Author: Nagendra Singh Tomar
Date: 09:32:23 10/25/02
Go up one level in this thread
On October 25, 2002 at 09:28:32, Peter Fendrich wrote: >On October 25, 2002 at 08:39:47, Nagendra Singh Tomar wrote: > >>Hi everybody! >> >>In my endeavour to understand the techniques of comp chess I am in the process >>of understanding the benefits (what/why/how) of a zero window search. Told >>superficially it can be said that since we have a very small window to search we >>search less nodes, but I want to understand it to the minute details. >> >>What I observe and understand is follows. >> >>score = -alphabeta(board, -beta, -alpha, depth-1) <-- 1st call (full width) >>. >>. update alpha if score more than alpha >>. >>score = -alphabeta(board, -alpha-1, -alpha, depth-1) <-- subsequent calls >> zero window >> >>Any saving whatsoever that can be achieved by reducing the window size, should >>be because of beta cutoffs (fail-highs) increasing. >>Here we are keeping beta as -alpha (same as we would have kept it for a full >>window search, though we have kept alpha as -alpha-1 in contrast to -beta for a >>full win search) >>But in the subsequent call (child nodes) this alpha (-alpha-1) will be negated >>and passed as beta, which will be very less. Is the saving in nodes searched, >>due to this beta being passed as a very small value in the child->child nodes >>searched. >>I know I have made it complex, but I hope you understand me. >> >>IOW changing the value of alpha cannot reduce the number of nodes searched. But >>the fact that this alpha will be passed as beta in the subsequent call to >>alphabeta, results in lot of beta cuts and hence reduced nodes searched. >> >> >>Am I thinking on the right lines ? >> >>tomar > >I think that you've mixed up the alpha and beta values from the calling nodes >view. -alpha-1 is a temporary beta value and not a new alpha. >I would explain it for myself like this: >I assume that the first search (the full width) is with the best move in the >node. Therfore I reduce the beta value to the most narrow window possible, >namely alpha+1. So the the rest of the searches will use (alpha, alpha+1) until >proved wrong (score >= alpha+1) and I have to re-search with a wider window. >Now beta replaced by alpha+1 negated is -aplha-1. > >Maybe that's what you said? > But the idea is that we have tried the best move first so no other move will be more than alpha and hence no move will "prove it wrong". The gain we r getting by getting fail-highs at the next lower ply in the tree >Peter
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.