The development of requirements specifications inevitably involves modification and evolution. To support modification while preserving particular requirements goals and properties, we propose the use of a cycle composed of two phases: analysis and revision. In the analysis phase, a desirable property of the system is checked against a partial specification. Should the property be violated, diagnostic information is provided. In the revision phase, the diagnostic information is used to help modify the specification in such a way that the new specification no longer violates the original property.

. Since we consider theory refinement (TR) as a possible key concept for a methodologically clear view of knowledge-base maintenance, we try to give a structured overview about the actual state-of-the-art in TR. This overview is arranged along the description of TR as a search problem. We explain the basic approach, show the variety of existing systems and try to give some hints about the direction future research should go. 1 Introduction Based upon experiences from previous projects and application studies concerning logic-based knowledge representation [6, 2, 13] and knowledge evolution [15, 11], the aim of the VEGA project (Knowledge Validation and Exploration by Global Analysis) is to contribute to a methodology of maintaining declarative knowledge [1, 12, 14]. This goal demands techniques and methods for all maintenance activities (e.g. debugging, adaptation, change) within the life-cycle of knowledge bases (KBs)---relying on elementary building blocks (exploration and validatio...

This paper describes and evaluates the types of machine learning techniques which are appropriate for the special needs of automatic target recognition. Artificial neural networks has been the most popular choice. However, they have several shortcomings, including that they do not readily allow the learned domain theory to be revised in response to new conditions, new targets, or new sensors. Recent advances in machine learning for other applications suggest that explanation-based learning and theory revision are promising candidates to overcome these deficiencies. Feature Extraction Classification Domain Theory identified target sensor observations feature vector models of targets Figure 1: Components of a typical ATR system. 1 Introduction Automatic target recognition (ATR) is a process by which a computer system extracts features from one or more sensors and then identifies the target. Human operators cannot reliably identify targets on a continuous basis due to the need for rap...

This paper describes a machine learning technique, fuzzy theory revision (FTR), that improves the performance of classifiers in the presence of background effects such as weather patterns, solar loading, etc. The technique allows existing knowledge bases to be adapted for use in new situations and revised to include new sensors and/or features without retraining over the original training set. Learning is based on a small training set taken from the current application. The resulting predicate logic representation can be understood and conceptually verified by an analyst unlike artificial neural networks. FTR extends traditional theory revision methods with 1) a fuzzy set representation appropriate for feature vector inputs (typical of automatic target recognition and medical diagnosis tasks) and 2) a cost modifier which specifies how radical a change to knowledge base is permitted. Experiments with 2109 simulated missile tracks from IR and visual satellite images resulted in a 23% imp...