Advancing ship automatic navigation strategy with prior knowledge and hierarchical penalty in irregular obstacles: a reinforcement learning approach to enhanced efficiency and safety

With the global wave of intelligence and automation, ship autopilot technology has become the key to improving the efficiency of marine transportation, reducing operating costs, and ensuring navigation safety. However, existing reinforcement learning (RL)–based autopilot methods still face challenges such as low learning efficiency, redundant invalid exploration, and limited obstacle avoidance capability. To this end, this research proposes a GEPA model that integrates prior knowledge and hierarchical reward and punishment mechanisms to optimize the autopilot strategy for unmanned vessels based on deep Q-network (DQN). The GEPA model introduces a priori knowledge to guide the decision-making of the intelligent agent, reduces invalid explorations, and accelerates the learning convergence, and combines with hierarchical composite reward and punishment mechanisms to improve the rationality and safety of autopilot by means of end-point incentives, path-guided rewards, and irregular obstacle avoidance penalties. The experimental results show that the GEPA model outperforms the existing methods in terms of navigating efficiency, training convergence speed, path smoothness, obstacle avoidance ability and safety, with the number of training rounds to complete the task reduced by 24.85%, the path length reduced by up to about 70 pixels, the safety distance improved by 70.6%, and the number of collisions decreased significantly. The research in this paper provides an effective reinforcement learning optimization strategy for efficient and safe autonomous navigating of unmanned ships in complex marine environments, and can provide important theoretical support and practical guidance for the development of future intelligent ship technology.