Author: Ed Schröder
Date: 23:46:31 08/28/00
Go up one level in this thread
On August 28, 2000 at 21:36:59, Christophe Theron wrote: >On August 28, 2000 at 17:42:59, Uri Blass wrote: > >>On August 28, 2000 at 14:15:57, Christophe Theron wrote: >> >><snipped> >>>So I think that no top program is a 100% root processor (not even 80% actually), >>>and I also believe that no top program is a 0% root processor. >> >>How can you calculate if it is 10%,20% or 30%? > > > >Uri, I'm very sorry to use numbers in such an inaccurate way, but I don't know >how to describe it better. > >A pity you can understand only mathematically sound sentences! :) > > > > > >>I know that a program is more than 0% root processor if I see cases when the >>evaluation is changed not because of deeper search. > > > >If you want to be picky about this, then let me give you one example: > >a null mover can be in some positions absolutely unable to see a win. Just play >the right move and let it think for the opponent and it sees the problem >instantly. There was an unexpected zugzwang. > >Problems in the search can produce such behaviours. > >I know this is not what you meant. It's just to be mathematically sound. :) > >This behaviour can also come from asymetry in the evaluation. It is the case >with Tiger for example (I have a little bit of asymetry for some evaluation >terms). > > > > >>Example: >>I remember cases when I analyzed a game with Genius3 and Genius gave me an >>advantage of 0.5 pawn. >>The program wanted to castle and expected castling from my opponent at all time >>controls. >>These moves were exactly the moves that were played in the game. >> >>After giving the moves to Genius3 it suddenly saw no advantage for me and not >>because of seeing deeper. >> >>When I see cases like this I know that the program does preprocessing. >>I know from my experience that i never saw things like this with chessmaster so >>it seems to me that chessmaster does not do preprocessing and if it does it is >>very small and does not change the evaluation by more than 0.1 pawn. > > > >The problem is just to play the right moves most of the time. Not to be right >with the evaluation most of the time. > >I understand what you mean, but if a little bit of preprocessing can improve the >search speed by 100%, that means an average elo improvement of 70 points. Being >accurate (more accurate) with the evaluation from move to move does not give as >much elo points. > >You may disklike preprocessing, but you would much more dislike a tactically >weak program. > > > > Christophe Good points Christophe. Let me add some comments too to clear this up. Preprocessing has a bad image and for good reasons. Preprocessing is a very bad thing if this is the BASE for a chess program. But preprocessing also has it advantages if you use it in a smart way mainly to "easy fix" some negative side effects you notice in eval and search (or whatever) as Christophe has pointed out with his null-move example. Zugzwang is the main pain of null-move and if it can be solved in the preprocessor than one big hurray. One Rebel example: when in the middle-game when white has castled on the king-side and there is an exchange of bishops on g3 Rebel tends to recapture with fxg3 instead of hxg3 which in most cases is not such a good idea. Now a programmer can do 2 things a) rewrite a major part of the king-safety that will fix the problem and make the program 5-10% slower because of the changes or b) fix the problem with preprocessing giving hxg3 in this specific case just the little push it needs. There is absolutely no harm if you use preprocessing in such exceptional cases, on the contrary it is just smart as solving it in eval would make the program weaker. There is absolutely no harm in preprocessing if you use it in a correct way to solve some MINOR things. Speaking for Rebel preprocessing cases are only a few and not more than 0.5% of its total knowledge. Ed
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.