New Criteria and a New Algorithm for Learning in Multi-Agent Systems