Online learning in multi-agent systems: Regret, equilibrium, and the road ahead
Panayotis Mertikopoulos  1@  
1 : Laboratoire dÍnformatique de Grenoble
Université Pierre Mendès France - Grenoble 2, Université Joseph Fourier - Grenoble 1, Institut National Polytechnique de Grenoble, Centre National de la Recherche Scientifique : UMR5217, Université Grenoble Alpes

From auctions to traffic planning and the training of adversarial neural nets, the goal of multi-agent learning is always the same : to make more informed decisions that lead to better rewards over time. But who benefits in the end ? Each agent unilaterally ? All agents concurrently ? Both / None of the above ? This tutorial is intended to explore these questions by providing an introduction to online optimization and learning from a multi-agent, gametheoretic viewpoint. We will cover the basics of learning in finite games (focusing on the multiplicative/exponential weights algorithm and its properties), and the mainstay algorithms for online optimization in continuous games (online gradient descent, follow the regularizer leader, etc.). We will also discuss the impact of the information available to the players, as well as a range of concrete applications to machine learning and operations research.


Personnes connectées : 5 Vie privée
Chargement...