22 ms·

Why chess programs find good moves, but barely understand chess

21 points by soundsop 18y ago

qqq 18y ago
Chess programs that only look 1 move ahead and make all their moves in under a second are better than most human players. Their evaluation function to decide what is a good position is actually quite smart and has a lot of knowledge of chess. 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...
- dfox 18y ago
  I think that the main problem here lies in meaning of "understand chess". 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.
  - qqq 18y ago
    The eval function contains more accurate knowledge of what matters in chess than most humans.
brentr 18y ago
I am not competent enough to comment on the design of computer chess programs, however, as a chess player, I can state that computers perform well in very tactical positions. When the positions are more strategic, such as a closed position arising from a large number of queen pawn openings (1.d4) and such openings like the Ruy Lopez Closed, the human's ability to take into account very subtle strategic ideas usually give the human the upper hand. That is why Kasparov tried to keep most of his games against Deep Blue closed. 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.
smanek 18y ago
This reminds me of Searle's Chinese Room paradox. 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.
rw 18y ago
Too speciesist.
dominik 18y ago
We don't understand how humans "understand" chess, either. 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.
- brentr 18y ago
  Much of chess is based on patterns. Grandmasters spend hours a day going over line after line in various openings. The real game doesn't usually start till around move 15. Every once in a while, an opening novelty is found before move 10 that dramatically changes opening theory. When the novelties finally are published (i.e., it was finally played in a game), it just becomes part of the theory that anyone who plays that line at the top levels must know. They will take that line and run it through Fritz, Shredder, Junior, and any other chess program they might have and work our preliminary strategies. Any promising lines will then be assessed by a team of weaker GM's helping a much stronger GM train for some important match.

21 points by soundsop 18y ago

qqq 18y ago

Chess programs that only look 1 move ahead and make all their moves in under a second are better than most human players. Their evaluation function to decide what is a good position is actually quite smart and has a lot of knowledge of chess. 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...

dfox 18y ago
I think that the main problem here lies in meaning of "understand chess". 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.
- qqq 18y ago
  The eval function contains more accurate knowledge of what matters in chess than most humans.

brentr 18y ago

I am not competent enough to comment on the design of computer chess programs, however, as a chess player, I can state that computers perform well in very tactical positions. When the positions are more strategic, such as a closed position arising from a large number of queen pawn openings (1.d4) and such openings like the Ruy Lopez Closed, the human's ability to take into account very subtle strategic ideas usually give the human the upper hand. That is why Kasparov tried to keep most of his games against Deep Blue closed. 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.

smanek 18y ago

This reminds me of Searle's Chinese Room paradox. 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.

rw 18y ago

Too speciesist.

dominik 18y ago

We don't understand how humans "understand" chess, either. 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.

brentr 18y ago
Much of chess is based on patterns. Grandmasters spend hours a day going over line after line in various openings. The real game doesn't usually start till around move 15. Every once in a while, an opening novelty is found before move 10 that dramatically changes opening theory. When the novelties finally are published (i.e., it was finally played in a game), it just becomes part of the theory that anyone who plays that line at the top levels must know. They will take that line and run it through Fritz, Shredder, Junior, and any other chess program they might have and work our preliminary strategies. Any promising lines will then be assessed by a team of weaker GM's helping a much stronger GM train for some important match.