6.1 Acknowledgments 5 Related Work

The contents of this paper beneetted from several discussions with Walter Ham-scher, who generously volunteered to run our light bulb example through one of his diagnosis systems, and from the constructive comments of two anonymous referees. References 1] K. Abbott. Robust operative diagnosis as problem solving in a hypothesis space. Poole 12, p. 1310] has noted the generality of the model-based paradigm and suggested its applicability to a large class of recognition problems including planning. There have also been extensions to Reiter's algorithm, such as the work of Ng 11], which extends the algorithm to handle time-varying, physical devices. However, we know of no attempts to extend model-based diagnosis to accommodate FDIR, with the possible exception of the work of Friedrich and colleagues 6], who deene a notion of \therapy" and sketch an algorithm for \the standard therapeutic approach." The latter can be characterized as a process of interleaving diagnosis and repair to suppress \undesired symptoms." This approach diiers from ours in that it eliminates or repairs only those components whose treatment causes the disappearance of the observed symptoms; it assumes that granularity of reconnguration is precisely that of diagnosis, i.e., the reconngurable units are the same as the diagnosable units; and it assumes that the level of acceptable system functionality remains constant from diagnosis to reconnguration. 6 Conclusions and Future Directions We have proposed a characterization of reconnguration as an extension of Reiter's theory of model-based diagnosis. Our contribution has been to recognize and exploit an analogy between the problem of model-based diagnosis and that of reconngura-tion. The simplicity of this analogy suggests that it should be possible to use an existing diagnosis engine to compute reconngurations. Hamscher's report 10] on the satisfactory mechanization of our light bulb example using an experimental diagnostic system is encouraging in this regard. In order to realize the beneets claimed in the introduction, we need to develop methods for interleaving diagnosis and reconnguration. The simplest approach requires massive iteration: for each candidate diagnosis, and for each acceptable behavior , compute the reconngurations which achieve that behavior. If all diagnoses yield to a single reconnguration, we are done; if not, we need methods for eliminating candidate diagnoses and reconngurations. We plan to explore focusing strategies and other techniques for avoiding this unacceptably large number of iterations and exploring the combined space of diagnoses and reconngurations more eeciently. We also plan to evaluate our approach …