New research group in Dresden at the TU Dresden creates link to the AI community. First practical project: navigation of cargo ships on the Lower Rhine.

Reinforcement learning is a method of machine learning and is becoming increasingly important in the age of digitalisation and artificial intelligence (AI). Its areas of application are very practical, such as the optimisation of traffic light circuits, which is why more and more researchers and companies are devoting themselves to the topic. A Reinforcement Learning (RL) Group is currently being set up at the Chair of Econometrics and Statistics, esp. in the Transport Sector at the "Friedrich List" Faculty of Transport and Traffic Sciences  at TU Dresden. "This is unique in Dresden so far", as Prof. Ostap Okhrin, head of the professorship, explains.

However, Ostap Okhrin and his team do not want to work on the topic alone "in a quiet room" and are specifically looking to join forces with others interested in the topic, such as the Machine Learning Community Dresden (MLC-Dresden). Fabian Hart, research assistant at the professorship, recently gave a lecture on RL research and related projects at the professorship.

At the beginning, he presented reinforcement learning as "the optimisation of an agent's actions over time given a certain reward". He used the development of an automated system for navigating cargo ships on the Lower Rhine in Germany as a practical application - a project co-financed by the Federal Waterways Engineering and Research Institute (BAW) at the chair. Tasks to be tackled with this technique include:

  •  Trajectory planning
  •  Decision-making for overtaking manoeuvres
  •  Ship following mode
  •  Safety distances

The agent, in Fabian's Hart's project a ship, is exposed to a range of input information that it must take into account when making decisions. These can be external, such as the river geometry ahead, the river dynamics or the flow depth. But they can also be internal to the vessel, such as engine dynamics, hydrodynamic effects reacting to the rudder, or vessel pitch as a function of speed. Of the many actions that such a ship must perform, Fabian Hart chose longitudinal steering as the field into which he introduced reinforcement learning.

The online seminar of the Machine Learning Community Dresden was attended by 60 people. According to MLC-Dresden, the question and answer session following the lecture was very lively. Nine questions were recorded in the interactive notes, which in turn each led to sub-discussions. "The colloquium that developed was super interesting even for people who are not familiar with reinforcement learning," says a blog post about the event. Community members also shared interesting papers that explore, for example, the interpretability of such agents. These were captured in "show notes".

More info: MLC-Dresden has shared the slides and show notes from the seminar and colloquium in the blog post via links.