One of our future aims is to extend the system so as to be able to improve the integrated theory after new information has become available. The system would then function as a 'closed loop system' which would be capable of using the integrated theory as input for further learning. But which part of the theory should we attempt to modify?
One rather obvious strategy tries to focus the attention on 'weak parts' of the theory so as not waste too much effort with unnecessary modifications. Identifying weak parts of the theory is somewhat similar to debugging, only it is more complex. In debugging one tries to identify a faulty step, by comparing, for example, the user's answer with the system's answer. Here one needs to consider how a particular step contributes to overall system<<s performance. If some particular step degrades the system's performance, it is a good candidate for further changes.
Murray and Porter [1989] have developed a system PROTOKI (a prototype of a larger system KI). The system is intended to provide support when new piece of information is integrated in the existing knowledge base. The process of integration involves three main steps: recognition, elaboration and adaptation. The third step is concerned with resolution of anomalies. These arise, for example, when conflicting solutions can be reached using different chains of reasoning. The prototype system inspects the explanation that lead to the anomaly and tries to determine the weakest premise. This premise is then modified.
On one hand Murray and Porter<<s work goes further than ours. So far we have not tried to integrated knowledge consisting of rules that are chained. On the other hand PROTOKI requires a great deal of domain dependent knowledge, and besides, is not capable of resolving certain language differences which we shall discuss below.