Generalized Behavior Decision-Making Model for Ship Collision Avoidance via Reinforcement Learning Method

Abstract

Due to the increasing number of transportation vessels, marine traffic has become more congested. According to the statistics, 89% to 95% of maritime accidents are related to human factors. In order to reduce marine incidents, ship automatic collision avoidance has become one of the most important research issues in the field of ocean engineering. A generalized behavior decision-making (GBDM) model, trained via a reinforcement learning (RL) algorithm, is proposed in this paper, and it can be used for ship autonomous driving in multi-ship encounter situations. Firstly, the obstacle zone by target (OZT) is used to calculate the area of future collisions based on the dynamic information of ships. Meanwhile, a virtual sensor called a grid sensor is taken as the input of the observation state. Then, International Regulations for Preventing Collision at Sea (COLREGs) is introduced into the reward function to make the decision-making fully comply with COLREGs. Different from the previous RL-based collision avoidance model, the interaction between the ship and the environment only works in the collision avoidance decision-making stage. Finally, 60 complex multi-ship encounter scenarios clustered by the COLREGs are taken as the ship’s GBDM model training environments. The simulation results show that the proposed GBDM model and training method has flexible scalability in solving the multi-ship collision avoidance problem complying with COLREGs in different scenarios.