Algorithmic Learning and Incentives: Synergies in Optimization and Mechanism Design

The goal of the proposed project is to study further the ongoing interplay between machine learning and algorithmic game theory.

To this end, we plan to focus on two main directions along this front. The first one concerns the design of learning algorithms for game-theoretic solution concepts in multi-agent learning. These questions generally boil down to solving appropriate optimization problems over continuous action spaces and find applications in the actual training of neural networks. As the current theoretical understanding of such problems is still lagging behind, our main objective is to work towards developing a unifying optimization framework for designing new training algorithms, and capturing the limitations of existing algorithms. Our second direction is to consider learning-oriented approaches in auction mechanisms. Following recent suggestions, we plan to study the design of mechanisms under data-driven paradigms, where the auctioneer can exploit access to data of past auctions, and under learning-augmented scenarios, where the goal is to exploit predictions on features of candidate solutions. These paradigms are motivated by practical considerations and create a whole new dimension on the classic models of mechanism design, with exciting new challenges.