This paper studies offline reinforcement learning with linear function approximation in a setting with decision-theoretic, but not estimation sparsity. The structural restrictions of the data-generating process presume that the transitions factor into a sparse component that affects the reward and could affect additional exogenous dynamics that do not affect the reward. Although the minimally sufficient adjustment set for estimation of full-state transition properties depends on the whole state, the optimal policy and therefore state-action value function depends only on the sparse component: we call this causal/decision-theoretic sparsity. We develop a method for reward-filtering the estimation of the state-action value function to the sparse component by a modification of thresholded lasso in least-squares policy evaluation. We provide theoretical guarantees for our reward-filtered linear fitted-Q-iteration, with sample complexity depending only on the size of the sparse component.

We introduce a computationally effective algorithm for a linear model selection consisting of three steps: screening--ordering--selection (SOS). Screening of predictors is based on the thresholded Lasso that is l_1 penalized least squares. The screened predictors are then fitted using least squares (LS) and ordered with respect to their t statistics. Finally, a model is selected using greedy generalized information criterion (GIC) that is l_0 penalized LS in a nested family induced by the ordering. We give non-asymptotic upper bounds on error probability of each step of the SOS algorithm in terms of both penalties. Then we obtain selection consistency for different (n, p) scenarios under conditions which are needed for screening consistency of the Lasso. For the traditional setting (n >p) we give Sanov-type bounds on the error probabilities of the ordering--selection algorithm. Its surprising consequence is that the selection error of greedy GIC is asymptotically not larger than of exhaustive GIC. We also obtain new bounds on prediction and estimation errors for the Lasso which are proved in parallel for the algorithm used in practice and its formal version.