Risk-sensitive learners in network selection games

We consider a network with finite number of users where each user observes only a numerical value of its measurement. The system is interactive in the sense that each user´s payoff is affected by the environment state and the choices of all the other users. This scenario can be modeled as dynamic robust game. We examine how risk-sensitive learners influence the convergence time of such a game in a specific network selection problem. Based on imitative combined fully distributed payoff and strategy learning (CODIPAS), we provide a simple class of network selection games in which a convergence to global optimum can be obtained with a very fast convergence rate. We show that the risk-sensitive index can be used to improve the convergence time in a wide range of parameters.