Efficient Belief Propagation for Utility Maximization and Repeated Inference