Student research projects and theses

Extend, apply and refine our AutoDL tool Auto-PyTorch for new areas such as outlier detection, maintenance prediction or time series prediction. We recommend a strong background in machine learning (especially Deep Learning) and their chosen application for this work. When applying, please indicate the direction you would like to work in and provide a rough plan of how you would implement AutoPytorch in your target area.

Contact: Difan Deng

Modelling, implementation and evaluation of DAC for any target algorithm. We recommend a strong background in RL, basic knowledge of DAC, and the target domain of your choice to be successful in this topic. Possible target domains include machine learning or reinforcement learning, MIPS or SAT solvers, and evolutionary algorithms.

Contact: Theresa Eimer

Scaling-up Reinforcement Learning to environments more complicated than games, such as chess, necessitates us to bias our learning methods to tackle many issues with such settings, such as large state and action spaces, complicated dynamics, etc. Behavioral similarity encompasses exploiting conditions that lead to similar policies, i.e., if provided with the same state, they produce identical action distributions. In this thesis, we want to study these methods in the contextual setting, wherein we will try to answer the following questions:

1. Are similarity methods robust to small environmental changes, such as contextual Reinforcement Learning? 
2. Can we improve these methods using this additional contextual information using loss augmentations or other procedures? 
3. Can we meta-learn such metrics across environment distributions? 

Ideally, for a bachelor thesis, this topic would entail ablating existing methods, while for master theses, we would take it a step further and try to develop a new approach.

Contact: Aditya Mohan

Meta-Policy gradients (MPGs) aim at learning a set of hyperparameters in a single lifetime. The key idea is to interleave policy iteration with hyperparameter updates, thus, learning an objective (Bellman or policy gradient objective) and then using it to learn a new parameter update.  Recent work has shown that contextual information about the environment (such as information about goals) can enrich meta-gradients.  

Pitch: Study the impact of contextual information on vanilla and bootstrapped meta-gradients for generalization to similar environments. This work entails two questions:

1. Do standard MPG techniques work well in contextual setting?
2. Does incorporating contextual information help with learning better hyperparameter schedules using MPGs in such settings?

Depending on your interest, we can scope out how to concretely answer these questions. 

Multi-Fidelity's primary goal is to achieve cost savings in the hyperparameter optimization process through approximated final performance, such as shorter training or training with less data. Since these approximations are typically gradual, it is advantageous to reduce the approximation error by selectively evaluating at multiple stages. The aim of this study is to transfer the perceived prior knowledge from one stage to the next and critically question which stage one should choose. The approach is to understand multi-fidelity as a continuous learning process across several similar tasks that need to be classified. We recommend a strong background in AutoML for this work.

Contact: Tim Ruhkopf

We can generate augmentation functions by meta-learning, something for the policy objective in PPO. However, it is open whether this is generally true for algorithm components in reinforcement learning, whether we could also learn augmentation ensembles, and how well these functions generalise. The goal of this work is to extend existing techniques to new algorithms and components.

Contact: Theresa Eimer

SMAC is a robust and flexible framework for Bayesian optimisation. It supports users in finding well-suited hyperparameter configurations for their algorithms, e.g. in machine learning. The core of SMAC consists of Bayesian optimisation combined with an "aggressive racing" mechanism that efficiently decides which configuration is more suitable.
Of course, SMAC is not yet complete and we have some possible final thesis topics, e.g.:

• Implementation and evaluation of state-of-the-art models that approximate the underlying function.
• Implementation and evaluation of input and output warping techniques
• Implementation and evaluation of multi-objective quality indicators

and many more!

Contact: Carolin Benjamins

In collaboration with Zoo Hannover, we aim to enhance object detection and tracking algorithms to monitor the maternal care of Thomson’s gazelles. We integrate Automated Machine Learning (AutoML) and Computer Vision to achieve this, focusing mainly on Hyperparameter Optimization (HPO). The project's primary objective is to improve the analysis of animal behavior using camera data. After identifying leading tracking algorithms, this master's thesis involves an in-depth examination of relevant hyperparameters specific to the corresponding tracking algorithms. The goal of this research is the strategic use of the AutoML tool SMAC to optimize the performance of the tracking algorithms, aiming for enhanced accuracy and efficiency in analyzing animal behavior.

Kontakt: Leona Hennig