Distributed Mechanism Learning

Mechanism design is a field concerned with finding rules for economic processes which incentivize self-interested agents to behave in a way, such that a common goal is reached. This project aims to build robust infrastructure for mechanism design via machine learning, to make theoretical results more applicable to practical networked deployments. We plan to do this by finding solutions for the following two problems and making them accessible to developers, while keeping the required domain knowledge to a minimum:

On the one hand, a trusted third party is often assumed to exist, which is supposed to learn and execute the mechanism. In practice, finding neutral trusted parties who do not stand to gain anything from cheating can be hard. To solve this problem, we distribute the computation of the trusted party over multiple computers, ideally controlled by different entities, using multiparty computation. This way, we get a more robust trust base with better alignment of incentives.

On the other hand, current models often assume prior knowledge about preference distributions of agents to learn optimal mechanisms. In practice, this knowledge is not always available. We exchange finding optimal solutions using prior information with finding approximate solutions using no prior information, by way of differentially private learning. This results in more general applicability, especially in settings with sparse information.