Deep reinforcement learning model for Multi-Ship collision avoidance decision making design implementation and performance analysis
Ruihong Pan, Wei Zhang, Shijie Wang, Shuhua Kang
Abstract
This paper proposes a novel multi-ship collision avoidance decision-making model based on deep reinforcement learning (DRL). The model addresses the critical challenge of preventing ship collisions while maintaining efficient navigation in complex maritime environments. Our innovation lies in the integration of a comprehensive state representation capturing key inter-ship relationships, a reward function that dynamically balances safety, efficiency, and COLREGs compliance, and an enhanced DQN architecture with dueling networks and double Q-learning specifically optimized for maritime scenarios. Experimental results demonstrate that our approach significantly outperforms state-of-the-art DRL methods, achieving a 30.8% reduction in collision rates compared to recent multi-agent DRL implementations, 20% improvement in safety distances, and enhanced regulatory compliance across diverse scenarios. The model shows superior scalability in high-density traffic, with only 12.6% performance degradation compared to 18.4-45.2% for baseline methods. These advancements provide a promising solution for autonomous ship navigation and maritime safety enhancement.