Edge-Compatible Reinforcement Learning for Recommendations