Roxane Koitz, Franz Wotawa
Proceedings of the International Workshop on Defeasible and Ampliative Reasoning (DARe-16)
Publication year: 2016

While abductive reasoning provides an intuitive approach to diagnosis, its computational complexity remains an obstacle. Even though certain model representations are tractable, computing solutions for instances of reasonable size and complexity persists to pose a challenge. Hence, the discovery of efficient methods to derive abductive explanations presents itself as appealing research area. In this paper, we investigate the structural properties inherent to formalizations suitable for abductive failure localization. Based on the features extracted we construct a meta-approach exploiting a machine learning classifier to predict the abductive reasoning technique yielding the “best” performance on a specific diagnosis scenario. To assess whether the proposed attributes are in fact sufficient for forecasting the appropriate abduction procedure and to evaluate the efficiency of our algorithm selection in comparison to traditional abductive reasoning approaches, we conducted an empirical experiment. The results obtained indicate that the trained model is capable of predicting the most efficient algorithm and further, we can show that the meta-approach is capable of outperforming each single abductive reasoning method investigated