Reinforcement learning with limited reinforcement: using Bayes risk for active learning in POMDPs