First learn inverse model, then fine tune with ffw error learning