Formal self-reference in artificial intelligence

Basic research project, 1996-1998, extended to 2001, Ministry of Science and Technology, Republic Of Slovenia

I think, therefore I ambleProposal summary: The structures of self-reference in reflexive theories will be used in the design of intelligent systems, as models of indexicals, performatives and self-awareness. Stronger formal models of reflection will be investigated, constructed as reflexive progressions of reflexive theories.

Hypothesis

The basic hypothesis of this research is that the formal structures of
self-reference in reflexive theories can be used in computational models of self-referential aspects of intelligence: the ways it refers to itself within language, through indexicals such as "I", "now", performatives such as "I promise", "I apologize") and besides language, as the consciousness which intelligence has of itself. The formal hypothesis of the research is that, besides reflexive theories, it is possible to define and implement, as a reflexive meta-circular interpreter, reflexive progressions of such theories, in which the operator which produces the progression takes into account its own effects. If the formal reflection which is manifested by reflexive theories of Gödel's kind is not sufficient as a model of the reflection of consciousness, a better approximation would be a reflexive progression of reflexive theories. The basis for the supposition that it is possible to define reflexive Gödel (or other) operators, which would also take into account their own effects, is the reflexivity of substituition in the traditional structure of arithmetical self-reference (its action on a representation of itself) and the reflexivity of the usual meta-circular Prolog interpreter, which can also use its own clauses (Bratko, Prolog programming for AI, p. 536), although only to produce statements of iterated provability. Thus, a reflexive Gödel operator would operate both on a (reflexive) theory and a representation of itself.

Proposed timetable and results

1996: improving the preliminary results of using formal self-referential structures in modelling the semantics of indexicals and performatives; analysis of the uses of Gödel's theorems in the theory of artificial intelligence

The main result was that algorithmic insight into one's own workings cannot be deepened by anticipating and incalculating its effects. That is, although there are Gödel operators of any degree of reflexivity, they are only "extensional", internally performing the steps of reflexion they appear to skip; there are no "intensional" operators which would only represent (as opposed to perform) the action of a Gödel operator and then improve upon it.
An interpretation of
self-reference as self-recognition was also constructed.
1997: analysis of the Gödel sentence as a logic meta-program and definition of a reflexive Gödel operator and the corresponding progressions of (reflexive) theories

Not yet completed; work continues on assesing the usefulness of Kleene's recursion theorems to these ends. As a side-effort, it was proved that Gödel's mapping also holds in non-standard models. On the analytical side, self-referential aspects of experience were used to analyse and reject a dualist structure of experience.
Team members were also engaged in (implementation) work on agents and in the publication of the book Mind versus Computer, IOS Press.
1998: definition and implementation of a reflexive meta-circular interpreter and tests of its usefulness in tasks requiring reflection (learning)

Not completed, in view of the preliminary negative results in 1996. Work on the implications of Kleene's theorem for computational reflection continued, partly relocated to the project Reflection in machine learning.
Work on modelling reflection in consciousness also continues.

Project team

Background

The background of the project is the growing role of self-referential structures in the development of (meta-level) computer architectures (Yonezawa & Smith 92), especially in logic programming (meta-circular interpretation) (Abramson & Rogers 89) and artificial intelligence (reflective systems, meta-knowledge, learning) (Maes & Nardi 88). In computer science in general, these structures appear in connection with the specifically computational problem of regulating the flow of control information (Smith 82, Yonezawa & Smith 92), and in artificial intelligence in the design of introspective systems, which can reason about themselves and their behavior (Maes 88). On the other hand, in artificial intelligence and cognitive science, discussions reappear of the relevance of Gödel's and others' limitative theorems for the
possibility of artificial intelligence, especially the possibility of developing formal models of consciousness (Sloman 92, Penrose 94). The structures of formal self-reference, familiar from Gödel's theorems (Smullyan 92) are also the more specific, unified logical background of this research, as opposed to the different forms of self-reference which appear in the problems of control regulation in various computational architectures (Silver 86). These formal structures (Gödel, Smullyan) can be used as models of various self-referential aspects of intelligence, within language (Lucy 93) eg. talk of oneself by using indexicals such as "I", "now", performatives such as "I promise", "I apologize", or outside of language (self-awareness, consciousness of oneself) (Baars 88).

