On the Complexity of Learning to Cooperate with Populations of Socially Rational Agents