A solution concept with an exploration bias for repeated stochastic coalitional games
Résumé
Classically, in coalition formation, agents know in advance the deterministic utilities they will obtain from coalitions. Relaxing these two assumptions (determinism and a priori knowledge) is important to deal with real-world applications. A way to do that is to consider the framework of repeated stochastic coalitional games. Here, agents decide at each time step which coalition to form on the basis of limited information. Then, their observations allow them to update their knowledge. We propose a solution concept that explicitly integrates an exploration bias to allow agents to sometimes form coalitions that have a low utility but that would be interesting to form to obtain more information. We compare this concept to a greedy approach and highlight its efficiency with respect to the structure of the real utilities, unknown to the agents.
Origine : Fichiers produits par l'(les) auteur(s)