Are exploration-based strategies of interest for repeated stochastic coalitional games?
Résumé
Coalitional games are models of cooperation where selfish agents must form groups (coalitions) to maximize their utility. In these models, it is generally assumed that the utility of a coalition is fixed and known. As these assumptions are not realistic in many applications, some works addressed this problem by considering repeated stochastic coalitional games. In such games, agents repeatedly form coalitions and observe their utility a posteriori in order to update their knowledge. However, it is generally assumed that agents have a greedy behavior: they always form the best coalitions they estimate at a given time step. In this article, we study if other strategies (behaviors) that allow agents to explore under-evaluated coalitions may be of interest. To this end, we propose a model of repeated stochastic coalitional game where agents use a neural network to estimate the utility of the coalitions. We compare different exploration strategies, and we show that, due to the structure of the coalitional games, the greedy strategy is the best despite the fact exploration-based strategies better estimate the utilities.
Origine : Fichiers produits par l'(les) auteur(s)