Section: New Results

Diagnostic and causal reasoning

Participants : Philippe Besnard, Louis Bonneau de Beaufort, Marie-Odile Cordier, Yves Moinard, Karima Sedki.

Stemming on [29] , [30] , [31] , [32] , [33] , we have designed an inference system based on causal statements. This is related to diagnosis (observed symptoms explained by faults). The aim is to produce possible explanations for some observed facts. Previously existing proposals were ad-hoc or, as in [36] , [47] , they were too close to standard logic in order to make a satisfactory diagnosis. A key issue for this kind of work is to distinguish logical implication from causal links and from ontological links. This is done by introducing a simple causal operator, and an is-A hierarchy. These two operators are added to a restricted first order logic of the Datalog kind (no function symbols). Then, our system produces elementary explanations for some set of observed facts. Each explanation links some facts to the considered observation, together with a set of atoms called the justifications: The observation is explained from these facts, provided the justifications are possible (not contradicted by the available data). This formalism has also been translated into answer set programming [57] , [58] ).

This year, we have extended our formalism in order to deal with more complex problems such as finding explanations for the hurricane Xynthia (2010, February 28). In such situations, there are many data and many possible elementary explanations can be examined. This involves an extension of our formalism, in order to deal with more complex chains of causations and is-A links. We are on the way to end this task. Our formalism makes precise what all these possible explanations are. Then, in order to deal with so many possible complex explanations, we integrate this causal formalism into an argumentation framework. Logic-based formalizations of argumentation [34] take pros and cons for some conclusion into account. These formalizations assume a set of formulae and then exhaustively lay out arguments and counterarguments. This involves providing an initiating argument for the inference and then providing undercuts to this argument, and then undercuts to undercuts. So here our causal formalism provides a (rather large) set of explanations, and the argumentation part allows to select the best ones, under various criteria.

Then, since answer set programming can easily deal with logical formalisms, the argumentation part will be incorporated into our already existing answers set programming translation of the causal formalism. Regarding answer set programming, we have also examined some more difficult examples [16] and participated to a chapter in the to be published "Panorama de l'intelligence artificielle. Ses bases méthodologiques, ses développements" [19] .