Meta-Reinforcement Learning Based on Self-Supervised Task Representation Learning