Tracking the Truth Requires a Non-wellfounded Prior!

Alexandru Baltag (ILLC Amsterdam) gives a talk at the MCMP Colloquium titled "Tracking the Truth Requires a Non-wellfounded Prior! A Study in the Learning Power (and Limits) of Bayesian (and Qualitative) Update". Abstract: The talk is about tracking "full truth" in the limit by iterated belief updates. Unlike Sonja's talk (which focused on finite models), we now allow the initial model (and thus the initial set of epistemic possibilities) to be infinite. We compare the truth-tracking power of various belief-revision methods, including probabilistic conditioning (also known as Bayesian update) and some of its qualitative, "plausibilistic" analogues (conditioning, lexicographic revision, minimal revision). We focus in particular on the question on whether any of these methods is "universal" (i.e. as good at tracking the truth as any other learning method). We show that this is not the case, as long as we keep the standard probabilistic (or belief-revision) setting. On the positive side, we show that if we consider appropriate generalizations of conditioning in a non-standard, non-wellfounded setting, then universality is achieved for some (though not all) of these learning methods. In the qualitative case, this means that we need to allow the prior plausibility relation to be a non-wellfounded (though total) preorder. In the probabilistic case, this means moving to a generalized conditional probability setting, in which the family of "cores" (or "strong beliefs") may be non-wellfounded (when ordered by inclusion or logical entailament). As a consequence, neither the family of classical probability spaces, nor lexicographic probability spaces, and not even the family of all countably additive (conditional) probability spaces, are rich enough to make Bayesian conditioning "universal", from a Learning Theoretic point of view! This talk is based on joint work with Nina Gierasimczuk and Sonja Smets.