This work comprise the project thesis, TMR4510, and masters thesis, TMR4930, at NTNU, fall 2019/spring 2020. It investigates different machine learning methods for finding control inputs to the azimuth thrusters of the ReVolt model ship from DNV GL. Each of the directories in src/ contains their own README files which explains how the code was used during the thesis, and how to run it yourself. The only requirement for using the code directly is access to the Cybersea simulator, which only DNV GL can grant access to. But the code is easily adjustable for any other simulator, and in src/rl/windows_workspace, you can find examples of how an environment was created of the simulator by using Python to communicate with a Java-based simulator in Windows.
Masters thesis: Going from the project thesis, it was expected that RL could improve the results from supervised learning due to difficulties of problem formulation using the latter, especially through the use of neural networks and Deep Learning, giving Deep Reinforcement Learning (DRL). The PPO algorithm from OpenAI was the chosen DRL algorithm. The outcome was a control scheme directly translating desired pose of the vessel into thruster commands, hence replacing both the traditional motion controller (like a PID with feedforward) and the thrust allocation scheme (based on e.g. the pseudoinverse or quadratic programming).
The performance was tested against these classic methods, and was found to be as good or better than the others in terms of station-keeping accuracy while simultaneously being more energy efficient during simulations. Real life sea trials was also done, with good results albeit battling some hardware issues of one of the thrusters. This demonstrated that Deep Learning based methods may perform well for control allocation, but that its robustness to unmodelled deviations between the system it was trained on vs. employed on should be focus for further research. My suggestion was to use the neural networks that was trained in simulation as initialization of network training in the real world, and continue the learning process once applied to the real world system.
The implementation of the PPO algorithm was built on OpenAI's Spinning Up repository (https://spinningup.openai.com/), and the work itself is summarized in the master's thesis "Dynamic Positioning using Deep Reinforcement Learning", submitted in June 2020. All the work has been made public for clarity, and for ease of use for others looking to work with Deep Reinforcement Learning for controlling vehicles.
Project thesis: A dataset was created, used for training a neural network using supervised learning in order to learn the physics between the forces/moments acting on ReVolt model ship and the commanded propeller RPMs and angles. A quadratic programming approach was also established as a benchmark (in src/ql/). The SL version resulted in a satisfactory thrust allocation scheme, but the quadratic programming approach was superior in terms of fuel consumption and more controlled motion of the vessel.
The supervised learning and reinforcement learning implementations uses different versions of e.g. tensorflow and keras. This was due to that the specific CPU on the computer used was optimized with tensorflow 1.10.0 and keras 2.2.0, used for training the supervised learning model with large amounts of data. But since the PPO algorithm was trained online, using another computer, using Windows (the same one running the ReVolt simulator), updated versions of the packages was used for better documentation and compatibility.
All in all, a virtual environment avoids ruining the other projects on your machine that might depend on other versions of packages used in this project.
virtualenv --system-site-packages -p python3 ./venvsource venv/bin/activate -> (venv) should pop up to the left on the command line.which python to confirm that the shell uses the python inside venv/. Use pip freeze to see the current package versions used in the virtual environment.deactivate