Adapting Virtual Embodiment through Reinforcement Learning