Training a Neural Network to Calculate the Closest Point of Approach

Abstract

The process of training a neural network on calculation of closest point of approach (CPA) between two ships, and testing its performance and accuracy is described in the paper. The architecture of the neural network, the type of input and output data, and creation of training data set are also described in the paper. Feed Forward Neural Networks with Backpropagation algorithm are used; training method is Supervised with Levenberg-Marquardt algorithm. The input data are positions, courses and speeds of vessels in a certain area, the output data are Closest Points of Approach (CPA) between them. The process of writing a script in MATLAB software environment is described. The script allows a user to generate training data with any number of vessels in an area. Comparison of the time spent on CPA calculation using formulas and using neural networks is carried out. It has been proven that when processing large data arrays, the CPA calculation with neural networks is much faster than by means of formulas. After neural networks training process and the calculations results comparison, one neural network with mean squared error of 0.21 is chosen. It can be used for CPA calculations in MATLAB-based simulations. In the future this network might become a base for a collision-avoidance neural network system, which will allow vessels to manoeuvre safely in order to avoid collisions in a certain area.