th_rl

Pytorch-based package for multi-agent reinforcement learning in an iterated prisonner dilemma setting.

Installation

• create python venv

python -m venv power_markets
power_markets\scripts\activate

• install in local editable mode:

git clone https://github.com/nikitcha/th_rl
pip install -e th_rl

• install as module from Github

pip install -U git+https://github.com/nikitcha/th_rl.git

• if you want to install just dependencies:

pip install -r th_rl/requirements.txt

Usage - training

• Create a set of training configs and store them somewhere, i.e. /some_path/configs
  • Configs should follow the structured laid out in 'example_config.json'
• Run training with the follwing command:
  • This would run each config 20 times
  • Result will be stored under /some_path/runs

python main.py --cdir=/some_path/configs --runs=20  

Or if installed:

python -m th_rl.main.py --cdir=/some_path/configs --runs=20  

Usage - plot results

• To plot a single trajectory:

python utils.py --dir=/some_path/runs/qtable_001/1 --fun=plot_experiment

• To plot the mean of all [20] runs:

python utils.py --dir=/some_path/runs/qtable_001 --fun=plot_mean_result

Or if installed:

python -m th_rl.utils --dir=/some_path/runs/qtable_001 --fun=plot_mean_result