The effects of self-reference in formal theories - the construction of sentences which "talk" about themselves and the theory to which they belong - are not limited to those theories themselves. If the effect of such a sentence, eg. Gödel's sentence, is incompleteness of the theory, this opens up the possibility of its extension, in different directions. The principles of extending reflexive theories, formulated so far (Gödel, Turing, Feferman) have been limited to incremental, linear advance along the progression of (transfinite) ordinals (Giunchiglia & Smaill 89). Such advance is, however, non-reflexive: the usual extension operator does not take into account its own role in the process of extension: it only repeatedly reproduces the basis for its application. A reflexive extension operator would not do away with the incompleteness of a reflexive theory, but it could extend it in longer leaps along the progression of ordinals. Such reflexive progressions of reflexive theories could be a better model of the kind of reflection which is peculiar to consciousness and which is usually considered to surpass the reflexivity of reflexive formal theories.

References

Abramson, H, Rogers, M.H. (ed), Meta-Programming in logic programming, MIT Press 1989

Baars, B.J., A cognitive theory of consciousness, Cambridge Univ. press 1988

Baars, B., Banks, W., Consciousness and cognition, an international journal, Academic press

Bartlett, S., Suber, P. (ed.), Self-reference - reflections on reflexivity, Martin Nijhoff 1987; includes A Bibliography of Works on Reflexivity

Bartlett, S. J. (ed.), Reflexivity: A Source-Book in Self-Reference, North-Holland 1992

Bojadziev, D., papers on self-reference.

Bratko, I., Prolog programming for AI, Addison-Wesley 1990

Case, J., Infinitary self-reference in learning theory, J. Exprt. Theor. AI 6, 1994

Feferman, S., Transfinite recursive progressions of axiomatic theories, J. Symbolic logic 27, 1962

George, F., Models of thinking, Allen & Unwin 1970

Giunchiglia, F., Smaill, A., Reflection in constructive and non-constructive automated reasoning, in H. Abramson, M.H. Rogers (ed), Meta-Programming in logic programming, MIT Press 1989

Kiczales, G., Rivieres, J., Bobrow, D., The art of meta-object protocol, MIT press 1992

Lucy, J.A. (ed.), Reflexive language, Cambridge Univ. press 1993

Maes, P., Nardi, D. (ed.), Meta-level Architectures and Reflection, North-Holland 1988

Penrose, R., The Emperor's New Mind, Oxford Univ. Press 1989

----, Shadows of the Mind, Oxford Univ. Press 1994

Pizzi, R., Self-observation in artificial reality, Note di software 48/49

Sanchis, L., Reflexive structures, Springer 1988, and Recursive functionals, Springer 1992

Silver, B., Meta-level inference, North-Holland 1986

Sloman, A.. The emperor's real mind: review of Roger Penrose's The Emperor's New Mind, Artificial Intelligence 56 (1992) 355-396

Smullyan, R.M., Gödel's incompleteness theorems, Oxford Univ. press 1992

Smith, B., Reflection and semantics in a procedural language, MIT TR-272, 1982

----, Varieties of Self-reference, in J.Y. Halpern (ed.), Theoretical Aspects of Reasoning about Knowledge, Proceedings of the 1986 Conference, Morgan Kaufmann 1986

Smorynsky, C., Self-reference in modal logic, Springer 1985

Stefik, M., Smoliar, S., Four reviews of "The Society of Mind", Artificial intelligence 48/3, 1991

Turing, A., Systems of logic based on ordinals, Proc. London math. soc. 45, 1939

Webb, J. (1980), Mechanism, Mentalism and Metamathematics - An Essay on Finitism, D. Reidel Publ. Co.

Yonezawa, A., Smith, B. (ed)., Reflection and Meta-level Architectures, Proc. Int. Worskhop on New Models for Software Architecture, Tokyo 1992