Mind

I have a question about "solved" games, and the significance of games to artificial intelligence. I take it games provide one way to assess artificial intelligence: if a computer is able to win at a certain game, such as chess, this provides evidence that the computer is intelligent. Suppose that in the future scientists manage to solve chess, and write an algorithm to play chess according to this solution. By hypothesis, then, a computer running this algorithm wins every game whenever possible. Would we conclude on this basis that the computer is intelligent? I have an intuition that intelligence cannot be reduced to any such algorithm, however complex. But that seems quite strange in a way, because it suggests that imperfect play might somehow demonstrate greater intelligence or creativity than perfect play. [If the notion of "solving" chess is problematic, another approach is to consider a computer which plays by exhaustively computing every possible sequence of moves. This is unfeasible with current technology, of course. But we can imagine a futuristic computer which has such great raw power that it's able to play effectively using this otherwise exceedingly "dumb" strategy.]

Comments

This is a very good question. It is reminiscent of the debate over the so-called "Turing Test", in particular, of an objection to the Turing Test made by Ned Block: his "Blockhead". See the SEP article on the Turing Test for more on this.

In the case of chess, it is generally believed that chess is solvable in principle. There are only finitely many possible moves at any stage, etc. So, in principle, a computer could check through all the possibilities and determine the optimum move at each stage. Practically, this is impossible at present, as there are too many moves. But if chess had been solved, and if a computer were simply programmed to make the best move at each stage, then it seems quite clear that no "intelligence" would be involved.

Of course, this does not by itself show that "intelligence cannot be reduced to any...algorithm", and the question whether it could be is hotly disputed. There are some famous (or infamous) arguments due to Lucas and Penrose that attempt to establish that conclusion. Have a look at the SEP article on The Computational Theory of Mind for more on that topic, esp. section 3.3.

The ability to play a game such as chess intelligently is certainly one partial measure of "intelligence", but an agent that could only play very good chess would hardly be considered "intelligent" in any general sense.

So I don't think that a winning chess program would be "intelligent".

There is also the question of how the program plays chess. Deep Blue and other computer chess programs usually do well by brute-force search through a game tree, not by simulating/mimicking/using the kinds of "intelligent" pattern-recognition behaviors that professional human chess players use.

But some might argue that that doesn't matter: If the external behavior is judged to be "intelligent", then perhaps it shouldn't matter how it is internally accomplished. For some illuminating remarks on this issue (not in the context of games, however), take a look at:

Dennett, Daniel (2009), "Darwin's 'Strange Inversion of Reasoning' ", Proc. Nat'l. Academy of Science 106, suppl. 1 (16 June): 10061-10065

http://www.pnas.org/content/106/Supplement_1/10061.full.pdf+html

Update: An interesting article about one of my computer science colleagues on the subject of cheating in chess and touching on the nature of "intelligence" in chess just appeared in Chess Life magazine; the link is here

Read another response by Richard Heck, William Rapaport

Read another response about Mind