Belief Dynamics under Iterated Revision: Cycles, Fixed Points and Truth-tracking

Sonja Smets (University of Groningen) gives a talk at the MCMP Colloquium titled "Belief Dynamics under Iterated Revision: Cycles, Fixed Points and Truth-tracking". Abstract: We investigate the long-term behavior of processes of learning by iterated belief-revision with new truthful information. In the case of higher-order doxastic sentences, the iterated revision can even be induced by repeated learning of the same sentence (which conveys new truths at each stage by referring to the agent's own current beliefs at that stage). For a number of belief-revision methods (conditioning, lexicographic revision and minimal revision), we investigate the conditions in which iterated belief revision with truthful information stabilizes: while the process of model-changing by iterated conditioning always leads eventually to a fixed point (and hence all doxastic attitudes, including conditional beliefs, strong beliefs, and any form of "knowledge", eventually stabilize), this is not the case for other belief-revision methods. We show that infinite revision cycles exist (even when the initial model is finite and even when in the case of repeated revision with one single true sentence), but we also give syntactic and semantic conditions ensuring that beliefs stabilize in the limit. Finally, we look at the issue of convergence to truth, giving both sufficient conditions ensuring that revision stabilizes on true beliefs, and (stronger) conditions ensuring that the process stabilizes on "full truth" (i.e. beliefs that are both true and complete). This talk is based on joint work with A. Baltag.