Author: Ernst A. Heinz
Date: 11:01:16 10/04/98
Go up one level in this thread
On October 04, 1998 at 13:33:30, Ed Schröder wrote: > >It is certainly years ago I worked on the ideas you brought up here so >I don't know all the details anymore however I will give it my best shot. Ed, First of all -- thank you very much for the detailed and informative answer. I greatly appreciate that you are willing to share some internals of "Rebel" with us! >First you have to understand Rebel is no "null-move" program (and probably) >never will be so "selective search" is done in a total different way as these >days is common using "null move". Yes, I know because you have already mentioned this before. >Rebel's selective search basically is the "static score" coming from the >standard evaluation function which is compared to ALPHA first. Next >a lot of exceptions (say chess knowledge) are checked. Based on all >the information the selective search decides to a complete prune (the >very bad ones) or to reduce the depth with 2 plies (the bad ones) or >to reduce the depth with 1 ply (bad, but give it a try anyway) or leave >the depth unchanged (good moves). Okay, so "Rebel" (at least up to version 9, I guess) is still a purely statically selective searcher in the old tradition of Shannon-Type B. >In the selective part I have tried the ideas you mention on the last 2-3 >plies and even on all plies of the selective search. It speed-up Rebel >tremendously in the sense of a much higher iteration depth but I always >had bad feelings about the algorithm (loss of positional understanding) >which later was confirmed by the results I got from auto232. And - / - >means drop the idea. > >The idea still has my attention, after all these are the days of Pc's on >450 Mhz and my last tries are from the ChessMachine 16 Mhz days >if memory serves me well. In my opinion, this explains much ... :-) My experimental results show that these statically selective forward-pruning techniques require minimum search depths of 9-10 plies in order to really take off in a state-of-the-art null-move searcher. At shallower depths (which you probably encountered on the 16MHz "ChessMachine") this kind of pruning is actually quite risky -- I would probably disable it in blitz games. Moreover, my limited and risk-assessed pruning at depths of 2 and 3 plies might surely combine better with null-move pruning than with a purely statically selective search as that of "Rebel". =Ernst=
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.