An Algorithmic Approach to Recover Inconsistent Knowledge-Bases

We consider an algorithmic approach for revising inconsistent data and restoring its consistency. This approach detects the \spoiled" part of the data (i.e., the set of assertions that cause inconsistency), deletes it from the knowledge-base, and then draws classical conclusions from the \recovered" information. The essence of this approach is its coherence with the original (possibly inconsistent) data: On one hand it is possible to draw classical conclusions from any data that is not related to the contradictory information, while on the other hand, the only inferences allowed by this approach are those that do not contradict any former conclusion. This method may therefore be used by systems that restore consistent information and are obliged to their resource of information. Common examples of this case are diagnostic procedures that analyse faulty components of malfunction devices, and database management systems that amalgamate distributed knowledge-bases. 1 Motivation In this paper we introduce an algorithmic approach to revise inconsistent information and restore its consistency. This approach (sometimes called \coherent" 5], or \conservative" 15]) considers contradictory data as useless, and uses only a consistent part of the original information for making inferences. To see the rationality behind this approach consider, for instance, the following set of pro-positional assertions: Since :p is true in KB, so is :p_q (even if q is false), and so a plausible inference mechanism should not apply here the Disjunctive Syllogism to p and :p_q. Intuitively, this is so since the information regarding p is contradictory, and so one should not rely on it for drawing inferences. On the other hand, applying the Disjunctive Syllogism to fr; :r_sg may be justiied by the fact that this subset of formulae should not be aaected by the inconsistency in KB, therefore inference rules that are classically valid can be applied to it. The two major goals of coherent approaches in general, and our formalism in particular, are therefore the following: