نوع مقاله : علمی پژوهشی
عنوان مقاله English
نویسندگان English
This study presents a comprehensive comparative evaluation of three routing algorithms in static network environments with complex obstacles. The algorithms include A* as a deterministic heuristic search method, Q-learning as a classical reinforcement learning approach, and Double Q-learning as an improved version designed specifically to reduce overestimation bias. Experiments were conducted in two distinct environments: a simple 25×25 grid network and a complex 40×40 urban environment. Algorithm performance was systematically assessed using multi-dimensional criteria including path length, execution time, success rate, path smoothness measured by number of turns, and learning stability across multiple episodes. To ensure statistical reliability of the results, each algorithm underwent 30 repeated independent runs with advanced statistical analyses including 95% confidence intervals, paired t-tests, and ANOVA to validate significance of observed differences. Findings revealed that A* consistently outperformed other algorithms across all quantitative metrics, generating optimal and smoothest paths with average length of 42 steps, minimal execution time of 0.0043 seconds, and only 7 turns. Although both reinforcement learning algorithms achieved 100% success rates in all experimental conditions, they required substantially longer execution times and produced lower-quality paths with considerably more turns (19 and 20 turns respectively). In the complex urban environment, performance disparities intensified significantly, with Double Q-learning optimality dropping from 91.3% to 90%. The results showed that reinforcement learning-based algorithms are better suited to unknown, dynamic, and uncertain environments due to their adaptive capabilities, whereas the A* algorithm achieves superior performance in static and well-defined environments.
کلیدواژهها English
