A logic of categorization

Abstract

The reasoning system known as NARS constitutes a model of categorization. NARS is designed to be an adaptive system that works under the constraint of insufficient knowledge and resources. It consists of a categorical language, an experience-grounded semantics, a set of syllogistic inference rules, a dynamic memory structure, and a control mechanism that manages asynchronized parallel inference. In the system, reasoning and categorization are two aspects of the same underlying process. As a model of categorization, NARS unifies several existing theories.

Keywords:

• Categorization
• Compositionality
• Experience
• Learning
• Reasoning

Acknowledgment

We thank Helga Keller for her inspirational administrative support.

Notes

†For more detailed descriptions and discussion of the other aspects of NARS, such as its working definition of intelligence, knowledge representation language, semantics, uncertainty management, inference rules, and control mechanism, see publications and demonstrations at http://www.cogsci.indiana.edu/farg/peiwang/papers.html. A comprehensive description of NARS will appear in Wang (2005b).

†For a more detailed discussion of the notion of truth-value in NARS, and its relationship with probability theory, see Wang (2001).

†A full description of Narsese, as well as comparisons between it and other knowledge representation approaches (such as semantic network, description logic, and so on), are in Wang (2006).

†In the future, Narsese will be extended to include goals, which are statements that will be turned into true (if they are not already true) through the execution of relevant operations.

‡In NARS, infallible inference can be carried out within an axiomatic sub-system. This issue is discussed by Wang (2006).

†A bag is implemented by a probabilistic priority queue with an associated hash table.

‡Please note that with syllogistic rules, every inference step takes place entirely within a concept.

†For a higher-order term (a statement), its meaning also includes statements that imply it and statements that it implies (Wang, 2006).

†At the current stage, the experience of NARS consists exclusively in communication with other cognitive systems, though this will not be the case when sensorimotor capacities are added to the system.

†For comparisons between NARS and other theories of intelligence, see (Wang 1995, 2004, 2006).

‡For a similar discussion about the unification of reasoning and learning, see Wang (2000).

§for a detailed description of the experience-grounded semantics, and how it is different from model-theoretic semantics, see Wang (2005).

†In this paper, we only compare NARS with other categorization models. See the other publications mentioned previously for comparisons between NARS and other approaches on definition of intelligence, knowledge representation, semantics, inference rules, uncertainty management, inference rules, learning processes, control mechanisms, and so on.

†Once again, given the complexity of NARS, in this paper we have not addressed many aspects of the system, but limit the discussion to categorization. For the same reason, we cannot discuss more complicated examples than the ones used above, because that would inevitably require the components of NARS that are beyond the scope of this paper.