Monitoring and diagnosis: stress from weakened environmental knowledge

A discussion is presented of the requirements for continuing monitoring and diagnostics in situations where knowledge of the world is unexpectedly weakened. There are advantages of conceptual and practical significance in using model-based generate-and-test approaches for handling failing sensors and/or missing data, as opposed to conventional (associational) representations. The representational complexity, expressed as numbers of rules, for an associational inference ranged from linear to exponential complexity, while a model-based approach required virtually no additional content or restructuring of its domain knowledge. Weakened data may prevent associational methods from asserting possible faults, and they may cause model-based methods to retain faults when data that might contradict them are not available. The model-based methods are found to make virtually optimum use of the remaining information. Advantages for sensor fusion and the construction of explanations are suggested