Author: Roy Eassa
Date: 09:29:33 10/05/01
Go up one level in this thread
On October 04, 2001 at 08:29:35, Torstein Hall wrote: >On October 03, 2001 at 18:55:30, Uri Blass wrote: > >>On October 03, 2001 at 18:33:10, Torstein Hall wrote: >> >>>On October 03, 2001 at 17:59:14, Uri Blass wrote: >>> >>>>On October 03, 2001 at 16:51:49, Torstein Hall wrote: >>>> >>>>>On October 03, 2001 at 16:10:43, Robert Hyatt wrote: >>>>> >>>>>>On October 03, 2001 at 16:02:26, Torstein Hall wrote: >>>>>> >>>>>>>What if you run two paralelle search/chess processes. One going very fast with >>>>>>>very little evaluation. The other going slow, with a big evaluation. The fast >>>>>>>one always start searching on the move calculated by the slow process with the >>>>>>>big evaluation, just checking for big materiall loss, tactical stupidities >>>>>>>further down the tree. If it find one, the fast process sends a message goes >>>>>>>back to the slow process and tells it do start work on the next best move. >>>>>>> >>>>>>>Then you perhaps can have the best from two "worlds". Intelligent search, with >>>>>>>no tactical blunders! >>>>>>> >>>>>>>Torstein >>>>>> >>>>>> >>>>>>Read Jonathan Schaeffer's reports on "Sun Phoenix". He did exactly that. >>>>>>But he did it because he was not getting a very good distributed speedup >>>>>>on larger numbers of processors. So some did a normal chess search together >>>>>>as a group, the rest ran a tactical searcher called "minix". Minix was used >>>>>>to refute moves chosen by the positional program. >>>>>> >>>>>>The problem is trying to rationalize the knowledgeable search vs the tactical >>>>>>search. If the tactical search says your positional move loses material, what >>>>>>do you do? Propose another move? And if _that_ loses material? The search >>>>>>becomes hugely inefficient... >>>>> >>>>>If the tactical search says you lose material, of course you have to change the >>>>>next best move, and so on. But that limit can perhaps be even more than a pawn? >>>>>And of course if all the first moves in your oredering are very bad, it do not >>>>>matter that much if you are not very effective. You are probably already >>>>>loosing! >>>>> >>>>>Anyway, its another approach, and I find the consept intriguing. >>>>>Can I find Jonathan Schaffer's report on Sun Phoenix on the net somewhere? >>>>> >>>>>Torstein >>>> >>>>I totally dislike the idea of using fast searchers for tactics. >>>> >>>>I believe that in theory a fast searcher should be worse in tactics because it >>>>has no idea which lines to prune or extend and has no idea about the right order >>>>of moves. >>>> >>>>It is possible to get more nodes per second by not using null move or extensions >>>>and not calculating order of moves but the program is certainly going to have a >>>>bigger branching factor and it is going to be weaker in tactics. >>>> >>>>My example is an extreme example but I believe that if you want an engine to be >>>>better in tactics then doing it slower in nodes per second may be a good idea. >>>> >>>>Uri >>> >>>I for one do not have the time or ability to build a chess program, so I can not >>>argue about, what teqnique to get deepest first. But the whole idea was to have >>>a fast tactical deep searcher to correct a slow, big "evaluator". >>> >>>Torstein >> >>I believe that if you care only about tactics then it is better to have a slow >>searcher and in this case adding a big evaluation is not going to be a big >>problem. >> >>Adding a big evaluation is a big problem if you have a fast searcher and in this >>case the program may be 10 times slower because of complex evaluation but if you >>add big search rules before evaluation then the program may be only 2 times >>slower thanks to the evaluation and adding the big evaluation is not a big >>problem. >> >>Note that I also did not build a chess program(I built only a move generator to >>calculate the number of legal games of fixed number of moves) and it is only my >>intuition. >> >>Uri > >The idea was for a two processor system to have a fast tactical "checker" >running in paralell with a slow large evaluator. If the fast tactical search >finds a refutation, then you give a message to the slow big evaluator to change >line. If this realy would work, is quite another question! :-) > >Torstein I think the idea is that a "smart" search is often actully _faster_ at finding deep tactics than a dumb search, because you more than gain back the time "lost" (due to smart evaluation) with intelligent move ordering and pruning of the tree.
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.