Meta-Reinforcement Learning of Structured Exploration Strategies