A Deep Hierarchical Variational Autoencoder for World Models in Complex Reinforcement Learning Environments