Author: Robert Hyatt
Date: 13:03:47 09/11/02
Go up one level in this thread
On September 11, 2002 at 13:20:55, Vincent Diepeveen wrote: >On September 11, 2002 at 12:31:44, Robert Hyatt wrote: > >>On September 10, 2002 at 20:45:43, Vincent Diepeveen wrote: >> >>>On September 10, 2002 at 18:06:01, Gian-Carlo Pascutto wrote: >>> >>>>On September 10, 2002 at 17:51:11, Vincent Diepeveen wrote: >>>> >>>>>On September 10, 2002 at 17:43:15, martin fierz wrote: >>>>> >>>>>>On September 10, 2002 at 17:18:24, Vincent Diepeveen wrote: >>>>>> >>>>>>>On September 10, 2002 at 17:10:38, martin fierz wrote: >>>>>>> >>>>>>>>On September 10, 2002 at 09:26:14, Eli Liang wrote: >>>>>>>> >>>>>>>>>(3) Reading Aske Plaat's search & re-search paper, it really seems like mtd(f) >>>>>>>>>is something of a magic bullet. But I note it seems that more programs don't >>>>>>>>>use it than do (for example Crafty). What is wrong with mtd(f) which Plaat >>>>>>>>>doesn't say? >>>>>>> >>>>>>>losing 1 bit is a problem for you? >>>>>> >>>>>>nope. losing 2 bytes is more like it... >>>>> >>>>>who stores a bound in 2 bytes? >>>>> >>>>>Why not in 1 bit? >>>> >>>>You want to store two actual values, not flags that indicate what >>>>kind of bound it is. >>> >>>did i implement it smarter then or what? >>>i used 2 bits in total. 'upperbound, lowerbound, truebound'. >>>the search result is based upon a single bound. So it IS the same, >>>it IS higher or it IS lower. >>> >>>What am i missing here? >>> >> >> >>A _lot_. >> >>This is a known issue with mtd(f) and fail-soft. If you bounce over the >>true score, to the "other side" then suddenly where you were storing >>upper bounds you are now storing lower bounds, and vice-versa. Now if >>you bounce back over the true score again, you have no useful upper bounds >>where you need them, you only have lower bounds. And you have no useful lower >>bounds where you need them you have only upper bounds. > >you never bounce over the true score. > >You go from a score X --> Y in a lineair way. _not_ necessarily. In pure mtd(f), maybe not, it depends on how you adjust the score between iterations. Using fail-soft _can_ make you bounce over the true score... > >If it's a big distance, bad luck, we need many researches. If it's >very close, lucky MTD. > >I do not see how i can get *anyhow* jump over the true score unless >there is search bugs in the program (like forward pruning or lazy eval >which is wrong implemented or other futility stuff). fail-soft stores various scores in the hash table that might not be correct for lots of reasons. null-move and other things already cause false fail highs. That happens in mtd(f) also... > >If i go from 0.001 0.002 0.003 0.004 0.005 0.006 0.007 > >then i reached my goal. > >If you go to -0.007 then it's simply a few searches more: > > 0.001 is first search then: 0.000 -0.001 -0.002 -0.003 -0.004 -0.005 -0.006 >-0.007 > >I do not see *anyhow* how you can jump over the score. See above. And you should start _above_ the score anyway and work down, because fail lows are easier to prove than fail highs at the root. > >Because if you *do*, then you aren't searching with MTD, but >with some badly implemented minimal window search. > >>if you store _both_ then you don't take as big a performance hit when you >>"bounce over" the true score. >> >>This is what I mentioned when I said "doing mtd(f) is _not_ just a few lines >>of code changes, it requires fundamental changes in places like hashing." >> >>>>-- >>>>GCP
This page took 0.01 seconds to execute
Last modified: Thu, 07 Jul 11 08:48:38 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.