Sparse Graphical Memory (SGM)

Sparse Graphical Memory (SGM) is a data structure for reinforcement-learning agents to solve long-horizon, sparse-reward navigation tasks. This codebase is a TensorFlow implementation of SGM accompanying the paper Sparse Graphical Memory for Robust Planning.

To install:

Create a new conda environment: conda create -n sgm python=3.6
Activate the conda environment: conda activate sgm
Install the requirements: pip install -r requirements.txt
Install the package: pip install -e .

To launch the experiments:

To launch SGM experiments, run all the cells in notebooks/SGM in Thinned FourRooms.ipynb
To launch SoRB experiments, run all the cells in notebooks/SoRB in Thinned FourRooms.ipynb

To visualize the results:

Call sgm/plot.py with the directories of the logs produced by the experiments followed by the --returns_v_cleanup_steps flag, e.g., python sgm/plot.py logs/thinned_fourrooms_sgm logs/thinned_fourrooms_sorb --returns_v_cleanup_steps
Find the visualized results in the plots directory, e.g., in plots/returns_v_cleanup_steps

Commands you may find useful

Enable rendering of environments when there is no display, e.g., on a server: xvfb-run -s "-screen 0 1400x900x24" jupyter notebook https://stackoverflow.com/questions/40195740/how-to-run-openai-gym-render-over-a-server

Create tunnel from local machine to remote machine: ssh -N -f -L localhost:8888:localhost:8888 user@remote.hostname.edu

Run Jupyter notebook with no time limit on cell execution and replace the notebook's contents with the new output: jupyter nbconvert --ExecutePreprocessor.timeout=-1 --execute --to notebook --inplace <notebook.ipynb> https://stackoverflow.com/questions/35545402/how-to-run-an-ipynb-jupyter-notebook-from-terminal

Credits

This code was built upon the code released by Eysenbach et al. under the Apache License, Version 2.0.

Reference

@inproceedings{laskin2020sparse,
  Author = {Laskin, Michael and Emmons,
  Scott and Jain, Ajay and Kurutach, Thanard
  and Abbeel, Pieter and Pathak, Deepak},
  Title = {Sparse Graphical
  Memory for Robust Planning},
  Booktitle = {arXiv preprint arXiv:2003.06417},
  Year = {2020}
}

MishaLaskin / sgm