Learning autonomous maritime navigation with offline reinforcement learning and marine traffic data
Westerlund, Jimmy (2021)
Westerlund, Jimmy
2021
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe2021060734637
https://urn.fi/URN:NBN:fi-fe2021060734637
Tiivistelmä
Autonomous shipping is a heavily researched topic, and currently, there are large amounts of ship traffic data available but unexploited. Autonomous ships have the potential to reduce costs and increase safety. The challenge is achieving the correct maritime navigation behavior according to the situation reliably, which may be possible by exploiting historical ship traffic data. This thesis explores the possibility of using offline reinforcement learning based on AIS data to learn autonomous maritime navigation.
The hypothesis that AIS data can be used for training a reinforcement learning agent is tested by implementing an offline reinforcement learning agent. For comparison, an online agent that learns without data is also implemented. Both agents are trained and evaluated in a simulator, and the goal of both agents is to learn to navigate to a destination, given a starting point.
The results suggest that offline reinforcement learning can be used for automating maritime navigation, but a more extensive and more diverse dataset is needed to conclude its effectiveness.
The hypothesis that AIS data can be used for training a reinforcement learning agent is tested by implementing an offline reinforcement learning agent. For comparison, an online agent that learns without data is also implemented. Both agents are trained and evaluated in a simulator, and the goal of both agents is to learn to navigate to a destination, given a starting point.
The results suggest that offline reinforcement learning can be used for automating maritime navigation, but a more extensive and more diverse dataset is needed to conclude its effectiveness.