Recent years have seen increased interest in non-parametric reinforcement learning. There are now practical kernel-based algorithms for approximating value functions; however, kernel regression requires that the underlying ...