For Problems Sufficiently Hard ... AI Needs CogSci

Is cognitive science relevant to AI problems? Yes — but only when these problems are sufficiently hard. When they qualify as such, the best move for the clever AI researcher is to turn not to yet another faster machine bestowed by Moore’s Law, and not to some souped-up version of an instrument in the AI toolbox, but rather to the human mind’s approach to the problem in question. Despite Turing’s (Turing 1950) prediction, made over half a century ago, that by now his test (the Turing test, of course) would be passed by machines, the best conversational computer can’t out-debate a sharp toddler. The mind is still the most powerful thinking thing in the known universe; a brute fact, this. But what’s a “sufficiently hard” problem? Well, one possibility is that it’s a problem in a set of those which are Turing-solvable, but which takes a lot of time to solve. As you know, this set can be analyzed into various subsets; complexity theorists do that for us. Some of these subsets contain only problems that most would say are very hard. For example, most would say that an NP-complete problem is very hard. But is it sufficiently hard, in our sense? No. Let P be such a problem, a decision problem for F associated with some finite alphabet A, say. We have an algorithm A that solves P . And odds are, A doesn’t correspond to human cognition. The best way to proceed in an attempt to get a computer to decide particular members of A is to rely on computational horsepower, and some form of pruning to allow decisions to be returned in relatively short order. What we have just described structurally, maps with surprising accuracy onto what was done in AI specifically for the problem of chess. In his famous “20 Questions” paper, written in the very early days of AI and CogSci (and arguably at the very dawn of a sub-field very relevant, for reasons touched upon later, to issues dealt with herein: computational cognitive modeling and cognitive architectures), Newell (Newell 1973) suggested that perhaps the nature of human cognition could be revealed by building a machine able to play good chess. But Deep Blue was assuredly not what Newell had in mind. Deep Blue was an experiment in harnessing horsepower to muscle through a Turing-