Author: Ross Boyd
Date: 02:36:25 12/20/05
Go up one level in this thread
On December 20, 2005 at 03:04:21, Uri Blass wrote: >On December 20, 2005 at 00:01:55, Mike Byrne wrote: > >>I had noticed that Rybka is very tough for me even at sd==1 (a level where I can >>usally can hold my own. > >Yes >There are 2 possibilities >1)Rybka searches to depth that is more than 1 at depth 1. >2)Rybka's evaluation detects a lot of tactical material including a fork in the >next ply > > >Here is an example > >[D]3Nk3/4pppp/8/8/4r1q1/8/3PPPP1/3KQ1R1 b - - 0 1 > >Every program that I know except rybka beta is going to play Kxd8 at depth 1. >The fork by f3 is too deep for them to see it. > >Rybka prefers Rb4 at depth 1 > >I have doubt if it see the fork or see another the threat to win the queen so I >composed another position > >[D]3Nk3/2pppppp/1p6/p7/P3r1q1/1P6/2PPPPP1/3KQ1R1 b - - 0 1 > >Here rybka prefers Rd4 at depth 1. > >It may be extensions or it may be that the evaluation ask question like can the >opponent do fork next ply. > >If this is evaluation then it is very smart evaluation because the evaluation >can see that the fork does not work because of check after it in the following >position when Rybka plays Kxd8 at depth 1. > >[D]3Nk3/2pppppp/1p6/p7/P3r1q1/1P6/2PPPPP1/3QK1R1 b - - 0 1 > > >Uri You have excellent powers of observations Uri. Very interesting. I have been playing with an eval that rewards tactical motifs (such as potential pawn forks). All I can say is it is really hard to write something that is sufficiently efficient and accurate without losing out in speed. Also, as you mention, Rybka's eval/qs clearly 'sees' that forks can be evaded by giving check. That is very powerful. On the other hand, if a search can prune heavily and intelligently enough then it can afford to include some very cpu-intensive elements in the eval. And its amazing to see two top engines using such dissimilar methods. Fruit has a (relatively) simple eval and clean search... but Rybka is clearly doing more work at each node (although GCP pointed out that possibly Rybka is not counting QS nodes... so the low nps may be exaggerated). Cheers, Ross
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.