Learning, inference, and decision making with probabilistic user models, including considerations of preferences about outcomes under uncertainty, may be infeasible on portable devices. The subject invention provides systems and methods for pre-computing and storing policies based on offline preference assessment, learning, and reasoning about ideal actions and interactions, given a consideration of uncertainties, preferences, and/or future states of the world. Actions include ideal real-time inquiries about a state, using pre-computed value-of-information analyses. In one specific example, such pre-computation can be applied to automatically generate and distribute call-handling policies for cell phones. The methods can employ learning of Bayesian network user models for predicting whether users will attend meetings on their calendar and the cost of being interrupted by incoming calls should a meeting be attended.