Improving Training Result of Partially Observable Markov Decision Process by Filtering Beliefs