We propose a new class of estimates of the autoregression parameter in AR _p models, based on autoregression rank scores. These estimators are based on linear programming algorithms, combined with a discrete numerical optimization step. They are shown to be asymptotically equivalent to the R-estimators of autoregressive parameters proposed by Koul and Saleh [l]. In contrast with the latter, however, our autoregression rank score estimators are autoregression invariant, so that each component of the parameter can be estimated separately. This property allows for substituting p one-dimensional discrete optimization steps for a unique pdimensional one, which is computationally simpler