Marcello Sanguineti: Bridging machine learning and game theory: Shapley value approximation in cooperative games

Thursday 20 November 2025 at 14:00 in room D208

Speaker: Marcello Sanguineti (DIBRIS, University of Genova)

Title: Bridging machine learning and game theory: Shapley value approximation in cooperative games

Summary: Since the very beginning, there has been a fruitful exchange between machine learning, optimization, and approximation. While machine learning exploits optimization and approximation models and algorithms, it simultaneously poses problems that often represent optimization and approximation challenges. Moreover, machine learning can be fruitfully combined with optimization algorithms to improve their performance. Here, the optimization model known as Cooperative Game Theory is considered, which studies strategic interactions among agents and has links to disciplines such as Economics, Engineering, Computer Science, Robotics, Political and Social Sciences, Biology, Medicine, etc. The focus is on preliminary results and open issues in the approximate computation via supervised learning of the Shapley Value - a well-established concept in Cooperative Games - which serves as a metric for assessing the relative significance of players. It has found many applications, among which gauging the importance of individual nodes or arcs within complex networks (e.g., logistic, social, telecommunication networks), in such a way to design them optimally and/or improve existing ones. However, typically the computation of the Shapley Value of nodes/arcs in extensive networks is computationally expensive. This talk addresses the challenge of approximating via Machine Learning the Shapley Value in Cooperative Games defined on large networks, which are parameterized by quantities of interest (e.g., traffic demand in a transportation network). The smoothness properties of the Shapley Value with respect to network parameters are investigated and exploited to develop supervised learning techniques for its approximate computation. Numerical results on test-beds are presented. Open problems, research directions, and challenges in the interplay between Machine Learning and Cooperative Games are discussed.