Study on Enhancing Dueling DQN Performance in Incomplete Observable Environments

This study introduces an innovative approach that integrates a Kalman Filter with Dueling DeepQ-Network (Dueling DQN) for navigating mazes under conditions of incomplete observability. Conventionalreinforcement learning techniques often struggle to navigate efficiently in environments where only incompleteinformation is available. By incorporating the Kalman Filter, our method aims to refine the state estimationprocess under uncertainty, thereby enhancing the agent’s decision-making capability. This integration allows theagent to perform more informed actions by leveraging improved state predictions, leading to enhanced explorationefficiency. Our experiments conducted in maze environments with incomplete observability show that this novelapproach not only surpasses the performance of standard Dueling DQN algorithms in terms of total rewards butalso achieves a reduction in the number of actions required to reach the goal. These findings underscore theeffectiveness of combining Kalman Filter with Dueling DQN in improving reinforcement learning outcomes inenvironments characterized by incomplete observability.