Recurrent policy gradients