Enhanced Q-Learning for Grid Navigation: Direction-Sensitive Priors, Annealed Exploration, and Potential-Based Reward Optimization, Улучшенное Q-обучение для навигации по сетке, использующее априорные приоритеты направлений, выбор маршрутов методом отжига и оптимизацию вознаграждений на основе потенциальной достижимости цели

<jats:p>his paper presents an enhanced Q-learning framework specifically designed for grid navigation in environments with obstacles. Building upon the foundation of standard tabular Q-learning, we propose three pivotal improvements: a direction-sensitive Q-table initialization method that achieves goal alignment without requiring prior knowledge of obstacles; an annealed Boltzmann exploration strategy augmented with upper confidence bound terms to facilitate balanced and adaptive sampling; and a potential-based continuous reward shaping function to deliver denser feedback and accelerate the learning process. These enhancements address common challenges in sparse-reward settings, such as inefficient exploration and sluggish value propagation. Experimental evaluations conducted on randomly generated grids demonstrate that, compared to baseline methods such as standard Q-learning and its variants, our approach achieves higher success rates, shorter optimal path lengths, and faster convergence speeds. By preserving the model-free generality of Q-learning while enhancing sample efficiency, this framework proves well-suited for practical applications in robotics and path planning domains.</jats:p>