Towards the Deep Learning-Based Autonomous Collision Avoidance

Abstract Reinforcement learning (RL) is considered as an effective method to avoid ship collision. In this study, a ship collision avoidance method based on MADDPG in RL is proposed. In algorithm, some indicators, including DCPA, TCPA, relative distance and relative velocity between ships, are calculated and then collision risk index and the best way point are calculated. In the implementation part, by gathering observational information, setting the reward of the agent and restricting the action of the agent, the agent can learn an effective strategy after training. When there is a risk of collision, the ships in the environment can make reasonable actions to avoid collision. The final experiment verifies the effectiveness of this method.