Creating simulation environments where AIs tackle very basic problems and develop their own strategies could generate empirical data about what approaches are successful, providing a laboratory to understand which strategies work, how evolution of strategies occurs, and testing what network architectures can solve simple problems.