Towards Boosting the Performance of Deep Reinforcement Learning for Partially Observable Markov Decision Processes

This is a co-authored work towards boosting the performance of deep reinforcement learning for POMDP. By revisiting the theoretical deduction process for belief state updating, we propose to explicitly incorporate the embedding of actions to complement the existing approach. The formed pair of <action, observation> are expected to provide more robust and informative cues for speeding up the training process and improving the testing results.