RLApps

RLlib-based general reinforcement learning algorithms collections (derived from John B. Lanier's implementation). I bumped it up to Ray v1.12 and fixed some errors. More general RL algorithms will be introduced.

NOTE: As a nice gRPC and RLlib derivation, I recommend our groupmates to use it as a good practice. In addition, please feel free to enrich this repo by submitting your PRs.

Installation

(Tested on Ubuntu 20.04)

Set up Conda environment and python modules

After installing Anaconda, enter the repo directory and create the new environment:

conda create -n rlapps python==3.8 -y
conda activate rlapps
bash ./install.sh

Advanced Installation Notes (Optional)

If you need to compile/recompile OpenSpiel without pip installing it, perform the following steps with your conda env active. (The conda env needs to be active so that OpenSpiel can find and compile against the python development headers in the env. Python version related issues may occur otherwise):

mkdir build
cd build
CC=clang CXX=clang++ cmake -DPython_TARGET_VERSION=3.6 -DCMAKE_CXX_COMPILER=${CXX} -DPython3_FIND_VIRTUALENV=FIRST -DPython3_FIND_STRATEGY=LOCATION ../open_spiel
make -j$(nproc)
cd ../../..

To import OpenSpiel without using pip, add OpenSpiel directories to your PYTHONPATH in your ~/.bashrc (more details here):

# Add the following lines to your ~/.bashrc:
# For the python modules in open_spiel.
export PYTHONPATH=$PYTHONPATH:/<path_to_open_spiel_submodule>
# For the Python bindings of Pyspiel
export PYTHONPATH=$PYTHONPATH:/<path_to_open_spiel_submodule>/build/python

Next Steps

See Running Experiments