Use of neural networks for safe collision avoidance for groups of autonomous vessels

Abstract

This article analyzes and compares the effectiveness of two algorithms for collision avoidance in groups of autonomous vessels: one based on geometric analysis of vessel approach and cost function minimization for calculating a safe maneuver (traditional algorithm), and an algorithm utilizing a neural network. Both algorithms assume external control of a group of vessels within a specific area using a vessel traffic management system and coordinated maneuvering of dangerously approaching vessels. Descriptions of these algorithms are provided along with their simplified block diagrams; to solve the task of safely maneuvering a group of vessels, a sequential analysis of all possible vessel pairs in the group and changes in their courses is proposed. The process of creating three test datasets is described, two of which were generated using a program and included 100 scenarios each, while the third was manually composed and included 30 scenarios for different vessel group approach variations. During the testing of the neural network algorithm, two neural networks trained to predict safe courses for vessel pairs were utilized. The neural network used in the algorithm, trained on 743671 samples, allowed the processing of test vessel approach scenarios with an accuracy comparable to the traditional algorithm. Depending on the number of dangerously approaching vessels in the area, the neural network algorithm processed test scenarios 2–14 times faster than the traditional algorithm. The paper highlights the limitations of the described algorithms and outlines planned improvements for subsequent research, including the optimization of the safe maneuver selection methodology and further training of the neural network on larger data volumes.