A Multi-step Loss Function for Robust Learning of the Dynamics in Model-based Reinforcement Learning