The paper provides nice near-optimal sample complexity results for a setting of feature-based MBRL. The results are nontrivial extensions of previous tabular results. On the other hand, it requires a pretty strong anchor-state assumption, which to some extent limits the significance of the results.