Author: Vasik Rajlich
Date: 15:03:01 06/17/05
Go up one level in this thread
On June 17, 2005 at 14:49:12, Dieter Buerssner wrote: >On June 16, 2005 at 18:41:09, Vasik Rajlich wrote: > >>On June 16, 2005 at 16:20:38, Dieter Buerssner wrote: >> >>>On June 16, 2005 at 05:22:26, Vasik Rajlich wrote: >>> >>>>Fail soft helps when you need to re-search, so it helps more in MTD (f) than >>>>PVS, and doesn't matter at all in pure alpha-beta. >>> >>>Vas, fail soft vs. fail hard will change the search tree in "pure alpha-beta". >>>It will change move ordering. Fail soft has the potential, to give you better >>>move ordering. Assume some fail low position. Some refutations on ply deeper >>>mate, others just give the bound back as score, or scores in between. With fail >>>hard, you may order moves, that are refuted by mate early for the next search >>>(it needs some luck, to see mate scores, of course). In a later search with >>>different bounds and fail hard, you may search those moves early (depending on >>>all the move sorting heurists you use, of course. Yace for example stores a best >>>move in HT even when failing low). With fail soft, you might start with a more >>>reasonable move first. >>> >> >>Aha, yes - that's true. >> >>BTW I also store fail-low moves based on the fail-soft score. I doubt it helps >>much, but it at least shouldn't hurt. > >Yes. However I remember what Ernst Heinz said to me once: "fail hard läuft >einfach besser". Seems not really logical - but I guess, he has a point. Perhaps >the point is, that it is much easier, to have some subtle bug in fail soft >search. alpha-beta is not really forgiving here - it will not crash with such >bugs, just not perform as good as it should (perhaps only in few positions). > Yeah, that's a pretty mysterious statement. I see no reason to ever fail hard. I don't even see how fail-soft is more bug-prone. > >>>This also reminds me to the typical move ordering statistics used: "How often do >>>I fail high in the first move relative to all node I failed high?". In many >>>positions, (almost) any move will fail high for the current beta. With a higher >>>beta, only one or a couple may do it. When you picked some stupid move in an >>>earlier search, instead of one really convincing one, it should be better for >>>later searches. >>> >> >>True - although fail soft won't help you with this. > >It can help through the hash table (for example). It is very similar to the >situation of PVS - where fail soft can avoid some researches. The situation >arises, when you get a fail high/fail low at the root. Now you have to research >anyway - and fail soft scores from the HT may give a cutoff in this research, >while fail hard scores wouldn't. > Yes, you'll save some nodes, but your HT moves at fail-high nodes should be the same for fail-soft as for fail-hard. (Or at least they shouldn't be systematically better.) I don't see how to solve the problem of weak fail-high moves except by actually failing low .. Vas >>Some time ago, I was taking some statistics about my fail-soft. What I found is >>that it wasn't very soft. Once you do a search of above 6 or 7 plies, the search >>manages to get very close to the fail-hard value. > >Same here. > >Regards, >Dieter
This page took 0.02 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.