A general criterion and an algorithmic framework for learning in multi-agent systems