Progressive rules: a method for representing and using real-time knowledge

This paper introduces progressive rules, a new approach to knowledge representation for time-limited tasks. Progressive rules are first-order Horn clauses augmented with a coefficient of relevance for each antecedent, and a quantum value used to control propagation of confidence. They are a generalization of Michalski and Winston´s (1986) variable precision logic, and present several improvements relative to it. Using progressive rules, inference is controlled by recursively spreading activation from a goal to its subgoals, according to their relevance, and attempting first to satisfy the goals that have accumulated the most activation. When answering a question, this allows provisional answers to be supplied with growing confidence before inference is complete, guaranteeing an optimal use of time given the information available. Theoretical and empirical studies of the new approach´s performance show promising results