Learning Continuous Control Policies for Information-Theoretic Active Perception