Possible principles for aligned structure learning agents