Kausales Denken, Bayes-Netze und die Markov-Bedingung

A dominant current framework to model everyday causal knowledge are causal Bayes nets, which represent causal knowledge as directed acyclic graphs. One central assumption of this approach is the Markov constraint. According to the Markov constraint, each variable is independent of all non-descending variables conditional upon its direct causes. Recent research, however, has questioned the Markov condition as part of a psychological theory of causal reasoning. In a common-cause structure, judgments about the presence of a target effect given the presence or absence of its cause depend strongly upon the states of collateral effects of this cause, which violates the Markov condition. In this thesis it is shown that causal inferences are influenced by additional knowledge, particularly knowledge about underlying causal processes; this is the reason for apparent Markov violations. A computational model is presented which extends classical causal Bayes nets by adding a preventive noise source which is attached to each cause. The reasoning process, then, is modeled as adaptive error attribution. This model is empirically tested in three contexts. Furthermore it is shown that inferences are influenced by the properties of the involved objects and therefore are dependent on possible categorizations. Based on these findings an extension of the model is developed which computes target inferences across all possible partitioning of the effects, and integrates over the uncertainty of cluster assignments. Finally, possible consequences of the findings are discussed and a broader computational model of causal reasoning is drafted, which separates the level of causal background knowledge from the processing of statistical events., publication\\\\_type = type, publication\\\_type = type, publication\\_type = type, publication\_type = type

0227039