On the Identification and Mitigation of Weaknesses in the Knowledge Gradient Policy for Multi-Armed Bandits.