The use of artificial neural networks for the intelligent optimal control of surface ships

Many conventional ship autopilots use proportional integral and derivative (PID) control algorithms to guide a ship on a fixed heading (course-keeping) or a new heading (course-changing). Such systems usually have a gyrocompass as a single sensory input. Modern sea going vessels have a range of navigation aids most of which may be interconnected to form integrated systems. It is possible to employ the navigational data to provide best estimates of state vectors (Kalman filter) and optimal guidance strategies. Such techniques require powerful computing facilities, particularly if the dynamic characteristics of the vessel are changing, as may be the case in a maneuvering situation or changes in forward speed. This paper investigates the possibility of training a neural network to behave in the same manner as an optimal ship guidance system, the objective being to provide a system that can adapt its parameters so that it provides optimal performance over a range of conditions, without incurring a large computational penalty. A series of simulation studies have been undertaken to compare the performance of a trained neural network with that of the original optimal guidance system over a range of forward speeds. It is demonstrated that a single network has comparable performance to a set of optimal guidance control laws, each computed for different forward speeds