Author: Dan Homan
Date: 20:57:13 07/20/99
Go up one level in this thread
On July 20, 1999 at 20:35:49, Vincent Diepeveen wrote: >On July 20, 1999 at 11:25:07, Dan Homan wrote: > >>On July 20, 1999 at 08:52:18, Vincent Diepeveen wrote: >> >>> >>>Let's however write down some facts why my prog is unhappy with MTD. >>>It's up to others to generalize it to their progs: >>> >>> - the huge number of researches needed. In DIEP my evaluation is nowadays >>> in 1/1000 of a pawn. For a long time i had 1/200 of a pawn (in the time >>> i experimented with MTD), but now i have 1/1000 of a pawn. So a drop >>> of 0.20 pawn, which is a normal drop in DIEP, is in fact a drop of 200 >>> points. Happy researching! >> >>Even if the score dropped a whole pawn (1000 points in DIEP), this would >>only be 10 or 11 zero-width searches (2^10 = 1024) assuming that you >>bound the score in an efficient manner. Also, I can understand having >>a high resolution within the eval routine itself, but does it really help >>to have the output of the eval be in 1/1000 pawn units? I wouldn't trust >>the sum total of any eval routine to 0.001 pawns! Maybe you could >>output the total eval in units of 1/100 of a pawn (or even less). >> >>Actually, that is an interesting question. Does anyone know what the >>optimum eval unit is for searching? I am talking here only about what >>the eval outputs - not the unit used internally for calculating the eval. >>I know that most programs use the same unit for both purposes, but I >>wonder if that is really optimum. > >Every time i get to a draughts tournament i have to laugh a little when >playing a certain program. It has a stone at 1000 points worth, >yet he has no positional eval in 1/1000 of a stone. Only in 1/10 of a >pawn. So he's either at +0.100 or -0.100 or a multiple of that. > >If you only count material in chess >you can of course simply use 10 for a pawn. >a pawn = 1, knight=3,bishop = 3, rook=5, queen=10 >and you're ready. MTD will rock i bet. >Start with window 0 from root, >and end the game with a window -100. > >Hope that answers, as your question is kind of retoric. My question was serious. I have a way of going from one idea to another - sorry about that. I started by pointing out that you didn't need to do as many researches as you thought, to wondering what benefit you got from using 1/1000 of a pawn during your search, to the general question of whether it is a good idea to use different eval resolutions internal to the eval routine and when you return a value to the search. I wasn't suggesting anything as crazy as setting the eval resolution as coarse as one pawn. What I was curious about was whether the search was improved by having you eval output results in 1/1000ths of a pawn. I can understand the benefit of having 1/1000 internal to the evaluation routine - this way small positional benefits can add up to something significant, but might is it useful to return the value to the search with less resolution than 1/1000th? Do you see what I mean here? Lets say the eval works in 1/1000 units. Add all the scores up for pawn structure, king safety, piece placement, etc... Now we have a total score in 1/1000 units of a pawn, but do we trust this total score to this accuracy? Should we return this raw score directly to the search or should we round it down to 1/100 of a pawn or perhaps even less? Is there benefit in keeping the 1/1000 resolution when we return the total score to the search? Is there benefit to the searching routines in rounding the score down to a coarser resolution? I don't know the answers to these questions. I tried it briefly with my program (which works in 1/100 of a pawn normally) by simply bit shifting the eval result so the resolution was 1/25 of a pawn. I noticed a (small) speed-up in the search, but I worried that this resolution was too coarse and took the changes out. - Dan > >Sorry to bother you, forgot the name of your program and the number of >nodes a second you get with it. > > > >> - Dan
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.