Author: Robert Hyatt
Date: 08:11:59 04/07/05
Go up one level in this thread
On April 07, 2005 at 06:57:15, Johannes wrote: >On April 06, 2005 at 11:40:11, Robert Hyatt wrote: > >>On April 06, 2005 at 07:37:55, Johannes wrote: >> >>>Hello! >>>Right now im working on a new Version of my Chess Engine and an idea has come to >>>my mind, but i dont know if its applicable: >>>When I'm in Q-Search i only want to deal with material values. I handle theses >>>incrementally, so theres is no need for any function call except the Q-search >>>recursion. >>>After returning from Q-Search (at the full-width horizon) i simply add the >>>positional eval score to the score returned from Q-search. This score is >>>returned "up the tree". >>>For the positions before the horizon this should be no problem concerning >>>alpha/beta values etc... . Also positions beyond the horizon should not suffer >>>from this, as long as the alpha/beta values were not passed to a Node before the >>>horizon before. If they were, the alpha/beta bounds would not be correct, >>>because they include positional scores which are not considered beyond the >>>horizon. >>>So my idea is to simply substract the local positional score from the alpha/beta >>>values (as long as they are not -INF or +INF) at the full-width horizon, before >>>entering Q-Search. I assume that this would result in correct cutoffs. >>>Is this idea correct? >>>I didnt find any drawbacks yet, but i only have a vague intuition about it. >>>I really would appreciate if anybody could help me with this. >>>greets >>>johannes >> >> >>You can "cheat the system" however you want. In this case, you are positionally >>evaluating the board, and then making captures that could greatly change that >>positional evaluation, but you won't realize it. >> >>For example, a capture that shreds the king-side pawn structure will not be >>scored correctly since you won't realize that the king-side is destroyed, >>because you use the positional evaluation before the capture is made and just >>factor in material only after that point. Ditto for endgame ideas such as a >>pawn running after all pieces are exchanged. If you evaluate the position >>before the exchanges, the pawn can't run, if you then make the exchanges and >>think all is ok, you will lose. >> >>It is a dangerous road... > >Hi there! >Thanks for the quick responses. Of course your right about the problems with >missing information in the Q-Search. The background to my idea is the following: >Q-Search was always intended to be a safety measure to prevent unstable board >positions from being evaluated. Therefore I'm trying to put as less effort >(CPU-Time) as possible in it, e.g. dealing with material scores only safes time >here. >This increases my full-width horizon. I think this the tradeoff im facing here. >The Q-Search is bad informed, but in my opinion it should safe me from the >Horizon Effect in the first place. >The problemes you mentioned are really the only ones i can think of. There are >positions where the win/loss of one piece can greatly influence the static >positional score. In this positions the score returned from full-width horizon >wont be correct, because a capture beyond the horizon would have this influence >(although the capture is already considered in the material term). >Im going into a bit more experimenting on this idea, because i think it would be >nice to find a way to perform the Q-Search as cheap as possible. I dont want to >be informed by Q-Search, i just want the safety of a stable material score. >greets >johannes I personally subscribe to your "Q-search is a safety net only and should be as quick as possible." I do this by limiting what I search, rather than searching more but mis-evaluating it. Any time you move a piece or pawn, the score can change significantly. You have to weigh the cost of detecting the score change, vs the cost of not detecting it and accepting the error(s) that will produce. If you do pretty reasonable pawn structure stuff, then the errors will be larger, as moves like Bxf6 gxf6 wrecks the pawns in front of the king if it is at g8/h8, and you want to know that. Or if the capture removes the last opponent piece so that a passed pawn for either side is free to run and promote. Etc. Every error can lose a game. Of course, a faster q-search can let you see deeper and win a game as well. Your only concern has to be that you make sure that when you add up gains and losses due to a change, that the gains are larger than the losses for a net positive effect.
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.