Dantzig Selector with an Approximately Optimal Denoising Matrix and its Application to Reinforcement Learning