Related article on similar hierarchical / compositional policies learned by maximum entropy optimization: http://bair.berkeley.edu/blog/2017/10/06/soft-q-learning/