On learning compressed diagnosis classifiers

We address the problem of embedding a model-based diagnostic system representation within a processor with limited memory (as is typical of most real-world aerospace systems). Given a Boolean diagnostic model f in which we have a probability distribution over fault likelihoods, we describe a method for approximately generating an embedded representation of f by learning a decision tree that encodes only the probabilistically most-likely diagnoses. If the set of possible diagnoses follows a power-law distribution, we show that we can create decision trees that contain the vast majority of the probability mass of the full decision tree, but require significantly less memory than the full decision tree.