Author: Volker Böhm
Date: 22:59:27 09/12/04
Go up one level in this thread
On September 12, 2004 at 19:39:10, Scott Gasch wrote: >On September 12, 2004 at 18:19:04, Volker Böhm wrote: > >>Hi Scott, >> >>the idea behind a check-extension is not only to find great attacks beginning >>with a checking move. One other major idea is to prevent horizont effects. >>Example: you give away a piece for a really good attack. But the opponent has >>the opportunity to give you some checks, pushing the really good attack out of >>sight. Extending the checks in positions below alpha helps to fix this problem. >> >>I never found a good idea to reduce check extension. > >I always extend a checking move 1 ply. What I'm talking about is deciding to >either extend or not extend the one legal response on the next ply. I agree >with you about checks: I've never found a way to safely not extend them. Even >if the checking side is way ahead it could be checking just to push something >really bad over the horizon. Hi, the 1 legal response extension is good to prevent horizont effects too. Maby that is why your extension-reduction code does not give good effects. > >>By the way why do you extend on promotion? I never heard anyone doing it. Thus >>maybe the reduction will work here (the more reduction the better :-) >> >>A reduction that gained a little was the pawn-pushed-to-row-7 extension that I >>only extend, if the pawn on row 7 cannot be taken without material loss (SEE >= >>0). > >I wonder if the price of running the SEE on such a move is worth it. Or are you >doing it with attack tables? I do the poor man's version of that -- I don't >extend a pawn push if there's a piece blocking it in its new position. Yes I am doing it with a "simple SEE" using attack-tables. But I think that a full SEE is not too expensive here. > >>Something that could be tested (I think) is the reduction of the >>take-back-extension that is perhaps not that usefull if far away from the >>search-window. >> >>On the other hand: maybe futility, extended futility and razoring is a better >>concept than reducing extensions by comparing against alpha and beta. Thus I >>think you should test your ideas only if you have at least futility. > >I have futlity in the qsearch but nothing else. The first post I made is about >how if you decide to extend or not based on bounds you could run into a >situation where you search with a minimal window, decide to extend, fail high, >research with a full window, decide not to extend, and fail low. > >I think the same thing is possible with all this futility stuff people like to >do: if you are deciding to prune moves or not based on alpha you can search >quite a different tree on the research of a node with a full window, right? Yes right. My problem with the idea of reducing extensions according the current window is that you don´t include the remaining search depth in this decision. If you are far away from horizont you will perhaps have much opportunities to get back to your window. If you are near the horizont you will have much less opportunities. What is the difference between not extending the depth and reducing the depth? For me there is no clear difference. The reduction (futility, ...) is better done near the leaf than the root of the search, thus not extending shoud take that into account too. Thus I think that reducing extensions based on score and window is not the right thing to do. Greetings Volker > >Scott
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.