Prof. Martin Bichler

Prof. Felix Brandt

Our current research explores the computation of market equilibria in game-theoretic settings via multi-agent machine learning techniques: To do so, market participants update their behavior according to data of past market outcomes. To achieve good precision, this method requires the simulation of millions of episodes and computationally intensive Monte-Carlo integration. 
In this thesis, you will focus on methods from numerical analysis (e.g. stratified sampling, quasirandom numbers, variance reduction techniques) to analyze and improve sample-efficiency in this setting. The thesis should comprise both theoretical analysis and empirical work using implementations in python/pytorch.

This thesis can be supervised for students in the MA or IN departments.

Required: Strong background in mathematics, esp. probability theory, and numerics. Previous experience with programming in python. Basic understanding of game theory and machine learning.

In Reinforcement Learning, an agent continuously updating its strategy is prone to “forget” previously discovered strategic behavior when only learning based on short term feedback. One type of approach to handle this problem is by replaying past experiences once in a while in order to validate the performance of the current strategy. In this project, multiple approaches to Experience Replay will be analyzed in a specific multi-agent setting. References include “Prioritized Experience Replay” by T. Schaul et al., 2015 and “Stabilising experience replay for deep multi-agent reinforcement learning” by J. Foerster et al., 2017.

This thesis can be supervised for students in the MA or IN departments.

Mandatory Requirements: Python, Deep Learning. Optional Requirements: RL, MARL, Game Theory, PyTorch.

The sharing economy depends on the development of the sharing platform. Different platforms (e.g., ride-hailing, freight exchange, kidney exchange, resource allocation, ... ) have different characteristics. We are committed to abstracting mathematical models from reality to simulate, analyze and provide theory. Research issues include but are not limited to matching strategies, pricing issues, and online prediction.

Required: advanced programming skills (e.g., Python, Matlab, at least one), mathematics, operation research.

Donghao Zhu

Modifying a bid language for
spectrum auctions

(BA Thesis)

The Flexible Use and Efficient Licensing (FUEL) bid language was proposed for conducting a radio spectrum auction in the US to allocate licenses for the new 5G network to telecommunication providers. In this thesis it shall be analyzed how small modifications of this bid language influence the runtime and efficiency of the allocation problem.

required: advanced C++ skills, helpful: knowledge in auction theory & operations research

We provide a simulation system that can be extended by further valuation models for airlines and/or bidding languages. In simulation experiments, the impact on prices, the welfare distribution, and computational costs should be analyzed for different payment rules.

required: advanced programming skills (python); helpful: auction theory, operations research