Neural-Network-Driven Reward Prediction as a Heuristic: Advancing Q-Learning for Mobile Robot Path Planning

While Q-learning is a staple in path planning, it frequently suffers from slow convergence in practical scenarios. To overcome this, we propose NDR-QL (Neural-Network-Driven Reward Prediction based Q-learning), a method that leverages neural network outputs as heuristic information to accelerate the learning process. Specifically, we optimized a dual-output neural network by introducing a start-end channel separation mechanism and enhancing feature fusion. The resulting model generates two outputs: a narrowly focused "guideline" distribution and a broader "region" distribution. We utilize the guideline to calculate a continuous reward function and the region to initialize the Q-table with a strategic bias. Experiments on public datasets demonstrate that our model improves prediction accuracy by 5% over previous methods. Furthermore, NDR-QL accelerates convergence by 90% and produces superior path quality compared to existing improved Q-learning baselines.