Author: Robert Hyatt
Date: 15:18:03 08/27/02
Go up one level in this thread
On August 27, 2002 at 16:09:35, Uri Blass wrote: >On August 27, 2002 at 15:51:38, Robert Hyatt wrote: > ><snipped> >>>The question is if they mean to the plies that they sacrificed for the singular >>>extensions. >> >>I believe so. the only comment to any sort of selective forward pruning that >>I have ever heard from then was the futility pruning comment regarding the >>chess processor. In the games vs the micros, they said that they turned this >>off which slowed them by a factor of 10 or so, somehow... > >The point is that they did not say in the paper that they sacrificed 2 plies for >singular extensions but that they sacrificed 2 plies for selective search >algorithms. > >singular extensions are not selective search. yes they are, as I have explained before. "selective" means "to choose between". You can choose to forward prune, or you can choose to selectively extend certain lines. Either can be used to produce the same search tree/ result... Their idea of selective search was _always_ to extend interesting lines, while the forward pruner's idea is to discard uninteresting lines early. >selective search is what computer did in the past when they were not fast >enough. Not necessarily. Forward pruning _is_ one definition of selective search. But it isn't the _only_ definition. See above. All that "selective search" requires is that some lines be searched more deeply than others. Forward pruning or selective extensions both do this... > >No doubt that it does not make sense to do selective search close to the root >but doing selective search in the last 6 plies is a possible idea to consider. > Yes... IFF you mean "forward pruning" when you say "selective search". If not, then extensions near the root are actually more common. tapering the extensions off as the tree drives deeper. ><snipped> >>>In Crafty the processors always work and in deeper blue it was different. >> >>Not exactly. When talking to Hsu on many different occasions, he clearly said >>that it was not hard to run the chess processors at 50-70% duty cycle. He >>explained why 100% was very difficult (because of the balance between host >>processor speed and chess processor speed), but he was pretty clear about >>the 50%-70%. Which ought to translate to 500M-700M RAW nps on DB. > >was not hard does not mean that it was done. >The question is how much time he needed to do it and what he considered as not >hard. > Since this was the topic of his dissertation, I presume that he had looked at it pretty carefully... >If he needed some weeks and he considered the evaluation as more important then >it is possible that he focussed on the evaluation. > >Uri
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.