Complexity theory allows to classify problems by their algorithmic hardness. The classical framework in which it applies is the one of a centralized algorithm that knows every informa- tion. With the development of networks and decentralized architectures, distributed dynamics was studied. In many problems, in optimization or economy, actions and computations are made by independant agents that don’t share the same objective whose realization depends on the actions of other agents. Game theory is a natural framework to study solutions of this kind of problem. It provides solution concepts such as the Nash equilibrium.A natural way to compute these solutions is to make the agents “react” ; if an agent sees the actions of the other player, or more generally the state of the game, he can decide to change his decision to reach his objective and updates the state of the game. We call �dynamics� this kind of algorithms.We know some dynamics converges to a stable solution. We are interested by the speed of convergence of these dynamics. Some solution concepts are even complete for some complexity classes which make unrealistic the existence of fast converging dynamics. We used three ways to obtain a fast convergence : improving dynamics (using random bits), finding simple subcases, and finding an approximate solution.We extent fast convergence results to an approximate Nash equilibria in negative congestion games. However, we proved that finding an approximate Nash equilibrium in a congestion games without sign restriction is PLS-complete. On matching game, we studied the speed of concurrent dynamics when players have partial information that depends on a social network. Especially, we improved natural dynamics for them to reach an equilibrium inO(log(n)) rounds (with n is the number of players).
