Learning Efficient Nash Equilibria in Distributed Systems