Author: Mike S.
Date: 20:43:22 12/10/02
Go up one level in this thread
On December 10, 2002 at 20:04:10, Bob Durrett wrote: >On December 10, 2002 at 15:47:19, Mike S. wrote: >(...) If there is a threshold value (like in Fritz) to be set, which >>controls how big the difference of the evaluations has to be to produce a >>commentary variant, then I'm afraid it would have to be set to a small value for >>positional things. I would expect that positional observations are - most often >>- about much smaller eval differences that i.e. a blunder which looses a pawn or >>the like. > >Wouldn't that depend on how the chess engine performed its evaluations? If the >chess engine only does tactical evaluations, then that would always be true. >But cannot someone create position evaluation software that examines positional >as well as tactical features of the position? I actually believe that modern >chess engines do that already. Sure they do (and always did somehow, ever since). >Anyway, it is interesting that you say that. I believe that was the essence of >what Rolf was saying too, before he became frustrated and ran away. As I >understood it, he saw "positional" positions as being those that a chess engine >would "think" to be equal and quiet. Any tactics would be beyond the chess >engine's horizon, for any reasonable analysis time. > >The idea that "positional" positions could not yield high position evaluation >scores is something that should be examined closely. It is not clear to me that >it is impossible or highly improbable that chess engines would assign >significant values to some "positional" positions. [I guess Uri would say I >should give a position to illustrate this point, but I cannot think of one.] I didn't mean that it is impossible. That evaluations of positional and "dynamical" factors (i.e. king safety) can sum up to +2 or even more, is a common thing. Years ago, Mchess was one of the first programs which showed such big evals when it was obvious that they were not based on material advantages in it's variations (yet :o). But that are very imbalanced positions already (exceptions). What I meant, is an analysis threshold for the automatical analysis of complete games, which may contain positional decisions a/b/c, where (b) might only be evaluated 0.25 less than (a), etc. - But from a human's viewpoint, it can be worth more or even decisive for the result of the game, which mean you'd want the program comment on that. If (b) was played in the game, you get only a comment by the Fritz GUI's Full Analysis for this example, when the threshold is 0.25 or less. >(...) Maybe "positional" positions necessarily contain too much >uncertainty. [?] If significant searching is required to determine the best >move, then my intuition tells me we are back to square one, but that could be >wrong. The advantage of using games to judge about positional abilities of engines is IMO, it is not necessary to search for the best positonal move or to see, if there are several continuations which are nearly equal etc., BUT: You basically have to evaluate the engine's decisions, the one moves it actually played (as a kind of "step 1" analysis). Like in the Bb2xNf6 example: I don't have to know what's the best move in that position, to see that this was a bad one... (in that example, the better alternatives are from opening theory). Also, when I discover an engines has played a tricky continuation for a positional advantage, I can see that no matter if a GM might spot an even better idea or not. "Step 2" would be to search for cases when positional chance were missed, IOW good moves not found... which is much more difficult. I'm not sure if I'd be strong enough for that, except very easy things. In general, I'd prefer a GM to do these judgements instead of myself doing them, and then he should explain everything on two or three pages in small size letters. :o)) But when the GM takes a close look at 100 games of prog A, and at 100 of prog B, and tells me he has seen 30 good positional observations for A and 15 bad, and 24 good and 9 bad for B, hm... I'd still not know which prog is better (30-15 = 24-9). 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.