If you are looking for a bachelor or master thesis topic in AI, algorithmic game theory, or strategic learning, feel free to contact me by email.
I am also open to student-proposed topics if they connect to AI, game theory, or strategic decision making.
Current reinforcement-learning methods usually assume that the environment is Markovian, which is a strong assumption that many realistic scenarios do not satisfy. In this project, we can study concrete failures of current approaches and investigate how to refine the algorithms under weaker assumptions.
In many situations, one does not just care about finding any equilibrium, but rather one that also optimizes a secondary objective such as fairness. RL algorithms are efficient and often converge to equilibria. In this project, we would study how to modify an RL algorithm so that it also optimizes such a secondary objective.
A standard multi-layer perceptron can be viewed, from the perspective of information theory, as a Markov chain. This gives well-defined bounds on how information about the output encoded in the input layer propagates through the network. The goal of this project is to measure the flow of information and relate it to the learning process, including the study of loss of plasticity, where a network loses the ability to adapt to new training distributions.
LLM coding agents can already improve the efficiency of code by iterating and experimenting, but this process usually does not involve changing model weights. In this project, we would use reinforcement learning to finetune an LLM on a chosen family of problems. For example, the goal could be to optimize the performance of a C++ solver on a class of problems of interest.