Deep Reinforcement Learning Algorithms for Ship Navigation in Restricted Waters

Abstract

Reinforcement Learning has not been fully explored for the automated control of ships maneuvering movements in restricted waters. Nevertheless, more robust and efficient control can be achieved with such algorithms. This paper presents the use of Deep Q Network and Deep Deterministic Policy Gradient methods with a numerical simulator for ship maneuvers to develop control laws. Both methods proved to be efficient in navigational control through a channel. A comparison of response and control behavior resulting from each of the methods is presented.