Incrementally Learning Functions of the Return