Learning to Control an Octopus Arm with Gaussian Process Temporal Difference Methods