Christopher Hitchcock (Caltech): Updating on the Credences of Others

Abstract: How should you update your credences upon learning the credences of others? Because of the complexity of Bayesian conditionalization in this context, there has been considerable interest in developing simple heuristics, the most popular being linear averaging. However, linear averaging has a number of drawbacks: it does not commute with itself, nor with conditionalization; it does not preserve independence; and it is not always compatible with conditionalization. In addition, we argue that a further drawback with linear averaging is that it lacks a property we call ‘synergy’. We propose a new heuristic that is just as simple as linear averaging but doesn’t have these drawbacks. (joint work with Kenny Easwaran, Luke Fenton-Glynn, and Joel Velasco)