Author: Fernando Villegas
Date: 10:48:46 11/19/99
Go up one level in this thread
On November 18, 1999 at 21:10:20, Dann Corbit wrote: >On November 18, 1999 at 20:43:20, Fernando Villegas wrote: > >>Hi dan: >>This issue of plys, querality of the game, etc, is tricky. Seems to me -but I am >>not programmer, so this is more a question than an aseveration- that deeper >>search it cannot never be disociated of greater knowledge or whatever the name >>you give to the criteria with which the program prunnes. More plys means more >>moves to analyze and so more prunning to get a choice, but then more prunning >>means -¿?- more parameters to do it decently. I suppose the package of ideas >>with which you get a fair selection in the area of, say, 4 to 7 plys, cannot be >>the same for the area of 8 to 12 and so on. You need -¿?- more refined criteria. >Actually, I pretty much agree with everything you say here. There are two kinds >of plies. One is an exhaustive search of a ply. Very few programs really do >this since most of them use NULL move pruning, which means that *all* plies are >really selective. Then there are extensions. For instance, if there are >repeated captures or checks or something interesting that makes the program peek >forward, it may be possible to see very deeply in a very short time. This kind >of ply is very selective. Selective plies means that we incorporate knowlege. >The smart searchers are (undoubtably) incorporating great amounts of knowlege in >order to figure out where to exert the energy of searching. > >>Other thing is the quality of them. They can be sophisticated but ineficient. >>You can go deep even with the utmost silly ideas, by example, just looking for >>exchanging pieces as in some games played by kids or patzers. So maybe for a >>certain deep you need more knowledge, BUT then it appeasr the problem that there >>is a number of posible different packages of ideas of very different quality. >Indeed, what you are talking about is the type of extension employed to tell the >program how to guess where to search. If two successive moves tell me that >first I lose my queen and then a rook, it seems likely that I will stop >exploring that pathway. Once in a great while, it may cause me to miss a >checkmate but most of the time it will mean that I spend my energy searching >where it is wiser to hunt. The type of speculation used to extend plies will >cause the quality of the extensions to be preferred or to suffer. > >If >>all this is true, then: >>a) An increment of plys with the same knowledge package produces a diminishing >>return or even beyond some threshold can produce an awful result. >I think the next ply always gives you the same bonus, but each bonus becomes >more and more costly to achieve so I believe that I agree with you. > >>b) An increment of knowledge not neccesarily produces better selection. >For each additional ply searched, there is a 17% chance to change your mind >about the move chosen [IIRC] according to the "Goes Deep" studies by Hyatt and >Heinz [again IIRC]. Anyway, if you just chose a move, it may turn out that the >next ply will reveal it is a disaster. So (if I understand your conjecture) I >believe I agree with you again. > >>By the way, I have the impression that Tiger has been developed on the ground of >> the following idea: the progam is divided in different nodules of knowledge >>that are activated according what is happennign in the game. Perhaps this is >>not not really a new idea, but a considerable greater development of it could be >>so... >You probably know better than me, because I suspect your knowledge of chess is >better than mine. > >>Hope not to disturb you with these musings of a non programmer amateur >I think we agreed on everything. A little scary, isn't it? Probably just means >that we're both wrong. >;-) Do you mean that to agree with me means a sure mistake or that to agree with you means the same? :-) F
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.