A reinforcement learning approach to route selection for ice-class vessels

• Main Authors: Tran, Trung Tien, Browne, Thomas, Peters, Dennis, Veitch, Brian
• Format: Article in Journal/Newspaper
• Language: English
• Published: IEEE 2020
• Subjects: reinforcement learning; POLARIS; path planning; Arctic; ice covered waters
• Online Access: https://nrc-publications.canada.ca/eng/view/accepted/?id=dca26c86-da58-4e64-bcb0-1b7c11545683; https://nrc-publications.canada.ca/eng/view/object/?id=dca26c86-da58-4e64-bcb0-1b7c11545683; https://nrc-publications.canada.ca/fra/voir/objet/?id=dca26c86-da58-4e64-bcb0-1b7c11545683

Identifying an optimal route for a vessel navigating through ice-covered waters is a challenging problem. Route selection requires consideration of vessel safety, economics, and maritime regulations. Vessels navigating in the Arctic must follow the operational criteria imposed by the Polar Operational Limit Assessment Risk Indexing System (POLARIS), recently introduced by the International Maritime Organization. This research investigates a framework to find an optimal route for different ice-class vessels using reinforcement learning. The system defines a Markov Decision Process and uses Q-learning to explore an environment generated from a Canadian Ice Service ice chart. Reward functions are formulated to achieve operational objectives, such as minimizing the distance travelled and the duration of the voyage, while adhering to POLARIS criteria. The experimental results show that reinforcement learning provides a means to identify an optimal route for ice-class vessels. Peer reviewed: Yes NRC publication: Yes