Author: Omid David Tabibi
Date: 03:52:57 11/22/02
Go up one level in this thread
On November 22, 2002 at 06:45:30, Gian-Carlo Pascutto wrote: >On November 22, 2002 at 01:48:06, Robert Hyatt wrote: > >>It isn't what he is claiming. He claims that R=3 + verification is close to >>R=2 in nodes, and has fewer null-move failures. His data seems to support >that. >>R=3 with a depth-1 verification ought to be fairly close to R=2, just based on >>pure math. I'll leave it to you to figure out why... > >I don't really agree. > >I'm assuming you do the fairly intuitive math of 2+1=3 but things >are not so simple :) > >R=3 verif. does a R=3 search, one depth reduction on fail high (which >makes it equivalent to R=1 without nullmoving at that ply, but it is >safe because you guaranteed your opponent has no serious threat), and >R=3 cutoffs everywhere below > >R=2 does, well, R=2 cutoffs > >It's not so obvious these are close in nodes. In fact, the paper itself >points of that the methods scale very differently. > Many things are not that obvious. Please read the "Conclusions" section for other algorithms I tried but were inferior to the presented algorithm. One interesting point is that at depth 8, the size of the tree constructed by vrfd R=2 was slightly larger than std R=2; at depth 9, vrfd constructed a smaller tree, and the gap widens as we search deeper (see Figure 4). So, I believe than on every program, starting from a certain depth, vrfd R=3 will construct much smaller trees in comparison to std R=2. And the benefit will increase as we search deeper. >-- >GCP
This page took 0.02 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.