Does learning converge to equilibrium in generic games?

Game theory models interacting biological and social systems. In a repeated game, players may converge or their dynamics oscillate. If the system is not designed to converge, which of these two behaviors can we expect? I will give an overview of research that addresses this question by studying convergence to equilibrium in generic games. I will consider simple 2-player, 2-action games, where players learn through Experience-Weighted Attraction learning. We are able to characterize convergence in the space of parameters and games. Games with a “best-reply cycle”, such as Matching Pennies, are the ones in which convergence is less likely. Further, I will show that the frequency of best-reply cycles predicts convergence of six learning rules in 2-player games with an arbitrary number of actions, when games are sampled at random given constraints.The larger the number of actions, or the more anti-correlated the payoffs, the more best-reply cycles become dominant, and convergence becomes less likely. Finally, I consider games with an arbitrary number of players and with a network structure. I show that more players and more dense networks increase the importance of best-reply cycles, making convergence unlikely. Overall, these results indicate that in many cases equilibrium is an unrealistic assumption, and one must explicitly model the dynamics of learning.

Detalles de contacto:

Tobias Galla

Contact form

Esta web utiliza cookies para la recolección de datos con un propósito estadístico. Si continúas navegando, significa que aceptas la instalación de las cookies.

Más información De acuerdo