What makes propositional abduction tractable

Abduction is a fundamental form of nonmonotonic reasoning that aims at finding explanations for observed manifestations. This process underlies many applications, from car configuration to medical diagnosis. We study here the computational com-plexity of deciding whether an explanation exists in the case when the application domain is described by a propositional knowledge base. Building on previous results, we classify the complexity for local restrictions on the knowledge base and under various restrictions on hypotheses and manifestations. In comparison to the many previous studies on the complexity of abduction we are able to give a much more detailed picture for the complexity of the basic problem of deciding the existence of an explanation. It turns out that depending on the restrictions, the problem in this framework is always polynomial-time solvable, NP-complete, coNP-complete, or ΣP2-complete. Based on these results, we give an a posteriori justification of what makes propo-sitional abduction hard even for some classes of knowledge bases which allow for efficient satisfiability testing and deduction. This justification is very simple and intuitive, but it reveals that no nontrivial class of abduction problems is tractable. Indeed, tractability essentially requires that the language for knowledge bases is un-able to express both causal links and conflicts between hypotheses. This generalizes a similar observation by Bylander et al. for set-covering abduction