Model Revision: A Theory of Incremental Model Learning

Abstract

We describe an adaptive method for qualitative modeling of physical devices in which the behaviors of a device are acquired by locally revising the structure-behavior model of a structurally similar device. The structure-behavior model of the known device explicitly represents its structure, its output behaviors, and its internal causal behaviors that specify how its structure produces its output behaviors. The model is revised by model-revision plans, where each plan accommodates a specific type of structural difference between the new and the known devices. The process of model revision is focused by the knowledge of the internal causal behaviors of the known device. The revised structure-behavior model for the new device is stored in a model memory for potential reuse in future. We call this incremental model learning.