A neural network-based approach to collision avoidance for maritime autonomous surface ships

A neural network-based approach to cooperative collision avoidance for Maritime Autonomous Surface Ships (MASS) is proposed. To implement this approach, the neural network architecture is defined. The coordinates of all vessels involved in collision avoidance, together with the northern and eastern components of their absolute velocities, are used as network inputs. The outputs are course over ground alterations that enable the vessels to pass at a distance not less than the specified value. The network has eight hidden layers, each containing 1,000 neurons. The algorithm for generating the training dataset involves modeling clearly hazardous reference collision situations derived from safe situations using a constrained optimization method. The training algorithm includes two iterations. At each outer iteration, a gradient optimization method is applied and the network’s generalization performance is monitored. Each inner iteration corresponds to an epoch of the gradient-based training method. The approach was implemented for three vessels using the Python programming language and the TensorFlow library. One million samples were generated for both the training and validation datasets. After 150 outer iterations and five inner iterations, the neural network system predicted course over ground alterations for the training samples with a mean error of approximately 2°, which indicates convergence of the training process. It is noted that the maximum prediction error of the network remains at the level of several tens of degrees for both the validation and training datasets throughout the entire training process.