A reinforcement learning algorithm for mobile robot path planning with dynamic Q-value adjustment

Path planning is essential for mobile robots to execute various tasks across different fields, including intelligent systems. It primarily focuses on the interaction between the agent and its environment, allowing the agent to maximise total reward by an optimal strategy. Many path-planning algorithms that are not agent-based struggle with effectively exploring entirely unknown environments. To address these issues, we propose the Adam deep Q-learning network (ADQN) to solve such problems. ADQN introduces an innovative approach to choosing action and reward functions, optimising Q-value updates dynamically based on temporal-difference error changes for enhanced model convergence and stability. Evaluated across four simulations in two maze environments of varying complexities, ADQN shows significant improvements: reduced steps, increased rewards, faster and stable loss convergence, and notably higher success rates compared to Munchausen reinforcement learning, prioritised experience replay-double duelling deep Q-networks, max-mean loss in deep Q-network algorithms in grid-based experiments.