Neural-Logic Belief Networks

(NLBN)

Boon Toh LOW Norman Y. FOO

Knowledge Systems Group
Basser Department of Computer Science
University of Sydney
NSW 2006, Australia

email : [email protected] [email protected]

Abstract

This technical report presents a hybrid symbolic-neural network for representing and reasoning about commonsense beliefs called a Neural-Logic Belief Network (NLBN). Concepts are represented by nodes and relations amongst the concepts are represented by numerical directed links from one node to another. The network computation functions are similar to those of a neural network and classical logical relations such as AND, OR, etc., as well as human biased relations can be modelled by the network. The belief network is four-valued, that is a proposition is either being believed, its negation is believed, unknown or contradictory. Unlike the material implication in logic systems, defeasible IF-THEN rules in this formalism are unidirectional and well behaved; they only express the intended declarative semantics. A set of network operators similar to that of the belief revision operators have been defined and they allow knowledge to be dynamically updated in the knowledge system. This formalism has been implemented using Prolog on a unix machine. Its hybrid network architecture makes inferencing deterministic and efficient.