe enough. But the check is so cheap (it gets buried in the super-scalar code anyway) that it can be ignored and then you have _zero_ to worry about in wild positions. > >I think that has to be some kind of a record if so, surely it would mean a tree >explosion or really bad extension rules. > >I also have the maximum number of pseudo legal moves set to 255, this must be a >potential bug too then? Unknown. No one has produced a position to blow that yet. 218 is the most I know of. I use 256 myself. >Well I'm not going to check before adding a move if there is still room, call me >careless if you must ;) > >These are the only known limitations I have, anything else shouldn't crash if >the position is technical legal. You can mathematically prove what limit you need to set, based on your extensions, extension limits, and max iteration depth... If you dimension everything beyond this limit, you are safe enough. > >-S. > >>Sargon
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.