Trading off the Costs of Inference vs. Probing in Diagnosis

Johan de Kleer, Olivier Raiman

Abstract: This paper proposes a new algorithm which when provided the relative costs of computation vs. probing minimizes the total cost of diagnosis. During the diagnosis process the decision of whether to probe or to compute is dependent on the expected costs and benefits of each alternative. It is unlikely that we will be able to find general analytic and simple-to-compute models for the costs and benefits. Therefore, we base our algorithm on simple empirically derived models of costs and benefits. With these models, our algorithm operates by continuously choosing the optimum action to make next. This algorithm will not blow up on the rare pathological cases and will always (on average) find diagnoses at equal to or better cost than a conventional GDE/Sherlock. When the cost of probing is high, then our algorithm behaves exactly the same as GDE/Sherlock. When the cost of computation is high, the algorithm performs the diagnosis at far lower cost than GDE/Sherlock.