Cluster-based curriculum learning for DAC
Im Rahmen des Projekts AutoML


Dynamic Algorithm Configuration adapts an algorithm's hyperparameters on the fly using reinforcement learning, allowing for better overall performance. For practical application, this requires good generalization capabilities in order to perform well across diverse sets of instances for a given algorithm (e.g. different SAT problems for a SAT-solver or datasets for a Machine Learning agent). Recent results in reinforcement learning have shown that curriculum learning can accelerate learning and generalization progress. Using previous results, the goal for this thesis is to use available cluster information for an instance set to construct a learning curriculum and thus enabling the agent to learn more efficiently. This includes:

  1. Finding and implementing a mechanism to choose instances for training depending on cluster information
  2. Generating different instance sets for artificial benchmarks to evaluate the method
  3. Applying the result to real world instances (e.g. AI Planning or Coherent Ising Machines)


We strongly recommend that students should have experience for the following topics:

  • computer science and AI in general
  • Machine Learning
  • Reinforcement Learning
  • Deep Learning
To apply for this project, please send us an email with the following information:
  1. Which ML-related courses have you taken?
  2. Your transcript of records
  3. Which projects have you done so far (in Hannover and elsewhere)?
  4. Since different projects require different skill sets, please also rate your skills in the following categories on a scale from ++ (very good) to -- (no knowledge/skill):
  5. Creativity / ideas for developing new algorithms
  6. Getting someone else's large code base to run
  7. Running comprehensive experimental studies / keeping track of results
  8. Self-motivation to push through even if things don't work for a while
  9. Coding skills
    1. Python
    2. Keras
    3. PyTorch
    4. RL Frameworks like RLLib or ChainerRL
  10. Ability to read a RL paper, implement it and get it to work
  11. Ability to read a DL paper, implement it and get it to work
  12. Contact person: Theresa Eimer