Hierarchical Relative Entropy Policy Search

The book explores the significance of hierarchical structures in enhancing scalability and performance in motor skill tasks. It introduces the concept of a "mixed option policy," where a gating network selects which option to execute, followed by an option-policy that determines the action. This hierarchical approach enables the learning of multiple solutions to problems. The algorithm is grounded in an innovative information theoretic policy search method that effectively manages the exploitation-exploration trade-off, minimizing information loss during policy updates.