Generalization in Reinforcement Learning

Master thesis

Research focus: Automation and Robotics, Knowledge and Information Services
Courses of study: Informatics, Information Engineering, Information Management, Related Disciplines

Job Description

Investigating and Benchmarking the Robustness of State-of-the-Art Reinforcement Learning Algorithms against Changes in the Environment
Deep Reinforcement Learning (RL) is a promising field of research with impressive results on many tasks such as playing Go or Chess. However, even state-of-the-art RL agents are limited in their ability to generalize acquired knowledge towards previously unseen tasks. This limits their applicability in real world, non-deterministic environments.

The purpose of this thesis is to compare how different RL agents perform if presented with previously unseen environments. In this context she/he develops non-deterministic OpenAI Gym environments and a systematic procedure to assess an agent's generalization performance. Hereby, the student acquires state-of-the-art knowledge of generalization in RL and provides the research community with a systematic evaluation framework for generalization.

Your Responsibilities

  • In-depth literature review about generalization techniques in reinforcement learning
  • Development of a test bench for the generalization of reinforcement learning algorithms
  • Experimental comparison of state-of-the-art RL algorithms in non-deterministic environments

Our Offer

  • Continuous and thorough mentoring of the student
  • A highly motivated and fun team
  • Constructive teamwork

Your Profile

  • Basic knowledge in python programming and reinforcement learning
  • Ability to plan and work independently
  • Very good knowledge of German or English


We are looking forward to your application to Patrick Philipp ( or Marco Heyden (
Please provide us with the following information:

  • Transcript of Records
  • CV

Job Description

  • Start: From now on
  • Responsible institute at KIT: AIFB | Prof. Dr. York Sure-Vetter