Much work in theory revision is framed as a companion to explanation-based learning. Since the latter typically requires a complete and correct theory, theory revision techniques rework an invalid theory into a form such algorithms accept. If an explanation module fails to generate an explanation of some examples, the theory revision module could inductively guess at a refinement or correction, allowing valid explanations and use of the theory on similar examples in the future.
There are several research projects in the literature that apply this framework of ``learning by failing to explain'' to particular theory representations and tasks. Ourston and Mooney's EITHER [9] uses theory revision on supervised learning problems. The system accepts a theory expressed in propositional Horn-clause notation and a number of labelled training examples. The theory evaluates the features and predicts a label. If the label is wrong, EITHER computes the minimal change to the theory that corrects the prediction, then continues making changes to the theory until all examples are correctly classified. Our problem differs in that it involves unsupervised learning, so the training procedure cannot estimate its own performance.
A common knowledge representation for the diagnosis of complex
machines is the fault hierarchy, which lets technicians proceed from a
high-level description of symptoms to the identification of likely
causes and malfunctions through a series of tests. Langley et
al. [7] describe the 
theory revision algorithm
for correcting the fault hierarchy in case of diagnostical errors.
Like EITHER, their system detects training cases that are mislabelled
by the existing knowledge base. The revision procedure generates all
possible transformations of the fault hierarchy and chooses the one
that reaches the most correct diagnoses, continuing until there is no
transformation that improves on the current fault hierarchy. Besides
depending on labelled training examples,
has little relevance
to our problem because it exhaustively generates theory
transformations, which is not practical in continuous domains.
Some planning-based systems that interact with an environment also demonstrate the use of theory revision to complete their tasks successfully. If Gil's EXPO [3], or Pearson's IMPROV [10] fail to achieve a goal expected from planning, they attempt to correct their plan knowledge through interaction with the environment. Both agents take a variety of actions in the world, then analyze the effects to determine how to perform better in the future. Our problem is fundamentally different because our system cannot to perform experiments to test hypotheses. Instead, it is forced to passively observe the environment and build knowledge structures. Although Wang's OBSERVER [14] also passively observes a series of expert execution traces, it also requires sensors to record the effects of the expert's actions on goal conditions. A final distinction from all these planning systems is that, rather than accomplish any specific goal, our system attempts to augment current knowledge about the driving environment.