Integrating rules and connectionism for robust reasoning: A connectionist architecture with dual representation.

摘要

One of the more difficult problems for artificial intelligence is the problem of modeling commonsense reasoning and alleviating the brittleness of traditional symbolic rule-based models. This work attempts to tackle this problem by combining rules with connectionism in an integrated framework. This idea leads to the development of a connectionist architecture with dual representation that combines symbolic and subsymbolic (feature-based) processing for evidential robust reasoning: CONSYDERR.;An important aspect of this research is that the architecture utilizes the synergy resulting from the interaction of the two different types of representation and processing, and is thus capable of handling a large number of difficult issues in one integrated framework, such as partial and inexact information, cumulative evidential combination, lack of exact match, similarity matching, inheritance, and representational interactions, all of which are important elements of commonsense reasoning. The results suggest that connectionism coupled with rule-based symbolic processing capabilities can be effective and efficient models of reasoning for both theoretical and practical purposes.;Collins' protocols are analyzed based on the notions of rules and similarity, and are modeled by the architecture which carries out rule application and similarity matching through the interaction of the two levels. In order to understand rule encoding in the architecture, a formal analysis of the model is performed, which shows that it handles a superset of (propositional) Horn clause logic and Shoham's logic. To further improve the rule-based reasoning capability of the architecture, a solution to the connectionist variable binding problem is proposed. This work also explores the notion of causality and shows that commonsense causal knowledge can be well represented by CONSYDERR. Several other aspects of the architecture are discussed to demonstrate how connectionist models can supplement, enhance, and integrate symbolic rule-based reasoning.