Jämförande analys av A* och Q-learning-algoritmer för robotbanplanering i dynamiska lagerlokaler

The purpose of this thesis was to compare the performance of the A* and Q-learning algorithms for robot path planning in dynamic warehouse environments. The primary metric was throughput, defined as the number of packages successfully transported per robot within a set number of simulation steps. Results showed distinct strengths for each algorithm depending on environmental complexity. The Q-learning algorithm, enhanced with Double Deep Q-Network (DDQN) and QMIX for multi-agent coordination, demonstrated superior adaptability and higher throughput in environments with dynamic obstacles such as humans. In contrast, the A* algorithm proved reliable primarily in static or less dynamic scenarios. These findings suggest reinforcement learning methods, when suitably enhanced and tuned, can outperform traditional heuristic algorithms in dynamic warehouse conditions.