Dyna planning is an efficient way of learning from real and imaginary experience. Existing tabular and linear Dyna algorithms are single-step, because an "imaginary" feature is predicted only one
