On Learning Intrinsic Rewards for Policy Gradient Methods