Author: José Carlos
Date: 14:06:47 01/05/01
Go up one level in this thread
On January 05, 2001 at 14:57:24, Severi Salminen wrote: >> There're many possibilities here, and the best one can do is try them all. >>What I do in Averno is: >> - The first time, I sort with my normal criteria (wining captures, etc...) >> - After that, every fail high at the root makes that move become the first of >>the list so, after, several fail highs, at the beggining of the list I have the >>last fail highs, the most recent first. > >Could you or have you compared this with the "node" method? Which one gave you >the better results? I like the node approach because it gives a priority number >for each move in root node "looks" like a sound method. I tried the node method a while back and, as I didn't notice any differenc (I mean, better for some positions and worse for others) I kept my method just because it seems more logical to. Maybe it's not the best, but try to think what happens when a human player finds his prefferred move is bad. I (2135 FIDE) usually continue searching my second-prefferred move, and I like my program to do the same :) >>>>What is the best way adjust alpha and beta at the root after each iteration? >>> >>>First I do a 1 ply search. Then I do search with 1 pawn wide window. >> >> Aspiration search usually uses the value from the previous search, istead of >>this 1 ply search. >> My AS window is 25 centipawn. > >I meant that first I do 1 ply search and the I allways use the score from >previous iteration. So the 1 ply search is done with infinite window. Ok, so you mean your first ply search is full-window, and only narrow it for the second iteration, right? >>>If I fail >>>high or low I do a research: score-1, INF (fail high) or -INF,score+1 (fail >>>low). The score is the score returned from the search that failed low or high >in >>>the first place. >> >> I do an intermediate try, adding/substracting 150 centipawns and only if that >>fails high/low again, use +INF/-INF. > >Maybe I should also not give an infinite window at first place but try a wider >window first. Well, I try to make my program play so that it improves its >position 0.50 pawns on every move so I don't have to re-search ;) Not bad idea :), but anyway you should try a somewhat wider window before going to -INF,+INF. >>>Notice those +1 and -1. Without those you'll get fail lows and >>>highs again. I use fail soft and this means that you have to be sure what to >>>return from search, qsearch and evaluate. A wrong value returned from (say) >>>futility pruned node means problems. The point is that if you fail high or low >>>in the first search you have to be sure that the score returned is the other >>>bound of the true score. >>> >>>>I notice I sometimes when I get a fail low at the root, I do a research, and >>>>then get a fail high! The most likely explanation is that I am a moron and it >is bug. >>> >>>Well, we are all morons, so that can't be an explanation. But yes, it must be a >>>bug. >> >> Hashing and null move can very well cause this behaviour, not necessarily a >>bug. > >Ok, I don't do hashing yet. > >Severi Null-move, maybe...? José C.
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.