Author: Christophe Theron
Date: 08:44:06 12/30/01
Go up one level in this thread
On December 30, 2001 at 11:14:19, Severi Salminen wrote: >>The first use of SEE (historically) is not to sort capture moves. >>The first use of SEE is to replace what we call now QSearch. > >Ok. > >>You arrive at a horizon node. Now you want an evaluation of this node. There are >>two main ways of doing this: > >Ok, I didn't remember what horizon node means. > >>1) you call your SEE, and it tells you how much material the side to move can >>win in this position (by doing a Static Evaluation of the possible Exchanges >>from this position). You return a score of eval+SEE (or eval-SEE depending on >>your conventions). >> >>2) you call the QSearch, which is going to try all the possible captures and >>return a minimaxed score. >> >> >>We are talking in this thread about the advantage of 1 on 2 (or vice-versa). So >>we want to compare the difference of the tree size of 1 or 2. >> >>While I understand that you want to count horizon nodes in the QSearch (after >>all each horizon node is the root of its QSearch), horizon nodes are also >>visited by the first method (SEE). >> >>So in order to compute the difference in visited nodes of the two methods, you >>need to count QSearch nodes that would not be visited by the SEE method. So you >>don't count horizon nodes (you don't count the root of the QSearches, only the >>moves or nodes which are generated BY the QSearch). > >And when you (or was it Ed) said you spend 15% of nodes in QS you exclude >horizon nodes? It was Ed, but yes I think he was excluding horizon nodes. > I think my search will look a lot better with this new counting >method :) If I count the way I used to do, I have 65% of all nodes are qnodes. >I'll test what the number is if I include horizon nodes to actual nodes and not >to qnodes. > >I think we should set some standards for these bloody nodes and qnodes... Let's say that a qnode is a node GENERATED BY the QSearch. A node is a node searched by the main search or the QSearch. qnode_count / node_count <= 0.30 if your QSearch is working reasonably well. Christophe
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.