Author: Robert Hyatt
Date: 20:15:44 03/09/03
Go up one level in this thread
On March 09, 2003 at 14:39:07, Sune Fischer wrote: >On March 09, 2003 at 00:49:38, Robert Hyatt wrote: > >>On March 06, 2003 at 08:27:26, Georg v. Zimmermann wrote: >> >>> >>>> >>>>I don't see any reason to extend a move I fail high on, unless there is a threat >>>>further down. >> >>The problem is that often you have _one_ good move. All others fail low, >>and this _one_ move seems to be good enough and fails high. But if you >>go deeper, you see it fail also and don't make a mistake. Hence the point >>for singular-extensions. When you have lots of good moves, if you discover >>one is bad, that's ok, you have plenty of others. But once you know there >>is only _one_ good move, it had _better_ be good or the entire path gets >>mis-evaluated. > >So you mean if one move just barely fails high, and all the others don't even >reach alpha, then extend to make sure we really are failing high here? > Not for SE, no. There is a proper "window" to use. If one move fails high, you search the remainder with a window offset downward, but since that would result in a minimax tree, you search to reduced depth also to control the cost. But after I think about it, your explanation makes a bit of sense, yes. IE if there is only one good move, it needs to be searched deeper. Because you might be entering a line where you win a pawn, but depend on that one good move to prevent the world from crashing in on you. If it fails to do so, you end up losing. Deep Blue actually modified this algorithm to extend (but not as much) of two moves were better than the rest. Same idea... > >>>>Extending singular moves like PxQ where there is no threat involved is a waste >>>>of time as far as I can tell. >> >>You can avoide such extensions pretty easily. IE gross winning captures don't >>need to be extended most of the time. Ditto for recaptures. > >I think happens in most of the cases, your opponent moved Qd5, you had a pawn on >c4, now you have one good singular move but the whole position is rendered trash >and you ought to fail high here instantly and not extend at all. However, you fail high here, and when you try all the other moves, even with the offset window, they all fail low, since there _is_ only one good move here. It's just that this move is good because the opponent's last move was horrible. But SE will be triggered if you don't figure out how to prevent this. > >That is unless you extend to make sure that eating the queen is not going to >mate you, so it becomes sort of a threat verification search, only I think it >would suck at this because most of the time there is no threat at all. Right. But the point here is that you might detect this by noticing that the pxq move fails really high (fail-soft required of course and the material score would be the indicator that something is going on) and all other moves might not fail low using the _original_ value for alpha, not the value after ripping the queen. > >Maybe if the fail high margin is too high, don't treat it like a singular move, >is that what your saying? Maybe this could be made to work, it just makes it >rather fuzzy what a singular move really is then, having a margin on both sides >(not non-singular and not over-obviously-singular either!:). What you want to know is this: For the normal "window", one move is better than beta (this is for the fail high part, the PV singular is easier to understand) while all the other moves are much worse. But much worse than the normal beta value, _not_ much worse than the fail high value itself as it is artificially inflated by hanging a queen. IE if the real expectation for the score is (so far) +1.0, because I have "won" a pawn, then we hit the critical position with alpha and beta as +1 and +1.01 or something similar. I try the critical move and get a fail high, and all others fail badly low as I lose material if I can't play the critical move. I extend it and search it again. But in the other case, the expected score might be 0, and the ply after I hang the queen, the alpha/beta values are still 0,.01 or something similar. PxQ will clearly blow out past beta, but will all the other moves be worse than -.5 or so? Probably not so we don't extend. If I mistakenly compare to the alpha value after PxQ, then every other move might be much worse and trigger the extension and waste time. > >-S.
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.