Revising Beliefs Through Arguments: Bridging the Gap Between Argumentation and Belief Revision in MAS

This paper compares within the MAS framework two separate threads in the formal study of epistemic change: belief revision and argumentation theories. Belief revision describes how an agent is supposed to change his own mind, while argumentation deals with persuasive strategies employed to change the mind of other agents. These are two sides (cognitive and social) of the same epistemic coin: argumentation theories are incomplete, if they cannot be grounded in belief revision models – and vice versa. Nonetheless, so far the formal treatment of belief revision mostly neglected any systematic comparison with argumentation theories. In MAS such problem becomes evident and inescapable: belief change is usually triggered by communication and persuasion from other agents, involving deception, trust, reputation, negotiation, conflict resolution (all typical issues faced by argumentation-based models). Therefore, a closer comparison between belief revision and argumentation is a necessary preliminary step towards an integrated model of epistemic change in MAS. 1 Belief Revision Without Argumentation Following the seminal work in [13], belief revision has recently become an extremely active area of research at the confluence between AI, logic, cognitive science, and philosophy. Notwithstanding the impressive amount and quality of studies devoted to this topic (including many researches in the MAS community, e.g. [1, 9, 11, 12, 29]), belief revision has been mainly addressed in a rather single-minded fashion, isolating the issue of belief change from other related features of cognitive processing. As remarked in [25], current theories of belief revision have been put forward and discussed in a sort of epistemological vacuum, without providing a more comprehensive account of epistemic states and dynamics. Moreover, the process of belief change has been usually conceived as an isolated activity, neglecting even the most obvious connections with other cognitive tasks: e.g. inferential reasoning, communication, argumentation (significant exceptions to this trend are in [10, 12]). On the contrary, we claim that belief revision should be investigated as a specific function (albeit a crucial one) in the cognitive processing of epistemic states, integrating formal models of belief change in a more comprehensive epistemological theory, and providing systematic connections with related cognitive tasks. 1.1 Limitations of Current Theories The AGM paradigm [13] has been the most influential model of belief revision so far, serving as a frame of reference for both refinements and criticisms of the original proposal. Roughly summarizing (see [20] for further discussion), this model was first conceived as an idealistic theory of rational belief change: belief states were characterized as sets of propositions (infinite and deductively closed), three basic types of change were described (expansion, contraction, revision), and rationality was expressed by a set of postulates binding these operators. To decide between different outcomes of the revision process (i.e. different sets of propositions consistent with the rationality postulates), an ordering criterion was introduced in the original belief state, ranking propositions for their importance (epistemic entrenchment). This approach to belief revision fails to integrate with argumentation theories for two reasons: (1) it does not make any predictions or assumptions about how and why some propositions come to be believed, rather than others; (2) there is a deliberate lack of structural properties in the characterization of epistemic states. Argumentation theories capture how a desired change in the audience’s beliefs is brought about by the arguer: therefore, without an explicit theory of the reasons to believe something, the whole point of argumentation is lost. AGM-style approaches to belief revision simply lack the necessary internal structure to describe argumentative strategies (for a philosophically oriented discussion of justification in belief revision, see [14]). In this respect, the so called foundation theories of belief revision fare better than AGM, since they provide a precise account of the reasons supporting a given belief, e.g. using Truth Maintenance Systems [8]. Similar proposals have also been advanced in the field of multi-agent systems [9, 11, 12, 17], and there are several analogies between the criticisms to the AGM approach discussed in this paper and objections raised within the TMS community (e.g. the need for detailed analysis of the reasons that support and determine the agent’s beliefs), although our approach is more cognitive-oriented, while TMS put greater emphasis on computational issues. Since a detailed comparison between our approach and TMS is beyond the aim of this work (cf. 4 on future developments in the direction), here we will provide only a short comment on belief change and argumentation in TMS. In spite of the richer framework outlined by TMS for belief revision in MAS, only few of these theories explicitly address argumentation and/or communication (e.g. [17]), and the structural properties of epistemic states are restricted to factual supports for the agent’s beliefs, to ensure an accurate weighting of unreliable and/or contrasting sources of information. Although such structures are essential to integrate belief revision and argumentation, they are not enough: a fairly rich picture of argumentative strategies must include motivational and emotional features [7, 15, 16], not only factual credibility. Since also belief revision is affected by similar considerations, a more comprehensive cognitive model of epistemic change must be devised (cf. 2.1-2.4). 2 A Cognitive Model of Data-oriented Belief Revision (DBR) The following sections provide a short outline of an alternative model of belief revision, i.e. Data-oriented Belief Revision (DBR): for further details, see [6, 20]. Although this model is still mainly theoretical and far from implementation in MAS, it is conceived as a realistic cognitive framework for understanding belief revision in agent-based social simulation. 2.1 Data and Beliefs: Properties and Interactions Two basic epistemic categories, data and beliefs, are put forward in this model, to account for the distinction between pieces of information that are simply gathered and stored by the agent (data), and pieces of information that the agent considers reliable bases for action, decision, and specific reasoning tasks, e.g. prediction and explanation (beliefs). Clearly, the latter are a subset of the former: the agent might well be aware of a datum that he does not believe (i.e. he does not consider reliable enough); on the other hand, the agent should not be forced to forget (i.e. to lose as a datum) a piece of information which he temporarily rejects as a belief [6]. Moreover, a rejected piece of information retains significant epistemic properties (e.g. its own unreliability, and the reasons for it) that will often be crucial in future revisions and should be preserved by a formal model of belief change [9, 25]. The distinction between data and beliefs yields a number of relevant consequences for the formal study of epistemic dynamics: to start with, it leads to conceive belief change as a two-step process. Let us consider external belief change (cf. 2.3), by way of example. Whenever a new piece of evidence is acquired, either through perception or communication, it affects directly the agent’s data structure, and only indirectly his belief set. In other words, the effects (if any) of the new datum on the agent’s beliefs depend (1) on its effects on the other data, and (2) on the process of belief selection applied by the agent over such data (cf. 2.2). We propose to call this procedure Dataoriented Belief Revision (DBR). More generally, data and beliefs define the two basic cognitive layers of the whole epistemic processing performed by the agent, as summarized in Figure 1. An exhaustive discussion of this general model is beyond the aim of this paper: here we will focus mainly on the treatment of data, with special reference to information update, data properties and assessment, and belief selection (cf. 2.2-2.3), since these are the features most directly involved in belief revision. However, it is important to keep in mind the overall epistemic processing, if we want to provide a formal model adequate to express belief change in cognitive agents. 1 Broader accounts of belief revision have been advocated also for epistemic change in communication [12] and in defeasible reasoning [10, 25]. Info update Data mapping Storing Belief selection Inferential reasoning