Originally posted by sonhouse
Yes, the rating quantifies wins but not specifically what number of moves you see or positions you evaluate and so forth. Obviously various tradeoffs in tactics V strategy will achieve the same end. Couldn't a computer program work that stuff out? Analyze a set of games of person A Vs person B and see what the lower rated is missing in terms of tactics AND strategy? And presumably what the lower rated needs to do to get to the next level.
the problem with the prevalent engine paradigm is programs can't recognize a 'strategy', they have absolutely no understanding of the position even in the most elementary way. everything they do regarding 'positional chess' is faked, simply by hard coding things into the evaluation function. when you push a castled pawn, the engine is not trying to UNDERSTAND in any depth whether that's actually weakening or not. it'll instantly assign it whatever value the programmer saw fit, completely regardless of the position. they just brute force their way through tactical search trees and wait for the opponent to 'step out of it'. they can only mechanically follow the programers pre-laid instruction, and not an atomic operation more. trying to implement positional chess into and alpha/beta pruning cruncher is like trying to teach physics to a dog. it's not equipped to grasp physics, because it fundamentally lacks the required building blocks. you can teach it to ACT as if it 'understands' in some degree, but it's really just a trick to fool the audience. an illusion of understanding.
now, that could be changed by discarding the whole alpha/beta pruning paradigm, which itself is pretty much a corollary of the structural programming paradigm. take up a whole new way of looking at it, and actually teach
the programs chess. there have actually been some tries with neural networking, but none of them very successfull so far. but that is The way, IF you wan't a program to do things that are characteristic for a human. in effect, you want a classifier instead of a search algorith.
and even if you were successfull at that, you'd still be left outside when you'd try to quantify the results of such a machine. it would understand the concepts, recognize them in totally unseen new positions, but it couldn't give you that specific kind of answers you appear to look for. it would basically say: "depends from the position" or something equally general. just like a human would.