Menu

Seminar: Automated Reinforcement Learning

You can register for the seminar via HISinOne.

Background

Hyperparameter optimization (HPO) is a powerful approach to achieve the best performance on many different problems. However, reinforcement learning (RL) offers unique challenges, such that classical HPO and automated machine learning approaches are often not directly applicable. In this seminar, we will discuss the unique challenges posed by the automated reinforcement learning (AutoRL) problem and discuss a variety of solution approaches, ranging from static to dynamic configuration methods. We further explore topics in reinforcement learning, such as curriculum learning, to understand how environment design can impact the learning performance of RL agents.

Requirements

We require that you have taken lectures on

  • Reinforcement Learning, and/or
  • Machine Learning

We strongly recommend that you have heard lectures on

  • Automated Machine Learning
  • Reinforcement Learning

Organization

The seminar is intended to be held as a block seminar, with sessions at the end of the semester in which all students will present their papers. All students are required to read the relevant literature to prepare for this session. After each presentation, we will have time for a question & discussion round and all participants are expected to take part in these. Each student has to write a short paper about their assigned topic which is to be handed in one week prior to their presentation.

During the semester, the students are expected to meet once with their assigned supervisor to discuss their paper and clarify open questions. Before this meeting however, students are required to meet with a "study buddy" whose role it is to be a discussion partner to clarify the first set of questions, and give feedback for the presentation.

Grading

  • Presentation: 60% (20min + 10min Q&A)
  • Paper: 20% (4 pages in AutoML Conf format, due one week prior to your presentation)
  • Participation in Discussions: 20%

Literature

Relevant literature can be found at https://autorl.org/ and https://autorlworkshop.github.io/. This list contains many recent papers and blog posts in the scope, though not all of the papers that we intend to cover in the seminar. For a general overview on the topic please refer to the survey on AutoRL (https://jair.org/index.php/jair/article/view/13596).

Kick-Off Slides

https://docs.google.com/presentation/d/1jt-RDAZfqnPqKS5YXyuRWuqgvX3wd-kHYnIZf8uylwU/edit?usp=sharing (best viewed in presentation mode due to overlapping animations)

Paper Pool

https://docs.google.com/document/d/1Dp7edapWlOFm7Irihgx8pLJSBNwfGa3nCkspeFVTAcA/edit?usp=sharing