Adaptive Exploration Deep Q-Network for AGV Static Path Planning with Improved Convergence and Optimization

This paper investigates the use of Deep Reinforcement Learning (DRL) algorithms, particularly Deep Q-Networks (DQN), for static path planning. Traditional methods like Dijkstra and A* struggle with complex obstacles, resulting in suboptimal solutions. This study improves the DQN-based path planning method by integrating an adaptive epsilon and a heuristic exploration strategy, resulting in the Adaptive Exploration Deep Q-Network (AEDQN) algorithm. These enhancements mitigate the exploration-exploitation dilemma and the challenge of sparse rewards in traditional methods, thereby accelerating convergence and improving path optimization. The paper models the AGV environment using a grid map and performs simulation experiments comparing the AEDQN algorithm with standard DQN and other improved variants. Experimental results demonstrate that AEDQN significantly outperforms existing methods, achieving shorter path lengths and faster convergence times.