PKB-POMDP : Prior Knowledge Based POMDP Deep Reinforcement Learning Method

In practical deep reinforcement learning (DRL) applications, partial observability often degrades the performance of standard algorithms. The partially observable Markov decision process (POMDP) provides a formal framework to handle this, and recurrent neural network (RNN) based methods have become the dominant approach by encoding observation histories into hidden states. However, due to the partial observability of POMDP, most of the currently proposed methods have performance bottleneck problems. To address this, we propose Prior Knowledge Based POMDP (PKB-POMDP), which enhances RNN-based training by incorporating prior task knowledge. Using joint distributions and conditional independence, our method introduces additional informative variables to alleviate the impact of partial observability. Experiments on Cartpole, Lunarlander, Hopper, and Ant show that PKB-POMDP consistently outperforms baseline methods. Notably, in complex environments such as Hopper and Ant, our method achieves a performance improvement of nearly 200% compared to the baseline methods.