Sven Seuken will attend the Dagstuhl seminar on "Game Theory Meets Computational Learning Theory" from June 18, 2017 till June 23, 2017 and deliver the following keynote talk:
Title: Design of Machine Learning-based Mechanisms
Abstract: In this talk, we present a new paradigm we call “designing machine learning-based mechanisms.” In contrast to most prior work, our paradigm uses machine learning (ML) directly on the agents’ reports, not to optimize some future mechanism, but to immediately make use of the learning outcome (for the current instance). We instantiate this new idea via combinatorial auctions (CAs), and show how using ML inside CAs can substantially simplify the interaction with the bidders. In our CAs, the bidders report their values (bids) to a proxy agent by answering a small number of value queries. The proxy agent then uses an ML algorithm to generalize from those bids to the whole value space, and the efficient allocation is computed based on the generalized valuations. We discuss that this new design leads to new challenges regarding allocative efficiency, individual rationality, and incentives. However, we show that an iterative auction design and an epsilon-expressive ML algorithm address these challenges. To instantiate our design, we use support vector regression (SVR) as the ML algorithm, which enables us to formulate the winner determination problem as a succinct integer program. Finally, we present some experimental results for two stylized spectrum auction domains. Our results demonstrate that even with a small number of bids, our ML-based auctions achieve high allocative efficiency.
Based on joint work with: Gianluca Brero and Benjamin Lubin.
Related paper: [pdf]