shaoeChen / sutton_rooms

Room world from Sutton et al (Between MDP and semi-MDP)

sutton_rooms

Environment

Room world from Sutton et al (Between MDP and semi-MDP)

Paper: Between MDPs and semi-MDPs

Experiment

Used the room environment to test out some ideas for hierarchical reinforcement learning and planning in HRL.

1. Plain flat Q-learning
   1. Just one version. [code]
2. Hierarchical Q-learning
   1. Basic version (s-MDP; two-layer hierarchy with predefined deterministic lower-level policy) [code]
   2. Intraoption-learning version (lower-level is trainable) [code]
3. Planning Hierarchical Q-learning
   1. Basic version (same as Hierarchical Q-learning but with a 2-step plan output from the upper level; No replanning) [code]
   2. Version with replanning [code]