Recent days, with the globalization of the world economy the ships count are increasing for marine transportation and the Waterways are growing increasingly overcrowded than previous. This situation may leads to problem of collision of ships which may cause losing of life and damage for property and to nature. Many automatically collision models of ships are implemented but many focused only on the ship-ship encounter situation only. By using a grid sensor which is virtual sensor, agents of Deep Reinforcement Learning (DRL) classify approach of a multiple ships. This framework introduces an automatically collision detection algorithm for ships using DRL in continuous action spaces. DRL is used for avoidance of collision with a maximum distance of safe passing between ships. A unique method is developed named inside Obstacle Zone by Target (OZT) used to change learning capability that expands the OZT. Using Bi-directional Long Short-Term Memory (BI-LSTM) cell, network is redesigned and continuous action spaces training is carried out to train a model with longer safe distance of ships. In collision detection model that the bow cross range is effective for COLREGs compliant collision avoidance that is proposed in this model. The propose model also validates a scenario that included more ships and have passed that Imazu problem. The proposed BI-LSTM model achieved 80.43 % of accuracy, 95.67% of precision, 85.89% of recall and 93.54% of f1 score values.
Avoidance of Ship Collision Using Deep Reinforcement Learning with Bi-Directional Long Short-Term Memory in Continuous Action Spaces
2024-07-26
269549 byte
Conference paper
Electronic Resource
English
Deep Learning Approach to Satellite Collision Avoidance Using Long Short-Term Memory
Springer Verlag | 2024
|Ship Collision Avoidance Autonomous Avoidance System using Deep Learning
European Patent Office | 2021
|Pedestrian Collision Avoidance Using Deep Reinforcement Learning
Springer Verlag | 2022
|Vehicles Control: Collision Avoidance using Federated Deep Reinforcement Learning
ArXiv | 2023
|