Theories of Artificial Intelligence
Meta-theoretical considerations
TheoriesOfAI - Pei Wang.pdf by Pei Wang, and comments by Joscha Bach
- The "theory of AI" is impossible by definition, since we cannot have a theory for "those processes we don't yet understand" — as soon as we have a good theory for such a process, it is no longer considered as AI anymore [Minsky (1985); Schank (1991)] [Page 2];
- Lack of Theory & Theoretical Nihilism: Fields of science or engineering are commonly guided by corresponding theories but AI seems to be an exception. A major obstacle is "theoretical nihilism" — facing the well-known difficulty in establishing a theory of AI, the research community as a whole has not made enough effort in this task.
- Look, ma, no hands syndrome: The "problem-oriented" attitude toward AI focuses on the problem-solving capability of a computer system, while does not care much for the underlying theory.
- If we have a theory, it is not considered AI anymore: The "theory of AI" is impossible by definition, since we cannot have a theory for "those processes we don’t yet understand" — as soon as we have a good theory for such a process, it is no longer considered as AI anymore.
- Theory of AI need to be both descriptive and normative: The theory of AI must cover certain mechanisms in the human mind (Human Intelligence; Descriptive), then generalize and abstract them (to be the General Intelligence; Descriptive and Normative both), and finally specify them in a computational form (to become the Computer Intelligence; Normative).
- Scientific Theory must have following properties:
- Correctness: A theory should be supported by available evidence.
- Concreteness: A theory should be instructive in problem solving.
- Compactness: A theory should be simple.