Section: New Results

Causal reasoning and argumentation

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

We have continued our work on reasoning (precisely search for explanations) from causal relations and ontology [48] . We resort to a well-known model [49] in computational argumentation in order to provide some structure to the collection of potential explanations given by our causal formalism. We have developed a case study, namely the Xynthia storm case, (February 2010, western France, trial September 2014) for which there exists a huge amount of data from various official reports. We have implemented an ASP program which thereby provides another application, besides those already mentioned: mining and landscape simulation, for ASP.

Cognitive map is a qualitative decision model which is frequently used in social science and decision making applications. This model allows to easily organize individuals’ judgments, thinking or beliefs about a given problem in a graphical representation containing different concepts and influences between them. However, cognitive maps cannot model uncertainty within the variables and provides only deductive reasoning (predicting an effect given a cause). In [37] , we show how to translate the knowledge represented in cognitive maps in the form of arguments and attack relations among them. Given a decision problem, we propose to build, first, a cognitive map by eliciting knowledge from experts and then to transform it into a weighted argumentation framework (WAF for short) for ensuring efficient reasoning. Another contribution concerns enriching the WAF obtained from a given cognitive map for dealing with dynamics through the consideration of a varying set of observations.

Cognitive maps and Bayesian networks are useful formalisms to address knowledge representation. Cognitive maps are powerful graphical models for gathering or displaying knowledge but while offering an easy means to express individuals judgments, drawing inferences remains a difficult task. Bayesian networks are widely used for decision making processes that face uncertain information or diagnosis but are difficult to elicitate. To take advantage of both formalisms and to overcome their drawbacks, Bayesian causal maps (BCM) were developed [75] . In [6] , we propose to start from a causal map to construct the model and then set the conditional probabilities. Once the common causal map (CM) is built we can transform it into a BCM which combines causal modeling techniques and bayesian probability theory. We have developed a complete framework and applied it on a real problem in an environmental context. The implemented decision facilitating tool enables the representation of different shellfish dredgers views about their activity as well as the test of different fishery management scenarios.