This is the official repository for [Int-HRL: Towards Intention-based Hierarchical Reinforcement Learning](https://collaborative-ai.org/publications/penzkofer24_ncaa/)<br>
Int-HRL uses eye gaze from human demonstration data on the Atari game Montezuma's Revenge to extract human player's intentions and converts them to sub-goals for Hierarchical Reinforcement Learning (HRL). For further details take a look at the corresponding paper.
## Dataset
Atari-HEAD: Atari Human Eye-Tracking and Demonstration Dataset available at [https://zenodo.org/record/3451402#.Y5chr-zMK3J](https://zenodo.org/record/3451402#.Y5chr-zMK3J) <br>
To pre-process the Atari-HEAD data run [Preprocess_AtariHEAD.ipynb](Preprocess_AtariHEAD.ipynb), yielding the `all_trials.pkl` file needed for the following steps.
## Sub-goal Extraction Pipeline
1. [RAM State Labeling](RAMStateLabeling.ipynb): annotate Atari-HEAD data with room id and level information, as well as agent and skull location
2. [Subgoals From Gaze](SubgoalsFromGaze.ipynb): run sub-goal proposal extraction by generating saliency maps
3. [Alignment with Trajectory](TrajectoryMatching.ipynb): run expert trajectory to get order of subgoals
The Int-HRL agent is based on the hierarchically guided Imitation Learning method (hg-DAgger/Q), where we adapted code from [https://github.com/hoangminhle/hierarchical_IL_RL](https://github.com/hoangminhle/hierarchical_IL_RL) <br>
Due to the novel sub-goal extraction pipeline, our agent does not require experts during training and is more than three times more sample efficient compared to hg-DAgger/Q. <br>
To run the full agent with 12 separate low-level agents for sub-goal execution, run [agent/run_experiment.py](agent/run_experiment.py), for single agents (one low-level agent for all sub-goals) run [agent/single_agent_experiment.py](agent/single_agent_experiment.py).
