Stabilizing Reinforcement Learning in Dynamic Environment with Application to Online Recommendation | Synapse