Skip to content

Latest commit

 

History

History
79 lines (62 loc) · 3.06 KB

README.md

File metadata and controls

79 lines (62 loc) · 3.06 KB
Raw

Adaptive Mechanism Design (AMD) in Sequential Social Dilemma (SSD)

This is the code for project of cource Foundations of Reinforcement Learning (FoRL) Spring Semester 2023. The major contributions of this project is

  • A full implementation of AMD algorithm [original paper] for arbitrary environments, using ray==2.3.1.
    • migrating to higher version of ray needs additional effort
  • Two RL environments, Wolfpack and Gathering. [original paper]

For further information, please refer to our report.

Instruction for local and Euler environment setup

I decide to use Python 3.10.4 and CUDA 11.8 as standard version. This is the default version on Euler.

These versions can be modifyed. Depending on the repo we are migrating.

Euler setup

On Euler, for each time you need to load the module.

module load gcc/8.2.0 python_gpu/3.10.4 cuda/11.8.0 git-lfs/2.3.0 git/2.31.1 eth_proxy

Package gcc/8.2.0 is necessary. Only this module is loaded, then can you search out result about python and cuda. You can search for the version of python and cuda you want by command

module avail ${package name: cuda, python, etc}

create a virtual env with venv

py_venv_dir="${SCRATCH}/.python_venv"
python -m venv ${py_venv_dir}/forl-proj --upgrade-deps
# To install python packages, run
${SCRATCH}/.python_venv/forl-proj/bin/pip install -r requirements.txt --cache-dir ${SCRATCH}/pip_cache
# actiavte
source "${SCRATCH}/.python_venv/forl-proj/bin/activate"
# deactivate
deactivate

Local setup

On local machine, to install this exact python version I use conda (you can also use venv).

conda create --name=forl-proj python=3.10
# activate
conda activate forl-proj
# deactivate
conda deactivate

Submit jobs on Euler/ Slurm

You can first edit the resources needed in start-ray.sbatch file, and submit jobs by

# command
echo "$(cat start-ray-nodes.sbatch; echo command_to_submit_jobs )" > temp && sbatch < temp && rm temp
# command from file
echo "$(cat start-ray-nodes.sbatch file_of_job )" > temp && sbatch < temp && rm temp

For example

echo "$(cat start-ray-nodes.sbatch exp_scripts/wolfpack/amd-qadj-delay_model-conv_assump-neural.bash)" > temp && sbatch < temp && rm temp

Tips for using LSF or Slurm

There are some links here

Most importantly, this interactive website can generate sbatch scripts.

Ray on SLURM

See the following link: