Author: Dave Gomboc
Date: 09:03:24 09/03/02
Go up one level in this thread
On September 03, 2002 at 11:49:42, Robert Hyatt wrote: >On September 03, 2002 at 11:30:35, Gian-Carlo Pascutto wrote: > >>On September 03, 2002 at 11:21:53, Robert Hyatt wrote: >> >>>Maybe or maybe not. If what you "pick up" is a potentially game-winning >>>or game-losing "think" I'd think you would want a significant score change >>>to reflect that? >>> >>>I do... >> >>You should have picked up a part of it in the last iteration. >> >>-- >>GCP > > >Frogs should have pockets so they could carry guns/knives and not have to >worry about snakes. > >I don't know about you, but the reason I do a depth+1 iteration is to pick >up something new that I didn't see in the previous iteration. Otherwise I >wouldn't waste the time. > >If you pick up "part" of it in the last iteration, then your score must have >changed significantly. mtd(f) would have had problems there. And now you >are going to pick up another "part" of it in this iteration. And again, mtd(f) >is going to have problems... > >So long as your score changes in any significant way, mtd(f) will be hurt by >it. And how much fail-soft is going to help in reducing the number of >researches depends on how good your move ordering is in a tree where things >suddenly went badly wrong to throw the score way outside the current alpha/beta >window... > >I would never say it won't work. But for the kind of evaluation I have, I >don't believe it will ever work well. For the kinds of eval swings I have >seen from previous versions of Fritz, as well as Tiger and others, I don't >believe it will work particularly well for them either... > > >Back to the basic idea. mtd(f) is supposedly a 10% improvement over PVS. > >Suppose your program changes its mind significantly once every 5 consecutive >moves. Four of those searches will be 10% more efficient with mtd(f). The >last search will take a big hit. Probably at least a factor of two. Is it >worth it to save 10% four times, then lose 100% once? It takes a pretty simple >evaluation, or at least one that is _very_ stable (which to me implies simple >still) to avoid this. > >I worked on trying it for about three months. I didn't go "all the way" and >add both upper/lower scores/moves into the hash table entries, as mtd(f) really >needs. But I had to do so many re-searches that I didn't deem it worthwhile to >worry with that... I did spend a lot of time trying to limit the number of >re-searches. But if you do three or more, you begin to lose to straight PVS >in the general case... > >and you are going to be forced to do at least two searches in the best case... Well, I do think you need to at least make that switch (to storing both upper and lower bounds) to give it a fair shot. If you don't, the algorithm can ping-pong like crazy with the "right" input :-) Dave
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.