Mirror Descent Search
This is the experimental code for Mirror Descent Search in our paper -- Mirror Descent Search and Acceleration.
Dependency
- Python3
- Numpy
- Matplotlib
- Scipy
- Pyglet
- PyQt4 / PyQt5
pip install -r requirements.txt
Usage
python main.py
The output directory is created in ./mds_out/ with timestamp.
Parameter
See parameter description.
python main.py -h
- Example: set agent AMDS
python main.py --agent AMDS
Correspondence between this code and our paper
- G-AMDS in our paper:
--agent AMDS
- G-MDS in our paper:
--agent GMDS
- PI2 in our paper:
--agent PI2
- REPS in our paper:
--agent REPS
Output file
Above files are created in the output directory:
- args.txt: Run arguments.
- command.txt: Run command.
- cost.csv: Cost per the number of rollouts.
- cost.pdf: Plot of correlation between cost and number of rollouts.
- cost.png: Plot of correlation between cost and number of rollouts.
- git-*.txt: Git command.
- trajectory.mp4: Movie of agent's trajectory.
- rollout.csv: The number of rollouts.
- x.csv: X-axis trajectory per time
- y.csv: Y-axis trajectory per time
Rendering
If your environment has $DISPLAY, rendering is enable.
python main.py --render
Reference