1: \begin{abstract}

2: We consider the capacitated selfish replication game with binary preferences. We first devise an algorithm which can reach a pure Nash equilibrium on trees after $n$ steps where $n$ is the number of nodes (players) in the game. Introducing an ordinal potential function for such a game, we determine the convergence rate of the best response dynamics to a pure Nash equilibrium based on the radii of the updating agents. Using this potential function, we propose a least best response dynamics. We prove polynomial time convergence of such dynamics to a pure Nash equilibrium for general graphs when the number of resources is limited or their edge density is high enough with respect to the number of resources. This partially answers a conjecture on the existence of PLS-hard algorithms for finding the equilibrium when the number of resources is more than 2. We show that the least best response dynamics improve the convergence speed of an earlier algorithm \cite{gopalakrishnan2012cache} in the case of two resources by a factor of at least $n^2$. We extend the results for the capacitated selfish replication games with arbitrary cache size and provide an effective allocation rule for extra capacities in the network in order to accelerate the least best response dynamics to an equilibrium in polynomial time. Finally, we make a comparison between graph coloring and the set of equilibria of the capacitated selfish replication game.

3: \end{abstract}