A multi-stage collision avoidance model for autonomous ship based on fuzzy set theory with TL-DDQN algorithm

Autonomous ship with intelligent collision avoidance and maneuvering is one of the most effective ways to solve the problem of maritime traffic accidents caused by human factors. This paper proposed a multi-stage collision avoidance model based on fuzzy set theory and Deep Reinforcement Learning(DRL). It can assist autonomous ships in achieving precise collision avoidance guidance in complex waterway traffic environments. When designing the reward function, the multi-stage collision avoidance model divides the collision avoidance process into different stages according to the relative locations and situations between ships, and adopts different fuzzy membership functions to reward the autonomous ship. When designing the training framework, Transfer Learning-Double Deep Q Network(TL-DDQN) transfers the trained neural network parameters and experience playback pool parameters to the test DRL according to the situation encountered. Simulations are conducted to verify the performance of the multi-stage collision avoidance model. The results show that the proposed model can handle the complex encounter situation well and make autonomous ship reach the target safely. •A multi stage collision avoidance decision-making strategy based on fuzzy set theory is proposed to better balance the straight-line navigation and collision avoidance operations of the ship.•A TL-DDQN deep reinforcement learning framework is proposed.•Several reward functions considering navigation characteristics, COLREGs, ship-to-ship distance are designed to ensure safety of ship navigation and avoid collision.