Game theory aims at predicting rational agents behavior in situations of interdependency. A particular attention has been given to the formal proofs of existence, uniqueness and properties of equilibria, that is situations in which no agent can unilaterally gain by deviating. However, when the system is large - due to the presence of many agents or many possible actions - computing such equilibria may just not be possible (NP-hardness). One approach is to consider a repeated scenario of the game where at each iteration, players try strategies with the aim of learning their best strategies and collectively converge to equilibria, based on trial-and-errors mechanism. In this talk we present some of the challenges of the theory of learning and some of our recent results.
Bienvenue à tous!