Daniel Herrmann, University of California at Irvine.
Talk: Deciding as Learning
FACULTY OF PHILOSOPHY BOARD ROOM, 4PM-530PM, 14 JULY 2022
Talk Abstract: Bayesian epistemology describes the optimal way of learning: conditionalization. Bayesian decision theory describes the optimal way of making decisions: maximize expected utility. Dynamic versions of decision theory involve an agent changing the probabilities she assigns to propositions in a way that is sensitive to their expected utility. This looks puzzling: why shouldn’t she apply normal, epistemic reasoning about these propositions? One answer might be to stipulate that different update rules apply to actions. But this just pushes the question up a level: what are actions, and why should agents learn about them differently? This question is especially pressing for an agent who reasons about herself as part of the world like anything else. I argue that we can resolve this tension by considering how agents reason under conditions of belief-state dependence. In particular, when a partition and an agent’s degrees of belief over that partition exhibit the right kind of dependence, it makes sense for an agent to think of the partition as under her control. Furthermore, under such conditions, an agent might take her desires to be evidence about that partition. With these conditions in place, deciding is revealed to be an instance of learning.
Everyone is welcome - for any queries contact Arif Ahmed, Faculty of Philosophy (ama24@cam.ac.uk)
