AKILES : An Approach to Automatic Knowledge Integration in Learning Expert Systems

Abstract

Knowledge integration is defined here as a machine learning task from a practical point of view—by identifying the requirements that a real-world complex application domain poses on the expert system in relation to a changing world. We present our current approach to knowledge integration in an expert system, required when the structure of the physical system, the world on which the expert system operates changes. Our exemplar domain task is technical diagnosis. We test our approach on the particular architecture of MOLTKE/3, our workbench for technical diagnosis1- which integrates second-generation expert system techniques in a unique framework. Knowledge integration is seen as the task of elaborating and accomodating new information (due to structural changes) in the expert system's knowledge, maintaining consistency in the knowledge base. The main focus is towards improving the adaptability of the expert system to the structural changes. The approach is based on three principles from the adaptation process: incrementality, extensive and intensive use of domain knowledge, and focus on strategic knowledge. We discuss how AKILES’ knowledge integration task can be used to complete the modeling cycle, i.e., covering the model-evaluation step in the layout-elaboration-evaluation cycle, as defined in [13].