Possible Applications to Artificial Intelligence

The foregoing discussion of concepts has been mainly concerned with how to translate into a suitable formal language certain sentences of ordinary language. The success of the formalization is measured by the extent to which the logical consequences of these sentences in the formal system agree with our intuitions of what these consequences should be. Another goal of the formalization is to develop an idea of what concepts really are, but the possible formalizations have not been explored enough to draw even tentative conclusions about that.

For artificial intelligence, the study of concepts has yet a different motivation. Our success in making computer programs with general intelligence has been extremely limited, and one source of the limitation is our inability to formalize what the world is like in general. We can try to separate the problem of describing the general aspects of the world from the problem of using such a description and the facts of a situation to discover a strategy for achieving a goal. This is called separating the epistemological and the heuristic parts of the artificial intelligence problem and is discussed in (McCarthy and Hayes 1969).

We see the following potential uses for facts about knowledge:

1. A computer program that wants to telephone someone must reason about who knows the number. More generally, it must reason about what actions will obtain needed knowledge. Knowledge in books and computer files must be treated in a parallel way to knowledge held by persons.
2. A program must often determine that it does not know something or that someone else doesn't. This has been neglected in the usual formalizations of knowledge, and methods of proving possibility have been neglected in modal logic. Christopher Goad (to be published) has shown how to prove ignorance by proving the existence of possible worlds in which the sentence to be proved unknown is false. Presumably proving one's own ignorance is a stimulus to looking outside for the information. In competitive situations, it may be important to show that a certain course of action will leave competitors ignorant.
3. Prediction of the behavior of others depends on determining what they believe and what they want.

It seems to me that AI applications will especially benefit from first order formalisms of the kind described above. First, many of the present problem solvers are based on first order logic. Morgan (1976) in discussing theorem proving in modal logic also translates modal logic into first order logic. Second, our formalisms leaves the syntax and semantics of statements not involving concepts entirely unchanged, so that if knowledge or wanting is only a small part of a problem, its presence doesn't affect the formalization of the other parts.