Intelligent decision-making system for multiple marine autonomous surface ships based on deep reinforcement learning

This study addresses the issue of multi-ship collision avoidance and the requirements for autonomous navigation of Marine Autonomous Surface Ships. To address this challenge, we propose a collision avoidance decisionmaking system for multi-ship encounters that integrates the Proximal Policy Optimization (PPO) algorithm with the ship domain while considering ship manoeuvring characteristics. Moreover, the reward function designed for the PPO algorithm can effectively ensure the trained collision avoidance decision-making model adheres to COLREGs and to prioritizes the utilization of small angle steering whenever possible. Additionally, a neural network based on Bidirectional Gated Circulation Unit (BiGRU) is established, incorporating Layer Normalization to handle high-dimensional observation data and enhancing the algorithm's efficiency. Through simulation experiments, the results demonstrate the flexibility and safety of the proposed decision-making method in multi-ship collision avoidance and surface obstacle avoidance scenarios.