As knowledge-based systems are brought to practical applications and knowledge bases are to be used over years, the problem of knowledge base evolution naturally comes up: the key issue is how to ensure that the knowledge base does always represent all the knowledge that is relevant for solving tasks, i.e. being 'complete', and does not become out of date or invalid, i.e. remaining 'sound' with respect to some specific situational context. Although this is a goal hard to achieve, it shows the direction in which knowledge base evolution research should work: to overcome the (always ``damned but nevertheless done'') accumulation of 'small local hacks' causing unforeseeable consequences and to find a compromise between this ad-hoc KB modification approach and the other extreme of restarting the whole knowledge engineering work ranging from the formal specification down to the concrete representation with each KB modification.
In this paper, we have shown that knowledge base evolution can be regarded as a theory revision process. Research in Inductive Logic Programming provides us with a set of techniques that can be applied to incorporate new knowledge into the knowledge base (knowledge base exploration), e.g. by generalization and abduction. On the other hand, techniques from deductive database research can be used for ensuring the integrity of the knowledge base, i.e., for solving the knowledge base verification and validation task.
For both tasks we have developed extensions and modifications motivated by the special characteristics of the application. The generalization techniques taken from ILP have been extended towards the incorporation of meta-knowledge for guiding the generalization process (plgg) and towards additional language features for representing generalization results (e.g., finite domain terms). Additionally, we have proposed an alternative to -subsumption based on an extension of Horn rules incorporating terminological knowledge representation and reasoning (TL-subsumption). In order to get efficient evolution techniques also for large sets of rules and facts we extended the rewriting techniques from deductive databases for abduction and integrity checking.
Further work on knowledge base evolution should not only consider developing more powerful exploration and verification methods, but should also focus on the knowledge representation language itself. It is obvious that a more powerful but still semantically clear representation formalism, as e.g. introduced for TL-subsumption, will be of great advantage for all kinds of knowledge evolution techniques. For example, introducing sorts or types as mentioned in several parts of this paper can be a first but only intermediate step: generalization within a sort lattice does already yield a more fine-grain clause ordering than simple -subsumption. However, extending the logic-based representation language by substituting or complementing constitutively given sorts by intensionally defined concepts and concept terms in the sense of terminological reasoning will be necessary for finding and expressing 'really least general' generalizations and thus being able to support knowledge base evolution over a long period of time.
Currently only little work is available on tailoring the knowledge representation formalism to knowledge base evolution needs [\protect\citeauthoryearBoley1993]. But being convinced that research on this will be a key issue for the success of knowledge base evolution in the future, we will also concentrate on further improving knowledge representation approaches like TL-subsumption besides developing the evolution techniques themselves.