Why chess programs find good moves, but barely understand chess
zenpawn.com
zenpawn.com
What was interesting is when Kasparov played Fritz and Deep Fritz, he actually chose some very tactical openings and didn't perform too poorly. Then again, it's Kasparov. Kramnik had equal results.
From what I've read, the latest research seems to indicate top chess players train their brain's facial recognition system to recognize tactical board positions. But it's not clear how a grandmaster finds the right move, especially considering simultaneous games where a GM plays on 20 or 30 boards at once against club players, often winning all of the games while only spending a few seconds, if that, on each move.
Imagine you had an English speaking person providing responses to mandarin stimuli in accordance with a function that mapped every input of mandarin words to a mandarin output. Most people would agree that he doesn't really 'understand' mandarin, even though he can 'converse' fluently in it.
Most of the article consists of saying that chess programs have certain weaknesses that humans do not have. That's definitely true, but it doesn't mean that computers "barely understand chess".
The article's author doesn't seem to be a serious chess player. His 19 move "forced line" contains no branches to show what happens with other defenses and, at a glance, looks like black used some moves inefficiently and it'd be hard to be confident there was nothing better without seeing a lot more analysis. I'd only trust it either with extensive branches and notes, or if he checked it with a strong computer...
You can say that computers understand chess because they are able to consistently beat best human players. Or you could say that they understand it because they are able to play it at all.
Or you can say that they do not understand chess, because they use some combination of hard coded heuristics (and the evaluation function is nothing but that) and brute force, which is at first glance too trivial approach to be called understanding.