Author: Vincent Diepeveen
Date: 10:01:51 06/20/05
Go up one level in this thread
On June 20, 2005 at 12:28:36, rasjid chan wrote: > >provided:- > R(mtd)=benefit-of-mtd/cost-of-eval-granularity-reduction > 1.0 MTD cannot work in well ordered trees. >I am no expert, but the idea of MTD don't seem appealing when we need to start >with a "guess" and then the prospect of being kicked and proded many times. >Most likely we need to wait for a genius if it can be made to work better then >PVS. Nature is not very generous. It has provided a round wheel of PVS search >and to get something more round don't help much. Maybe... Not even a genius can something that needs theta( c * x^y ) where c = huge and for sure bigger than 10 but usually more like 20 make faster than theta( 2*x^y) >Most who have an elaborate eval() assume their eval() is sensitive to greater >than 1/10 pawn but it may not be. If there are 30 critical eval() factor, then >missing 5 may be serious. Further, who can have perfect eval() tuning. So >we may say very few can say their eval() is sensitive to 1/10 pawn. pawn = 1000 >If this is the case then mtd may have a chance. Between strong programs, it >is like a waiting game of who make the first blunder when score goes from 0.2 >to -0.4, then to -0.8..-1.3... MTD will be searching plies less deeply when score starts falling. So worst case is horrible indeed. >MTD, when we have eval()/10 is like maximizing on it's opportunistic nature >and it waits at 0.2 and remain inactive from 0.11 to 0.29 but pounced when it >goes to 0.3/0.4...0.7... 1.4 when then the weakness of mtd do not matter any >more. > >I doubt any one tested with eval()/10. I can't test at the moment as I have >disabled research. besides that division instruction is one of the slowest instruction on your computers chip, it won't work of course. This is so trivial. Any resolution you lose will take care you play worse. Note that dividing by 10 doesn't solve the problem if pawn = 1000. Crucial moves played by diep at world champs 2004 were selected at resolutions of 0.024 of a pawn fail highed in last seconds very regurarly. I have seen in games also fail highs of 0.001 of a pawn which selected a BETTER move. Obviously, a more rude resolution will be the same like throwing away hundreds of rating points. >Best Regards >Rasjid
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.