Iterative Model-Based Reinforcement Learning Using Simulations in the Differentiable Neural Computer