Next: Selected Methods for Up: Knowledge Base Evolution Previous: The Knowledge Base

Theory Revision

The problem of building up a knowledge base (knowledge acquisition) can be seen as a two-phase process (cf. [\protect\citeauthoryearGinsberg et al.1988]): In the first phase the knowledge engineer builds an initial model (i.e. the seeding of the knowledge base). In the second phase this initial knowledge base is refined or revised into a high performance knowledge base. During the further practical use of the knowledge-base, the dynamically changing world may cause the knowledge base to become invalid in one of the following senses:

In the first situation we have a new application case (i.e. a positive example) that is not yet derivable from the knowledge base. In the second situation, we can derive a specific solution from the knowledge base which is no longer admissable (e.g. because of new environmental protection laws). This is consequently called a negative example.

From a more formal point of view, this means that a given knowledge-base KB has to be revised using positive examples and/or negative examples such that all the positive examples but none of the negative examples are covered by the resulting knowledge base KB'.

Taking the knowledge base as a Horn theory consisting of facts and rules and satisfying a set of integrity constraints IC, the exploration task of theory revision is to change into such that

and

The resulting theory must, of course, still satisfy the given integrity constraints, i.e. IC T' must be consistent. This integrity checking represents the verification task of theory revision and thus again demonstrates the interleaved exploration and verification principle.

The main task, however, remains how to obtain the revised theory . In principle, there are two possibilities:

In the following section we will discuss some selected techniques from the fields of inductive logic programming and deductive databases, which could be applied within the proposed theory revision framework.



Next: Selected Methods for Up: Knowledge Base Evolution Previous: The Knowledge Base


Harold Boley & Stefani Possner (possner@dfki.uni-kl.de)