Reference: Hewett, R. Model Aggregation on Model-Based Approach using Prime Models - A Preliminary Report. Knowledge Systems Laboratory, February, 1991.
Abstract: Models are central in reasoning about physical systesm. While many studies of model based reasoning in Artificial Intelligence have focused on fundamental reasoning techniques, only recently have attempts been made to study the principles of model construction. The tradeoff between quality and efficiency in reasoning produces one important issue in model development - how to build models that contain enough information (good approximation) for the reasoning and yet do not contain so much detail (good abstraction) that searching would become intractable. The problem becomes even more pronounced when dealing with large and complex systems. One common practice is to decompose systems into subdivisions where problem-solving can be broken down into smaller problems. However, decomposability alone is not enough. The organization of models must provide integration of various reasoning results to allow coherent reasoning about the overall system. Another approach to help us manage complex systems with a high computational cost is by reasoning with multiple appropriate models. This allows us to reason at multiple levels of abstraction. For example, different models of a device correspond to different abstractions of it. Each model is well suited to a particular class of uses - both in terms of the answers it can provide and in terms of efficiency of the inference. However, as the systems get larger and more complicated, complete instantiation becomes both undesirable and impossible. To handle this problem one needs to know what is relevant and what models to select. All of these topics are currently under research in model-based reasoning aobut large and complex systems.