Author: Tord Romstad
Date: 13:30:05 09/27/05
Go up one level in this thread
On September 27, 2005 at 15:05:08, Steve Maughan wrote: >You're saying that Fruit 2.1 has a "bug" since it cannot solve Fine 70. I'm >saying that you cannot say there is a bug in Fruit 2.1 just because it does not >solve one position. If Fruit was using a conventional hashing approach then >Fine 70 would normally be solved quickly but since it's using something that is >really quite different (is anyone else using double bounds?) then I think all >the conventional tests go out of the window - Fine 70 included. Using double bounds is not at all unusual. I am fairly sure almost all MTD engines use them, including both of mine (the old Gothmog and my most recent Glaurung development version, which also uses MTD). Both of them solve Fine70 instantly. Double bounds are less common in PVS engines, but I know there are a few who use them (I don't remember any names, though). You make it sound like using two bounds is a very radical and fundamentally different way to use a transposition table, but it isn't. The information found in a transposition table with two bounds is a superset of the information in a table with a single bound. Assuming an equal number of transposition table entries, a program with two bounds should normally search slightly smaller trees than an equivalent program with a single bound, because of a bigger number of transposition table cutoffs. On the other hand, each entry will of course consume more space in a two-bound transposition table, and the number of entries you can fit in a given number of MBs will therefore be smaller. Whether using two bounds is worth the price depends on the program, but most of the evidence I have seen indicates that it doesn't make a big difference for most PVS engines, but that two bounds are clearly superior in MTD engines. One minor annoyance when using two bounds is that it occasionally happens that an entry has an upper bound and a lower bound with the same depth, and the upper bound is *smaller* than the lower bound. This is an unavoidable side-effect of allowing transposition table cutoffs. In practise, this doesn't seem to cause any noticable problems. Whether this phenomenon could be part of the explanation for Fruit's problems with Fine70 is an interesting question (although thinking about it makes my head hurt really badly). I personally think Fruit's problems are entirely unrelated to using two bounds, but I wouldn't bet on it. 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.