Structured World Belief for Reinforcement Learning in POMDP