Independence and d-separation in Abstract Argumentation[go to overview]
Efficient inference requires the ability to distinguish parts of a knowledge base that are relevant and irrelevant to a given query. Indeed, various notions of (ir)relevance have been studied in AI. A well-known example is conditional independence in probabilistic reasoning. In this talk we define a notion of conditional independence in a new setting, namely abstract argumentation. This is defined relative to an argumentation framework and semantics and deals with the status of arguments. More precisely: two sets A and B of arguments are independent given a third set C whenever, once the status of the elements of C is known, then knowing the status of the elements of A tells us nothing about the status of the elements of B and vice versa. We then show that questions about conditional independence can be answered (without computing labelings or extensions) by examining the structure of the argumentation framework. We make use of d-separation, which is known as a criterion used in a similar way to answer questions about probabilistic independence by examining the structure of a Bayesian network. Our method may be used to perform various types of inference in abstract argumentation in a more efficient way.
13.02.20 - 10:15