Author: Bob Durrett
Date: 16:33:21 12/06/02
Go up one level in this thread
On December 06, 2002 at 18:37:14, Mike S. wrote: >On December 06, 2002 at 15:34:58, Bob Durrett wrote: > >>On December 06, 2002 at 13:18:31, Mike S. wrote: >>(...) >>>I thought, if the positional ideas are related to a *sacrifice*, these could be >>>especially suitable as computer test positions, because they may have the >>>necessary "testing character" of the solution move. > >>A true positional sacrifice [a sacrifice intended only to obtain positional >>advantage] is a very interesting special case. True sacrifices to obtain >>non-positional advantages, i.e. tactical, might also be considered a special >>case, too. How chess engines handle positions involving either kind of true >>sacrifice is a very interesting question. What is needed is a quantitative >>measurement of the performance of the top chess engines when faced with such >>positions. (...) > >Actually most "modern" test suites should contain such positions, IOW things are >a bit easier than you seem to suppose. "True" sacs are pretty normal nowadays. >There's no other problem than to find good test positions (and that has been >done many time already, so there's no demand for any new especially >sophisticated approaches IMO). Just find good test positons, and that' it... Well, a good computer programmer should be able to write a program to search a large database of high-quality games for the desired positions. At least, I assume that to be true. It should be just a matter of imposing the right search criteria. > >>>I'm afraid I can't explain >>>that properly... but I'm convinced that "normal" moves are not good for test >>>positions, neither for tactics nor for positional tests. >> >>Why not good? The non-sacrificial positions need not be positions having >>several equally good moves. I suspect that positions of this type could be >>found, where one move was much better than the rest. > >The criteria that there's only one best move, is *not sufficient*. It has to be >a move which tells "by itself" that the engine has spotted something special. A >sacrifice (simply a blunder if there wouldn't be the special pointe) can provide >that. That's what I mean by "testing character". > >With "normal" moves, which don't sac anything or don't refute material offered, >it's very difficult to tell if the engine has chosen this move for the intended >reason or by luck. This is the point. An engine won't sac a queen without very >good reasons, but a normal silent move like can remain doubtfull (in terms of >testing clarity). I believe I have found a solution to that problem: If the engine is selecting "non-tactical" moves which it doesn't understand, then it should select the right move only half the time, more or less, assuming there are two seemingly equal moves. If there are several good moves, the engine might stumble on the right move only 1/3 the time, or 1/4 the time, etc. So . . . Simply use a large test sample. If an engine under test finds the correct move most of the time, then it can be said with a high degree of certainty that the given engine is capable of solving such "positional" positions. Incidentally, this suggests that good test positions should have two or more equal moves based on tactical evaluation. [Ideally, the correct move should be a tactically "quiet" move, although I'm not sure about that.] Only the engine capable of distinguishing between moves on the basis of positional merit would consistently pick out the correct positional move. That would be a very nice engine to own. Bob D. > >>Note that I am thinking that design of engines to do well with positional >>positions would *primarily* be a problem for design of the "position evaluation >>software," as opposed to the "searching software." That may not be a valid >>assumption. I don't know. > >Positional evaluation is part of computer chess since Shannon's papers (= ever >since); there is no basic problem or demand for anything new about that. Every >engine does it somehow (of course the quality and number of factors evaulated >differs). > >The following position is one of the best examples I know (although I'm not 100% >sure if Black can't *delay* the idea; in-between-moves can be a big problem for >positional tests): > >[D]1r1r2k1/2b1qp1p/b1p3p1/p1p1p3/2P1P3/1PN1BP2/P1Q3PP/R2R2K1 b - - 0 1 >From Liublinsky - Botvinnik, Moskva 1943 > >Some engines will find 1...Rd4!, transforming the doubled c-pawn into an >protected passer (if White accepts; it's typical that engines which are planning >Rd4, don't expect White to take the rook). > >Regards, >M.Scheidl
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.