Did you ever want to know when and where your ship arrives? Worry no more ... 😎

This is our project of a distributed machine learning application to predict the routes and ETAs of ships based on their AIS data, built as the summer term assignment for the Software Development module at Hamburg University of Technology.

The application we develeped uses a broker-agent architecture. The broker is the centerpiece of this architecture. It servers a web interface where the user can upload the input data to run predictions on (as a .arff file) and then see a visual representation of the predicted route on a map. In order to predict an accurate route for the given AIS data sample of a ship, the broker sends out requests to multiple agents - other servers, each an "expert" for its own sector of a ny of the two routes considered in this project. The sectors are shown in the map below.

For each agent, we selected the corresponding data, cleaned and preprocessed it, and then trained three seperate models on the data - a Random Forrest, an Artificial Neural Network, and A K-Nearest-Neighbour regressor. The agent returns to the agent the output of the model that performed best on the evaluation on the test data. Once all agents have returned their predictions, the broker shows the predicted route in the web interface. See below for a diagram of this communication scheme.

All implementations are written in Python. Data preprocessing and training of the models was done using the Scikit-learn and Pandas packages inside of Jupyter notebooks. The final product is running Flask web servers inside of Docker containers. For an overview of the complete tech stack, see the image below.

The easiest way to try is to use the version we hosted ourselves on Heroku. Just visit [ais-predictor.herokuapp.com], upload a correctly sructured .arff file of AIS data and predict away. The servers might take a moment to wake up because Heroku puts them into sleep mode after sitting idle for 30 minutes. Note that, unfortunately, we cannot provie our original .arff files for copyright reasons.

Alternatively, you may run the Docker images for the broker and all agents yourself. Keep in mind that you will probably need to change the agent URLs given to the broker.

• IAmTotallyBiased (ChaoJJ)
• Jan Kaiser (jank324)
• Vincent Hanke (VincentHanke)